Add Batch 7735fa54-f98d-471c-b6aa-9cc17dd12944 data

.gitattributes

@@ -5539,3 +5539,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Language[[:space:]]Model[[:space:]]Analysis[[:space:]]for[[:space:]]Ontology[[:space:]]Subsumption[[:space:]]Inference/6fa31862-b1ea-4de8-8da1-a4f2eba6ef5f_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Language[[:space:]]Modeling[[:space:]]with[[:space:]]Latent[[:space:]]Situations/916ebf86-4821-46f4-a642-81f73035dbd7_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Language[[:space:]]Models[[:space:]]for[[:space:]]German[[:space:]]Text[[:space:]]Simplification_[[:space:]]Overcoming[[:space:]]Parallel[[:space:]]Data[[:space:]]Scarcity[[:space:]]through[[:space:]]Style-specific[[:space:]]Pre-training/79849c60-d2cf-40cf-8f0a-9db2151769d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Language[[:space:]]acquisition_[[:space:]]do[[:space:]]children[[:space:]]and[[:space:]]language[[:space:]]models[[:space:]]follow[[:space:]]similar[[:space:]]learning[[:space:]]stages_/6b7349d3-6e5d-4155-ae82-6dcafa1b591d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]Are[[:space:]]Partially[[:space:]]Primed[[:space:]]in[[:space:]]Pronoun[[:space:]]Interpretation/60280f41-b321-4144-822d-79720ae15459_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]Can[[:space:]]be[[:space:]]Lazy[[:space:]]Learners_[[:space:]]Analyze[[:space:]]Shortcuts[[:space:]]in[[:space:]]In-Context[[:space:]]Learning/e5cccaf8-0a87-470a-bb95-d56e7d2ebccc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Built-in[[:space:]]Autoregressive[[:space:]]Search[[:space:]]Engines/5ad13cc6-938f-4faa-8252-ed3a3053bd31_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Controllable[[:space:]]Working[[:space:]]Memory/df593dd7-917d-44ac-ab6c-8a0670ece26b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Layerwise[[:space:]]universal[[:space:]]adversarial[[:space:]]attack[[:space:]]on[[:space:]]NLP[[:space:]]models/6e17acd0-02c0-4f9f-8520-e21ab1c219b9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learn[[:space:]]to[[:space:]]Not[[:space:]]Link_[[:space:]]Exploring[[:space:]]NIL[[:space:]]Prediction[[:space:]]in[[:space:]]Entity[[:space:]]Linking/980725a0-e77c-4bea-bb00-a61ff06440a3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learned[[:space:]]Adapters[[:space:]]Are[[:space:]]Better[[:space:]]Than[[:space:]]Manually[[:space:]]Designed[[:space:]]Adapters/55f2333f-7821-49a3-9f81-0d94e944171c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Event-aware[[:space:]]Measures[[:space:]]for[[:space:]]Event[[:space:]]Coreference[[:space:]]Resolution/74a798ae-3644-4b8d-971d-d0e003ce1a90_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Joint[[:space:]]Structural[[:space:]]and[[:space:]]Temporal[[:space:]]Contextualized[[:space:]]Knowledge[[:space:]]Embeddings[[:space:]]for[[:space:]]Temporal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Completion/75d1f7d8-9259-45ff-9159-6adaa62b7d87_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Query[[:space:]]Adaptive[[:space:]]Anchor[[:space:]]Representation[[:space:]]for[[:space:]]Inductive[[:space:]]Relation[[:space:]]Prediction/c5a83937-655e-490b-b048-9a9ecb713b9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]by[[:space:]]Analogy_[[:space:]]Diverse[[:space:]]Questions[[:space:]]Generation[[:space:]]in[[:space:]]Math[[:space:]]Word[[:space:]]Problem/447c53b5-536e-40fc-a981-719a2fa2b65c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]from[[:space:]]Children_[[:space:]]Improving[[:space:]]Image-Caption[[:space:]]Pretraining[[:space:]]via[[:space:]]Curriculum/57e78277-b0f3-4ea8-adb1-c97f9f6a94b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]from[[:space:]]a[[:space:]]Friend_[[:space:]]Improving[[:space:]]Event[[:space:]]Extraction[[:space:]]via[[:space:]]Self-Training[[:space:]]with[[:space:]]Feedback[[:space:]]from[[:space:]]Abstract[[:space:]]Meaning[[:space:]]Representation/f44688f3-20a7-4c11-be17-a7fc58ba20ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Generalize[[:space:]]for[[:space:]]Cross-domain[[:space:]]QA/06da328c-3693-43bd-a99d-e97084fdd1de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Leverage[[:space:]]High-Order[[:space:]]Medical[[:space:]]Knowledge[[:space:]]Graph[[:space:]]for[[:space:]]Joint[[:space:]]Entity[[:space:]]and[[:space:]]Relation[[:space:]]Extraction/6a40fbbf-56bf-49e6-8124-4bc0f22ff171_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Predict[[:space:]]Persona[[:space:]]Information[[:space:]]for[[:space:]]Dialogue[[:space:]]Personalization[[:space:]]without[[:space:]]Explicit[[:space:]]Persona[[:space:]]Description/f0fee746-cc54-4fcd-99b0-7a50975d57b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Rank[[:space:]]Utterances[[:space:]]for[[:space:]]Query-Focused[[:space:]]Meeting[[:space:]]Summarization/33cc5cd7-624d-47c2-b5d3-66683b229bd6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Lego-MT_[[:space:]]Learning[[:space:]]Detachable[[:space:]]Models[[:space:]]for[[:space:]]Massively[[:space:]]Multilingual[[:space:]]Machine[[:space:]]Translation/51c725c6-9464-4e5f-b279-06b59feba491_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Less[[:space:]]Likely[[:space:]]Brainstorming_[[:space:]]Using[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Generate[[:space:]]Alternative[[:space:]]Hypotheses/4df00d29-4775-4f6a-96f6-99e29515a8c1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]Denoised[[:space:]]Abstract[[:space:]]Meaning[[:space:]]Representation[[:space:]]for[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction/954ac98b-3627-402e-ba0f-8545768a7058_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]Explicit[[:space:]]Procedural[[:space:]]Instructions[[:space:]]for[[:space:]]Data-Efficient[[:space:]]Action[[:space:]]Prediction/b978f616-c755-4f80-be8a-049326b0dd52_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]Synthetic[[:space:]]Targets[[:space:]]for[[:space:]]Machine[[:space:]]Translation/c679735b-5a99-4bf4-87b8-c1c384540cb8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]Training[[:space:]]Data[[:space:]]in[[:space:]]Few-Shot[[:space:]]Prompting[[:space:]]for[[:space:]]Numerical[[:space:]]Reasoning/fb282b11-871a-47e7-a4b1-643549fab2d6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Lexical[[:space:]]Translation[[:space:]]Inconsistency-Aware[[:space:]]Document-Level[[:space:]]Translation[[:space:]]Repair/8de74fd8-7dcb-4bf1-87d1-f3e1c9e62ba8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LightFormer_[[:space:]]Light-weight[[:space:]]Transformer[[:space:]]Using[[:space:]]SVD-based[[:space:]]Weight[[:space:]]Transfer[[:space:]]and[[:space:]]Parameter[[:space:]]Sharing/39835909-3feb-4f77-b618-495253273076_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Listen,[[:space:]]Decipher[[:space:]]and[[:space:]]Sign_[[:space:]]Toward[[:space:]]Unsupervised[[:space:]]Speech-to-Sign[[:space:]]Language[[:space:]]Recognition/d2c5b0fb-3a18-4e37-b1ae-eb8e0f52deaf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Local[[:space:]]Temperature[[:space:]]Beam[[:space:]]Search_[[:space:]]Avoid[[:space:]]Neural[[:space:]]Text[[:space:]]DeGeneration[[:space:]]via[[:space:]]Enhanced[[:space:]]Calibration/d2d26b82-599b-4ac9-84e5-4ac148f3fc4a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Logical[[:space:]]Transformers_[[:space:]]Infusing[[:space:]]Logical[[:space:]]Structures[[:space:]]into[[:space:]]Pre-Trained[[:space:]]Language[[:space:]]Models/e1df4d4e-0b48-471c-a883-3469d1f653bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Long[[:space:]]to[[:space:]]reign[[:space:]]over[[:space:]]us_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]of[[:space:]]Machine[[:space:]]Translation[[:space:]]and[[:space:]]a[[:space:]]New[[:space:]]Monarch/0a31e81f-1fb9-406a-8486-d4299b7ca79c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Low-Rank[[:space:]]Updates[[:space:]]of[[:space:]]pre-trained[[:space:]]Weights[[:space:]]for[[:space:]]Multi-Task[[:space:]]Learning/5b90eb3d-78b3-4f8e-8fec-c3318da4668e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MISMATCH_[[:space:]]Fine-grained[[:space:]]Evaluation[[:space:]]of[[:space:]]Machine-generated[[:space:]]Text[[:space:]]with[[:space:]]Mismatch[[:space:]]Error[[:space:]]Types/028eaee1-224f-4ff5-98ba-f898d5b5e42b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ML-LMCL_[[:space:]]Mutual[[:space:]]Learning[[:space:]]and[[:space:]]Large-Margin[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Improving[[:space:]]ASR[[:space:]]Robustness[[:space:]]in[[:space:]]Spoken[[:space:]]Language[[:space:]]Understanding/9a1716be-cb2a-4761-b1e2-62232bca611a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MMSD2.0_[[:space:]]Towards[[:space:]]a[[:space:]]Reliable[[:space:]]Multi-modal[[:space:]]Sarcasm[[:space:]]Detection[[:space:]]System/6ebcabc1-8fc5-4ae6-a8b0-7b5d87702d09_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MTCue_[[:space:]]Learning[[:space:]]Zero-Shot[[:space:]]Control[[:space:]]of[[:space:]]Extra-Textual[[:space:]]Attributes[[:space:]]by[[:space:]]Leveraging[[:space:]]Unstructured[[:space:]]Context[[:space:]]in[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/9f16f026-a6f7-49fa-a734-d26fd4f0f93f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MTGP_[[:space:]]Multi-turn[[:space:]]Target-oriented[[:space:]]Dialogue[[:space:]]Guided[[:space:]]by[[:space:]]Generative[[:space:]]Global[[:space:]]Path[[:space:]]with[[:space:]]Flexible[[:space:]]Turns/9be2d9ab-1328-48a2-b99c-2c96420959ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MTR_[[:space:]]A[[:space:]]Dataset[[:space:]]Fusing[[:space:]]Inductive,[[:space:]]Deductive,[[:space:]]and[[:space:]]Defeasible[[:space:]]Reasoning/cff8496e-17bb-43f4-b56d-9e88a45acac1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MURMUR_[[:space:]]Modular[[:space:]]Multi-Step[[:space:]]Reasoning[[:space:]]for[[:space:]]Semi-Structured[[:space:]]Data-to-Text[[:space:]]Generation/0b06c6ac-0b91-4924-96b1-26569658b341_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MVP_[[:space:]]Multi-task[[:space:]]Supervised[[:space:]]Pre-training[[:space:]]for[[:space:]]Natural[[:space:]]Language[[:space:]]Generation/58a594ee-b15d-46dd-a4f8-51d15c8e39ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Making[[:space:]]Better[[:space:]]Use[[:space:]]of[[:space:]]Training[[:space:]]Corpus_[[:space:]]Retrieval-based[[:space:]]Aspect[[:space:]]Sentiment[[:space:]]Triplet[[:space:]]Extraction[[:space:]]via[[:space:]]Label[[:space:]]Interpolation/5844c40d-588a-4a37-a9d7-bb39595c6b9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Making[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models[[:space:]]Better[[:space:]]Learn[[:space:]]Few-Shot[[:space:]]Spoken[[:space:]]Language[[:space:]]Understanding[[:space:]]in[[:space:]]More[[:space:]]Practical[[:space:]]Scenarios/8cc027e1-57bc-4a36-9cc1-ba59cd317495_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Making[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models[[:space:]]both[[:space:]]Task-solvers[[:space:]]and[[:space:]]Self-calibrators/cfc26661-8bc8-400e-b584-1da3e3710596_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mapping[[:space:]]Brains[[:space:]]with[[:space:]]Language[[:space:]]Models_[[:space:]]A[[:space:]]Survey/1784e28c-3a14-4eb8-8e7f-d985ac0a0fe5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mars_[[:space:]]Modeling[[:space:]]Context[[:space:]]&[[:space:]]State[[:space:]]Representations[[:space:]]with[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]End-to-End[[:space:]]Task-Oriented[[:space:]]Dialog/03c375d0-3487-43bd-b228-7e3ed7a3d04d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Masked[[:space:]]Audio[[:space:]]Text[[:space:]]Encoders[[:space:]]are[[:space:]]Effective[[:space:]]Multi-Modal[[:space:]]Rescorers/329a6e25-6cb8-49e4-9d16-70b9faf4e482_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Masked[[:space:]]Latent[[:space:]]Semantic[[:space:]]Modeling_[[:space:]]an[[:space:]]Efficient[[:space:]]Pre-training[[:space:]]Alternative[[:space:]]to[[:space:]]Masked[[:space:]]Language[[:space:]]Modeling/d50cf3b9-a830-43a4-8f3f-c86ab21879f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Maximum[[:space:]]Entropy[[:space:]]Loss,[[:space:]]the[[:space:]]Silver[[:space:]]Bullet[[:space:]]Targeting[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]in[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models/d7c050e4-5ec5-444a-b5f6-99de5b792e2b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Measuring[[:space:]]Intersectional[[:space:]]Biases[[:space:]]in[[:space:]]Historical[[:space:]]Documents/41995865-156c-477a-bf24-49b70b31fce2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Measuring[[:space:]]Your[[:space:]]ASTE[[:space:]]Models[[:space:]]in[[:space:]]The[[:space:]]Wild_[[:space:]]A[[:space:]]Diversified[[:space:]]Multi-domain[[:space:]]Dataset[[:space:]]For[[:space:]]Aspect[[:space:]]Sentiment[[:space:]]Triplet[[:space:]]Extraction/1693153d-9e5d-4ec5-b40d-ce877c323b7d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Measuring[[:space:]]and[[:space:]]Mitigating[[:space:]]Local[[:space:]]Instability[[:space:]]in[[:space:]]Deep[[:space:]]Neural[[:space:]]Networks/cb31d90e-2a3e-4650-8054-536cb84c467c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MedNgage_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Understanding[[:space:]]Engagement[[:space:]]in[[:space:]]Patient-Nurse[[:space:]]Conversations/69cf41d9-bf3d-469b-99c9-1898901f3948_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Medical[[:space:]]Dialogue[[:space:]]Generation[[:space:]]via[[:space:]]Dual[[:space:]]Flow[[:space:]]Modeling/8fda489c-5416-4008-a3ad-4707c6afbd45_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Membership[[:space:]]Inference[[:space:]]Attacks[[:space:]]against[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Neighbourhood[[:space:]]Comparison/24c94e80-52b2-422d-835f-2268953fe120_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Meta-training[[:space:]]with[[:space:]]Demonstration[[:space:]]Retrieval[[:space:]]for[[:space:]]Efficient[[:space:]]Few-shot[[:space:]]Learning/85a0b17b-7f36-4465-896d-e9b766f916d5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mind[[:space:]]the[[:space:]]Biases_[[:space:]]Quantifying[[:space:]]Cognitive[[:space:]]Biases[[:space:]]in[[:space:]]Language[[:space:]]Model[[:space:]]Prompting/868caa0a-78d9-4953-81e2-c8db2e89964b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mini-Model[[:space:]]Adaptation_[[:space:]]Efficiently[[:space:]]Extending[[:space:]]Pretrained[[:space:]]Models[[:space:]]to[[:space:]]New[[:space:]]Languages[[:space:]]via[[:space:]]Aligned[[:space:]]Shallow[[:space:]]Training/3347fd23-0f0a-4722-84f5-ffebba963b86_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Misleading[[:space:]]Relation[[:space:]]Classifiers[[:space:]]by[[:space:]]Substituting[[:space:]]Words[[:space:]]in[[:space:]]Texts/ee9cd3e7-592e-4fc5-8c00-153ca2023e70_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mitigating[[:space:]]the[[:space:]]Learning[[:space:]]Bias[[:space:]]towards[[:space:]]Repetition[[:space:]]by[[:space:]]Self-Contrastive[[:space:]]Training[[:space:]]for[[:space:]]Open-Ended[[:space:]]Generation/656dd1bc-aa11-47f7-8759-296706036421_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MixPAVE_[[:space:]]Mix-Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Few-shot[[:space:]]Product[[:space:]]Attribute[[:space:]]Value[[:space:]]Extraction/1a449d90-f610-483f-b6e0-ed9b54e64167_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MoNET_[[:space:]]Tackle[[:space:]]State[[:space:]]Momentum[[:space:]]via[[:space:]]Noise-Enhanced[[:space:]]Training[[:space:]]for[[:space:]]Dialogue[[:space:]]State[[:space:]]Tracking/e05b432d-56a1-4dae-8f6e-8bbade3c6cff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Model[[:space:]]Analysis[[:space:]]&[[:space:]]Evaluation[[:space:]]for[[:space:]]Ambiguous[[:space:]]Question[[:space:]]Answering/0f137193-35e3-4656-bf39-057f1f40764c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Model[[:space:]]Interpretability[[:space:]]and[[:space:]]Rationale[[:space:]]Extraction[[:space:]]by[[:space:]]Input[[:space:]]Mask[[:space:]]Optimization/33582781-e8cd-4e14-89e9-e0e39c59158b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Model[[:space:]]Intrinsic[[:space:]]Features[[:space:]]of[[:space:]]Fine-tuning[[:space:]]based[[:space:]]Text[[:space:]]Summarization[[:space:]]Models[[:space:]]for[[:space:]]Factual[[:space:]]Consistency/e520d862-13ea-46c6-b75f-aff70f95af1c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Modeling[[:space:]]Adversarial[[:space:]]Attack[[:space:]]on[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models[[:space:]]as[[:space:]]Sequential[[:space:]]Decision[[:space:]]Making/bbd2b0f7-11b9-41df-84c2-1146c3dd5d97_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Language acquisition_ do children and language models follow similar learning stages_/6b7349d3-6e5d-4155-ae82-6dcafa1b591d_content_list.json

@@ -0,0 +1,2191 @@
+[
+    {
+        "type": "text",
+        "text": "Language acquisition: do children and language models follow similar learning stages?",
+        "text_level": 1,
+        "bbox": [
+            136,
+            86,
+            867,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Linnea Evanson",
+        "bbox": [
+            178,
+            139,
+            322,
+            154
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Meta AI Paris;",
+        "bbox": [
+            203,
+            158,
+            295,
+            170
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Laboratoire des systèmes perceptifs",
+        "bbox": [
+            139,
+            174,
+            359,
+            187
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ecole normale supérieure",
+        "bbox": [
+            171,
+            187,
+            327,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "PSL University",
+        "bbox": [
+            201,
+            198,
+            297,
+            210
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "linnea.evanson8@gmail.com",
+        "bbox": [
+            134,
+            214,
+            364,
+            227
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yair Lakretz*",
+        "bbox": [
+            438,
+            139,
+            559,
+            154
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Cognitive Neuroimaging Unit",
+        "bbox": [
+            406,
+            162,
+            591,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "CEA,INSERM",
+        "bbox": [
+            448,
+            175,
+            547,
+            186
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Université Paris-Saclay",
+        "bbox": [
+            426,
+            186,
+            569,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "NeuroSpin Center",
+        "bbox": [
+            442,
+            198,
+            554,
+            210
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "yair.lakretz@gmail.com",
+        "bbox": [
+            396,
+            214,
+            600,
+            227
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jean-Rémi King*",
+        "bbox": [
+            677,
+            139,
+            826,
+            155
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Meta AI Paris;",
+        "bbox": [
+            702,
+            158,
+            794,
+            168
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Laboratoire des systèmes perceptifs",
+        "bbox": [
+            638,
+            174,
+            858,
+            187
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ecole normale supérieure",
+        "bbox": [
+            670,
+            187,
+            828,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "PSL University",
+        "bbox": [
+            700,
+            198,
+            796,
+            210
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "jeanremi@meta.com",
+        "bbox": [
+            672,
+            215,
+            828,
+            227
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            341,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "During language acquisition, children follow a typical sequence of learning stages, whereby they first learn to categorize phonemes before they develop their lexicon and eventually master increasingly complex syntactic structures. However, the computational principles that lead to this learning trajectory remain largely unknown. To investigate this, we here compare the learning trajectories of deep language models to those of children. Specifically, we test whether, during its training, GPT-2 exhibits stages of language acquisition comparable to those observed in children aged between 18 months and 6 years. For this, we train 48 GPT-2 models from scratch and evaluate their syntactic and semantic abilities at each training step, using 96 probes curated from the BLiMP, Zorro and BIG-Bench benchmarks. We then compare these evaluations with the behavior of 54 children during language production. Our analyses reveal three main findings. First, similarly to children, the language models tend to learn linguistic skills in a systematic order. Second, this learning scheme is parallel: the language tasks that are learned last improve from the very first training steps. Third, some – but not all – learning stages are shared between children and these language models. Overall, these results shed new light on the principles of language acquisition, and highlight important divergences in how humans and modern algorithms learn to process natural language.",
+        "bbox": [
+            141,
+            279,
+            460,
+            734
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            746,
+            260,
+            762
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Language acquisition is marked by a series of successive stages (Dupoux, 2018; Kuhl, 2004; Werker, 2018). Within their first year of existence, humans infants successively acquire prosody contours (Mehler et al., 1988), phonetic categories (Werker and Tees, 1984; Kuhl et al., 1992; Mazuka et al., 2011) and frequent words (Tincoff and Jusczyk, 1999; Bergelson and Swingley, 2012).",
+        "bbox": [
+            112,
+            769,
+            490,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "They then learn to produce basic syntactic structures (e.g. \"The boy sang\" or \"The boy fell\"), questions (e.g. \"What sound does a cow make?\") and nested syntactic structures (e.g. \"The boy that I saw sang\"), at approximately 12, 30, and 42 months, respectively (Friedmann et al., 2021). Even though some children may take slightly longer to learn than others, there is a set order in which children acquire various syntactic structures (Friedmann and Reznick, 2021).",
+        "bbox": [
+            507,
+            253,
+            884,
+            413
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Our understanding of the entire learning trajectory of children remains very coarse, however. This partly stems from the difficulty of measuring linguistic skills in young children. In babies, experimenters typically measure eye gaze and sucking rate while children process linguistic stimuli, as these reflexive behaviors are known to increase during surprising events. Such \"implicit\" approaches have successfully been used to assess whether nonspeaking infants detect linguistic violations (Zamuner, 2006), distinguish lexical from grammatical words (Shi et al., 1999) or discriminate their native language from a foreign language (Mehler et al., 1988; Kuhl et al., 2006; Nazzi et al., 2000). In older children, linguistic skills can also be more explicitly measured from spontaneous speech and sentence repetition. For example, a recent study by Friedmann et al. (2021), to which we compare our work in this paper, quantified the extent to which 18 month to 6 year-old children produce variably complex syntactic structures. For both of these approaches, however, the measures from children at such early ages can be noisy and fragmented.",
+        "bbox": [
+            507,
+            417,
+            884,
+            788
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Interestingly, these issues do not apply to modern language models. Deep learning architectures trained to predict words from their proximal contexts have proved immensely effective at learning to process natural language (Radford et al., 2019; Devlin et al., 2019). Unlike humans, these algorithms can be easily probed during training, at any time point and rate, and with unlimited number of",
+        "bbox": [
+            507,
+            790,
+            885,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal Contribution",
+        "bbox": [
+            141,
+            904,
+            270,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12205",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 12205-12218",
+        "bbox": [
+            220,
+            945,
+            778,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            959,
+            700,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "test stimuli, without interfering with their language acquisition (Jawahar et al., 2019; Manning et al., 2020; Bowman and Dahl, 2021). Furthermore, high-performing deep nets have been shown to implicitly (Lakretz et al., 2019; Gulordava et al., 2018) or explicitly learn to represent and use syntactic structures (Manning et al., 2020), as well as to use features such as concreteness and lexical class to learn language (Chang and Bergen, 2022). Finally, and importantly, these deep neural networks have recently been shown to represent lexical, syntactic and compositional representations similarly to the adult brain (Jain and Huth, 2018; Toneva and Wehbe, 2019; Caucheteux and King, 2022; Pasquiou et al., 2022, 2023; Caucheteux et al., 2023). Evidencing similar learning trajectories in children and language models could thus provide an invaluable framework to better understand the computational principles underlying language acquisition.",
+        "bbox": [
+            110,
+            84,
+            492,
+            390
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Here, we compare the trajectory of language acquisition between human children and modern language models. We focus on three main questions. First, do these models learn linguistic skills in a systematic order? Second, is this trajectory sequential or parallel? Third, is this trajectory similar to that of children? These hypotheses are illustrated in Figure 1.",
+        "bbox": [
+            112,
+            391,
+            489,
+            518
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we train 48 GPT-2 architectures (Radford et al., 2019) from scratch, using a standard next-word prediction objective. We then evaluate, at each training step, their linguistic abilities with 96 semantic and syntactic probes curated from the BLiMP, Zorro and BIG-Bench benchmarks (Warstadt et al., 2020; Huebner et al., 2021; Srivastava et al., 2022). Finally, we compare a subset of these probes to the behavior of 54 children aged, between 18 months and 6 years (Friedmann et al., 2021).",
+        "bbox": [
+            112,
+            519,
+            489,
+            697
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Approach",
+        "text_level": 1,
+        "bbox": [
+            112,
+            703,
+            235,
+            720
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Language models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            727,
+            297,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We consider two main language models. First, we use a pretrained language model – GPT-2 – as provided by HuggingFace and pretrained on 40 GB of data (Radford et al., 2019). Second, we separately train 48 versions of a 12-layer GPT-2 model from scratch. We train each model on WikiText103 (Merit et al., 2016) with a distinct random seed to set its initial parameters and data-loader. Each model is evaluated on all linguistic probes every",
+        "bbox": [
+            112,
+            749,
+            489,
+            894
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "100 training steps. Further training details are provided in Appendix B.",
+        "bbox": [
+            507,
+            84,
+            884,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Zero-shot linguistic probes",
+        "text_level": 1,
+        "bbox": [
+            507,
+            123,
+            764,
+            139
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Zero-shot linguistic probes are sentences or phrases crafted to evaluate whether a model has learned a particular linguistic skill, without training or finetuning the model on that particular skill. In practice, a zero-shot probe consists of comparing the estimated probability of a grammatical sentence with that of a matched ungrammatical sentence. This two-alternative forced-choice approach can be compared to \"acceptability judgements\", classically used in linguistics (Warstadt et al., 2019).",
+        "bbox": [
+            505,
+            146,
+            884,
+            306
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our models on 96 different linguistic probes, curated from three open source benchmarks, the details of which are presented in Appendix C.",
+        "bbox": [
+            507,
+            307,
+            884,
+            370
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we compare the probability of each sentence in a grammatical/ungrammatical pair by evaluating the sum of the logarithm of the loss output by the softmax layer:",
+        "bbox": [
+            507,
+            370,
+            882,
+            436
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\sum_ {i = 0} ^ {n _ {g}} \\log (f (X _ {g}) _ {i}) <   \\sum_ {j = 0} ^ {n _ {u}} \\log (f (X _ {u}) _ {j}) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            556,
+            463,
+            880,
+            506
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "with  $f$  the softmax layer of the language model,  $X_{g}$  and  $X_{u}$  the grammatical and ungrammatical sentences, respectively, and  $n_g$  and  $n_u$ , the number of tokens in the grammatical and ungrammatical sentences, respectively.",
+        "bbox": [
+            507,
+            519,
+            882,
+            599
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The accuracy of a given probe is the percentage of pairs where the estimated probability of the grammatical sentence is higher than that of the ungrammatical sentence.",
+        "bbox": [
+            507,
+            600,
+            882,
+            663
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 Assessing learning trajectory",
+        "text_level": 1,
+        "bbox": [
+            507,
+            671,
+            781,
+            686
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To evaluate whether the trajectory of language acquisition is shared across models, we rank the probes by their \"acquisition time\", i.e. the number of steps taken by a model to reach  $90\\%$  of its final accuracy on a particular probe, for each model independently. We then assess the correlation of ranks between all pairs of the 48 models and take the average of these correlations. To estimate the statistical significance of this average correlation we redo this calculation for all possible model pairs after shuffling the ranks of one of the models in each pair. We repeat this permutation 1,000 times, getting 1,000 values for this shuffled correlation. If in all cases this shuffled correlation is lower than",
+        "bbox": [
+            505,
+            694,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>https://huggingface.co/gpt2",
+        "bbox": [
+            136,
+            903,
+            310,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12206",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the true average correlation, then the order of acquisition time is shared across models with  $p < 0.001$ .",
+        "bbox": [
+            112,
+            84,
+            489,
+            131
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.4 Parallel versus Sequential learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            139,
+            433,
+            155
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Language acquisition may be characterized by a \"sequential\" or a \"parallel\" learning scheme (Figure 1). \"Sequential\" learning designates the case where a complex skill does not start to be learned before simpler skills are mastered. By contrast, \"Parallel\" learning designates the case where all skills are acquired simultaneously, but at different speeds. The null hypothesis is that the order in which an agent learns linguistic skills varies across agents. To determine the learning scheme of language models, we consider whether the probes have a positive derivative in the first three checkpoints (parallel learning) or not (sequential learning), and whether they have statistically different learning rates (by performing a one-way ANOVA test) across the three groups.",
+        "bbox": [
+            112,
+            162,
+            489,
+            420
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.5 Assessing linguistic skill from children's behavior",
+        "text_level": 1,
+        "bbox": [
+            112,
+            426,
+            473,
+            456
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Friedmann et al. (2021) studied 54 Hebrew-speaking children between the ages of 18 - 71 months and investigated the emergence of 11 linguistic phenomena, which the authors propose to organize into three stages (details in Appendix A). For our analysis we select the following tests, one from each stage:",
+        "bbox": [
+            112,
+            464,
+            489,
+            577
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Stage 1: Simple sentences in subject-verb (SV) order",
+            "- Stage 2: Wh Questions",
+            "- Stage 3: Relative Clauses"
+        ],
+        "bbox": [
+            136,
+            590,
+            485,
+            680
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Data collection consisted of spontaneous speech samples produced by each child at home. Each sample was then manually annotated to detect the presence of each of the linguistic phenomena. A linguistic phenomenon was considered learned if and only if it was present in the speech sample. Speech samples had a mean length of 151 utterances per sample and standard deviation of 37. The aggregated data was made available directly in the original paper (under Creative Commons Attribution 4.0 International License), and here used for comparison with our language models. In Table 1 we show which probes in the models matched with these tests.",
+        "bbox": [
+            112,
+            694,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Results",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            608,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We aim to compare the learning trajectories of deep language models to those observed in 54 children aged between 18 months and 6 years. For this, we trained variants of GPT-2 models (Radford et al., 2019) from 48 different random seeds with the WikiText103 dataset (Merit et al., 2016) and evaluated each model on 96 linguistic probes every 100 steps.",
+        "bbox": [
+            507,
+            105,
+            884,
+            233
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "At the end of this training, 64 probes (66%) were achieved above chance level (50% accuracy) by all models. In comparison, a pretrained version of GPT-2 large (Radford et al., 2019) provided by Hugging Face $^2$ , and trained on a much larger dataset $^3$ , achieves above-chance performance on 93 of the 96 probes.",
+        "bbox": [
+            507,
+            235,
+            884,
+            347
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 A systematic learning trajectory",
+        "text_level": 1,
+        "bbox": [
+            507,
+            353,
+            808,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For clarity, we focus on the learning dynamics of the probes that ultimately achieve above-chance performance in our training. Figure 2 lists all probes learned above chance level, ordered by their average acquisition time. We perform the permutation analysis outlined in 2.3, to evaluate whether the order of acquisition is shared between models, and find that their order of acquisition is correlated with  $R = 0.743$  and  $p < 0.001$ . These results suggest that there is a systematic learning trajectory among models.",
+        "bbox": [
+            507,
+            375,
+            884,
+            551
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Learning is parallel across linguistic tasks",
+        "text_level": 1,
+        "bbox": [
+            507,
+            557,
+            880,
+            571
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Are these linguistic skills learned sequentially or in parallel (Figure 1)? To address this question, we evaluate whether each linguistic probe starts to improve from the very first training steps but with different rates (i.e. a \"parallel\" learning scheme) or, on the contrary, whether some probes only start to improve once others have reached a particular performance (i.e. a \"sequential\" learning scheme). As the individual learning trajectories of each probe were noisy, we group the 64 linguistic probes into three categories: early, middle and late acquisition time (Figure 3).",
+        "bbox": [
+            507,
+            577,
+            882,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Overall, we observe parallel learning between the three groups: their performances all increase from the beginning of training:  $95\\%$  of tests in all three groups have a positive derivative within the first three hundred steps. However, they have different learning rates, as evaluated with a one-way ANOVA test on the learning rate (i.e. change",
+        "bbox": [
+            507,
+            772,
+            882,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://huggingface.co/tftransformers/gpt2-large",
+        "bbox": [
+            529,
+            890,
+            831,
+            904
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "${}^{3}{40}\\mathrm{{GB}}$  compared to the  ${181}\\mathrm{{MB}}$  of WikiText103",
+        "bbox": [
+            529,
+            904,
+            833,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12207",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d0dea45aa26fe24dd3b78ce9112237b6059833c376ac00595f37540c4834e8e0.jpg",
+        "image_caption": [
+            "Figure 1: Hypotheses. Skill performance (y-axis) as a function of training (x-axis) illustrated on three tasks (colors) and two agents (model or children). Sequential learning implies that the learning of a complex skill (e.g. C, shown in red) does not start before simplest skills (e.g. A in green and B in blue) are fully learned. Parallel learning implies that all skills are acquired simultaneously, but at different speeds. Sequential and parallel learning may cross an arbitrary performance threshold at the same training step. The frequency at which we can probe artificial networks during learning is much greater than what is realistically possible in children, giving us the timescale granularity to distinguish sequential and parallel learning trajectories. We also present a null hypothesis, that artificial networks with different random seeds may learn skills in a different order."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            84,
+            378,
+            230
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f27c71387747471f90818c4ef27dfce6ff3da95f965a0e7c9e24ba7e83f2d63e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            83,
+            630,
+            230
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/35d1246d79b6f047f0f4d3fb0e916b43e5ea80c505b8139ad91a8c9b5598a3b4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            630,
+            83,
+            880,
+            225
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/9da4eafe1260db2bc5382933dd3306e9f3a9477c453ba70fb06d01fc73f8eccf.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Stage</td><td>Children</td><td>Language Model</td></tr><tr><td>1</td><td>Simple sentences in Subject-Verb (SV) order</td><td>SV agreement across simple sentences</td></tr><tr><td>2</td><td>Wh-questions</td><td>SV agreement in questions</td></tr><tr><td>3</td><td>Relative Clauses (RCs)</td><td>SV agreement across object RCs</td></tr></table>",
+        "bbox": [
+            139,
+            370,
+            858,
+            441
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Linguistic phenomena in children selected for comparison with probes in the language models.",
+        "bbox": [
+            144,
+            449,
+            850,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "of accuracy over time) obtained in each group and in each model  $(p < 10^{-23})$ .",
+        "bbox": [
+            112,
+            489,
+            487,
+            521
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Comparison with children",
+        "text_level": 1,
+        "bbox": [
+            112,
+            529,
+            368,
+            545
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Do these learning trajectories match the behavior of human children? For the three probes that correspond to the three stages identified in children's language acquisition (Table 1), we observe that the order in which these three probes are learned by the language models is the same as those of children (Figure 4). This effect is robust across random seeds: 46 of our 48 GPT-2 models follows this order, where chance level is  $(1/3!)^{46} = 1.60e^{-36}$ . For this subset of language abilities, models and children thus seem to acquire syntactic skills in a similar order.",
+        "bbox": [
+            112,
+            551,
+            489,
+            743
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Learning of theoretically-defined stages is parallel",
+        "text_level": 1,
+        "bbox": [
+            112,
+            751,
+            487,
+            784
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In part 3.1, we showed that GPT-2 learns its language abilities in parallel. Does this learning scheme also characterize the three syntactic skills investigated in children? To address this question, we now look at the learning curves of the skills defined in Table 1, as well as an additional probe: Nounpp, as it can be separated into congruent and incongruent cases which is important for the anal",
+        "bbox": [
+            112,
+            790,
+            490,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ysis in section 3.5. Overall, we observe that these probes are all learned in parallel in the model (Figure 5A).",
+        "bbox": [
+            507,
+            489,
+            884,
+            537
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.5 Models use both syntax and heuristics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            552,
+            853,
+            569
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Both an understanding of syntactic rules and a superficial heuristics can lead to above-chance performances on these probes. Indeed, in many sentences (e.g. The cat [subject] of the lady [attractor] is [verb] hungry.), the number of the verb is congruent with the number of the adjacent attractor, even if the two are not related syntactically. To verify that the GPT-2 models effectively learn the syntactic rules, we thus separately examine congruent and incongruent cases. Incongruent cases require knowledge of the syntax of the sentence as the correct verb number is different from the number of the attractor. Empirically, we observe that the models do not learn the incongruent case in stage three above chance level, and just barely reach chance level on the incongruent case in stage two (Figure 5B), indicating that our models are using heuristics rather than syntactic rules to achieve high accuracy on the congruent cases (leading to above chance performance on the probe overall in Figure 5A). On the contrary, the pretrained GPT",
+        "bbox": [
+            505,
+            581,
+            885,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12208",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/b010ed78004a7b83855311dec96637769c4214e3054ba0dba02e8742b0e7c893.jpg",
+        "image_caption": [
+            "Figure 2: The performance of the models on each linguistic probe over time is smoothed using a moving average filter with window size of 6 checkpoints then the number of steps required to reach  $90\\%$  of final accuracy (acquisition time) is calculated. Probes are ordered by increasing average acquisition time. Results shown for 48 models. Only probes which have final accuracy greater than chance  $(50\\%)$  are shown. This demonstrates that probes tend to be learned in the same order by all agents with  $\\mathrm{R} = 0.743$ ,  $\\mathrm{p} < 0.001$ , disproving the null hypothesis."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            83,
+            880,
+            694
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2 large achieves above  $75\\%$  accuracy also on the incongruent cases of these probes. Thus for the models trained on the WikiText103, syntax is only learned for stages one and two, and heuristics seem to explain the above chance accuracy in stage three. A larger training dataset is required to learn syntax and not only heuristics for the most difficult",
+        "bbox": [
+            112,
+            804,
+            490,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "examples.",
+        "bbox": [
+            509,
+            804,
+            588,
+            820
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.6 Impact of number biases in congruent and incongruent sentences",
+        "text_level": 1,
+        "bbox": [
+            509,
+            829,
+            880,
+            862
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In previous work, it was found that a variety of language models have a bias towards plural English verbs, and several studies (Jumelet et al., 2019;",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12209",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/2a909119396e6065cd8fb6ecc488cab23f1cbcbd9896e111eccb313d3bf103b9.jpg",
+        "image_caption": [
+            "Figure 3: Linguistic probes grouped into 3 groups to avoid plotting one line per probe. The probes in each group correspond to the colours in Figure 2. Probes which have final accuracy less than chance (50%) are placed in their own group, and tend to zero likely due to biases towards plural verbs in English (c.f. 3.6). Shading is standard error of the mean across probes in the group. This figure demonstrates that linguistic skills are learned in parallel not in sequence."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            84,
+            480,
+            300
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Lakretz et al., 2021a,b) determined that LSTM-based models have a default number and gender prediction preference. To examine whether number bias has a significant effect on our analysis, we compare congruent sentences with only singular or only plural verbs and incongruent sentences with a plural or a singular verb. Accuracy on predicting plural verbs increases sharply from the start of the training and then drops. By contrast, the accuracy of singular cases first drops and then rises (Figure 5C), indicating that the models are biased towards plural verbs at the beginning of training. This bias is overcome for the stage one probe but for stage two and three it remains throughout training. This explains the initial downward trend in Group 3 and why the unlearned probes tend toward zero in Figure 3.",
+        "bbox": [
+            112,
+            469,
+            489,
+            743
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            750,
+            240,
+            766
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The stages followed by children to acquire language has been the topic of intense research (Dupoux, 2018; Kuhl, 2004; Werker, 2018). While this learning trajectory is becoming clearer for sublexical representations (Dupoux, 2018), the acquisition of higher-level syntactic and semantic processing remains largely unclear. Here, we approach this long-lasting question through the lens of a deep language architecture, GPT-2 (Radford et al., 2019),",
+        "bbox": [
+            112,
+            774,
+            490,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "to test whether this model follows a learning trajectory similar to children.",
+        "bbox": [
+            507,
+            84,
+            884,
+            118
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Language acquisition: similarities and differences between humans and GPT-2",
+        "text_level": 1,
+        "bbox": [
+            507,
+            131,
+            868,
+            162
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "First, we show that GPT-2 models tend to learn a battery of linguistic phenomena (Warstadt et al., 2020; Lakretz et al., 2019; Huebner et al., 2021) in a consistent order. It is the reliability of the acquisition trajectory that allows a direct comparison with the learning trajectory of children (Friedmann et al., 2021). However, this consistency in GPT-2 models may result from two non-mutually exclusive factors, that remain to be disentangled: either the acquisition time of each linguistic phenomenon relates to its frequency in natural language (e.g. Simple subject-verb-complement are more frequent in natural language than nested syntactic structures; Karlsson 2007), and/or it relates to their intrinsic complexity (e.g. sentences with nested structure require more operations to be composed than simple sentences). Future work systematically controlling for these relative frequencies is thus necessary to distinguish these two possibilities, and would build upon work by Weber et al. (2021) who found that less frequent linguistic phenomena can be learned from fewer examples, though later in training.",
+        "bbox": [
+            507,
+            174,
+            884,
+            529
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Second, we show that the order in which linguistic skills are acquired is similar between children and GPT-2 – at least on the syntactic phenomena that were evaluated in these two cohorts, and with the limitation of using number agreement as a proxy to whether the models acquire the corresponding syntactic structure. Similarly to children, GPT-2 models master subject-verb agreement in SV sentences before they master it in questions, or across nested center-embedded clauses (object-relative clauses). This result thus complements a series of studies comparing modern language models and humans. For example, a recent study showed that transformers trained on child-directed data can achieve comparable accuracy on linguistic probes to large pre-trained models (Huebner et al., 2021). Similarly, several studies have recently shown that the representations of GPT-2 become increasingly similar to those of the adult human brain during its training (Caucheteux and King, 2022). Finally, Lavechin et al. (2022) showed that models trained on audio in a self-supervised fashion learn phoneme and lexical abilities in a similar trajectory to children.",
+        "bbox": [
+            507,
+            533,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12210",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/b52612123aa11f556fd821244d49acef99ea1664d766992b21fa9b4f64e0feca.jpg",
+        "image_caption": [
+            "Figure 4: Comparing language model performance on linguistic probes to children's performance. Example sentences observed in children were originally in Hebrew (Friedmann et al., 2021). Non-white indicates the phenomena is learned by the agent. The threshold for considering a probe learned by the model is performance above  $55\\%$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            84,
+            880,
+            181
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/781b658bfc78f7ed0cfd5ca39e7257e9e00e41fa315e3713a7e983d9647b09f4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            265,
+            485,
+            458
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/da214a664f165da850401a424511f74092afc15d53a881a9a745c6cfcbe30486.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            273,
+            875,
+            456
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f69bb2d9586d1e9e2088e0e72e3684b35d63a5659bce91acb74b82ef3c4f313.jpg",
+        "image_caption": [
+            "Figure 5: Subject verb agreement. Shading is standard error of the mean across seeds. S: Singular. P: Plural. A. Parallel learning is observed in the three stages defined by Friedmann et al. (2021), when results are averaged across congruent and incongruent cases. B. Subject verb agreement on incongruent cases which indicate whether the model understands syntax. The networks do not learn some syntax structures as the incongruent case of Short Nested Outer does not reach above chance level. C. Development trajectories of the bias towards plural number in English."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            464,
+            371,
+            640
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9f2fe0c53b9a8f161d38e2674a786a7947db551734aa4365370e19ba64859ed.jpg",
+        "image_caption": [
+            "C"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            376,
+            480,
+            615,
+            642
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9a08e1eaf97814dbb46825830657e249397e179eab8e5ca1883d404625c23eef.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            618,
+            464,
+            875,
+            642
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 A work in progress",
+        "text_level": 1,
+        "bbox": [
+            112,
+            774,
+            310,
+            790
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "It is important to stress that significant work remains to be done before drawing any definitive conclusions about the similarities between language acquisition in humans and algorithms.",
+        "bbox": [
+            112,
+            802,
+            487,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "First, we only consider a single architecture (GPT-2, Radford et al. (2019)) with a unique textual corpus (WikiText103). Testing whether our",
+        "bbox": [
+            112,
+            871,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "results hold true independently of the model and training corpus remains an important milestone for future research.",
+        "bbox": [
+            507,
+            774,
+            882,
+            820
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Second, linguistic abilities are not tested with the same protocols in children and in the models: the models are explicitly tested on next word prediction, with a two-alternative forced-choice metric, whereas children were implicitly evaluated on their ability to spontaneously use specific syntactic struc",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12211",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "tures during natural speech.",
+        "bbox": [
+            112,
+            84,
+            321,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Third, there were only three linguistic features that were directly comparable between the model probes and the data in children, and all were syntactic. This leaves a significant margin of progress to modulate our conclusion, and investigate whether the lexicon, narrative structures, pragmatics and world-knowledge are acquired in the same order in humans and algorithms.",
+        "bbox": [
+            112,
+            101,
+            487,
+            228
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Fourth, the order in which some linguistic skills were learned by GPT-2 does not trivially fit with linguistic theory. For example, the probe \"Simple\", which examines subject-verb agreement in a simple sentence, was one of the last probes to be learned by GPT-2 (it is part of group three in Figure 2). By contrast, \"Wh Questions Subject Gap Long Distance\" was among the first probes to be to be learned, even though it would be expected to be much harder than \"Simple\". This unexpected result may be due to the way we approximate \"Acquisition Time\", namely, the moment when the probes reaches  $90\\%$  of the final accuracy. Consequently, probes with very low final accuracy could end up with a shorter Acquisition Time, because noise may lead to crossing the  $90\\%$  threshold relatively quickly.",
+        "bbox": [
+            115,
+            230,
+            489,
+            502
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Finally, we show that our models appear to use heuristics rather than a deep understanding of syntax for the most difficult linguistic probes (incongruent numbers between verbs and their attractors) and were biased towards plural English verbs. While our models learn only  $66\\%$  of tasks to above chance level, a larger GPT-2 pretrained on considerably more texts successfully perform on  $97\\%$  of the tasks, and has an accuracy above  $75\\%$  on the incongruent examples, meaning this bias and reliance on heuristics could potentially be solved by training on a larger dataset.",
+        "bbox": [
+            112,
+            505,
+            489,
+            697
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In sum, additional work remains necessary to identify the exact elements of convergence and divergence between the acquisition of language in models and in children.",
+        "bbox": [
+            112,
+            697,
+            489,
+            760
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3 Fueling the debate between nativism versus empiricism",
+        "text_level": 1,
+        "bbox": [
+            112,
+            768,
+            448,
+            800
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The present study fuels a long-lasting debate on the acquisition of language. While \"empiricists\" argue that language can be acquired with a statistical approach (Clark, 2002; Kolodny et al., 2015; Chater and Christiansen, 2018; McCauley and Christiansen, 2019), \"nativists\" maintain that this ability depends on a core and innate operation, spe",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "cific to humans (Chomsky, 1959, 1971).",
+        "bbox": [
+            507,
+            84,
+            808,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The present study shows how modern language models may contribute to resolving this debate, by systematically studying which components of a model (e.g. architecture) or properties of the training data (e.g., frequency of sentence structures) contribute to shape the trajectory of language acquisition. Claims about an innate Universal Grammar could be understood as an inductive bias of a language model, implemented in its architecture and dynamics, which tightly constrains learning trajectories across models. If this bias is hierarchical (rather than linear) then this could lead to learning trajectories that follow the structure of the syntactic tree, consistently with the hypothesis of three linguistic stages presented by Friedmann et al. (2021) in humans and what we find in this study in language models. Indeed, neural language models have been previously shown to have a weak inductive bias towards hierarchical processing (McCoy et al., 2020; Kharitonov and Chaabouni, 2020), which can partially explain our results.",
+        "bbox": [
+            507,
+            101,
+            884,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This result echos the recent observation that syntactic trees spontaneously emerge in the middle layers of neural language models (Hewitt and Manning, 2019). Together, these elements thus suggest that modern neural networks provide fruitful models of language acquisition and could reconcile or settle the confronting theories of language acquisition (Warstadt and Bowman, 2022).",
+        "bbox": [
+            507,
+            439,
+            884,
+            565
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            573,
+            643,
+            587
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Overall, the similarities identified between children and GPT-2 suggest that there may be a small set of means by which to efficiently acquire language. This result is anything but trivial: humans and deep neural networks have extraordinarily different architectures, training, and language exposure. If generalized, this systematic learning trajectory would support the existence of an intrinsic hierarchy of linguistic structures that both machines and humans must climb, be that through inductive biases or properties of the training data, to master the faculty of language. And while these hypotheses remain open, the path to resolve them has never been clearer.",
+        "bbox": [
+            507,
+            594,
+            884,
+            818
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12212",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            285,
+            101
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We would like to thank Dieuwke Hupkes, Naama Friedmann, Marco Baroni and the attendees of the EviL meetings for their comments and suggestions.",
+        "bbox": [
+            112,
+            105,
+            489,
+            153
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This project has received funding from the European Union's Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 945304, for L.E for her work at PSL. This work was funded in part by FrontCog grant ANR-17-EURE-0017 for the work of L.E. and J.R.K. for their work at PSL.",
+        "bbox": [
+            112,
+            155,
+            490,
+            267
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            288,
+            213,
+            304
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Elika Bergelson and Daniel Swingley. 2012. At 6-9 months, human infants know the meanings of many common nouns. Proceedings of the National Academy of Sciences, 109:3253-3258.",
+            "Samuel R. Bowman and George Dahl. 2021. What will it take to fix benchmarking in natural language understanding? Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4843-4855.",
+            "Charlotte Caucheteux, Alexandre Gramfort, and Jean-Rémi King. 2023. Evidence of a predictive coding hierarchy in the human brain listening to speech. Nature Human Behaviour, pages 1-12.",
+            "Charlotte Caucheteux and Jean Rémi King. 2022. Brains and algorithms partially converge in natural language processing. Communications Biology, 5(1).",
+            "Tyler A. Chang and Benjamin K. Bergen. 2022. Word acquisition in neural language models. Transactions of the Association for Computational Linguistics, 10:1-16.",
+            "Nick Chater and Morten H Christiansen. 2018. Language acquisition as skill learning. Current opinion in behavioral sciences, 21:205-208.",
+            "Noam Chomsky. 1959. Review of verbal behavior. 35(1):26-58. Publisher: Linguistic Society of America.",
+            "Noam Chomsky. 1971. *Problems of Knowledge and Freedom*. New York,: W.W. Norton.",
+            "Alexander Clark. 2002. Unsupervised language acquisition: Theory and practice. arXiv preprint cs/0212024.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova Google, and AI Language. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186."
+        ],
+        "bbox": [
+            115,
+            309,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Emmanuel Dupoux. 2018. Cognitive science in the era of artificial intelligence: A roadmap for reverse-engineering the infant language-learner. Cognition, 173:43-59.",
+            "Naama Friedmann, Adriana Belletti, and Luigi Rizzi. 2021. Growing trees: The acquisition of the left periphery. Glossa: a journal of general linguistics, 39(1).",
+            "Naama Friedmann and Julia Reznick. 2021. Stages rather than ages in the acquisition of movement structures: Data from sentence repetition and 27696 spontaneous clauses. Glossa: a journal of general linguistics, 39(1).",
+            "Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195-1205, New Orleans, Louisiana. Association for Computational Linguistics.",
+            "John Hewitt and Christopher D Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129-4138.",
+            "Philip Huebner. 2022. Unmasked. https://github.com/phueb/UnMasked. (Accessed 2023/05/24).",
+            "Philip A Huebner, Elior Sulem, Cynthia Fisher, and Dan Roth. 2021. BabyBERTa : Learning More Grammar With Small-Scale Child-Directed Language. Proceedings of the 25th Conference on Computational Natural Language Learning, pages 624-646.",
+            "Shailee Jain and Alexander Huth. 2018. Incorporating context into language encoding models for fmri. In Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc.",
+            "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does bert learn about the structure of language? Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651-3657.",
+            "Jaap Jumelet, Willem Zuidema, and Dieuwke Hupkes. 2019. Analysing neural language models: Contextual decomposition reveals default reasoning in number and gender assignment. Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 1-11.",
+            "Fred Karlsson. 2007. Constraints on multiple center-embedding of clauses. Journal of Linguistics, 43(2):365-392."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12213",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Eugene Kharitonov and Rahma Chaabouni. 2020. What they do when in doubt: a study of inductive biases in seq2seq learners. arXiv:2006.14953.",
+            "Oren Kolodny, Arnon Lotem, and Shimon Edelman. 2015. Learning a generative probabilistic grammar of experience: A process-level model of language acquisition. Cognitive Science, 39(2):227-267.",
+            "Patricia K. Kuhl. 2004. Early language acquisition: cracking the speech code. Nature Reviews Neuroscience, 5:831-843.",
+            "Patricia K Kuhl, Erica Stevens, Akiko Hayashi, Toshisada Deguchi, Shigeru Kiritani, and Paul Iverson. 2006. Fast-track report infants show a facilitation effect for native language phonetic perception between 6 and 12 months. Developmental Science, 9:13-21.",
+            "Patricia K. Kuhl, Karen A. Williams, Francisco Lacerda, Kenneth N. Stevens, and Bjorn Lindblom. 1992. Linguistic experience alters phonetic perception in infants by 6 months of age. Science, 255:606-608.",
+            "Yair Lakretz, Théo Desbordes, Dieuwke Hupkes, and Stanislas Dehaene. 2021a. Causal transformers perform below chance on recursive nested constructions, unlike humans. arXiv:2110.07240.",
+            "Yair Lakretz, Dieuwke Hupkes, Alessandra Vergallito, Marco Marelli, Marco Baroni, and Stanislas Dehaene. 2021b. Mechanisms for handling nested dependencies in neural-network language models and humans. Cognition, 213:104699. Special Issue in Honour of Jacques Mehler, Cognition's founding editor.",
+            "Yair Lakretz, German Kruszewski, Theo Desbordes, Dieuwke Hupkes, Stanislas Dehaene, and Marco Baroni. 2019. The emergence of number and syntax units in LSTM language models. Association for Computational Linguistics, pages 11-20.",
+            "Marvin Lavechin, Maureen De Seyssel, Hadrien Titeux, Hervé Bredin, Guillaume Wisniewski, Alejandrina Cristia, and Emmanuel Dupoux. 2022. Statistical learning bootstraps early language acquisition. PsyArXiv.",
+            "Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg. 2016. Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies. Transactions of the Association for Computational Linguistics, 4:521-535.",
+            "Christopher D. Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, and Omer Levy. 2020. Emergent linguistic structure in artificial neural networks trained by self-supervision. Proceedings of the National Academy of Sciences of the United States of America, 117:30046-30054.",
+            "Reiko Mazuka, Yvonne Cao, Emmanuel Dupoux, and Anne Christophe. 2011. The development of a phonological illusion: A cross-linguistic study with japanese and french infants. Developmental Science, 14:693-699."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            916
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Stewart M McCauley and Morten H Christiansen. 2019. Language learning as language use: A cross-linguistic model of child language development. *Psychological review*, 126(1):1.",
+            "R Thomas McCoy, Robert Frank, and Tal Linzen. 2020. Does syntax need to grow on trees? sources of hierarchical inductive bias in sequence-to-sequence networks. Transactions of the Association for Computational Linguistics, 8:125-140.",
+            "Jacques Mehler, Peter Jusczyk, Ghislaine Lambertz, Nilofar Halsted, Josiane Bertoncini, and Claudine Amiel-Tison. 1988. A precursor of language acquisition in young infants. Cognition, 29:143-178.",
+            "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models. arXiv:1609.07843.",
+            "Thierry Nazzi, Peter W. Jusczyk, and Elizabeth K. Johnson. 2000. Language discrimination by english-learning 5-month-olds: Effects of rhythm and familiarity. Journal of Memory and Language, 43:1-19.",
+            "Alexandre Pasquiou, Yair Lakretz, John Hale, Bertrand Thirion, and Christophe Pallier. 2022. Neural language models are not born equal to fit brain data, but training helps. In ICML 2022-39th International Conference on Machine Learning, page 18.",
+            "Alexandre Pasquiou, Yair Lakretz, Bertrand Thirion, and Christophe Pallier. 2023. Information-restricted neural language models reveal different brain regions' sensitivity to semantics, syntax and context. arXiv:2302.14389.",
+            "Yada Pruksachatkun, Phil Yeres, Haokun Liu, Jason Phang, Phu Mon Htut, Alex Wang, Ian Tenney, and Samuel R. Bowman. 2020. jiant: A software toolkit for research on general-purpose text understanding models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 109-117, Online. Association for Computational Linguistics.",
+            "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. Semantic Scholar. (Accessed 2023-05-04).",
+            "Rushen Shi, Janet F Werker, and James L Morgan. 1999. Newborn infants' sensitivity to perceptual cues to lexical and grammatical words. Cognition, 72:B11-B21.",
+            "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv 2206.04615.",
+            "Ruth Tincoff and Peter W. Jusczyk. 1999. Some beginnings of word comprehension in 6-month-olds. Psychological Science, 10:172-175."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12214",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mariya Toneva and Leila Wehbe. 2019. Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain). In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.",
+            "Alex Warstadt and Samuel R Bowman. 2022. What artificial neural networks can tell us about human language acquisition. arXiv:2208.07998.",
+            "Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng Fu Wang, and Samuel R. Bowman. 2020. Erratum: \"blimp: The benchmark of linguistic minimal pairs for english\". Transactions of the Association for Computational Linguistics, 8:867-868.",
+            "Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625-641.",
+            "Lucas Weber, Jaap Jumelet, Elia Bruni, and Dieuwke Hupkes. 2021. Language modelling as a multi-task problem. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2049-2060, Online. Association for Computational Linguistics.",
+            "Janet F. Werker. 2018. Perceptual beginnings to language acquisition. Applied Psycholinguistics, 39:703-728.",
+            "Janet F. Werker and Richard C. Tees. 1984. Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. *Infant Behavior and Development*, 7:49–63.",
+            "Tania S. Zamuner. 2006. Sensitivity to word-final phonotactics in 9- to 16-month-old infants. Infancy, 10:77-95."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            605
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            115,
+            620,
+            203,
+            636
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Tests in Children",
+        "text_level": 1,
+        "bbox": [
+            115,
+            642,
+            302,
+            657
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Detailed description of tests available in children, in the three linguistic stages defined by (Friedmann et al., 2021):",
+        "bbox": [
+            112,
+            665,
+            487,
+            712
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Stage 1: Subject-Verb Simple, Subject-Verb Unaccusative, Verb-Subject Unaccusative",
+            "- Stage 2: Root WH-Argument, WH-Adjunct Excluding Why, Preposed Adverb, Root y/n",
+            "- Stage 3: Why, Relative Clause, Topicalisation, Embedding"
+        ],
+        "bbox": [
+            136,
+            725,
+            487,
+            843
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The probes chosen for comparison (stated in Table 1), were the only probes that matched well with one of the test available in children. In addition Nounpp was examined in the models, as it fits into",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "the linguistic stage 2, and, as it is part of the BIG-Bench probes, could be separated into congruent and incongruent sentences.",
+        "bbox": [
+            507,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Model Training",
+        "text_level": 1,
+        "bbox": [
+            509,
+            139,
+            684,
+            155
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "To evaluate linguistic abilities of a high-performance language model, we first use the HuggingFace pretrained GPT-2 large which has 774M parameters and is trained on 40GB of data. This model has one-shot perplexity of 22 on WikiText103 (Radford et al., 2019).",
+        "bbox": [
+            507,
+            162,
+            884,
+            256
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Then, to evaluate how linguistic abilities vary with language acquisition, we separately trained 48 models (each with a distinct random seed which set the model's initial parameters and the seed of the datalocator) using the 12-layer GPT-2 architecture (Radford et al., 2019) provided by HuggingFace on WikiText103 (Merit et al., 2016) with a learning rate of  $10^{-5}$  and a batch size of 16 distributed over 8 GPUs, making a total batch size of 64 and a context size of 128. Training was stopped when then validation loss plateaued, reaching final perplexity of 28 after 10 epochs. This is lower perplexity than the one-shot performance of the HuggingFace pretrained 12-layer GPT-2 which was 37.5, which is logical as our model was trained specifically on this dataset.",
+        "bbox": [
+            507,
+            258,
+            882,
+            514
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In all cases we used the pretrained tokenizer which has vocabulary size of 50,257. All other parameters were the default training arguments for the transformer provided by HuggingFace. The HuggingFace architectures are publicly available under an MIT license, and WikiText103 is available under Creative Commons Attribution-ShareAlike License.",
+        "bbox": [
+            507,
+            516,
+            882,
+            642
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Linguistic Probe Benchmarks",
+        "text_level": 1,
+        "bbox": [
+            507,
+            650,
+            805,
+            665
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We use three different zero-shot benchmarks. The first benchmark, 'BLiMP' (The Benchmark of Linguistic Minimal Pairs for English) (Warstadt et al., 2020) contains 67 different probes, each in the form of 1,000 pairs of grammatical and ungrammatical sentences designed to isolate a specific linguistic phenomenon. Adult human performance on BLiMP is  $96.4\\%$  (Warstadt et al., 2020). The second benchmark, 'Zorro'  $^{5}$ , was developed with a vocabulary frequent in child-directed corpora. Zorro contains 13 probes, each consisting of 2,000 pairs of sentences. Finally, the third benchmark is the Subject-Verb Agreement Task of BIG",
+        "bbox": [
+            507,
+            673,
+            884,
+            882
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "4 https://huggingface.co/gpt2",
+        "bbox": [
+            529,
+            891,
+            707,
+            904
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{5}$ https://github.com/phueb/Zorro",
+        "bbox": [
+            529,
+            904,
+            731,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "12215",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Bench (Srivastava et al., 2022; Lakretz et al., 2019, 2021b; Linzen et al., 2016; Gulordava et al., 2018). We focus on the syntactic probes, namely: \"Simple English\" which contains 600 pairs, \"NounPP\" which contains 2,400 pairs, and \"Short Nested Inner\", \"Short Nested Outer\", \"Long Nested Inner\" and \"Long Nested Outer\" which each contain 4,096 pairs of grammatical and ungrammatical sentences.",
+        "bbox": [
+            112,
+            84,
+            489,
+            212
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Accuracy on a linguistic probe is evaluated with the Jiant (Pruksachatkun et al., 2020) and UnMasked method (Huebner, 2022). In practice, sentences are input to the model in batches of 300, with padding on the left to make all sentences the length of the longest sentence in the batch. The logit values of punctuation are discarded when estimating the probability of a sentence.",
+        "bbox": [
+            112,
+            212,
+            489,
+            340
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Zorro, Jiant and UnMasked are publicly available under the MIT License, BLiMP under a CC BY License, and BIG-Bench under the Apache License 2.0.",
+        "bbox": [
+            112,
+            342,
+            489,
+            404
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "12216",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "text_level": 1,
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            127,
+            127,
+            532,
+            143
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            143,
+            231,
+            159
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            127,
+            170,
+            552,
+            186
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            187,
+            231,
+            200
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            127,
+            212,
+            695,
+            229
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            230,
+            231,
+            244
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "□ A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            127,
+            255,
+            668,
+            272
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "text_level": 1,
+        "bbox": [
+            114,
+            300,
+            490,
+            316
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            215,
+            336
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            127,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            363,
+            231,
+            379
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            127,
+            390,
+            778,
+            406
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            407,
+            231,
+            422
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            127,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            498,
+            231,
+            513
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            127,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            573,
+            231,
+            588
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            127,
+            599,
+            880,
+            631
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            632,
+            231,
+            646
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            127,
+            658,
+            880,
+            739
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            740,
+            231,
+            753
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "text_level": 1,
+        "bbox": [
+            114,
+            765,
+            495,
+            781
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            787,
+            215,
+            802
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            127,
+            813,
+            880,
+            845
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            846,
+            231,
+            860
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "12217",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank."
+        ],
+        "bbox": [
+            127,
+            84,
+            880,
+            282
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "text_level": 1,
+        "bbox": [
+            114,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank."
+        ],
+        "bbox": [
+            127,
+            340,
+            880,
+            640
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "12218",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/Language acquisition_ do children and language models follow similar learning stages_/6b7349d3-6e5d-4155-ae82-6dcafa1b591d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c55f47a8f538e1823176b0b3677bfdd40e0d1bc18e3940810c1996eb8e1f33c
+size 609174

2023/Language acquisition_ do children and language models follow similar learning stages_/full.md

@@ -0,0 +1,389 @@
+# Language acquisition: do children and language models follow similar learning stages?
+
+Linnea Evanson
+
+Meta AI Paris;
+
+Laboratoire des systèmes perceptifs
+
+Ecole normale supérieure
+
+PSL University
+
+linnea.evanson8@gmail.com
+
+Yair Lakretz*
+
+Cognitive Neuroimaging Unit
+
+CEA,INSERM
+
+Université Paris-Saclay
+
+NeuroSpin Center
+
+yair.lakretz@gmail.com
+
+Jean-Rémi King*
+
+Meta AI Paris;
+
+Laboratoire des systèmes perceptifs
+
+Ecole normale supérieure
+
+PSL University
+
+jeanremi@meta.com
+
+# Abstract
+
+During language acquisition, children follow a typical sequence of learning stages, whereby they first learn to categorize phonemes before they develop their lexicon and eventually master increasingly complex syntactic structures. However, the computational principles that lead to this learning trajectory remain largely unknown. To investigate this, we here compare the learning trajectories of deep language models to those of children. Specifically, we test whether, during its training, GPT-2 exhibits stages of language acquisition comparable to those observed in children aged between 18 months and 6 years. For this, we train 48 GPT-2 models from scratch and evaluate their syntactic and semantic abilities at each training step, using 96 probes curated from the BLiMP, Zorro and BIG-Bench benchmarks. We then compare these evaluations with the behavior of 54 children during language production. Our analyses reveal three main findings. First, similarly to children, the language models tend to learn linguistic skills in a systematic order. Second, this learning scheme is parallel: the language tasks that are learned last improve from the very first training steps. Third, some – but not all – learning stages are shared between children and these language models. Overall, these results shed new light on the principles of language acquisition, and highlight important divergences in how humans and modern algorithms learn to process natural language.
+
+# 1 Introduction
+
+Language acquisition is marked by a series of successive stages (Dupoux, 2018; Kuhl, 2004; Werker, 2018). Within their first year of existence, humans infants successively acquire prosody contours (Mehler et al., 1988), phonetic categories (Werker and Tees, 1984; Kuhl et al., 1992; Mazuka et al., 2011) and frequent words (Tincoff and Jusczyk, 1999; Bergelson and Swingley, 2012).
+
+They then learn to produce basic syntactic structures (e.g. "The boy sang" or "The boy fell"), questions (e.g. "What sound does a cow make?") and nested syntactic structures (e.g. "The boy that I saw sang"), at approximately 12, 30, and 42 months, respectively (Friedmann et al., 2021). Even though some children may take slightly longer to learn than others, there is a set order in which children acquire various syntactic structures (Friedmann and Reznick, 2021).
+
+Our understanding of the entire learning trajectory of children remains very coarse, however. This partly stems from the difficulty of measuring linguistic skills in young children. In babies, experimenters typically measure eye gaze and sucking rate while children process linguistic stimuli, as these reflexive behaviors are known to increase during surprising events. Such "implicit" approaches have successfully been used to assess whether nonspeaking infants detect linguistic violations (Zamuner, 2006), distinguish lexical from grammatical words (Shi et al., 1999) or discriminate their native language from a foreign language (Mehler et al., 1988; Kuhl et al., 2006; Nazzi et al., 2000). In older children, linguistic skills can also be more explicitly measured from spontaneous speech and sentence repetition. For example, a recent study by Friedmann et al. (2021), to which we compare our work in this paper, quantified the extent to which 18 month to 6 year-old children produce variably complex syntactic structures. For both of these approaches, however, the measures from children at such early ages can be noisy and fragmented.
+
+Interestingly, these issues do not apply to modern language models. Deep learning architectures trained to predict words from their proximal contexts have proved immensely effective at learning to process natural language (Radford et al., 2019; Devlin et al., 2019). Unlike humans, these algorithms can be easily probed during training, at any time point and rate, and with unlimited number of
+
+test stimuli, without interfering with their language acquisition (Jawahar et al., 2019; Manning et al., 2020; Bowman and Dahl, 2021). Furthermore, high-performing deep nets have been shown to implicitly (Lakretz et al., 2019; Gulordava et al., 2018) or explicitly learn to represent and use syntactic structures (Manning et al., 2020), as well as to use features such as concreteness and lexical class to learn language (Chang and Bergen, 2022). Finally, and importantly, these deep neural networks have recently been shown to represent lexical, syntactic and compositional representations similarly to the adult brain (Jain and Huth, 2018; Toneva and Wehbe, 2019; Caucheteux and King, 2022; Pasquiou et al., 2022, 2023; Caucheteux et al., 2023). Evidencing similar learning trajectories in children and language models could thus provide an invaluable framework to better understand the computational principles underlying language acquisition.
+
+Here, we compare the trajectory of language acquisition between human children and modern language models. We focus on three main questions. First, do these models learn linguistic skills in a systematic order? Second, is this trajectory sequential or parallel? Third, is this trajectory similar to that of children? These hypotheses are illustrated in Figure 1.
+
+Specifically, we train 48 GPT-2 architectures (Radford et al., 2019) from scratch, using a standard next-word prediction objective. We then evaluate, at each training step, their linguistic abilities with 96 semantic and syntactic probes curated from the BLiMP, Zorro and BIG-Bench benchmarks (Warstadt et al., 2020; Huebner et al., 2021; Srivastava et al., 2022). Finally, we compare a subset of these probes to the behavior of 54 children aged, between 18 months and 6 years (Friedmann et al., 2021).
+
+# 2 Approach
+
+# 2.1 Language models
+
+We consider two main language models. First, we use a pretrained language model – GPT-2 – as provided by HuggingFace and pretrained on 40 GB of data (Radford et al., 2019). Second, we separately train 48 versions of a 12-layer GPT-2 model from scratch. We train each model on WikiText103 (Merit et al., 2016) with a distinct random seed to set its initial parameters and data-loader. Each model is evaluated on all linguistic probes every
+
+100 training steps. Further training details are provided in Appendix B.
+
+# 2.2 Zero-shot linguistic probes
+
+Zero-shot linguistic probes are sentences or phrases crafted to evaluate whether a model has learned a particular linguistic skill, without training or finetuning the model on that particular skill. In practice, a zero-shot probe consists of comparing the estimated probability of a grammatical sentence with that of a matched ungrammatical sentence. This two-alternative forced-choice approach can be compared to "acceptability judgements", classically used in linguistics (Warstadt et al., 2019).
+
+We evaluate our models on 96 different linguistic probes, curated from three open source benchmarks, the details of which are presented in Appendix C.
+
+Specifically, we compare the probability of each sentence in a grammatical/ungrammatical pair by evaluating the sum of the logarithm of the loss output by the softmax layer:
+
+$$
+\sum_ {i = 0} ^ {n _ {g}} \log (f (X _ {g}) _ {i}) <   \sum_ {j = 0} ^ {n _ {u}} \log (f (X _ {u}) _ {j}) \tag {1}
+$$
+
+with  $f$  the softmax layer of the language model,  $X_{g}$  and  $X_{u}$  the grammatical and ungrammatical sentences, respectively, and  $n_g$  and  $n_u$ , the number of tokens in the grammatical and ungrammatical sentences, respectively.
+
+The accuracy of a given probe is the percentage of pairs where the estimated probability of the grammatical sentence is higher than that of the ungrammatical sentence.
+
+# 2.3 Assessing learning trajectory
+
+To evaluate whether the trajectory of language acquisition is shared across models, we rank the probes by their "acquisition time", i.e. the number of steps taken by a model to reach  $90\%$  of its final accuracy on a particular probe, for each model independently. We then assess the correlation of ranks between all pairs of the 48 models and take the average of these correlations. To estimate the statistical significance of this average correlation we redo this calculation for all possible model pairs after shuffling the ranks of one of the models in each pair. We repeat this permutation 1,000 times, getting 1,000 values for this shuffled correlation. If in all cases this shuffled correlation is lower than
+
+the true average correlation, then the order of acquisition time is shared across models with  $p < 0.001$ .
+
+# 2.4 Parallel versus Sequential learning
+
+Language acquisition may be characterized by a "sequential" or a "parallel" learning scheme (Figure 1). "Sequential" learning designates the case where a complex skill does not start to be learned before simpler skills are mastered. By contrast, "Parallel" learning designates the case where all skills are acquired simultaneously, but at different speeds. The null hypothesis is that the order in which an agent learns linguistic skills varies across agents. To determine the learning scheme of language models, we consider whether the probes have a positive derivative in the first three checkpoints (parallel learning) or not (sequential learning), and whether they have statistically different learning rates (by performing a one-way ANOVA test) across the three groups.
+
+# 2.5 Assessing linguistic skill from children's behavior
+
+Friedmann et al. (2021) studied 54 Hebrew-speaking children between the ages of 18 - 71 months and investigated the emergence of 11 linguistic phenomena, which the authors propose to organize into three stages (details in Appendix A). For our analysis we select the following tests, one from each stage:
+
+- Stage 1: Simple sentences in subject-verb (SV) order  
+- Stage 2: Wh Questions  
+- Stage 3: Relative Clauses
+
+Data collection consisted of spontaneous speech samples produced by each child at home. Each sample was then manually annotated to detect the presence of each of the linguistic phenomena. A linguistic phenomenon was considered learned if and only if it was present in the speech sample. Speech samples had a mean length of 151 utterances per sample and standard deviation of 37. The aggregated data was made available directly in the original paper (under Creative Commons Attribution 4.0 International License), and here used for comparison with our language models. In Table 1 we show which probes in the models matched with these tests.
+
+# 3 Results
+
+We aim to compare the learning trajectories of deep language models to those observed in 54 children aged between 18 months and 6 years. For this, we trained variants of GPT-2 models (Radford et al., 2019) from 48 different random seeds with the WikiText103 dataset (Merit et al., 2016) and evaluated each model on 96 linguistic probes every 100 steps.
+
+At the end of this training, 64 probes (66%) were achieved above chance level (50% accuracy) by all models. In comparison, a pretrained version of GPT-2 large (Radford et al., 2019) provided by Hugging Face $^2$ , and trained on a much larger dataset $^3$ , achieves above-chance performance on 93 of the 96 probes.
+
+# 3.1 A systematic learning trajectory
+
+For clarity, we focus on the learning dynamics of the probes that ultimately achieve above-chance performance in our training. Figure 2 lists all probes learned above chance level, ordered by their average acquisition time. We perform the permutation analysis outlined in 2.3, to evaluate whether the order of acquisition is shared between models, and find that their order of acquisition is correlated with  $R = 0.743$  and  $p < 0.001$ . These results suggest that there is a systematic learning trajectory among models.
+
+# 3.2 Learning is parallel across linguistic tasks
+
+Are these linguistic skills learned sequentially or in parallel (Figure 1)? To address this question, we evaluate whether each linguistic probe starts to improve from the very first training steps but with different rates (i.e. a "parallel" learning scheme) or, on the contrary, whether some probes only start to improve once others have reached a particular performance (i.e. a "sequential" learning scheme). As the individual learning trajectories of each probe were noisy, we group the 64 linguistic probes into three categories: early, middle and late acquisition time (Figure 3).
+
+Overall, we observe parallel learning between the three groups: their performances all increase from the beginning of training:  $95\%$  of tests in all three groups have a positive derivative within the first three hundred steps. However, they have different learning rates, as evaluated with a one-way ANOVA test on the learning rate (i.e. change
+
+![](images/d0dea45aa26fe24dd3b78ce9112237b6059833c376ac00595f37540c4834e8e0.jpg)  
+Figure 1: Hypotheses. Skill performance (y-axis) as a function of training (x-axis) illustrated on three tasks (colors) and two agents (model or children). Sequential learning implies that the learning of a complex skill (e.g. C, shown in red) does not start before simplest skills (e.g. A in green and B in blue) are fully learned. Parallel learning implies that all skills are acquired simultaneously, but at different speeds. Sequential and parallel learning may cross an arbitrary performance threshold at the same training step. The frequency at which we can probe artificial networks during learning is much greater than what is realistically possible in children, giving us the timescale granularity to distinguish sequential and parallel learning trajectories. We also present a null hypothesis, that artificial networks with different random seeds may learn skills in a different order.
+
+![](images/f27c71387747471f90818c4ef27dfce6ff3da95f965a0e7c9e24ba7e83f2d63e.jpg)
+
+![](images/35d1246d79b6f047f0f4d3fb0e916b43e5ea80c505b8139ad91a8c9b5598a3b4.jpg)
+
+<table><tr><td>Stage</td><td>Children</td><td>Language Model</td></tr><tr><td>1</td><td>Simple sentences in Subject-Verb (SV) order</td><td>SV agreement across simple sentences</td></tr><tr><td>2</td><td>Wh-questions</td><td>SV agreement in questions</td></tr><tr><td>3</td><td>Relative Clauses (RCs)</td><td>SV agreement across object RCs</td></tr></table>
+
+Table 1: Linguistic phenomena in children selected for comparison with probes in the language models.
+
+of accuracy over time) obtained in each group and in each model  $(p < 10^{-23})$ .
+
+# 3.3 Comparison with children
+
+Do these learning trajectories match the behavior of human children? For the three probes that correspond to the three stages identified in children's language acquisition (Table 1), we observe that the order in which these three probes are learned by the language models is the same as those of children (Figure 4). This effect is robust across random seeds: 46 of our 48 GPT-2 models follows this order, where chance level is  $(1/3!)^{46} = 1.60e^{-36}$ . For this subset of language abilities, models and children thus seem to acquire syntactic skills in a similar order.
+
+# 3.4 Learning of theoretically-defined stages is parallel
+
+In part 3.1, we showed that GPT-2 learns its language abilities in parallel. Does this learning scheme also characterize the three syntactic skills investigated in children? To address this question, we now look at the learning curves of the skills defined in Table 1, as well as an additional probe: Nounpp, as it can be separated into congruent and incongruent cases which is important for the anal
+
+ysis in section 3.5. Overall, we observe that these probes are all learned in parallel in the model (Figure 5A).
+
+# 3.5 Models use both syntax and heuristics
+
+Both an understanding of syntactic rules and a superficial heuristics can lead to above-chance performances on these probes. Indeed, in many sentences (e.g. The cat [subject] of the lady [attractor] is [verb] hungry.), the number of the verb is congruent with the number of the adjacent attractor, even if the two are not related syntactically. To verify that the GPT-2 models effectively learn the syntactic rules, we thus separately examine congruent and incongruent cases. Incongruent cases require knowledge of the syntax of the sentence as the correct verb number is different from the number of the attractor. Empirically, we observe that the models do not learn the incongruent case in stage three above chance level, and just barely reach chance level on the incongruent case in stage two (Figure 5B), indicating that our models are using heuristics rather than syntactic rules to achieve high accuracy on the congruent cases (leading to above chance performance on the probe overall in Figure 5A). On the contrary, the pretrained GPT
+
+![](images/b010ed78004a7b83855311dec96637769c4214e3054ba0dba02e8742b0e7c893.jpg)  
+Figure 2: The performance of the models on each linguistic probe over time is smoothed using a moving average filter with window size of 6 checkpoints then the number of steps required to reach  $90\%$  of final accuracy (acquisition time) is calculated. Probes are ordered by increasing average acquisition time. Results shown for 48 models. Only probes which have final accuracy greater than chance  $(50\%)$  are shown. This demonstrates that probes tend to be learned in the same order by all agents with  $\mathrm{R} = 0.743$ ,  $\mathrm{p} < 0.001$ , disproving the null hypothesis.
+
+2 large achieves above  $75\%$  accuracy also on the incongruent cases of these probes. Thus for the models trained on the WikiText103, syntax is only learned for stages one and two, and heuristics seem to explain the above chance accuracy in stage three. A larger training dataset is required to learn syntax and not only heuristics for the most difficult
+
+examples.
+
+# 3.6 Impact of number biases in congruent and incongruent sentences
+
+In previous work, it was found that a variety of language models have a bias towards plural English verbs, and several studies (Jumelet et al., 2019;
+
+![](images/2a909119396e6065cd8fb6ecc488cab23f1cbcbd9896e111eccb313d3bf103b9.jpg)  
+Figure 3: Linguistic probes grouped into 3 groups to avoid plotting one line per probe. The probes in each group correspond to the colours in Figure 2. Probes which have final accuracy less than chance (50%) are placed in their own group, and tend to zero likely due to biases towards plural verbs in English (c.f. 3.6). Shading is standard error of the mean across probes in the group. This figure demonstrates that linguistic skills are learned in parallel not in sequence.
+
+Lakretz et al., 2021a,b) determined that LSTM-based models have a default number and gender prediction preference. To examine whether number bias has a significant effect on our analysis, we compare congruent sentences with only singular or only plural verbs and incongruent sentences with a plural or a singular verb. Accuracy on predicting plural verbs increases sharply from the start of the training and then drops. By contrast, the accuracy of singular cases first drops and then rises (Figure 5C), indicating that the models are biased towards plural verbs at the beginning of training. This bias is overcome for the stage one probe but for stage two and three it remains throughout training. This explains the initial downward trend in Group 3 and why the unlearned probes tend toward zero in Figure 3.
+
+# 4 Discussion
+
+The stages followed by children to acquire language has been the topic of intense research (Dupoux, 2018; Kuhl, 2004; Werker, 2018). While this learning trajectory is becoming clearer for sublexical representations (Dupoux, 2018), the acquisition of higher-level syntactic and semantic processing remains largely unclear. Here, we approach this long-lasting question through the lens of a deep language architecture, GPT-2 (Radford et al., 2019),
+
+to test whether this model follows a learning trajectory similar to children.
+
+# 4.1 Language acquisition: similarities and differences between humans and GPT-2
+
+First, we show that GPT-2 models tend to learn a battery of linguistic phenomena (Warstadt et al., 2020; Lakretz et al., 2019; Huebner et al., 2021) in a consistent order. It is the reliability of the acquisition trajectory that allows a direct comparison with the learning trajectory of children (Friedmann et al., 2021). However, this consistency in GPT-2 models may result from two non-mutually exclusive factors, that remain to be disentangled: either the acquisition time of each linguistic phenomenon relates to its frequency in natural language (e.g. Simple subject-verb-complement are more frequent in natural language than nested syntactic structures; Karlsson 2007), and/or it relates to their intrinsic complexity (e.g. sentences with nested structure require more operations to be composed than simple sentences). Future work systematically controlling for these relative frequencies is thus necessary to distinguish these two possibilities, and would build upon work by Weber et al. (2021) who found that less frequent linguistic phenomena can be learned from fewer examples, though later in training.
+
+Second, we show that the order in which linguistic skills are acquired is similar between children and GPT-2 – at least on the syntactic phenomena that were evaluated in these two cohorts, and with the limitation of using number agreement as a proxy to whether the models acquire the corresponding syntactic structure. Similarly to children, GPT-2 models master subject-verb agreement in SV sentences before they master it in questions, or across nested center-embedded clauses (object-relative clauses). This result thus complements a series of studies comparing modern language models and humans. For example, a recent study showed that transformers trained on child-directed data can achieve comparable accuracy on linguistic probes to large pre-trained models (Huebner et al., 2021). Similarly, several studies have recently shown that the representations of GPT-2 become increasingly similar to those of the adult human brain during its training (Caucheteux and King, 2022). Finally, Lavechin et al. (2022) showed that models trained on audio in a self-supervised fashion learn phoneme and lexical abilities in a similar trajectory to children.
+
+![](images/b52612123aa11f556fd821244d49acef99ea1664d766992b21fa9b4f64e0feca.jpg)  
+Figure 4: Comparing language model performance on linguistic probes to children's performance. Example sentences observed in children were originally in Hebrew (Friedmann et al., 2021). Non-white indicates the phenomena is learned by the agent. The threshold for considering a probe learned by the model is performance above  $55\%$ .
+
+![](images/781b658bfc78f7ed0cfd5ca39e7257e9e00e41fa315e3713a7e983d9647b09f4.jpg)
+
+![](images/da214a664f165da850401a424511f74092afc15d53a881a9a745c6cfcbe30486.jpg)
+
+![](images/8f69bb2d9586d1e9e2088e0e72e3684b35d63a5659bce91acb74b82ef3c4f313.jpg)  
+Figure 5: Subject verb agreement. Shading is standard error of the mean across seeds. S: Singular. P: Plural. A. Parallel learning is observed in the three stages defined by Friedmann et al. (2021), when results are averaged across congruent and incongruent cases. B. Subject verb agreement on incongruent cases which indicate whether the model understands syntax. The networks do not learn some syntax structures as the incongruent case of Short Nested Outer does not reach above chance level. C. Development trajectories of the bias towards plural number in English.
+
+![](images/d9f2fe0c53b9a8f161d38e2674a786a7947db551734aa4365370e19ba64859ed.jpg)  
+C
+
+![](images/9a08e1eaf97814dbb46825830657e249397e179eab8e5ca1883d404625c23eef.jpg)
+
+# 4.2 A work in progress
+
+It is important to stress that significant work remains to be done before drawing any definitive conclusions about the similarities between language acquisition in humans and algorithms.
+
+First, we only consider a single architecture (GPT-2, Radford et al. (2019)) with a unique textual corpus (WikiText103). Testing whether our
+
+results hold true independently of the model and training corpus remains an important milestone for future research.
+
+Second, linguistic abilities are not tested with the same protocols in children and in the models: the models are explicitly tested on next word prediction, with a two-alternative forced-choice metric, whereas children were implicitly evaluated on their ability to spontaneously use specific syntactic struc
+
+tures during natural speech.
+
+Third, there were only three linguistic features that were directly comparable between the model probes and the data in children, and all were syntactic. This leaves a significant margin of progress to modulate our conclusion, and investigate whether the lexicon, narrative structures, pragmatics and world-knowledge are acquired in the same order in humans and algorithms.
+
+Fourth, the order in which some linguistic skills were learned by GPT-2 does not trivially fit with linguistic theory. For example, the probe "Simple", which examines subject-verb agreement in a simple sentence, was one of the last probes to be learned by GPT-2 (it is part of group three in Figure 2). By contrast, "Wh Questions Subject Gap Long Distance" was among the first probes to be to be learned, even though it would be expected to be much harder than "Simple". This unexpected result may be due to the way we approximate "Acquisition Time", namely, the moment when the probes reaches  $90\%$  of the final accuracy. Consequently, probes with very low final accuracy could end up with a shorter Acquisition Time, because noise may lead to crossing the  $90\%$  threshold relatively quickly.
+
+Finally, we show that our models appear to use heuristics rather than a deep understanding of syntax for the most difficult linguistic probes (incongruent numbers between verbs and their attractors) and were biased towards plural English verbs. While our models learn only  $66\%$  of tasks to above chance level, a larger GPT-2 pretrained on considerably more texts successfully perform on  $97\%$  of the tasks, and has an accuracy above  $75\%$  on the incongruent examples, meaning this bias and reliance on heuristics could potentially be solved by training on a larger dataset.
+
+In sum, additional work remains necessary to identify the exact elements of convergence and divergence between the acquisition of language in models and in children.
+
+# 4.3 Fueling the debate between nativism versus empiricism
+
+The present study fuels a long-lasting debate on the acquisition of language. While "empiricists" argue that language can be acquired with a statistical approach (Clark, 2002; Kolodny et al., 2015; Chater and Christiansen, 2018; McCauley and Christiansen, 2019), "nativists" maintain that this ability depends on a core and innate operation, spe
+
+cific to humans (Chomsky, 1959, 1971).
+
+The present study shows how modern language models may contribute to resolving this debate, by systematically studying which components of a model (e.g. architecture) or properties of the training data (e.g., frequency of sentence structures) contribute to shape the trajectory of language acquisition. Claims about an innate Universal Grammar could be understood as an inductive bias of a language model, implemented in its architecture and dynamics, which tightly constrains learning trajectories across models. If this bias is hierarchical (rather than linear) then this could lead to learning trajectories that follow the structure of the syntactic tree, consistently with the hypothesis of three linguistic stages presented by Friedmann et al. (2021) in humans and what we find in this study in language models. Indeed, neural language models have been previously shown to have a weak inductive bias towards hierarchical processing (McCoy et al., 2020; Kharitonov and Chaabouni, 2020), which can partially explain our results.
+
+This result echos the recent observation that syntactic trees spontaneously emerge in the middle layers of neural language models (Hewitt and Manning, 2019). Together, these elements thus suggest that modern neural networks provide fruitful models of language acquisition and could reconcile or settle the confronting theories of language acquisition (Warstadt and Bowman, 2022).
+
+# 4.4 Conclusion
+
+Overall, the similarities identified between children and GPT-2 suggest that there may be a small set of means by which to efficiently acquire language. This result is anything but trivial: humans and deep neural networks have extraordinarily different architectures, training, and language exposure. If generalized, this systematic learning trajectory would support the existence of an intrinsic hierarchy of linguistic structures that both machines and humans must climb, be that through inductive biases or properties of the training data, to master the faculty of language. And while these hypotheses remain open, the path to resolve them has never been clearer.
+
+# Acknowledgements
+
+We would like to thank Dieuwke Hupkes, Naama Friedmann, Marco Baroni and the attendees of the EviL meetings for their comments and suggestions.
+
+This project has received funding from the European Union's Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 945304, for L.E for her work at PSL. This work was funded in part by FrontCog grant ANR-17-EURE-0017 for the work of L.E. and J.R.K. for their work at PSL.
+
+# References
+
+Elika Bergelson and Daniel Swingley. 2012. At 6-9 months, human infants know the meanings of many common nouns. Proceedings of the National Academy of Sciences, 109:3253-3258.  
+Samuel R. Bowman and George Dahl. 2021. What will it take to fix benchmarking in natural language understanding? Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4843-4855.  
+Charlotte Caucheteux, Alexandre Gramfort, and Jean-Rémi King. 2023. Evidence of a predictive coding hierarchy in the human brain listening to speech. Nature Human Behaviour, pages 1-12.  
+Charlotte Caucheteux and Jean Rémi King. 2022. Brains and algorithms partially converge in natural language processing. Communications Biology, 5(1).  
+Tyler A. Chang and Benjamin K. Bergen. 2022. Word acquisition in neural language models. Transactions of the Association for Computational Linguistics, 10:1-16.  
+Nick Chater and Morten H Christiansen. 2018. Language acquisition as skill learning. Current opinion in behavioral sciences, 21:205-208.  
+Noam Chomsky. 1959. Review of verbal behavior. 35(1):26-58. Publisher: Linguistic Society of America.  
+Noam Chomsky. 1971. *Problems of Knowledge and Freedom*. New York,: W.W. Norton.  
+Alexander Clark. 2002. Unsupervised language acquisition: Theory and practice. arXiv preprint cs/0212024.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova Google, and AI Language. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186.
+
+Emmanuel Dupoux. 2018. Cognitive science in the era of artificial intelligence: A roadmap for reverse-engineering the infant language-learner. Cognition, 173:43-59.  
+Naama Friedmann, Adriana Belletti, and Luigi Rizzi. 2021. Growing trees: The acquisition of the left periphery. Glossa: a journal of general linguistics, 39(1).  
+Naama Friedmann and Julia Reznick. 2021. Stages rather than ages in the acquisition of movement structures: Data from sentence repetition and 27696 spontaneous clauses. Glossa: a journal of general linguistics, 39(1).  
+Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195-1205, New Orleans, Louisiana. Association for Computational Linguistics.  
+John Hewitt and Christopher D Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129-4138.  
+Philip Huebner. 2022. Unmasked. https://github.com/phueb/UnMasked. (Accessed 2023/05/24).  
+Philip A Huebner, Elior Sulem, Cynthia Fisher, and Dan Roth. 2021. BabyBERTa : Learning More Grammar With Small-Scale Child-Directed Language. Proceedings of the 25th Conference on Computational Natural Language Learning, pages 624-646.  
+Shailee Jain and Alexander Huth. 2018. Incorporating context into language encoding models for fmri. In Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc.  
+Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does bert learn about the structure of language? Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651-3657.  
+Jaap Jumelet, Willem Zuidema, and Dieuwke Hupkes. 2019. Analysing neural language models: Contextual decomposition reveals default reasoning in number and gender assignment. Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 1-11.  
+Fred Karlsson. 2007. Constraints on multiple center-embedding of clauses. Journal of Linguistics, 43(2):365-392.
+
+Eugene Kharitonov and Rahma Chaabouni. 2020. What they do when in doubt: a study of inductive biases in seq2seq learners. arXiv:2006.14953.  
+Oren Kolodny, Arnon Lotem, and Shimon Edelman. 2015. Learning a generative probabilistic grammar of experience: A process-level model of language acquisition. Cognitive Science, 39(2):227-267.  
+Patricia K. Kuhl. 2004. Early language acquisition: cracking the speech code. Nature Reviews Neuroscience, 5:831-843.  
+Patricia K Kuhl, Erica Stevens, Akiko Hayashi, Toshisada Deguchi, Shigeru Kiritani, and Paul Iverson. 2006. Fast-track report infants show a facilitation effect for native language phonetic perception between 6 and 12 months. Developmental Science, 9:13-21.  
+Patricia K. Kuhl, Karen A. Williams, Francisco Lacerda, Kenneth N. Stevens, and Bjorn Lindblom. 1992. Linguistic experience alters phonetic perception in infants by 6 months of age. Science, 255:606-608.  
+Yair Lakretz, Théo Desbordes, Dieuwke Hupkes, and Stanislas Dehaene. 2021a. Causal transformers perform below chance on recursive nested constructions, unlike humans. arXiv:2110.07240.  
+Yair Lakretz, Dieuwke Hupkes, Alessandra Vergallito, Marco Marelli, Marco Baroni, and Stanislas Dehaene. 2021b. Mechanisms for handling nested dependencies in neural-network language models and humans. Cognition, 213:104699. Special Issue in Honour of Jacques Mehler, Cognition's founding editor.  
+Yair Lakretz, German Kruszewski, Theo Desbordes, Dieuwke Hupkes, Stanislas Dehaene, and Marco Baroni. 2019. The emergence of number and syntax units in LSTM language models. Association for Computational Linguistics, pages 11-20.  
+Marvin Lavechin, Maureen De Seyssel, Hadrien Titeux, Hervé Bredin, Guillaume Wisniewski, Alejandrina Cristia, and Emmanuel Dupoux. 2022. Statistical learning bootstraps early language acquisition. PsyArXiv.  
+Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg. 2016. Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies. Transactions of the Association for Computational Linguistics, 4:521-535.  
+Christopher D. Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, and Omer Levy. 2020. Emergent linguistic structure in artificial neural networks trained by self-supervision. Proceedings of the National Academy of Sciences of the United States of America, 117:30046-30054.  
+Reiko Mazuka, Yvonne Cao, Emmanuel Dupoux, and Anne Christophe. 2011. The development of a phonological illusion: A cross-linguistic study with japanese and french infants. Developmental Science, 14:693-699.
+
+Stewart M McCauley and Morten H Christiansen. 2019. Language learning as language use: A cross-linguistic model of child language development. *Psychological review*, 126(1):1.  
+R Thomas McCoy, Robert Frank, and Tal Linzen. 2020. Does syntax need to grow on trees? sources of hierarchical inductive bias in sequence-to-sequence networks. Transactions of the Association for Computational Linguistics, 8:125-140.  
+Jacques Mehler, Peter Jusczyk, Ghislaine Lambertz, Nilofar Halsted, Josiane Bertoncini, and Claudine Amiel-Tison. 1988. A precursor of language acquisition in young infants. Cognition, 29:143-178.  
+Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models. arXiv:1609.07843.  
+Thierry Nazzi, Peter W. Jusczyk, and Elizabeth K. Johnson. 2000. Language discrimination by english-learning 5-month-olds: Effects of rhythm and familiarity. Journal of Memory and Language, 43:1-19.  
+Alexandre Pasquiou, Yair Lakretz, John Hale, Bertrand Thirion, and Christophe Pallier. 2022. Neural language models are not born equal to fit brain data, but training helps. In ICML 2022-39th International Conference on Machine Learning, page 18.  
+Alexandre Pasquiou, Yair Lakretz, Bertrand Thirion, and Christophe Pallier. 2023. Information-restricted neural language models reveal different brain regions' sensitivity to semantics, syntax and context. arXiv:2302.14389.  
+Yada Pruksachatkun, Phil Yeres, Haokun Liu, Jason Phang, Phu Mon Htut, Alex Wang, Ian Tenney, and Samuel R. Bowman. 2020. jiant: A software toolkit for research on general-purpose text understanding models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 109-117, Online. Association for Computational Linguistics.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. Semantic Scholar. (Accessed 2023-05-04).  
+Rushen Shi, Janet F Werker, and James L Morgan. 1999. Newborn infants' sensitivity to perceptual cues to lexical and grammatical words. Cognition, 72:B11-B21.  
+Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv 2206.04615.  
+Ruth Tincoff and Peter W. Jusczyk. 1999. Some beginnings of word comprehension in 6-month-olds. Psychological Science, 10:172-175.
+
+Mariya Toneva and Leila Wehbe. 2019. Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain). In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.  
+Alex Warstadt and Samuel R Bowman. 2022. What artificial neural networks can tell us about human language acquisition. arXiv:2208.07998.  
+Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng Fu Wang, and Samuel R. Bowman. 2020. Erratum: "blimp: The benchmark of linguistic minimal pairs for english". Transactions of the Association for Computational Linguistics, 8:867-868.  
+Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625-641.  
+Lucas Weber, Jaap Jumelet, Elia Bruni, and Dieuwke Hupkes. 2021. Language modelling as a multi-task problem. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2049-2060, Online. Association for Computational Linguistics.  
+Janet F. Werker. 2018. Perceptual beginnings to language acquisition. Applied Psycholinguistics, 39:703-728.  
+Janet F. Werker and Richard C. Tees. 1984. Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. *Infant Behavior and Development*, 7:49–63.  
+Tania S. Zamuner. 2006. Sensitivity to word-final phonotactics in 9- to 16-month-old infants. Infancy, 10:77-95.
+
+# Appendix
+
+# A Tests in Children
+
+Detailed description of tests available in children, in the three linguistic stages defined by (Friedmann et al., 2021):
+
+- Stage 1: Subject-Verb Simple, Subject-Verb Unaccusative, Verb-Subject Unaccusative  
+- Stage 2: Root WH-Argument, WH-Adjunct Excluding Why, Preposed Adverb, Root y/n  
+- Stage 3: Why, Relative Clause, Topicalisation, Embedding
+
+The probes chosen for comparison (stated in Table 1), were the only probes that matched well with one of the test available in children. In addition Nounpp was examined in the models, as it fits into
+
+the linguistic stage 2, and, as it is part of the BIG-Bench probes, could be separated into congruent and incongruent sentences.
+
+# B Model Training
+
+To evaluate linguistic abilities of a high-performance language model, we first use the HuggingFace pretrained GPT-2 large which has 774M parameters and is trained on 40GB of data. This model has one-shot perplexity of 22 on WikiText103 (Radford et al., 2019).
+
+Then, to evaluate how linguistic abilities vary with language acquisition, we separately trained 48 models (each with a distinct random seed which set the model's initial parameters and the seed of the datalocator) using the 12-layer GPT-2 architecture (Radford et al., 2019) provided by HuggingFace on WikiText103 (Merit et al., 2016) with a learning rate of  $10^{-5}$  and a batch size of 16 distributed over 8 GPUs, making a total batch size of 64 and a context size of 128. Training was stopped when then validation loss plateaued, reaching final perplexity of 28 after 10 epochs. This is lower perplexity than the one-shot performance of the HuggingFace pretrained 12-layer GPT-2 which was 37.5, which is logical as our model was trained specifically on this dataset.
+
+In all cases we used the pretrained tokenizer which has vocabulary size of 50,257. All other parameters were the default training arguments for the transformer provided by HuggingFace. The HuggingFace architectures are publicly available under an MIT license, and WikiText103 is available under Creative Commons Attribution-ShareAlike License.
+
+# C Linguistic Probe Benchmarks
+
+We use three different zero-shot benchmarks. The first benchmark, 'BLiMP' (The Benchmark of Linguistic Minimal Pairs for English) (Warstadt et al., 2020) contains 67 different probes, each in the form of 1,000 pairs of grammatical and ungrammatical sentences designed to isolate a specific linguistic phenomenon. Adult human performance on BLiMP is  $96.4\%$  (Warstadt et al., 2020). The second benchmark, 'Zorro'  $^{5}$ , was developed with a vocabulary frequent in child-directed corpora. Zorro contains 13 probes, each consisting of 2,000 pairs of sentences. Finally, the third benchmark is the Subject-Verb Agreement Task of BIG
+
+Bench (Srivastava et al., 2022; Lakretz et al., 2019, 2021b; Linzen et al., 2016; Gulordava et al., 2018). We focus on the syntactic probes, namely: "Simple English" which contains 600 pairs, "NounPP" which contains 2,400 pairs, and "Short Nested Inner", "Short Nested Outer", "Long Nested Inner" and "Long Nested Outer" which each contain 4,096 pairs of grammatical and ungrammatical sentences.
+
+Accuracy on a linguistic probe is evaluated with the Jiant (Pruksachatkun et al., 2020) and UnMasked method (Huebner, 2022). In practice, sentences are input to the model in batches of 300, with padding on the left to make all sentences the length of the longest sentence in the batch. The logit values of punctuation are discarded when estimating the probability of a sentence.
+
+Zorro, Jiant and UnMasked are publicly available under the MIT License, BLiMP under a CC BY License, and BIG-Bench under the Apache License 2.0.
+
+# A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Left blank.
+
+A2. Did you discuss any potential risks of your work?
+
+Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Left blank.
+
+□ A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+# B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+Left blank.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Left blank.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Left blank.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Left blank.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Left blank.
+
+# C Did you run computational experiments?
+
+Left blank.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Left blank.
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank.
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank.

2023/Language acquisition_ do children and language models follow similar learning stages_/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4d2b5676014cd0180d57c905f4832cdf960804e37eff1bb8dae048cac5cc2ae
+size 566271

2023/Large Language Models Are Partially Primed in Pronoun Interpretation/60280f41-b321-4144-822d-79720ae15459_content_list.json

@@ -0,0 +1,2067 @@
+[
+    {
+        "type": "text",
+        "text": "Large Language Models Are Partially Primed in Pronoun Interpretation",
+        "text_level": 1,
+        "bbox": [
+            121,
+            89,
+            875,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Suet-Ying Lam†* Qingcheng Zeng‡* Kexun Zhang◇* Chenyu You* Rob Voigt‡",
+        "bbox": [
+            119,
+            123,
+            880,
+            143
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "† Department of Linguistics, UMass Amherst",
+        "bbox": [
+            315,
+            156,
+            685,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$\\ddagger$  Department of Linguistics, Northwestern University",
+        "bbox": [
+            282,
+            174,
+            719,
+            191
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$\\diamond$  Department of Computer Science, UCSB",
+        "bbox": [
+            324,
+            192,
+            677,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$\\clubsuit$  Department of Electrical Engineering, Yale University",
+        "bbox": [
+            270,
+            208,
+            729,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While a large body of literature suggests that large language models (LLMs) acquire rich linguistic representations, little is known about whether they adapt to linguistic biases in a human-like way. The present study probes this question by asking whether LLMs display human-like referential biases using stimuli and procedures from real psycholinguistic experiments. Recent psycholinguistic studies suggest that humans adapt their referential biases with recent exposure to referential patterns; closely replicating three relevant psycholinguistic experiments from Johnson and Arnold (2022) in an in-context learning (ICL) framework, we found that InstructGPT adapts its pronominal interpretations in response to the frequency of referential patterns in the local discourse, though in a limited fashion: adaptation was only observed relative to syntactic but not semantic biases. By contrast, FLAN-UL2 fails to generate meaningful patterns. Our results provide further evidence that contemporary LLMs discourse representations are sensitive to syntactic patterns in the local context but less so to semantic patterns. Our data and code are available at https://github.com/zkx06111/1lm_priming.",
+        "bbox": [
+            141,
+            281,
+            460,
+            665
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            680,
+            258,
+            695
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While neural network models, and particularly pretrained large language models have shown excellent performance at particular language processing tasks, many questions remain about the extent to which models optimized for such performance encode, as a side effect, human-like linguistic knowledge and cognitive biases. We know that they do to some extent; existing work has shown, for example, that neural models encode aspects of human-like long-distance number agreement (Gulordava et al., 2018), incremental syntactic state (Futrell et al.,",
+        "bbox": [
+            112,
+            707,
+            490,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2019), and syntactic generalization more broadly (Hu et al., 2020). In this paper, we examine whether FLAN-UL2 (Tay et al., 2023) and InstructGPT (Ouyang et al., 2022), two representative LLMs, display adaptation in pronoun interpretation when exposed to consistent referential patterns in the local discourse context.",
+        "bbox": [
+            507,
+            252,
+            884,
+            363
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Compared with syntactic or lexical knowledge, representing referential knowledge is possibly more complex; we know from psycholinguistic studies that human referential interpretation integrates multiple levels of linguistic structure. Humans do not interpret ambiguous pronouns at random but are guided by both syntactic and semantic information. It is well-established that absent other cues humans prefer a syntactic subject in choosing the antecedent of the ambiguous pronoun, i.e., subject bias (Ariel, 1990; Brennan, 1995). In example (1), she is more likely to be interpreted as the subject Ada than the non-subject Eva  $\\mathbb{I}$  , even though both referents are possible antecedents for the pronoun.",
+        "bbox": [
+            507,
+            366,
+            884,
+            607
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(1)  $\\mathrm{Ada}_1$  talked with  $\\mathrm{Eva}_2$ . She...",
+        "bbox": [
+            515,
+            620,
+            774,
+            636
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "People are also sensitive to the semantic structure of the sentence when choosing an antecedent for an ambiguous pronoun, in addition to syntactic information. In a transfer event that depicts a transfer-of-possession from one entity (the source) to another (the goal), they prefer the goal referent (Ada in (2), Eva in (3)) over the source referent (Eva in (2), Ada in (3)) to be the antecedent (Arnold, 2001, 1998), i.e., goal bias:",
+        "bbox": [
+            507,
+            649,
+            884,
+            793
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(2) Goal-source (gs) verb:",
+        "bbox": [
+            515,
+            806,
+            715,
+            821
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ada<sub>1</sub> received a letter from Eva<sub>2</sub>. She<sub>1</sub>...",
+        "bbox": [
+            546,
+            822,
+            848,
+            838
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(3) Source-goal (sg) verb:",
+        "bbox": [
+            515,
+            851,
+            714,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ada1 sent a letter to Eva2. She2...",
+        "bbox": [
+            546,
+            868,
+            796,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "Indices of pronouns in examples indicate the preferred referent.",
+        "bbox": [
+            507,
+            892,
+            880,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "9493",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 9493-9506",
+        "bbox": [
+            228,
+            945,
+            768,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            959,
+            700,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In sum, people exhibit sensitivity to both syntax (subject bias) and semantics (goal bias) during pronoun interpretation. Importantly, these levels of linguistic structure are frequently entwined since both can influence referential interpretation. At times these influences may push in different directions: for instance, Ada in [3] is both the syntactic subject and the semantic source.",
+        "bbox": [
+            112,
+            84,
+            489,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Building on a long tradition investigating preferences in pronoun interpretation, recent psycholinguistic studies have probed into the deeper question of the origin of these biases. One hypothesis is that referential biases come from linguistic experience: when a bias occurs very frequently, people will tend to adapt this more frequent referential pattern, both in the immediate exposure as well as more large-scale past experience (Arnold, 1998, 2001). Recent evidence has provided support to this idea by demonstrating that recent exposure to certain referential patterns did change people's referential biases. In a series of psycholinguistic studies, Johnson and Arnold (2022) show that after reading numerous stories that consistently show a particular referential bias, e.g., always referring to the non-subject or source referent, people did have a stronger preference for these primed referents.",
+        "bbox": [
+            115,
+            217,
+            490,
+            507
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given this line of psycholinguistic research, the current study investigates the extent to which LLMs adapt and vary referential biases in pronoun interpretation through exposure to referential patterns in the local context. To do so we replicated actual psycholinguistic experiments from Johnson and Arnold (2022) in LLMs using an ICL framework and asked whether the responses of LLMs display adaptation from exposure to referential patterns like human experimental participants. Comparing syntactically-motivated to semantically-motivated exposure conditions will allow us to first examine whether LLMs display human-like subject bias and goal bias and further understand the extent to which LLMs make use of local frequency changes in discourse representations in these categories.",
+        "bbox": [
+            112,
+            512,
+            489,
+            770
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In-context learning refers to LLMs' ability to learn from demonstrations written in natural language prompts. Compared to previous work that has largely examined the encoding of such discourse knowledge using zero-shot inference (Upadhye et al., 2020), ICL is particularly suitable for experimental simulation since it replicates the naturalistic context in which later responses draw upon exposure to previous examples.",
+        "bbox": [
+            112,
+            774,
+            489,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Foreshadowing our results, we find that InstructGPT can adapt and thus vary its syntactic bias from exposure to referential patterns in the local context, but the same is not true for semantic bias. Given that InstructGPT still exhibits a goal bias in spite of local discourse priming, we argue this suggests LLMs only encode partial semantic knowledge in referential processing. To sum up, our contributions can be summarized as follows:",
+        "bbox": [
+            507,
+            84,
+            884,
+            229
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1) We extended a discourse understanding evaluation to state-of-the-art LLMs from a new perspective, asking whether LLM's referential bias can be modified by exposure to particular referential patterns, like how humans adapt referential bias from experience.",
+            "2) To the best of our knowledge, we are the first study that replicates actual psycholinguistic experiments using the ICL framework and compares LLMs' behaviors with real human participants.",
+            "3) We present results in this context showing further evidence that InstructGPT can acquire abstract syntactic knowledge in referential interpretation to some extent, but not semantic knowledge."
+        ],
+        "bbox": [
+            522,
+            237,
+            885,
+            514
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            523,
+            665,
+            539
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A growing body of literature has suggested that LLMs encode rich representations of linguistic structure at various levels, including aspects of syntax, semantics, and reference encoded throughout their representations (Tenney et al., 2019). One of the most well-documented lines of this work demonstrates that these models can acquire diverse elements of syntactic knowledge (Gulordava et al., 2018; Futrell et al., 2019; Hu et al., 2020).",
+        "bbox": [
+            507,
+            549,
+            884,
+            692
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This capacity for encoding linguistic understanding extends to priming effects with psycholinguistic analogies to humans. At the syntactic level, Sinclair et al. (2022) explored structural priming in various autoregressive LLMs and found priming effects despite a clear dependence on semantic information. At the semantic level, using English lexical stimuli in BERT (Devlin et al., 2019), Misra et al. (2020) found that BERT does display evidence of sensitivity to semantic priming, though this is localized to more unconstrained contexts and only certain semantic relations.",
+        "bbox": [
+            507,
+            694,
+            884,
+            885
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In analyses of LLMs' linguistic understanding modeled on psycholinguistic experiments, however,",
+        "bbox": [
+            507,
+            887,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "9494",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the question of discourse knowledge remains relatively under-explored. Recent existing work has presented partially conflicting accounts in this area, in particular with regard to how LLMs may or may not exhibit human-like biases in pronoun interpretation. For example, Davis and van Schijndel (2020) compared LSTM LMs and Transformer LMs behaviors and internal representations in dealing with implicit causality verbs, finding that surprisingly (contrary to humans) implicit causality only influences Transformer LMs' behavior for reference, but influences neither model for syntactic attachment. Sorodoc et al. (2020) also compared LSTM LMs and Transformer LMs in coreference resolution corpora, finding that although LMs are much better at grammar, they also captured referential aspects to some extent.",
+        "bbox": [
+            115,
+            84,
+            489,
+            356
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This existing work replicates sets of individual stimuli from psycholinguistic experiments in isolation; by contrast, our work takes a more behaviorally-oriented approach by replicating the stimuli, procedure, and even experimental design of a full set of human psycholinguistic studies. We do this to ask a further question: beyond exhibiting baseline human-like referential biases in pronoun interpretation, do LLMs display adaptation to the frequency of referential patterns in the local context like humans? This question is important because adapting to a referential bias over the course of an experiment requires a sustained representation of the frequency of the pattern, which involves a higher-level understanding of the discourse structure. Humans have shown abilities of this kind of adaptation at multiple linguistic levels. For instance, exposure to syntactic structures consecutively affects humans' choice of structures in both the short term and long term via priming effects (e.g., Branigan et al. 2005; Chang et al. 2000). Exposure to phonological patterns can also guide humans' segmentation patterns (e.g., Saffran et al. 1996a,b). At the semantic and pragmatic levels, humans can also adapt as listeners to speakers' variable choices of uncertainty expressions (Schuster and Degen, 2020).",
+        "bbox": [
+            115,
+            359,
+            489,
+            791
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Methods",
+        "text_level": 1,
+        "bbox": [
+            115,
+            810,
+            223,
+            824
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, we aim to replicate three experiments from Johnson and Arnold (2022), transferring their designs, procedures, and stimuli as faithfully as possible to the LLM context using in-context learning.",
+        "bbox": [
+            115,
+            839,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Source Experiments",
+        "text_level": 1,
+        "bbox": [
+            512,
+            85,
+            712,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first briefly summarize the experimental setup employed by Johnson and Arnold (2022). In each experimental setting, participants heard a series of two-sentence stories in which the first sentence contained two characters with gender-marked first names (Matt or Will for men, Liz or Ana for women). For each story, participants answered a content question to check comprehension, and then a reference question to check pronominal interpretation. In order to lower the ceiling and keep participants from falling into a pattern of simply answering \"yes,\" reference questions were equally split between default and non-default phrasings (e.g. between the subject/non-subject interpretations in Experiment 1a and 2a and between the source/goal interpretations in Experiment 2b). Figure 1 illustrates sample stimuli and the procedure of the three experiments conducted.",
+        "bbox": [
+            512,
+            108,
+            882,
+            395
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In each experiment participants were first shown a series of stories with exposure reference questions to establish a referential pattern; in these, the characters had different genders so pronoun interpretation was unambiguous. After the exposure phase (20 stories in Experiment 1, 12 stories in Experiment 2), further stories with unambiguous exposure questions were intermixed with 12 stories accompanied by critical reference questions; in these stories, the characters had the same gender, so pronoun interpretation was ambiguous and required reliance on discourse cues.",
+        "bbox": [
+            512,
+            399,
+            882,
+            590
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Across experiments, there were two key conditions that were manipulated for exposures. Under a subject exposure condition, the unambiguous intended referents of all exposure questions are subjects of the preceding clauses; in the corresponding non-subject exposure condition, they are the objects of the preceding clauses. Similarly, under a goal exposure condition the unambiguous intended referents of all exposure questions are goal referents while in a source exposure condition they are source referents.",
+        "bbox": [
+            512,
+            595,
+            882,
+            769
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Aiming to transfer these experiments as faithfully as possible to the LLM context, we used this experimental paradigm to evaluate the LLM by providing the model with the full text of each story prompt and content/reference question, then generated tokens in response which we interpreted as answers.",
+        "bbox": [
+            512,
+            772,
+            882,
+            883
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For clarity, we use the same experiment numbers and identifiers as in Johnson and Arnold (2022).",
+        "bbox": [
+            512,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "9495",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/13ba743cefbd27c4c0e3f7227491bf50163a4ae77168074819d8901207107d2e.jpg",
+        "image_caption": [
+            "Figure 1: Illustrated Experimental Procedure. Closely following the setup with human participants in Johnson and Arnold (2022), LLMs were primed via exposure to story/question pairs with unambiguous referents, and tested for their responses on ambiguous target pairs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            89,
+            877,
+            379
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Note that we did not replicate experiments 1b and 1c, which investigated whether people are still sensitive to referential patterns using different types of referential expressions (e.g., third-person names and first- and second-person pronouns). While these experiments provided insights into the linguistic structure at which people generalize referential biases, they were less relevant to the objective of this study and thus were not included.",
+        "bbox": [
+            112,
+            464,
+            487,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Experiment 1a In this experiment, all story prompts contain \"joint-action\" verbs using \"with\" in the form \"X did something with Y.\" Since these verbs lack a clear semantic bias (e.g., Arnold et al. 2018), this context allows us to evaluate LLMs' sensitivity to syntax-based biases in discourse by asking whether exposure to only subject-bias or object-bias examples will influence following answers on the ambiguous critical items. If LLMs are sensitive to syntax-based referential patterns, we expect more subject responses under the subject exposure condition and more non-subject responses under the non-subject exposure condition.",
+        "bbox": [
+            112,
+            620,
+            489,
+            829
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Experiment 2a Experiment 2a forms a bridge between adaptation to syntactic and semantic constraints. Are LLMs able to track patterns in both categories, for instance learning an exposure bias in one category while ignoring variation in the",
+        "bbox": [
+            112,
+            839,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "other? In this experiment, all story prompts contain source/goal verbs like \"give\" and \"receive,\" but these are distributed equally throughout exposures. The manipulation remains the same as Experiment 1a, in which LLMs are exposed to consistent and unambiguous subject interpretations in the subject exposure condition and non-subject interpretations in the non-subject exposure condition.",
+        "bbox": [
+            507,
+            464,
+            884,
+            594
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Experiment 2b This experiment focuses solely on source/goal biases, in which all story prompts contain source/goal words, but the unambiguous exposure items are manipulated to contain only source references in the source exposure condition and only goal references in the goal exposure condition.",
+        "bbox": [
+            507,
+            607,
+            884,
+            719
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 In-context Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            734,
+            714,
+            751
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We propose that since these experiments rely on short-term learning effects of exposure in an experimental context, they can only be effectively simulated with LLMs by using in-context learning recursively. Specifically, for each question, the model is provided access to all previous items in answering a new question, including its own previous responses. This is intended to mirror the process of human experimental participants making judgments in the light of recent exposure to input and",
+        "bbox": [
+            507,
+            758,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "9496",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/7eac8738a3744d17f8a5c6e9927d8cc4f4b3c45a032fa5cbeb0d11ac19659747.jpg",
+        "image_caption": [
+            "Figure 2: The ICL Simulation Framework. Black text is provided to the models as prompts, red text is generated by the model. With the last step's output appended to previous prompts, we ask LLMs the next question."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            84,
+            875,
+            363
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "their own past responses.",
+        "bbox": [
+            112,
+            430,
+            302,
+            445
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We manually checked the correctness of LLM responses to content questions intended to check comprehension, and as in the human experimentation context removed answers for which the LLMs provided incorrect answers. The answers to critical target questions were recorded for further statistical analysis. Our ICL procedure is shown in Figure 2.",
+        "bbox": [
+            112,
+            447,
+            485,
+            558
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Models and Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            573,
+            430,
+            589
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We used text-davinci-003 from the OpenAI API and open-sourced FLAN-UL2 as the LLMs of interest. Though these models are of course not exhaustive of the current landscape of LLMs, they provide some diversity since they differ in both structure (decoder-only vs. encoder-decoder) and parameter count (175B vs. 20B). To introduce more randomness into the experiment and allow enough sample sizes for statistical comparisons across conditions, we made slight modifications to the temperature hyperparameter on each run. Specifically, we assigned each run a random and unique temperature value between 0.2 to 1.0.",
+        "bbox": [
+            112,
+            596,
+            487,
+            803
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We also attempted to simulate 'participants' using a natural language prompt to approach different speaker identities following a similar methodology to Aher et al. (2022). We developed prompts with slots for titles, names, and country of origin to establish different character backgrounds simulating native English speakers from the United",
+        "bbox": [
+            112,
+            806,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "States, Britain, and Australia, following the participants' demographics in Johnson and Arnold (2022). However, these prompts did not induce greater diversity in responses than temperature modification, so only results using temperature modification are presented below. We present a further analysis for both methodologies in Appendix B.",
+        "bbox": [
+            507,
+            430,
+            884,
+            542
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the end, we simulated 24 'participants' each in Experiment 1a and Experiment 2a, and 60 'participants' in Experiment 2b. We included more in Experiment 2b because this experiment has a lower response variability and thus needs more data points for the statistical analysis.",
+        "bbox": [
+            507,
+            543,
+            882,
+            640
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Measures",
+        "text_level": 1,
+        "bbox": [
+            507,
+            653,
+            633,
+            669
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Following the analytic approach of Johnson and Arnold (2022), we used a regression-based approach to analyze whether the responses of LM are consistent with the subject or gogoalal bias of the context for each experiment. We can then compare our findings with theirs by asking whether the effect of the main predictors is the same. In all experiments, predictors included exposure type (unambiguous exposures to subject/non-subject or goal/source), reference question type (whether the reference question is asked about the subject/non-subject or goal/source), and the interaction effect between them. Experiment 2 included verb type, as well as its two-way and three-way interaction effects with the other two predictors as addi",
+        "bbox": [
+            507,
+            677,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "9497",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/b713dc53c9188dfb9abdb356f8e484704ba80437250132cd8b2c76b4f04ec6af.jpg",
+        "image_caption": [
+            "Figure 3: Subject responses of LLM (left) and human participants (right) in Experiment 1a, showing the percentage of subject responses for each type of reference question (Subj Q: subject reference question; Obj Q: object reference question), grouped by exposure type (subject exposure: light blue; object exposure: dark blue)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            85,
+            495,
+            350
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1af2893557bc50ad7b7367e24ba0d91c24fde990f14be4dfa9721e0516a3eb0d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            497,
+            85,
+            880,
+            350
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "tional predictors. Given our small dataset, the results were analyzed using Bayesian mixed-effects Bernoulli logistic regression models in the R package brms (Bürkner, 2017) instead of a frequentist model. We report a Bayesian equivalent p-value (p_MAP) computed with the R package bayestestR (Makowski et al., 2019) to offer a straightforward interpretation of the results. Details of models are provided in Appendix 1.",
+        "bbox": [
+            110,
+            432,
+            487,
+            576
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            593,
+            213,
+            607
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 FLAN-UL2",
+        "text_level": 1,
+        "bbox": [
+            112,
+            621,
+            252,
+            636
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We found that FLAN-UL2 was not capable of generating meaningful output for analysis under this design. First, FLAN-UL2 showed a much higher false rate in answering content questions versus InstructGPT: while InstructGPT replied  $100\\%$  correctly to these questions, FLAN-UL2 replied to only  $57\\%$  correctly. Second, for the ambiguous target items, FLAN-UL2 answered 'yes'  $100\\%$  of the time, indicating an extremely strong bias towards simply answering 'yes' and producing no meaningful variation. By contrast, InstructGPT answered 'yes' to target questions  $68\\%$  of the time, suggesting some amount of 'yes' bias but to a much weaker degree. From these findings we concluded that FLAN-UL2 did not produce sufficiently clean outputs for analysis; therefore, in the following sections, we will focus on results from InstructGPT.",
+        "bbox": [
+            112,
+            645,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Experiment 1a",
+        "text_level": 1,
+        "bbox": [
+            507,
+            432,
+            672,
+            447
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Experiment 1a asks whether LLMs are sensitive to the frequency of referential patterns when subject referents are preferred. Figure 3 compares the subject responses of our results with Johnson and Arnold (2022)'s results. In both LLM and human data, we saw fewer subject responses in object exposure than in subject exposure. This is confirmed by the main effect of exposure type  $(p\\_MAP < .001)$  revealed in the statistical analyses. There was also a main effect of response question type  $(p\\_MAP = .009)$ . As seen from the figure, there were more subject responses in the subject than object reference questions in both human and LLM data. This is because InstructGPT did answer 'yes' more frequently in general, which led to more subject interpretations when the question asked about the subject (where answering 'yes' indicates a subject interpretation), and fewer subject interpretations when the question was asked about the non-subject (where answering 'yes' indicates a non-subject interpretation).",
+        "bbox": [
+            505,
+            463,
+            884,
+            800
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "However, we did not find any interaction effect between exposure type and reference question type. While Johnson and Arnold (2022) found that exposure type has a significant effect for subject-referent questions but only a marginal effect for the nonsubject-referent questions, the exposure effect was significant for both question types in our",
+        "bbox": [
+            507,
+            806,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "9498",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/085f18e929dd4217fa850e839c7f55771276ed37d3200866b45e5b0bc40d304c.jpg",
+        "image_caption": [
+            "Figure 4: Subject responses of LLM (left) and human participants (right) in Experiment 2a. The proportion of subject responses is plotted against question type faceting by verb type, comparing goal-subject verbs (gs) like \"receive,\" where the subject is the goal referent) to subject-goal verbs (sg) like \"send,\" where the subject is the source referent."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            104,
+            495,
+            350
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2a343e7c521da318b156bebc78be20759d8c7cde2c27856af3ef2c79d08ed9e0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            105,
+            877,
+            350
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "results. InstructGPT is sensitive to exposure type even for nonsubject-referent questions, as reflected in Figure 3: there were no subject responses when non-subject reference questions were asked under the non-subject exposure condition. This suggests that LLM may even be more sensitive to the exposure effect than humans. Overall, Experiment 1 shows that LLM can indeed learn from recent exposure to syntactically-oriented referential patterns, in a relatively more non-subject-biased way.",
+        "bbox": [
+            112,
+            445,
+            489,
+            607
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Experiment 2a",
+        "text_level": 1,
+        "bbox": [
+            112,
+            621,
+            275,
+            638
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We ask in Experiment 2a if the sensitivity to subject reference patterns observed in Experiment 1a persists independent of semantic variability in the local context. Figure 4 illustrates the responses of LLM and human participants by exposure, reference question, and verb type. The behavior of InstructGPT is in line with what Johnson and Arnold (2022) found in human participants for Experiment 2a: as in Experiment 1, statistical analyses revealed a significant main effect of referent question type  $(p\\_ MAP = .034)$ , despite a marginally significant effect of exposure type  $(p\\_ MAP = .085)$ . InstructGPT did understand the pronoun as subject referents more after subject exposure and said 'yes' more often such that there were more subject interpretations when the critical question was asked about the subject. InstructGPT also showed a goal",
+        "bbox": [
+            112,
+            645,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "bias: there was a main effect of verb type ( $p\\_MAP = .001$ ), such that it referred to the subject more with gs verbs (where the subject is the goal referent) than sg verbs (where the subject is the source referent). We observed no interaction effect.",
+        "bbox": [
+            507,
+            445,
+            884,
+            525
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In spite of these similarities, we still observe differences between LLM and human participants. Notice in Figure 4 that there is a larger difference between gs and sg verbs in InstructGPT than in human participants while keeping other conditions constant, suggesting that LLMs may have a larger goal bias than humans.",
+        "bbox": [
+            507,
+            527,
+            882,
+            639
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Experiment 2b",
+        "text_level": 1,
+        "bbox": [
+            507,
+            653,
+            672,
+            670
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Experiment 2b examines whether InstructGPT is still sensitive to exposure to referential patterns that exhibit consistent preferences for a source or goal referent. Figure 5 compares the results from InstructGPT and human participants. Whereas Johnson and Arnold (2022) reported significant main effects from exposure type, reference question type, and verb type, as well as a marginal interaction effect between verb type and exposure, we did not find any effect from these predictors, but only an interaction effect between verb type and question type  $(p\\_ MAP < .001)$ . We further examined the effect of question type and exposure type for each verb type, but no significant effect was found for either predictor. These results suggest that exposure",
+        "bbox": [
+            505,
+            677,
+            884,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "header",
+        "text": "exposure subject object",
+        "bbox": [
+            415,
+            87,
+            581,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "9499",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7b07beb2cfb69b66120318acf4ae405d8c038ba1de7b9b4c1419c4cb6a66fb68.jpg",
+        "image_caption": [
+            "Figure 5: Goal responses of LLM (left) and human participants (right) in Experiment 2b."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            85,
+            499,
+            350
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/06201297b118d78ad080523bf1b86e5e6b261e1f6f037b5c0ba200c47b7f648f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            85,
+            878,
+            350
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "did not necessarily change InstructGPT's behaviors: for goal questions, both the goal and source exposure conditions led to  $100\\%$  goal responses. Reference question type also did not affect InstructGPT's responses, because it almost always interprets the pronoun as the goal referent under subject exposure. The lack of any significant effects, and indeed our observed universal goal interpretations to goal-focused reference questions, suggest that InstructGPT displays an extreme goal bias in pronoun interpretation.",
+        "bbox": [
+            112,
+            403,
+            489,
+            580
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            596,
+            240,
+            613
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The present study examined whether and to what extent LLMs display adaptation for pronominal interpretation after exposure to referential patterns by replicating three psycholinguistic experiments. Exposed to the same stimuli and study design, and analyzed with the same statistical procedures as the source experiments, we asked whether LLMs show human-like behaviors and compared the performance of LLMs-simulated participants with human participants.",
+        "bbox": [
+            112,
+            627,
+            489,
+            788
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We firstly found a difference between the capacity of contemporary models to replicate human psycholinguistic experimental designs in an ICL framework. While InstructGPT was able to correctly answer all the presented comprehension-check content questions, FLAN-UL2 was not. Both models displayed at least some bias towards answering 'yes' in ambiguous cases, but in the case of",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "FLAN-UL2 this bias resulted in  $100\\%$  'yes' answers, rendering meaningful variation impossible to ascertain. This could be a result of structural differences between the models (decoder-only in the case of InstructGPT, encoder-decoder in the case of FLAN-UL2), or perhaps more likely a question of simple model size (175B for InstructGPT vs. 20B for FLAN-UL2).",
+        "bbox": [
+            507,
+            403,
+            884,
+            531
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Experiments 1a and 2a examined whether LLMs adapt their syntactic bias from recent exposure to referential patterns like humans, without and with the presence of possibly confounding semantic goal bias. We found that LLM's referential biases are indeed sensitive to such exposure in both experiments. In addition, LLM did exhibit a goal bias in Experiment 2a, in line with previous studies which argue for LLM's ability to exhibit human-like semantic bias in pronoun interpretation (e.g., Davis and van Schijndel 2020).",
+        "bbox": [
+            507,
+            533,
+            884,
+            709
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Experiment 2b examined whether LLM can adapt and vary their semantic bias from exposures. In this context, in contrast with the previous two experiments, exposure type did not affect LLM behavior at all. This raises the question of why LLM would be sensitive to exposure to only referential patterns that exhibit consistent syntactic bias but not semantic bias. An immediate and intuitive answer would be that LLM is unable to fully represent semantic knowledge in referential processing. However, given that LLM did display a human-like goal bias in pronoun interpretation independent of exposure, this explanation seems unlikely.",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "9500",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We suggest that our work provides evidence that LLM only partially represents the semantic knowledge involved in referential processing for two reasons.",
+        "bbox": [
+            112,
+            84,
+            489,
+            145
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "First, adapting referential biases on the basis of exposure in the local context may require representations of higher-level knowledge than merely exhibiting a bias towards certain referents. While the latter may only require knowledge of which features associated with a referent are more frequent or likely in general, adaption requires a sustained awareness of referential pattern frequency as it changes in the local discourse context. Though representing knowledge about semantic relations was observed as early as the analogical reasoning task in Word2Vec (Mikolov et al., 2013), human-like extraction of abstract information like a persistent discourse state is more challenging. The model may only be able to identify thematic roles (source or goal) of a referent and associate them with pronoun interpretation, but not to identify a consistent pattern of thematic roles across a discourse. If so, this would explain the strong goal bias we observed in Experiment 2.",
+        "bbox": [
+            112,
+            149,
+            489,
+            470
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Second, it is possible that LLM has an extremely strong goal bias that masks the influence of exposures. If so, this suggests that LLM over-represents the semantic knowledge encoded in pronoun interpretation.",
+        "bbox": [
+            112,
+            473,
+            489,
+            552
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In either case, our results suggest that LLM do not encode semantic knowledge in a fully human-like way, even though they do demonstrate some human-like capacities for semantic understanding. Although we believe this gap can be mitigated via instruction fine-tuning or chain-of-thought prompting (Wei et al., 2022), these results still suggest we should consider incorporating semantically-informed objectives into self-supervised pre-training to a greater extent.",
+        "bbox": [
+            112,
+            554,
+            489,
+            715
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusions and Future Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            730,
+            410,
+            745
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "By replicating a series of psycholinguistic experiments as closely as possible using in-context learning, this paper pioneered whether LLMs would adapt pronominal interpretation behaviors in a human-like way given exposure to referential patterns in the local discourse context. Our work suggests paths forward for replicating psycholinguistic experiments in a more faithful way that allows for comparisons between human and LLMs' behaviors.",
+        "bbox": [
+            112,
+            758,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            615,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Several main limitations exist in our study in its current form. First, our reported results only simulated experimental participants by manipulating the temperature hyperparameter. We compared this approach with natural language prompting for Experiment 1, but that prompting did not increase \"participant\" diversity, so it was abandoned. Moreover, approaches for simulating psycholinguistic experimental \"participants\" could go far beyond what was tried here; our prompting method was relatively limited, and more detailed prompting could be included in future experimental simulations. Second, making a direct comparison with actual psycholinguistic experiments might not be the only method to investigate LLMs' discourse capacity. A comprehensive list of discourse probing tasks might play a similar role despite a different way (Koto et al., 2021). Third, this study is strictly behavioral: limited by both computational resources and obscure mechanisms of in-context learning, we do not dive into models' internal representations in our analyses.",
+        "bbox": [
+            507,
+            111,
+            884,
+            464
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            480,
+            672,
+            495
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We are very grateful to Jennifer Arnold for sharing the stimuli and design of Johnson and Arnold (2022) with us.",
+        "bbox": [
+            507,
+            508,
+            882,
+            555
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            586,
+            608,
+            601
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Gati Aher, Rosa I. Arriaga, and Adam Tauman Kalai. 2022. Using large language models to simulate multiple humans.",
+            "Mira Ariel. 1990. Accessing noun-phrase antecedents. Routledge.",
+            "Jennifer E Arnold. 1998. Reference form and discourse patterns. Stanford University.",
+            "Jennifer E Arnold. 2001. The effect of thematic roles on pronoun use and frequency of reference continuation. Discourse processes, 31(2):137-162.",
+            "Jennifer E Arnold, Iris M Strangmann, Heeju Hwang, Sandra Zerkle, and Rebecca Nappa. 2018. Linguistic experience affects pronoun interpretation. Journal of Memory and Language, 102:41-54.",
+            "Dale J Barr, Roger Levy, Christoph Scheepers, and Harry J Tily. 2013. Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of memory and language, 68(3):255-278."
+        ],
+        "bbox": [
+            509,
+            609,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9501",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Holly P Branigan, Martin J Pickering, and Janet F McLean. 2005. Priming prepositional-phrase attachment during comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31(3):468.",
+            "Susan E Brennan. 1995. Centering attention in discourse. Language and Cognitive processes, 10(2):137-167.",
+            "Paul-Christian Bürkner. 2017. brms: An r package for bayesian multilevel models using stan. Journal of statistical software, 80:1-28.",
+            "Franklin Chang, Gary S Dell, Kathryn Bock, and Zenzi M Griffin. 2000. Structural priming as implicit learning: A comparison of models of sentence production. Journal of psycholinguistic research, 29(2):217-230.",
+            "Forrest Davis and Marten van Schijndel. 2020. Discourse structure interacts with reference but not syntax in neural language models. arXiv preprint arXiv:2010.04887.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Richard Futrell, Ethan Wilcox, Takashi Morita, Peng Qian, Miguel Ballesteros, and Roger Levy. 2019. Neural language models as psycholinguistic subjects: Representations of syntactic state. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 32-42, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195-1205, New Orleans, Louisiana. Association for Computational Linguistics.",
+            "Jennifer Hu, Jon Gauthier, Peng Qian, Ethan Wilcox, and Roger Levy. 2020. A systematic assessment of syntactic generalization in neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1725-1744, Online. Association for Computational Linguistics.",
+            "Elyce D Johnson and Jennifer E Arnold. 2022. The frequency of referential patterns guides pronoun comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Fajri Koto, Joy Han Lau, and Timothy Baldwin. 2021. Discourse probing of pretrained language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3849-3864, Online. Association for Computational Linguistics.",
+            "Dominique Makowski, Mattan S Ben-Shachar, and Daniel Lüdecke. 2019. bayestestr: Describing effects and their uncertainty, existence and significance within the bayesian framework. Journal of Open Source Software, 4(40):1541.",
+            "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc.",
+            "Kanishka Misra, Allyson Ettinger, and Julia Rayz. 2020. Exploring BERT's sensitivity to lexical cues using tests from semantic priming. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4625-4635, Online. Association for Computational Linguistics.",
+            "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback.",
+            "Jenny R Saffran, Richard N Aslin, and Elissa L Newport. 1996a. Statistical learning by 8-month-old infants. Science, 274(5294):1926-1928.",
+            "Jenny R Saffran, Elissa L Newport, and Richard N Aslin. 1996b. Word segmentation: The role of distributional cues. Journal of memory and language, 35(4):606-621.",
+            "Sebastian Schuster and Judith Degen. 2020. I know what you're probably going to say: Listener adaptation to variable use of uncertainty expressions. Cognition, 203:104285.",
+            "Arabella Sinclair, Jaap Jumelet, Willem Zuidema, and Raquel Fernández. 2022. Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations. Transactions of the Association for Computational Linguistics, 10:1031-1050.",
+            "Ionut-Teodor Sorodoc, Kristina Gulordava, and Gemma Boleda. 2020. Probing for referential information in language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4177–4189, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "9502",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. 2023. UI2: Unifying language learning paradigms.",
+        "bbox": [
+            115,
+            85,
+            487,
+            165
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. CoRR, abs/1905.05950.",
+        "bbox": [
+            114,
+            173,
+            489,
+            212
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Shiva Upadhye, Leon Bergen, and Andrew Kehler. 2020. Predicting reference: What do language models learn about discourse models? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 977-982, Online. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            222,
+            489,
+            302
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models.",
+        "bbox": [
+            114,
+            310,
+            489,
+            365
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Statistical Information",
+        "text_level": 1,
+        "bbox": [
+            114,
+            375,
+            347,
+            390
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "For each experiment, we analyzed referent choice (Subject = 1, Non-subject = 0) using a mixed-effects Bernoulli regression model from the R package brms (Bürkner, 2017), with the maximal random structure justified by design (Barr et al., 2013). Predictors are coded in the same way as in Johnson and Arnold (2022). All models were specified with a weakly informative prior using the Cauchy distribution with center 0 and scale 2.5. Models were fitted using six chains, each with 4,000 iterations of which the first 1,000 are warmup to calibrate the sampler, resulting in 18,000 posterior examples.",
+        "bbox": [
+            112,
+            400,
+            489,
+            593
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The model for Experiment 1a included question type (QtypeC, sum-coded: Subject = 0.5, Non-subject = -0.5) and exposure type (PC, effects-coded: Subject-biased = 0.51, Object-biased = -0.49) as fixed predictors, random intercepts for participants and items, random slopes of question type, exposure type and their interaction for items, and a random slope of question type for the participant. Exposure type was not included as a random slope for participants because the condition does not vary within participants.",
+        "bbox": [
+            112,
+            593,
+            489,
+            770
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "brm (Rc ~ QtypeC*PC +  \n(1+QtypeC*PC|ID)+  \n(1+QtypeC|Subject),  \ndata=(e1a, Exposure!= \"None\"),  \nfamily=\"bernoulli\",  \nchains=6,  \niter=4000,  \nwarmup=1000,  \ncontrol =",
+        "guess_lang": "txt",
+        "bbox": [
+            112,
+            775,
+            448,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "list(adapt delta = 0.95),  \nprior=  \nc(set_prior (\"cauchy(0,2.5)\"))",
+        "guess_lang": "txt",
+        "bbox": [
+            581,
+            84,
+            880,
+            133
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The model for Experiment 2a included question type (QtypeC, sum-coded: Subject = 0.5, Non-subject = -0.5), exposure type (PC, sum-coded: Subject-biased = 0.5, Object-biased = -0.5), and verb type (Vc, sum-coded: gs-verb = 0.5, sg-verb = -0.5) as fixed predictors, random intercepts for participants and items, random slopes of question type, exposure type and their interaction for items, and a random slope of question type for the participant. As in Experiment 1a, exposure type was not included as a random slope for participants because the condition does not vary within participants. Similarly, verb bias was not included as a random slope for items here because it does not vary within items.",
+        "bbox": [
+            507,
+            145,
+            884,
+            387
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "brm (Rc ~ QtypeC*PC*Vc +  \n(1+PC*Qtypec|Item)+  \n(1+QtypeC*Vc|Subject),  \ndata=e2a,  \nfamily=\"bernoulli\", chains=6,  \niter=4000, warmup=1000,  \ncontrol = list(adapt delta = 0.98),  \ncores = 6,  \nprior=  \nc(set_prior (\"cauchy(0,2.5)\"))",
+        "guess_lang": "txt",
+        "bbox": [
+            544,
+            404,
+            884,
+            565
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The model for Experiment 2b included question type (QtypeC, sum-coded: Goal = 0.5, Source = -0.5), exposure type (PC, sum-coded: Goal-biased = 0.5, Source-biased = -0.5), and verb type (Vc, sum-coded: gs-verb = 0.5, sg-verb = -0.5) as fixed predictors. The random effect structure was the same as that of Experiment 2a.",
+        "bbox": [
+            505,
+            577,
+            882,
+            690
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "brm (Rc ~ PC\\*Vc\\*QtypeC  \n+ (1+PC\\*QtypeC|Item)  \n+ (1+Vc\\*QtypeC|Subject),  \ndata=e2b,  \nfamily=\"bernoulli\",  \nchains=6,  \niter=4000,  \nwarmup=1000,  \ncontrol =  \n    list(adapt delta = 0.999),  \ncores = 6,  \nprior=  \n    c(set_prior (\"cauchy(0,2.5)\"))",
+        "guess_lang": "python",
+        "bbox": [
+            554,
+            709,
+            882,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "9503",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Temperature vs. Prompt",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            364,
+            101
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Our prompt-based simulation of multiple participants embedded names, countries, prefixes, and genders into a carrier sentence:  $\\{Prefix + Name\\}$  is a native English speaker living in {Country}. {Gender} is asked in a psycholinguistic experiment to answer the following questions. For example, Mr. Smith is a native English speaker living in England. He is asked in a psycholinguistic experiment to answer the following questions.",
+        "bbox": [
+            112,
+            109,
+            489,
+            254
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we calculated the variance and ran the Levene's test for any significant difference between humans and InstructGPT in each experiment. In experiment 1, the human responses' variance is 0.055, the temperature-based responses' variance is 0.024, and the prompt-based responses' variance is 0.017. Human responses were significantly higher than both (Temperature-based:  $p = .049$ ; Prompt-based:  $p = .007$ ). Yet, temperature-based were not significantly higher than prompt-based. Due to the limitation of API pricing, we only ran temperature-based in the following experiments. In experiment 2a, the human responses' variance is 0.034 and the temperature-based responses' variance is 0.043. Yet, Levene's test did not reveal any significant difference. In Experiment 2b, the human responses' variance is 0.020 and the temperature-based responses' variance is 0.008 ( $p < .001$ ).",
+        "bbox": [
+            115,
+            254,
+            489,
+            543
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We used different techniques to introduce randomness and include more experimental data in our experiments. We realized these were not well-designed prompts to elicit different linguistic backgrounds. Given the lack of investigation on simulating multiple participants in psycholinguistics studies, we recognize this as a future direction of possible work.",
+        "bbox": [
+            112,
+            545,
+            489,
+            671
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "9504",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            129,
+            126,
+            532,
+            143
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Last section without the number",
+        "bbox": [
+            149,
+            143,
+            391,
+            158
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            127,
+            170,
+            552,
+            186
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            186,
+            349,
+            200
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            129,
+            212,
+            695,
+            229
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The introduction is in the first section.",
+        "bbox": [
+            149,
+            230,
+            431,
+            244
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            129,
+            256,
+            668,
+            272
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            215,
+            336
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            127,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            363,
+            248,
+            379
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            127,
+            390,
+            778,
+            406
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            407,
+            248,
+            422
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            127,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            498,
+            248,
+            513
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            127,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            573,
+            248,
+            588
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            127,
+            599,
+            880,
+            631
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            632,
+            248,
+            646
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            127,
+            658,
+            880,
+            739
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            740,
+            248,
+            753
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            494,
+            781
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Experiments were described in section 3.",
+        "bbox": [
+            132,
+            787,
+            436,
+            801
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            812,
+            880,
+            845
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "In the method section, we described our parameters.",
+        "bbox": [
+            149,
+            846,
+            539,
+            860
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_footnote",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "9505",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            127,
+            84,
+            878,
+            115
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. We did not do a hyperparameter search.",
+        "bbox": [
+            149,
+            117,
+            564,
+            131
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            129,
+            142,
+            882,
+            190
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "We provide detailed results in the results section.",
+        "bbox": [
+            149,
+            192,
+            510,
+            206
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            129,
+            217,
+            882,
+            265
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "We provide existing packages' information in the methods section.",
+        "bbox": [
+            149,
+            267,
+            636,
+            282
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            127,
+            340,
+            882,
+            372
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            373,
+            248,
+            388
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            127,
+            399,
+            882,
+            447
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            449,
+            248,
+            464
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            127,
+            475,
+            882,
+            521
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            524,
+            248,
+            539
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            127,
+            549,
+            873,
+            564
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            565,
+            248,
+            581
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            127,
+            592,
+            880,
+            623
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            626,
+            248,
+            640
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "9506",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/Large Language Models Are Partially Primed in Pronoun Interpretation/60280f41-b321-4144-822d-79720ae15459_model.json

@@ -0,0 +1,2373 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.122,
+                0.09,
+                0.877,
+                0.112
+            ],
+            "angle": 0,
+            "content": "Large Language Models Are Partially Primed in Pronoun Interpretation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.12,
+                0.124,
+                0.882,
+                0.144
+            ],
+            "angle": 0,
+            "content": "Suet-Ying Lam†* Qingcheng Zeng‡* Kexun Zhang◇* Chenyu You* Rob Voigt‡"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.316,
+                0.158,
+                0.686,
+                0.175
+            ],
+            "angle": 0,
+            "content": "† Department of Linguistics, UMass Amherst"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.284,
+                0.175,
+                0.72,
+                0.192
+            ],
+            "angle": 0,
+            "content": "\\(\\ddagger\\) Department of Linguistics, Northwestern University"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.325,
+                0.193,
+                0.678,
+                0.209
+            ],
+            "angle": 0,
+            "content": "\\(\\diamond\\) Department of Computer Science, UCSB"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.272,
+                0.209,
+                0.73,
+                0.227
+            ],
+            "angle": 0,
+            "content": "\\(\\clubsuit\\) Department of Electrical Engineering, Yale University"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.282,
+                0.461,
+                0.666
+            ],
+            "angle": 0,
+            "content": "While a large body of literature suggests that large language models (LLMs) acquire rich linguistic representations, little is known about whether they adapt to linguistic biases in a human-like way. The present study probes this question by asking whether LLMs display human-like referential biases using stimuli and procedures from real psycholinguistic experiments. Recent psycholinguistic studies suggest that humans adapt their referential biases with recent exposure to referential patterns; closely replicating three relevant psycholinguistic experiments from Johnson and Arnold (2022) in an in-context learning (ICL) framework, we found that InstructGPT adapts its pronominal interpretations in response to the frequency of referential patterns in the local discourse, though in a limited fashion: adaptation was only observed relative to syntactic but not semantic biases. By contrast, FLAN-UL2 fails to generate meaningful patterns. Our results provide further evidence that contemporary LLMs discourse representations are sensitive to syntactic patterns in the local context but less so to semantic patterns. Our data and code are available at https://github.com/zkx06111/1lm_priming."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.681,
+                0.26,
+                0.696
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.708,
+                0.491,
+                0.886
+            ],
+            "angle": 0,
+            "content": "While neural network models, and particularly pretrained large language models have shown excellent performance at particular language processing tasks, many questions remain about the extent to which models optimized for such performance encode, as a side effect, human-like linguistic knowledge and cognitive biases. We know that they do to some extent; existing work has shown, for example, that neural models encode aspects of human-like long-distance number agreement (Gulordava et al., 2018), incremental syntactic state (Futrell et al.,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.253,
+                0.885,
+                0.364
+            ],
+            "angle": 0,
+            "content": "2019), and syntactic generalization more broadly (Hu et al., 2020). In this paper, we examine whether FLAN-UL2 (Tay et al., 2023) and InstructGPT (Ouyang et al., 2022), two representative LLMs, display adaptation in pronoun interpretation when exposed to consistent referential patterns in the local discourse context."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.367,
+                0.885,
+                0.608
+            ],
+            "angle": 0,
+            "content": "Compared with syntactic or lexical knowledge, representing referential knowledge is possibly more complex; we know from psycholinguistic studies that human referential interpretation integrates multiple levels of linguistic structure. Humans do not interpret ambiguous pronouns at random but are guided by both syntactic and semantic information. It is well-established that absent other cues humans prefer a syntactic subject in choosing the antecedent of the ambiguous pronoun, i.e., subject bias (Ariel, 1990; Brennan, 1995). In example (1), she is more likely to be interpreted as the subject Ada than the non-subject Eva \\(\\mathbb{I}\\) , even though both referents are possible antecedents for the pronoun."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.516,
+                0.621,
+                0.776,
+                0.637
+            ],
+            "angle": 0,
+            "content": "(1) \\(\\mathrm{Ada}_1\\) talked with \\(\\mathrm{Eva}_2\\). She..."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.65,
+                0.885,
+                0.794
+            ],
+            "angle": 0,
+            "content": "People are also sensitive to the semantic structure of the sentence when choosing an antecedent for an ambiguous pronoun, in addition to syntactic information. In a transfer event that depicts a transfer-of-possession from one entity (the source) to another (the goal), they prefer the goal referent (Ada in (2), Eva in (3)) over the source referent (Eva in (2), Ada in (3)) to be the antecedent (Arnold, 2001, 1998), i.e., goal bias:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.516,
+                0.807,
+                0.717,
+                0.822
+            ],
+            "angle": 0,
+            "content": "(2) Goal-source (gs) verb:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.547,
+                0.824,
+                0.85,
+                0.839
+            ],
+            "angle": 0,
+            "content": "Ada<sub>1</sub> received a letter from Eva<sub>2</sub>. She<sub>1</sub>..."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.516,
+                0.852,
+                0.715,
+                0.867
+            ],
+            "angle": 0,
+            "content": "(3) Source-goal (sg) verb:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.547,
+                0.869,
+                0.797,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Ada1 sent a letter to Eva2. She2..."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Indices of pronouns in examples indicate the preferred referent."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9493"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.229,
+                0.946,
+                0.769,
+                0.959
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2023, pages 9493-9506"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.297,
+                0.96,
+                0.702,
+                0.973
+            ],
+            "angle": 0,
+            "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.213
+            ],
+            "angle": 0,
+            "content": "In sum, people exhibit sensitivity to both syntax (subject bias) and semantics (goal bias) during pronoun interpretation. Importantly, these levels of linguistic structure are frequently entwined since both can influence referential interpretation. At times these influences may push in different directions: for instance, Ada in [3] is both the syntactic subject and the semantic source."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.218,
+                0.492,
+                0.508
+            ],
+            "angle": 0,
+            "content": "Building on a long tradition investigating preferences in pronoun interpretation, recent psycholinguistic studies have probed into the deeper question of the origin of these biases. One hypothesis is that referential biases come from linguistic experience: when a bias occurs very frequently, people will tend to adapt this more frequent referential pattern, both in the immediate exposure as well as more large-scale past experience (Arnold, 1998, 2001). Recent evidence has provided support to this idea by demonstrating that recent exposure to certain referential patterns did change people's referential biases. In a series of psycholinguistic studies, Johnson and Arnold (2022) show that after reading numerous stories that consistently show a particular referential bias, e.g., always referring to the non-subject or source referent, people did have a stronger preference for these primed referents."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.513,
+                0.49,
+                0.771
+            ],
+            "angle": 0,
+            "content": "Given this line of psycholinguistic research, the current study investigates the extent to which LLMs adapt and vary referential biases in pronoun interpretation through exposure to referential patterns in the local context. To do so we replicated actual psycholinguistic experiments from Johnson and Arnold (2022) in LLMs using an ICL framework and asked whether the responses of LLMs display adaptation from exposure to referential patterns like human experimental participants. Comparing syntactically-motivated to semantically-motivated exposure conditions will allow us to first examine whether LLMs display human-like subject bias and goal bias and further understand the extent to which LLMs make use of local frequency changes in discourse representations in these categories."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.775,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In-context learning refers to LLMs' ability to learn from demonstrations written in natural language prompts. Compared to previous work that has largely examined the encoding of such discourse knowledge using zero-shot inference (Upadhye et al., 2020), ICL is particularly suitable for experimental simulation since it replicates the naturalistic context in which later responses draw upon exposure to previous examples."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.23
+            ],
+            "angle": 0,
+            "content": "Foreshadowing our results, we find that InstructGPT can adapt and thus vary its syntactic bias from exposure to referential patterns in the local context, but the same is not true for semantic bias. Given that InstructGPT still exhibits a goal bias in spite of local discourse priming, we argue this suggests LLMs only encode partial semantic knowledge in referential processing. To sum up, our contributions can be summarized as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.238,
+                0.887,
+                0.335
+            ],
+            "angle": 0,
+            "content": "1) We extended a discourse understanding evaluation to state-of-the-art LLMs from a new perspective, asking whether LLM's referential bias can be modified by exposure to particular referential patterns, like how humans adapt referential bias from experience."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.344,
+                0.885,
+                0.424
+            ],
+            "angle": 0,
+            "content": "2) To the best of our knowledge, we are the first study that replicates actual psycholinguistic experiments using the ICL framework and compares LLMs' behaviors with real human participants."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.433,
+                0.887,
+                0.515
+            ],
+            "angle": 0,
+            "content": "3) We present results in this context showing further evidence that InstructGPT can acquire abstract syntactic knowledge in referential interpretation to some extent, but not semantic knowledge."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.523,
+                0.238,
+                0.887,
+                0.515
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.524,
+                0.666,
+                0.54
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.55,
+                0.885,
+                0.693
+            ],
+            "angle": 0,
+            "content": "A growing body of literature has suggested that LLMs encode rich representations of linguistic structure at various levels, including aspects of syntax, semantics, and reference encoded throughout their representations (Tenney et al., 2019). One of the most well-documented lines of this work demonstrates that these models can acquire diverse elements of syntactic knowledge (Gulordava et al., 2018; Futrell et al., 2019; Hu et al., 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.695,
+                0.885,
+                0.886
+            ],
+            "angle": 0,
+            "content": "This capacity for encoding linguistic understanding extends to priming effects with psycholinguistic analogies to humans. At the syntactic level, Sinclair et al. (2022) explored structural priming in various autoregressive LLMs and found priming effects despite a clear dependence on semantic information. At the semantic level, using English lexical stimuli in BERT (Devlin et al., 2019), Misra et al. (2020) found that BERT does display evidence of sensitivity to semantic priming, though this is localized to more unconstrained contexts and only certain semantic relations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.888,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In analyses of LLMs' linguistic understanding modeled on psycholinguistic experiments, however,"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9494"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.085,
+                0.49,
+                0.357
+            ],
+            "angle": 0,
+            "content": "the question of discourse knowledge remains relatively under-explored. Recent existing work has presented partially conflicting accounts in this area, in particular with regard to how LLMs may or may not exhibit human-like biases in pronoun interpretation. For example, Davis and van Schijndel (2020) compared LSTM LMs and Transformer LMs behaviors and internal representations in dealing with implicit causality verbs, finding that surprisingly (contrary to humans) implicit causality only influences Transformer LMs' behavior for reference, but influences neither model for syntactic attachment. Sorodoc et al. (2020) also compared LSTM LMs and Transformer LMs in coreference resolution corpora, finding that although LMs are much better at grammar, they also captured referential aspects to some extent."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.36,
+                0.49,
+                0.793
+            ],
+            "angle": 0,
+            "content": "This existing work replicates sets of individual stimuli from psycholinguistic experiments in isolation; by contrast, our work takes a more behaviorally-oriented approach by replicating the stimuli, procedure, and even experimental design of a full set of human psycholinguistic studies. We do this to ask a further question: beyond exhibiting baseline human-like referential biases in pronoun interpretation, do LLMs display adaptation to the frequency of referential patterns in the local context like humans? This question is important because adapting to a referential bias over the course of an experiment requires a sustained representation of the frequency of the pattern, which involves a higher-level understanding of the discourse structure. Humans have shown abilities of this kind of adaptation at multiple linguistic levels. For instance, exposure to syntactic structures consecutively affects humans' choice of structures in both the short term and long term via priming effects (e.g., Branigan et al. 2005; Chang et al. 2000). Exposure to phonological patterns can also guide humans' segmentation patterns (e.g., Saffran et al. 1996a,b). At the semantic and pragmatic levels, humans can also adapt as listeners to speakers' variable choices of uncertainty expressions (Schuster and Degen, 2020)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.117,
+                0.811,
+                0.224,
+                0.825
+            ],
+            "angle": 0,
+            "content": "3 Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.84,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this work, we aim to replicate three experiments from Johnson and Arnold (2022), transferring their designs, procedures, and stimuli as faithfully as possible to the LLM context using in-context learning."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.086,
+                0.713,
+                0.1
+            ],
+            "angle": 0,
+            "content": "3.1 Source Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.109,
+                0.884,
+                0.397
+            ],
+            "angle": 0,
+            "content": "We first briefly summarize the experimental setup employed by Johnson and Arnold (2022). In each experimental setting, participants heard a series of two-sentence stories in which the first sentence contained two characters with gender-marked first names (Matt or Will for men, Liz or Ana for women). For each story, participants answered a content question to check comprehension, and then a reference question to check pronominal interpretation. In order to lower the ceiling and keep participants from falling into a pattern of simply answering \"yes,\" reference questions were equally split between default and non-default phrasings (e.g. between the subject/non-subject interpretations in Experiment 1a and 2a and between the source/goal interpretations in Experiment 2b). Figure 1 illustrates sample stimuli and the procedure of the three experiments conducted."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.4,
+                0.884,
+                0.592
+            ],
+            "angle": 0,
+            "content": "In each experiment participants were first shown a series of stories with exposure reference questions to establish a referential pattern; in these, the characters had different genders so pronoun interpretation was unambiguous. After the exposure phase (20 stories in Experiment 1, 12 stories in Experiment 2), further stories with unambiguous exposure questions were intermixed with 12 stories accompanied by critical reference questions; in these stories, the characters had the same gender, so pronoun interpretation was ambiguous and required reliance on discourse cues."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.596,
+                0.884,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Across experiments, there were two key conditions that were manipulated for exposures. Under a subject exposure condition, the unambiguous intended referents of all exposure questions are subjects of the preceding clauses; in the corresponding non-subject exposure condition, they are the objects of the preceding clauses. Similarly, under a goal exposure condition the unambiguous intended referents of all exposure questions are goal referents while in a source exposure condition they are source referents."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.773,
+                0.884,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Aiming to transfer these experiments as faithfully as possible to the LLM context, we used this experimental paradigm to evaluate the LLM by providing the model with the full text of each story prompt and content/reference question, then generated tokens in response which we interpreted as answers."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.888,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "For clarity, we use the same experiment numbers and identifiers as in Johnson and Arnold (2022)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.929,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9495"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.123,
+                0.09,
+                0.878,
+                0.38
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.398,
+                0.884,
+                0.442
+            ],
+            "angle": 0,
+            "content": "Figure 1: Illustrated Experimental Procedure. Closely following the setup with human participants in Johnson and Arnold (2022), LLMs were primed via exposure to story/question pairs with unambiguous referents, and tested for their responses on ambiguous target pairs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.466,
+                0.489,
+                0.61
+            ],
+            "angle": 0,
+            "content": "Note that we did not replicate experiments 1b and 1c, which investigated whether people are still sensitive to referential patterns using different types of referential expressions (e.g., third-person names and first- and second-person pronouns). While these experiments provided insights into the linguistic structure at which people generalize referential biases, they were less relevant to the objective of this study and thus were not included."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.621,
+                0.49,
+                0.83
+            ],
+            "angle": 0,
+            "content": "Experiment 1a In this experiment, all story prompts contain \"joint-action\" verbs using \"with\" in the form \"X did something with Y.\" Since these verbs lack a clear semantic bias (e.g., Arnold et al. 2018), this context allows us to evaluate LLMs' sensitivity to syntax-based biases in discourse by asking whether exposure to only subject-bias or object-bias examples will influence following answers on the ambiguous critical items. If LLMs are sensitive to syntax-based referential patterns, we expect more subject responses under the subject exposure condition and more non-subject responses under the non-subject exposure condition."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Experiment 2a Experiment 2a forms a bridge between adaptation to syntactic and semantic constraints. Are LLMs able to track patterns in both categories, for instance learning an exposure bias in one category while ignoring variation in the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.466,
+                0.885,
+                0.595
+            ],
+            "angle": 0,
+            "content": "other? In this experiment, all story prompts contain source/goal verbs like \"give\" and \"receive,\" but these are distributed equally throughout exposures. The manipulation remains the same as Experiment 1a, in which LLMs are exposed to consistent and unambiguous subject interpretations in the subject exposure condition and non-subject interpretations in the non-subject exposure condition."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.608,
+                0.885,
+                0.72
+            ],
+            "angle": 0,
+            "content": "Experiment 2b This experiment focuses solely on source/goal biases, in which all story prompts contain source/goal words, but the unambiguous exposure items are manipulated to contain only source references in the source exposure condition and only goal references in the goal exposure condition."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.736,
+                0.715,
+                0.752
+            ],
+            "angle": 0,
+            "content": "3.2 In-context Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.759,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We propose that since these experiments rely on short-term learning effects of exposure in an experimental context, they can only be effectively simulated with LLMs by using in-context learning recursively. Specifically, for each question, the model is provided access to all previous items in answering a new question, including its own previous responses. This is intended to mirror the process of human experimental participants making judgments in the light of recent exposure to input and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9496"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.122,
+                0.085,
+                0.877,
+                0.365
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.376,
+                0.883,
+                0.406
+            ],
+            "angle": 0,
+            "content": "Figure 2: The ICL Simulation Framework. Black text is provided to the models as prompts, red text is generated by the model. With the last step's output appended to previous prompts, we ask LLMs the next question."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.431,
+                0.303,
+                0.446
+            ],
+            "angle": 0,
+            "content": "their own past responses."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.448,
+                0.487,
+                0.56
+            ],
+            "angle": 0,
+            "content": "We manually checked the correctness of LLM responses to content questions intended to check comprehension, and as in the human experimentation context removed answers for which the LLMs provided incorrect answers. The answers to critical target questions were recorded for further statistical analysis. Our ICL procedure is shown in Figure 2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.574,
+                0.431,
+                0.59
+            ],
+            "angle": 0,
+            "content": "3.3 Models and Experimental Settings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.597,
+                0.489,
+                0.804
+            ],
+            "angle": 0,
+            "content": "We used text-davinci-003 from the OpenAI API and open-sourced FLAN-UL2 as the LLMs of interest. Though these models are of course not exhaustive of the current landscape of LLMs, they provide some diversity since they differ in both structure (decoder-only vs. encoder-decoder) and parameter count (175B vs. 20B). To introduce more randomness into the experiment and allow enough sample sizes for statistical comparisons across conditions, we made slight modifications to the temperature hyperparameter on each run. Specifically, we assigned each run a random and unique temperature value between 0.2 to 1.0."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We also attempted to simulate 'participants' using a natural language prompt to approach different speaker identities following a similar methodology to Aher et al. (2022). We developed prompts with slots for titles, names, and country of origin to establish different character backgrounds simulating native English speakers from the United"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.431,
+                0.885,
+                0.543
+            ],
+            "angle": 0,
+            "content": "States, Britain, and Australia, following the participants' demographics in Johnson and Arnold (2022). However, these prompts did not induce greater diversity in responses than temperature modification, so only results using temperature modification are presented below. We present a further analysis for both methodologies in Appendix B."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.544,
+                0.884,
+                0.641
+            ],
+            "angle": 0,
+            "content": "In the end, we simulated 24 'participants' each in Experiment 1a and Experiment 2a, and 60 'participants' in Experiment 2b. We included more in Experiment 2b because this experiment has a lower response variability and thus needs more data points for the statistical analysis."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.655,
+                0.634,
+                0.67
+            ],
+            "angle": 0,
+            "content": "3.4 Measures"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.678,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Following the analytic approach of Johnson and Arnold (2022), we used a regression-based approach to analyze whether the responses of LM are consistent with the subject or gogoalal bias of the context for each experiment. We can then compare our findings with theirs by asking whether the effect of the main predictors is the same. In all experiments, predictors included exposure type (unambiguous exposures to subject/non-subject or goal/source), reference question type (whether the reference question is asked about the subject/non-subject or goal/source), and the interaction effect between them. Experiment 2 included verb type, as well as its two-way and three-way interaction effects with the other two predictors as addi"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9497"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.086,
+                0.496,
+                0.351
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.499,
+                0.086,
+                0.881,
+                0.351
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.364,
+                0.884,
+                0.409
+            ],
+            "angle": 0,
+            "content": "Figure 3: Subject responses of LLM (left) and human participants (right) in Experiment 1a, showing the percentage of subject responses for each type of reference question (Subj Q: subject reference question; Obj Q: object reference question), grouped by exposure type (subject exposure: light blue; object exposure: dark blue)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.433,
+                0.489,
+                0.577
+            ],
+            "angle": 0,
+            "content": "tional predictors. Given our small dataset, the results were analyzed using Bayesian mixed-effects Bernoulli logistic regression models in the R package brms (Bürkner, 2017) instead of a frequentist model. We report a Bayesian equivalent p-value (p_MAP) computed with the R package bayestestR (Makowski et al., 2019) to offer a straightforward interpretation of the results. Details of models are provided in Appendix 1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.594,
+                0.214,
+                0.608
+            ],
+            "angle": 0,
+            "content": "4 Results"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.623,
+                0.253,
+                0.637
+            ],
+            "angle": 0,
+            "content": "4.1 FLAN-UL2"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.646,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We found that FLAN-UL2 was not capable of generating meaningful output for analysis under this design. First, FLAN-UL2 showed a much higher false rate in answering content questions versus InstructGPT: while InstructGPT replied \\(100\\%\\) correctly to these questions, FLAN-UL2 replied to only \\(57\\%\\) correctly. Second, for the ambiguous target items, FLAN-UL2 answered 'yes' \\(100\\%\\) of the time, indicating an extremely strong bias towards simply answering 'yes' and producing no meaningful variation. By contrast, InstructGPT answered 'yes' to target questions \\(68\\%\\) of the time, suggesting some amount of 'yes' bias but to a much weaker degree. From these findings we concluded that FLAN-UL2 did not produce sufficiently clean outputs for analysis; therefore, in the following sections, we will focus on results from InstructGPT."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.433,
+                0.673,
+                0.448
+            ],
+            "angle": 0,
+            "content": "4.2 Experiment 1a"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.464,
+                0.885,
+                0.801
+            ],
+            "angle": 0,
+            "content": "Experiment 1a asks whether LLMs are sensitive to the frequency of referential patterns when subject referents are preferred. Figure 3 compares the subject responses of our results with Johnson and Arnold (2022)'s results. In both LLM and human data, we saw fewer subject responses in object exposure than in subject exposure. This is confirmed by the main effect of exposure type \\((p\\_MAP < .001)\\) revealed in the statistical analyses. There was also a main effect of response question type \\((p\\_MAP = .009)\\). As seen from the figure, there were more subject responses in the subject than object reference questions in both human and LLM data. This is because InstructGPT did answer 'yes' more frequently in general, which led to more subject interpretations when the question asked about the subject (where answering 'yes' indicates a subject interpretation), and fewer subject interpretations when the question was asked about the non-subject (where answering 'yes' indicates a non-subject interpretation)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "However, we did not find any interaction effect between exposure type and reference question type. While Johnson and Arnold (2022) found that exposure type has a significant effect for subject-referent questions but only a marginal effect for the nonsubject-referent questions, the exposure effect was significant for both question types in our"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9498"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.416,
+                0.089,
+                0.582,
+                0.101
+            ],
+            "angle": 0,
+            "content": "exposure subject object"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.105,
+                0.496,
+                0.351
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.501,
+                0.106,
+                0.878,
+                0.351
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.364,
+                0.885,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Figure 4: Subject responses of LLM (left) and human participants (right) in Experiment 2a. The proportion of subject responses is plotted against question type faceting by verb type, comparing goal-subject verbs (gs) like \"receive,\" where the subject is the goal referent) to subject-goal verbs (sg) like \"send,\" where the subject is the source referent."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.447,
+                0.49,
+                0.608
+            ],
+            "angle": 0,
+            "content": "results. InstructGPT is sensitive to exposure type even for nonsubject-referent questions, as reflected in Figure 3: there were no subject responses when non-subject reference questions were asked under the non-subject exposure condition. This suggests that LLM may even be more sensitive to the exposure effect than humans. Overall, Experiment 1 shows that LLM can indeed learn from recent exposure to syntactically-oriented referential patterns, in a relatively more non-subject-biased way."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.623,
+                0.277,
+                0.639
+            ],
+            "angle": 0,
+            "content": "4.3 Experiment 2a"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.646,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We ask in Experiment 2a if the sensitivity to subject reference patterns observed in Experiment 1a persists independent of semantic variability in the local context. Figure 4 illustrates the responses of LLM and human participants by exposure, reference question, and verb type. The behavior of InstructGPT is in line with what Johnson and Arnold (2022) found in human participants for Experiment 2a: as in Experiment 1, statistical analyses revealed a significant main effect of referent question type \\((p\\_ MAP = .034)\\), despite a marginally significant effect of exposure type \\((p\\_ MAP = .085)\\). InstructGPT did understand the pronoun as subject referents more after subject exposure and said 'yes' more often such that there were more subject interpretations when the critical question was asked about the subject. InstructGPT also showed a goal"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.447,
+                0.885,
+                0.526
+            ],
+            "angle": 0,
+            "content": "bias: there was a main effect of verb type (\\(p\\_MAP = .001\\)), such that it referred to the subject more with gs verbs (where the subject is the goal referent) than sg verbs (where the subject is the source referent). We observed no interaction effect."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.529,
+                0.884,
+                0.64
+            ],
+            "angle": 0,
+            "content": "In spite of these similarities, we still observe differences between LLM and human participants. Notice in Figure 4 that there is a larger difference between gs and sg verbs in InstructGPT than in human participants while keeping other conditions constant, suggesting that LLMs may have a larger goal bias than humans."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.655,
+                0.673,
+                0.671
+            ],
+            "angle": 0,
+            "content": "4.4 Experiment 2b"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.678,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Experiment 2b examines whether InstructGPT is still sensitive to exposure to referential patterns that exhibit consistent preferences for a source or goal referent. Figure 5 compares the results from InstructGPT and human participants. Whereas Johnson and Arnold (2022) reported significant main effects from exposure type, reference question type, and verb type, as well as a marginal interaction effect between verb type and exposure, we did not find any effect from these predictors, but only an interaction effect between verb type and question type \\((p\\_ MAP < .001)\\). We further examined the effect of question type and exposure type for each verb type, but no significant effect was found for either predictor. These results suggest that exposure"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9499"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.086,
+                0.5,
+                0.351
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.501,
+                0.086,
+                0.88,
+                0.351
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.187,
+                0.364,
+                0.808,
+                0.379
+            ],
+            "angle": 0,
+            "content": "Figure 5: Goal responses of LLM (left) and human participants (right) in Experiment 2b."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.404,
+                0.49,
+                0.581
+            ],
+            "angle": 0,
+            "content": "did not necessarily change InstructGPT's behaviors: for goal questions, both the goal and source exposure conditions led to \\(100\\%\\) goal responses. Reference question type also did not affect InstructGPT's responses, because it almost always interprets the pronoun as the goal referent under subject exposure. The lack of any significant effects, and indeed our observed universal goal interpretations to goal-focused reference questions, suggest that InstructGPT displays an extreme goal bias in pronoun interpretation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.598,
+                0.241,
+                0.614
+            ],
+            "angle": 0,
+            "content": "5 Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.628,
+                0.49,
+                0.789
+            ],
+            "angle": 0,
+            "content": "The present study examined whether and to what extent LLMs display adaptation for pronominal interpretation after exposure to referential patterns by replicating three psycholinguistic experiments. Exposed to the same stimuli and study design, and analyzed with the same statistical procedures as the source experiments, we asked whether LLMs show human-like behaviors and compared the performance of LLMs-simulated participants with human participants."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We firstly found a difference between the capacity of contemporary models to replicate human psycholinguistic experimental designs in an ICL framework. While InstructGPT was able to correctly answer all the presented comprehension-check content questions, FLAN-UL2 was not. Both models displayed at least some bias towards answering 'yes' in ambiguous cases, but in the case of"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.404,
+                0.885,
+                0.532
+            ],
+            "angle": 0,
+            "content": "FLAN-UL2 this bias resulted in \\(100\\%\\) 'yes' answers, rendering meaningful variation impossible to ascertain. This could be a result of structural differences between the models (decoder-only in the case of InstructGPT, encoder-decoder in the case of FLAN-UL2), or perhaps more likely a question of simple model size (175B for InstructGPT vs. 20B for FLAN-UL2)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.534,
+                0.885,
+                0.71
+            ],
+            "angle": 0,
+            "content": "Experiments 1a and 2a examined whether LLMs adapt their syntactic bias from recent exposure to referential patterns like humans, without and with the presence of possibly confounding semantic goal bias. We found that LLM's referential biases are indeed sensitive to such exposure in both experiments. In addition, LLM did exhibit a goal bias in Experiment 2a, in line with previous studies which argue for LLM's ability to exhibit human-like semantic bias in pronoun interpretation (e.g., Davis and van Schijndel 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.71,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Experiment 2b examined whether LLM can adapt and vary their semantic bias from exposures. In this context, in contrast with the previous two experiments, exposure type did not affect LLM behavior at all. This raises the question of why LLM would be sensitive to exposure to only referential patterns that exhibit consistent syntactic bias but not semantic bias. An immediate and intuitive answer would be that LLM is unable to fully represent semantic knowledge in referential processing. However, given that LLM did display a human-like goal bias in pronoun interpretation independent of exposure, this explanation seems unlikely."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9500"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.085,
+                0.49,
+                0.146
+            ],
+            "angle": 0,
+            "content": "We suggest that our work provides evidence that LLM only partially represents the semantic knowledge involved in referential processing for two reasons."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.15,
+                0.49,
+                0.472
+            ],
+            "angle": 0,
+            "content": "First, adapting referential biases on the basis of exposure in the local context may require representations of higher-level knowledge than merely exhibiting a bias towards certain referents. While the latter may only require knowledge of which features associated with a referent are more frequent or likely in general, adaption requires a sustained awareness of referential pattern frequency as it changes in the local discourse context. Though representing knowledge about semantic relations was observed as early as the analogical reasoning task in Word2Vec (Mikolov et al., 2013), human-like extraction of abstract information like a persistent discourse state is more challenging. The model may only be able to identify thematic roles (source or goal) of a referent and associate them with pronoun interpretation, but not to identify a consistent pattern of thematic roles across a discourse. If so, this would explain the strong goal bias we observed in Experiment 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.474,
+                0.49,
+                0.553
+            ],
+            "angle": 0,
+            "content": "Second, it is possible that LLM has an extremely strong goal bias that masks the influence of exposures. If so, this suggests that LLM over-represents the semantic knowledge encoded in pronoun interpretation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.555,
+                0.49,
+                0.716
+            ],
+            "angle": 0,
+            "content": "In either case, our results suggest that LLM do not encode semantic knowledge in a fully human-like way, even though they do demonstrate some human-like capacities for semantic understanding. Although we believe this gap can be mitigated via instruction fine-tuning or chain-of-thought prompting (Wei et al., 2022), these results still suggest we should consider incorporating semantically-informed objectives into self-supervised pre-training to a greater extent."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.731,
+                0.411,
+                0.746
+            ],
+            "angle": 0,
+            "content": "6 Conclusions and Future Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.759,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": "By replicating a series of psycholinguistic experiments as closely as possible using in-context learning, this paper pioneered whether LLMs would adapt pronominal interpretation behaviors in a human-like way given exposure to referential patterns in the local discourse context. Our work suggests paths forward for replicating psycholinguistic experiments in a more faithful way that allows for comparisons between human and LLMs' behaviors."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.084,
+                0.616,
+                0.099
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.112,
+                0.885,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Several main limitations exist in our study in its current form. First, our reported results only simulated experimental participants by manipulating the temperature hyperparameter. We compared this approach with natural language prompting for Experiment 1, but that prompting did not increase \"participant\" diversity, so it was abandoned. Moreover, approaches for simulating psycholinguistic experimental \"participants\" could go far beyond what was tried here; our prompting method was relatively limited, and more detailed prompting could be included in future experimental simulations. Second, making a direct comparison with actual psycholinguistic experiments might not be the only method to investigate LLMs' discourse capacity. A comprehensive list of discourse probing tasks might play a similar role despite a different way (Koto et al., 2021). Third, this study is strictly behavioral: limited by both computational resources and obscure mechanisms of in-context learning, we do not dive into models' internal representations in our analyses."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.481,
+                0.673,
+                0.497
+            ],
+            "angle": 0,
+            "content": "Acknowledgement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.509,
+                0.884,
+                0.556
+            ],
+            "angle": 0,
+            "content": "We are very grateful to Jennifer Arnold for sharing the stimuli and design of Johnson and Arnold (2022) with us."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.587,
+                0.61,
+                0.602
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.611,
+                0.884,
+                0.652
+            ],
+            "angle": 0,
+            "content": "Gati Aher, Rosa I. Arriaga, and Adam Tauman Kalai. 2022. Using large language models to simulate multiple humans."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.665,
+                0.885,
+                0.693
+            ],
+            "angle": 0,
+            "content": "Mira Ariel. 1990. Accessing noun-phrase antecedents. Routledge."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.705,
+                0.885,
+                0.734
+            ],
+            "angle": 0,
+            "content": "Jennifer E Arnold. 1998. Reference form and discourse patterns. Stanford University."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.746,
+                0.885,
+                0.787
+            ],
+            "angle": 0,
+            "content": "Jennifer E Arnold. 2001. The effect of thematic roles on pronoun use and frequency of reference continuation. Discourse processes, 31(2):137-162."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.799,
+                0.885,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Jennifer E Arnold, Iris M Strangmann, Heeju Hwang, Sandra Zerkle, and Rebecca Nappa. 2018. Linguistic experience affects pronoun interpretation. Journal of Memory and Language, 102:41-54."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.865,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Dale J Barr, Roger Levy, Christoph Scheepers, and Harry J Tily. 2013. Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of memory and language, 68(3):255-278."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.51,
+                0.611,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9501"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Holly P Branigan, Martin J Pickering, and Janet F McLean. 2005. Priming prepositional-phrase attachment during comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31(3):468."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.16,
+                0.488,
+                0.2
+            ],
+            "angle": 0,
+            "content": "Susan E Brennan. 1995. Centering attention in discourse. Language and Cognitive processes, 10(2):137-167."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.208,
+                0.488,
+                0.247
+            ],
+            "angle": 0,
+            "content": "Paul-Christian Bürkner. 2017. brms: An r package for bayesian multilevel models using stan. Journal of statistical software, 80:1-28."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.255,
+                0.488,
+                0.32
+            ],
+            "angle": 0,
+            "content": "Franklin Chang, Gary S Dell, Kathryn Bock, and Zenzi M Griffin. 2000. Structural priming as implicit learning: A comparison of models of sentence production. Journal of psycholinguistic research, 29(2):217-230."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.329,
+                0.488,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Forrest Davis and Marten van Schijndel. 2020. Discourse structure interacts with reference but not syntax in neural language models. arXiv preprint arXiv:2010.04887."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.389,
+                0.488,
+                0.507
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.515,
+                0.488,
+                0.633
+            ],
+            "angle": 0,
+            "content": "Richard Futrell, Ethan Wilcox, Takashi Morita, Peng Qian, Miguel Ballesteros, and Roger Levy. 2019. Neural language models as psycholinguistic subjects: Representations of syntactic state. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 32-42, Minneapolis, Minnesota. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.641,
+                0.488,
+                0.757
+            ],
+            "angle": 0,
+            "content": "Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195-1205, New Orleans, Louisiana. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.766,
+                0.488,
+                0.858
+            ],
+            "angle": 0,
+            "content": "Jennifer Hu, Jon Gauthier, Peng Qian, Ethan Wilcox, and Roger Levy. 2020. A systematic assessment of syntactic generalization in neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1725-1744, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.866,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Elyce D Johnson and Jennifer E Arnold. 2022. The frequency of referential patterns guides pronoun comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.178
+            ],
+            "angle": 0,
+            "content": "Fajri Koto, Joy Han Lau, and Timothy Baldwin. 2021. Discourse probing of pretrained language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3849-3864, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.192,
+                0.883,
+                0.257
+            ],
+            "angle": 0,
+            "content": "Dominique Makowski, Mattan S Ben-Shachar, and Daniel Lüdecke. 2019. bayestestr: Describing effects and their uncertainty, existence and significance within the bayesian framework. Journal of Open Source Software, 4(40):1541."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.272,
+                0.883,
+                0.336
+            ],
+            "angle": 0,
+            "content": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.351,
+                0.883,
+                0.43
+            ],
+            "angle": 0,
+            "content": "Kanishka Misra, Allyson Ettinger, and Julia Rayz. 2020. Exploring BERT's sensitivity to lexical cues using tests from semantic priming. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4625-4635, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.443,
+                0.883,
+                0.547
+            ],
+            "angle": 0,
+            "content": "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.562,
+                0.883,
+                0.601
+            ],
+            "angle": 0,
+            "content": "Jenny R Saffran, Richard N Aslin, and Elissa L Newport. 1996a. Statistical learning by 8-month-old infants. Science, 274(5294):1926-1928."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.615,
+                0.883,
+                0.666
+            ],
+            "angle": 0,
+            "content": "Jenny R Saffran, Elissa L Newport, and Richard N Aslin. 1996b. Word segmentation: The role of distributional cues. Journal of memory and language, 35(4):606-621."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.682,
+                0.883,
+                0.733
+            ],
+            "angle": 0,
+            "content": "Sebastian Schuster and Judith Degen. 2020. I know what you're probably going to say: Listener adaptation to variable use of uncertainty expressions. Cognition, 203:104285."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.748,
+                0.883,
+                0.825
+            ],
+            "angle": 0,
+            "content": "Arabella Sinclair, Jaap Jumelet, Willem Zuidema, and Raquel Fernández. 2022. Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations. Transactions of the Association for Computational Linguistics, 10:1031-1050."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.84,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Ionut-Teodor Sorodoc, Kristina Gulordava, and Gemma Boleda. 2020. Probing for referential information in language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4177–4189, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9502"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.086,
+                0.488,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. 2023. UI2: Unifying language learning paradigms."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.174,
+                0.49,
+                0.214
+            ],
+            "angle": 0,
+            "content": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. CoRR, abs/1905.05950."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.223,
+                0.49,
+                0.303
+            ],
+            "angle": 0,
+            "content": "Shiva Upadhye, Leon Bergen, and Andrew Kehler. 2020. Predicting reference: What do language models learn about discourse models? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 977-982, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.311,
+                0.49,
+                0.366
+            ],
+            "angle": 0,
+            "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.376,
+                0.349,
+                0.391
+            ],
+            "angle": 0,
+            "content": "A Statistical Information"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.401,
+                0.49,
+                0.594
+            ],
+            "angle": 0,
+            "content": "For each experiment, we analyzed referent choice (Subject = 1, Non-subject = 0) using a mixed-effects Bernoulli regression model from the R package brms (Bürkner, 2017), with the maximal random structure justified by design (Barr et al., 2013). Predictors are coded in the same way as in Johnson and Arnold (2022). All models were specified with a weakly informative prior using the Cauchy distribution with center 0 and scale 2.5. Models were fitted using six chains, each with 4,000 iterations of which the first 1,000 are warmup to calibrate the sampler, resulting in 18,000 posterior examples."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.594,
+                0.49,
+                0.771
+            ],
+            "angle": 0,
+            "content": "The model for Experiment 1a included question type (QtypeC, sum-coded: Subject = 0.5, Non-subject = -0.5) and exposure type (PC, effects-coded: Subject-biased = 0.51, Object-biased = -0.49) as fixed predictors, random intercepts for participants and items, random slopes of question type, exposure type and their interaction for items, and a random slope of question type for the participant. Exposure type was not included as a random slope for participants because the condition does not vary within participants."
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.114,
+                0.776,
+                0.449,
+                0.919
+            ],
+            "angle": 0,
+            "content": "brm (Rc ~ QtypeC*PC +  \n(1+QtypeC*PC|ID)+  \n(1+QtypeC|Subject),  \ndata=(e1a, Exposure!= \"None\"),  \nfamily=\"bernoulli\",  \nchains=6,  \niter=4000,  \nwarmup=1000,  \ncontrol ="
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.583,
+                0.085,
+                0.882,
+                0.134
+            ],
+            "angle": 0,
+            "content": "list(adapt delta = 0.95),  \nprior=  \nc(set_prior (\"cauchy(0,2.5)\"))"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.146,
+                0.885,
+                0.388
+            ],
+            "angle": 0,
+            "content": "The model for Experiment 2a included question type (QtypeC, sum-coded: Subject = 0.5, Non-subject = -0.5), exposure type (PC, sum-coded: Subject-biased = 0.5, Object-biased = -0.5), and verb type (Vc, sum-coded: gs-verb = 0.5, sg-verb = -0.5) as fixed predictors, random intercepts for participants and items, random slopes of question type, exposure type and their interaction for items, and a random slope of question type for the participant. As in Experiment 1a, exposure type was not included as a random slope for participants because the condition does not vary within participants. Similarly, verb bias was not included as a random slope for items here because it does not vary within items."
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.546,
+                0.405,
+                0.885,
+                0.567
+            ],
+            "angle": 0,
+            "content": "brm (Rc ~ QtypeC*PC*Vc +  \n(1+PC*Qtypec|Item)+  \n(1+QtypeC*Vc|Subject),  \ndata=e2a,  \nfamily=\"bernoulli\", chains=6,  \niter=4000, warmup=1000,  \ncontrol = list(adapt delta = 0.98),  \ncores = 6,  \nprior=  \nc(set_prior (\"cauchy(0,2.5)\"))"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.578,
+                0.884,
+                0.691
+            ],
+            "angle": 0,
+            "content": "The model for Experiment 2b included question type (QtypeC, sum-coded: Goal = 0.5, Source = -0.5), exposure type (PC, sum-coded: Goal-biased = 0.5, Source-biased = -0.5), and verb type (Vc, sum-coded: gs-verb = 0.5, sg-verb = -0.5) as fixed predictors. The random effect structure was the same as that of Experiment 2a."
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.556,
+                0.711,
+                0.883,
+                0.921
+            ],
+            "angle": 0,
+            "content": "brm (Rc ~ PC\\*Vc\\*QtypeC  \n+ (1+PC\\*QtypeC|Item)  \n+ (1+Vc\\*QtypeC|Subject),  \ndata=e2b,  \nfamily=\"bernoulli\",  \nchains=6,  \niter=4000,  \nwarmup=1000,  \ncontrol =  \n    list(adapt delta = 0.999),  \ncores = 6,  \nprior=  \n    c(set_prior (\"cauchy(0,2.5)\"))"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9503"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.084,
+                0.366,
+                0.102
+            ],
+            "angle": 0,
+            "content": "B Temperature vs. Prompt"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.11,
+                0.49,
+                0.255
+            ],
+            "angle": 0,
+            "content": "Our prompt-based simulation of multiple participants embedded names, countries, prefixes, and genders into a carrier sentence: \\(\\{Prefix + Name\\}\\) is a native English speaker living in {Country}. {Gender} is asked in a psycholinguistic experiment to answer the following questions. For example, Mr. Smith is a native English speaker living in England. He is asked in a psycholinguistic experiment to answer the following questions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.255,
+                0.49,
+                0.544
+            ],
+            "angle": 0,
+            "content": "Specifically, we calculated the variance and ran the Levene's test for any significant difference between humans and InstructGPT in each experiment. In experiment 1, the human responses' variance is 0.055, the temperature-based responses' variance is 0.024, and the prompt-based responses' variance is 0.017. Human responses were significantly higher than both (Temperature-based: \\( p = .049 \\); Prompt-based: \\( p = .007 \\)). Yet, temperature-based were not significantly higher than prompt-based. Due to the limitation of API pricing, we only ran temperature-based in the following experiments. In experiment 2a, the human responses' variance is 0.034 and the temperature-based responses' variance is 0.043. Yet, Levene's test did not reveal any significant difference. In Experiment 2b, the human responses' variance is 0.020 and the temperature-based responses' variance is 0.008 (\\( p < .001 \\))."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.546,
+                0.49,
+                0.673
+            ],
+            "angle": 0,
+            "content": "We used different techniques to introduce randomness and include more experimental data in our experiments. We realized these were not well-designed prompts to elicit different linguistic backgrounds. Given the lack of investigation on simulating multiple participants in psycholinguistics studies, we recognize this as a future direction of possible work."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9504"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.133,
+                0.085,
+                0.435,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ACL 2023 Responsible NLP Checklist"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.108,
+                0.323,
+                0.123
+            ],
+            "angle": 0,
+            "content": "A For every submission:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.127,
+                0.533,
+                0.145
+            ],
+            "angle": 0,
+            "content": "A1. Did you describe the limitations of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.145,
+                0.393,
+                0.159
+            ],
+            "angle": 0,
+            "content": "Last section without the number"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.171,
+                0.553,
+                0.187
+            ],
+            "angle": 0,
+            "content": "A2. Did you discuss any potential risks of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.187,
+                0.351,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.213,
+                0.697,
+                0.23
+            ],
+            "angle": 0,
+            "content": "A3. Do the abstract and introduction summarize the paper's main claims?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.231,
+                0.433,
+                0.246
+            ],
+            "angle": 0,
+            "content": "The introduction is in the first section."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.257,
+                0.67,
+                0.273
+            ],
+            "angle": 0,
+            "content": "A4. Have you used AI writing assistants when working on this paper?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.274,
+                0.233,
+                0.288
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.301,
+                0.49,
+                0.317
+            ],
+            "angle": 0,
+            "content": "B Did you use or create scientific artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.322,
+                0.216,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.348,
+                0.531,
+                0.364
+            ],
+            "angle": 0,
+            "content": "B1. Did you cite the creators of artifacts you used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.365,
+                0.249,
+                0.38
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.391,
+                0.779,
+                0.407
+            ],
+            "angle": 0,
+            "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.408,
+                0.249,
+                0.423
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.434,
+                0.882,
+                0.497
+            ],
+            "angle": 0,
+            "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.499,
+                0.249,
+                0.514
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.525,
+                0.882,
+                0.573
+            ],
+            "angle": 0,
+            "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.574,
+                0.249,
+                0.589
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.6,
+                0.882,
+                0.632
+            ],
+            "angle": 0,
+            "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.633,
+                0.249,
+                0.648
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.659,
+                0.882,
+                0.74
+            ],
+            "angle": 0,
+            "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.741,
+                0.249,
+                0.755
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.765,
+                0.495,
+                0.782
+            ],
+            "angle": 0,
+            "content": "C Did you run computational experiments?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.788,
+                0.437,
+                0.802
+            ],
+            "angle": 0,
+            "content": "Experiments were described in section 3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.813,
+                0.882,
+                0.846
+            ],
+            "angle": 0,
+            "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.847,
+                0.54,
+                0.861
+            ],
+            "angle": 0,
+            "content": "In the method section, we described our parameters."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.869,
+                0.878,
+                0.893
+            ],
+            "angle": 0,
+            "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9505"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.085,
+                0.88,
+                0.116
+            ],
+            "angle": 0,
+            "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.118,
+                0.566,
+                0.133
+            ],
+            "angle": 0,
+            "content": "Not applicable. We did not do a hyperparameter search."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.143,
+                0.884,
+                0.191
+            ],
+            "angle": 0,
+            "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.193,
+                0.511,
+                0.208
+            ],
+            "angle": 0,
+            "content": "We provide detailed results in the results section."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.218,
+                0.884,
+                0.266
+            ],
+            "angle": 0,
+            "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.268,
+                0.638,
+                0.283
+            ],
+            "angle": 0,
+            "content": "We provide existing packages' information in the methods section."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.294,
+                0.878,
+                0.31
+            ],
+            "angle": 0,
+            "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.315,
+                0.215,
+                0.33
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.341,
+                0.883,
+                0.373
+            ],
+            "angle": 0,
+            "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.374,
+                0.249,
+                0.39
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.4,
+                0.884,
+                0.448
+            ],
+            "angle": 0,
+            "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.45,
+                0.25,
+                0.465
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.476,
+                0.883,
+                0.523
+            ],
+            "angle": 0,
+            "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.525,
+                0.25,
+                0.54
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.55,
+                0.875,
+                0.565
+            ],
+            "angle": 0,
+            "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.567,
+                0.25,
+                0.582
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.593,
+                0.881,
+                0.624
+            ],
+            "angle": 0,
+            "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.627,
+                0.25,
+                0.642
+            ],
+            "angle": 0,
+            "content": "No response."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9506"
+        }
+    ]
+]

2023/Large Language Models Are Partially Primed in Pronoun Interpretation/60280f41-b321-4144-822d-79720ae15459_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c98b4be6185cb33951c696c1acab0844e22d011d107ecb293c96426781d44b02
+size 561111

2023/Large Language Models Are Partially Primed in Pronoun Interpretation/full.md

@@ -0,0 +1,386 @@
+# Large Language Models Are Partially Primed in Pronoun Interpretation
+
+Suet-Ying Lam†* Qingcheng Zeng‡* Kexun Zhang◇* Chenyu You* Rob Voigt‡
+
+† Department of Linguistics, UMass Amherst
+
+$\ddagger$  Department of Linguistics, Northwestern University
+
+$\diamond$  Department of Computer Science, UCSB
+
+$\clubsuit$  Department of Electrical Engineering, Yale University
+
+# Abstract
+
+While a large body of literature suggests that large language models (LLMs) acquire rich linguistic representations, little is known about whether they adapt to linguistic biases in a human-like way. The present study probes this question by asking whether LLMs display human-like referential biases using stimuli and procedures from real psycholinguistic experiments. Recent psycholinguistic studies suggest that humans adapt their referential biases with recent exposure to referential patterns; closely replicating three relevant psycholinguistic experiments from Johnson and Arnold (2022) in an in-context learning (ICL) framework, we found that InstructGPT adapts its pronominal interpretations in response to the frequency of referential patterns in the local discourse, though in a limited fashion: adaptation was only observed relative to syntactic but not semantic biases. By contrast, FLAN-UL2 fails to generate meaningful patterns. Our results provide further evidence that contemporary LLMs discourse representations are sensitive to syntactic patterns in the local context but less so to semantic patterns. Our data and code are available at https://github.com/zkx06111/1lm_priming.
+
+# 1 Introduction
+
+While neural network models, and particularly pretrained large language models have shown excellent performance at particular language processing tasks, many questions remain about the extent to which models optimized for such performance encode, as a side effect, human-like linguistic knowledge and cognitive biases. We know that they do to some extent; existing work has shown, for example, that neural models encode aspects of human-like long-distance number agreement (Gulordava et al., 2018), incremental syntactic state (Futrell et al.,
+
+2019), and syntactic generalization more broadly (Hu et al., 2020). In this paper, we examine whether FLAN-UL2 (Tay et al., 2023) and InstructGPT (Ouyang et al., 2022), two representative LLMs, display adaptation in pronoun interpretation when exposed to consistent referential patterns in the local discourse context.
+
+Compared with syntactic or lexical knowledge, representing referential knowledge is possibly more complex; we know from psycholinguistic studies that human referential interpretation integrates multiple levels of linguistic structure. Humans do not interpret ambiguous pronouns at random but are guided by both syntactic and semantic information. It is well-established that absent other cues humans prefer a syntactic subject in choosing the antecedent of the ambiguous pronoun, i.e., subject bias (Ariel, 1990; Brennan, 1995). In example (1), she is more likely to be interpreted as the subject Ada than the non-subject Eva  $\mathbb{I}$  , even though both referents are possible antecedents for the pronoun.
+
+(1)  $\mathrm{Ada}_1$  talked with  $\mathrm{Eva}_2$ . She...
+
+People are also sensitive to the semantic structure of the sentence when choosing an antecedent for an ambiguous pronoun, in addition to syntactic information. In a transfer event that depicts a transfer-of-possession from one entity (the source) to another (the goal), they prefer the goal referent (Ada in (2), Eva in (3)) over the source referent (Eva in (2), Ada in (3)) to be the antecedent (Arnold, 2001, 1998), i.e., goal bias:
+
+(2) Goal-source (gs) verb:
+
+Ada<sub>1</sub> received a letter from Eva<sub>2</sub>. She<sub>1</sub>...
+
+(3) Source-goal (sg) verb:
+
+Ada1 sent a letter to Eva2. She2...
+
+In sum, people exhibit sensitivity to both syntax (subject bias) and semantics (goal bias) during pronoun interpretation. Importantly, these levels of linguistic structure are frequently entwined since both can influence referential interpretation. At times these influences may push in different directions: for instance, Ada in [3] is both the syntactic subject and the semantic source.
+
+Building on a long tradition investigating preferences in pronoun interpretation, recent psycholinguistic studies have probed into the deeper question of the origin of these biases. One hypothesis is that referential biases come from linguistic experience: when a bias occurs very frequently, people will tend to adapt this more frequent referential pattern, both in the immediate exposure as well as more large-scale past experience (Arnold, 1998, 2001). Recent evidence has provided support to this idea by demonstrating that recent exposure to certain referential patterns did change people's referential biases. In a series of psycholinguistic studies, Johnson and Arnold (2022) show that after reading numerous stories that consistently show a particular referential bias, e.g., always referring to the non-subject or source referent, people did have a stronger preference for these primed referents.
+
+Given this line of psycholinguistic research, the current study investigates the extent to which LLMs adapt and vary referential biases in pronoun interpretation through exposure to referential patterns in the local context. To do so we replicated actual psycholinguistic experiments from Johnson and Arnold (2022) in LLMs using an ICL framework and asked whether the responses of LLMs display adaptation from exposure to referential patterns like human experimental participants. Comparing syntactically-motivated to semantically-motivated exposure conditions will allow us to first examine whether LLMs display human-like subject bias and goal bias and further understand the extent to which LLMs make use of local frequency changes in discourse representations in these categories.
+
+In-context learning refers to LLMs' ability to learn from demonstrations written in natural language prompts. Compared to previous work that has largely examined the encoding of such discourse knowledge using zero-shot inference (Upadhye et al., 2020), ICL is particularly suitable for experimental simulation since it replicates the naturalistic context in which later responses draw upon exposure to previous examples.
+
+Foreshadowing our results, we find that InstructGPT can adapt and thus vary its syntactic bias from exposure to referential patterns in the local context, but the same is not true for semantic bias. Given that InstructGPT still exhibits a goal bias in spite of local discourse priming, we argue this suggests LLMs only encode partial semantic knowledge in referential processing. To sum up, our contributions can be summarized as follows:
+
+1) We extended a discourse understanding evaluation to state-of-the-art LLMs from a new perspective, asking whether LLM's referential bias can be modified by exposure to particular referential patterns, like how humans adapt referential bias from experience.  
+2) To the best of our knowledge, we are the first study that replicates actual psycholinguistic experiments using the ICL framework and compares LLMs' behaviors with real human participants.  
+3) We present results in this context showing further evidence that InstructGPT can acquire abstract syntactic knowledge in referential interpretation to some extent, but not semantic knowledge.
+
+# 2 Related Work
+
+A growing body of literature has suggested that LLMs encode rich representations of linguistic structure at various levels, including aspects of syntax, semantics, and reference encoded throughout their representations (Tenney et al., 2019). One of the most well-documented lines of this work demonstrates that these models can acquire diverse elements of syntactic knowledge (Gulordava et al., 2018; Futrell et al., 2019; Hu et al., 2020).
+
+This capacity for encoding linguistic understanding extends to priming effects with psycholinguistic analogies to humans. At the syntactic level, Sinclair et al. (2022) explored structural priming in various autoregressive LLMs and found priming effects despite a clear dependence on semantic information. At the semantic level, using English lexical stimuli in BERT (Devlin et al., 2019), Misra et al. (2020) found that BERT does display evidence of sensitivity to semantic priming, though this is localized to more unconstrained contexts and only certain semantic relations.
+
+In analyses of LLMs' linguistic understanding modeled on psycholinguistic experiments, however,
+
+the question of discourse knowledge remains relatively under-explored. Recent existing work has presented partially conflicting accounts in this area, in particular with regard to how LLMs may or may not exhibit human-like biases in pronoun interpretation. For example, Davis and van Schijndel (2020) compared LSTM LMs and Transformer LMs behaviors and internal representations in dealing with implicit causality verbs, finding that surprisingly (contrary to humans) implicit causality only influences Transformer LMs' behavior for reference, but influences neither model for syntactic attachment. Sorodoc et al. (2020) also compared LSTM LMs and Transformer LMs in coreference resolution corpora, finding that although LMs are much better at grammar, they also captured referential aspects to some extent.
+
+This existing work replicates sets of individual stimuli from psycholinguistic experiments in isolation; by contrast, our work takes a more behaviorally-oriented approach by replicating the stimuli, procedure, and even experimental design of a full set of human psycholinguistic studies. We do this to ask a further question: beyond exhibiting baseline human-like referential biases in pronoun interpretation, do LLMs display adaptation to the frequency of referential patterns in the local context like humans? This question is important because adapting to a referential bias over the course of an experiment requires a sustained representation of the frequency of the pattern, which involves a higher-level understanding of the discourse structure. Humans have shown abilities of this kind of adaptation at multiple linguistic levels. For instance, exposure to syntactic structures consecutively affects humans' choice of structures in both the short term and long term via priming effects (e.g., Branigan et al. 2005; Chang et al. 2000). Exposure to phonological patterns can also guide humans' segmentation patterns (e.g., Saffran et al. 1996a,b). At the semantic and pragmatic levels, humans can also adapt as listeners to speakers' variable choices of uncertainty expressions (Schuster and Degen, 2020).
+
+# 3 Methods
+
+In this work, we aim to replicate three experiments from Johnson and Arnold (2022), transferring their designs, procedures, and stimuli as faithfully as possible to the LLM context using in-context learning.
+
+# 3.1 Source Experiments
+
+We first briefly summarize the experimental setup employed by Johnson and Arnold (2022). In each experimental setting, participants heard a series of two-sentence stories in which the first sentence contained two characters with gender-marked first names (Matt or Will for men, Liz or Ana for women). For each story, participants answered a content question to check comprehension, and then a reference question to check pronominal interpretation. In order to lower the ceiling and keep participants from falling into a pattern of simply answering "yes," reference questions were equally split between default and non-default phrasings (e.g. between the subject/non-subject interpretations in Experiment 1a and 2a and between the source/goal interpretations in Experiment 2b). Figure 1 illustrates sample stimuli and the procedure of the three experiments conducted.
+
+In each experiment participants were first shown a series of stories with exposure reference questions to establish a referential pattern; in these, the characters had different genders so pronoun interpretation was unambiguous. After the exposure phase (20 stories in Experiment 1, 12 stories in Experiment 2), further stories with unambiguous exposure questions were intermixed with 12 stories accompanied by critical reference questions; in these stories, the characters had the same gender, so pronoun interpretation was ambiguous and required reliance on discourse cues.
+
+Across experiments, there were two key conditions that were manipulated for exposures. Under a subject exposure condition, the unambiguous intended referents of all exposure questions are subjects of the preceding clauses; in the corresponding non-subject exposure condition, they are the objects of the preceding clauses. Similarly, under a goal exposure condition the unambiguous intended referents of all exposure questions are goal referents while in a source exposure condition they are source referents.
+
+Aiming to transfer these experiments as faithfully as possible to the LLM context, we used this experimental paradigm to evaluate the LLM by providing the model with the full text of each story prompt and content/reference question, then generated tokens in response which we interpreted as answers.
+
+For clarity, we use the same experiment numbers and identifiers as in Johnson and Arnold (2022).
+
+![](images/13ba743cefbd27c4c0e3f7227491bf50163a4ae77168074819d8901207107d2e.jpg)  
+Figure 1: Illustrated Experimental Procedure. Closely following the setup with human participants in Johnson and Arnold (2022), LLMs were primed via exposure to story/question pairs with unambiguous referents, and tested for their responses on ambiguous target pairs.
+
+Note that we did not replicate experiments 1b and 1c, which investigated whether people are still sensitive to referential patterns using different types of referential expressions (e.g., third-person names and first- and second-person pronouns). While these experiments provided insights into the linguistic structure at which people generalize referential biases, they were less relevant to the objective of this study and thus were not included.
+
+Experiment 1a In this experiment, all story prompts contain "joint-action" verbs using "with" in the form "X did something with Y." Since these verbs lack a clear semantic bias (e.g., Arnold et al. 2018), this context allows us to evaluate LLMs' sensitivity to syntax-based biases in discourse by asking whether exposure to only subject-bias or object-bias examples will influence following answers on the ambiguous critical items. If LLMs are sensitive to syntax-based referential patterns, we expect more subject responses under the subject exposure condition and more non-subject responses under the non-subject exposure condition.
+
+Experiment 2a Experiment 2a forms a bridge between adaptation to syntactic and semantic constraints. Are LLMs able to track patterns in both categories, for instance learning an exposure bias in one category while ignoring variation in the
+
+other? In this experiment, all story prompts contain source/goal verbs like "give" and "receive," but these are distributed equally throughout exposures. The manipulation remains the same as Experiment 1a, in which LLMs are exposed to consistent and unambiguous subject interpretations in the subject exposure condition and non-subject interpretations in the non-subject exposure condition.
+
+Experiment 2b This experiment focuses solely on source/goal biases, in which all story prompts contain source/goal words, but the unambiguous exposure items are manipulated to contain only source references in the source exposure condition and only goal references in the goal exposure condition.
+
+# 3.2 In-context Learning
+
+We propose that since these experiments rely on short-term learning effects of exposure in an experimental context, they can only be effectively simulated with LLMs by using in-context learning recursively. Specifically, for each question, the model is provided access to all previous items in answering a new question, including its own previous responses. This is intended to mirror the process of human experimental participants making judgments in the light of recent exposure to input and
+
+![](images/7eac8738a3744d17f8a5c6e9927d8cc4f4b3c45a032fa5cbeb0d11ac19659747.jpg)  
+Figure 2: The ICL Simulation Framework. Black text is provided to the models as prompts, red text is generated by the model. With the last step's output appended to previous prompts, we ask LLMs the next question.
+
+their own past responses.
+
+We manually checked the correctness of LLM responses to content questions intended to check comprehension, and as in the human experimentation context removed answers for which the LLMs provided incorrect answers. The answers to critical target questions were recorded for further statistical analysis. Our ICL procedure is shown in Figure 2.
+
+# 3.3 Models and Experimental Settings
+
+We used text-davinci-003 from the OpenAI API and open-sourced FLAN-UL2 as the LLMs of interest. Though these models are of course not exhaustive of the current landscape of LLMs, they provide some diversity since they differ in both structure (decoder-only vs. encoder-decoder) and parameter count (175B vs. 20B). To introduce more randomness into the experiment and allow enough sample sizes for statistical comparisons across conditions, we made slight modifications to the temperature hyperparameter on each run. Specifically, we assigned each run a random and unique temperature value between 0.2 to 1.0.
+
+We also attempted to simulate 'participants' using a natural language prompt to approach different speaker identities following a similar methodology to Aher et al. (2022). We developed prompts with slots for titles, names, and country of origin to establish different character backgrounds simulating native English speakers from the United
+
+States, Britain, and Australia, following the participants' demographics in Johnson and Arnold (2022). However, these prompts did not induce greater diversity in responses than temperature modification, so only results using temperature modification are presented below. We present a further analysis for both methodologies in Appendix B.
+
+In the end, we simulated 24 'participants' each in Experiment 1a and Experiment 2a, and 60 'participants' in Experiment 2b. We included more in Experiment 2b because this experiment has a lower response variability and thus needs more data points for the statistical analysis.
+
+# 3.4 Measures
+
+Following the analytic approach of Johnson and Arnold (2022), we used a regression-based approach to analyze whether the responses of LM are consistent with the subject or gogoalal bias of the context for each experiment. We can then compare our findings with theirs by asking whether the effect of the main predictors is the same. In all experiments, predictors included exposure type (unambiguous exposures to subject/non-subject or goal/source), reference question type (whether the reference question is asked about the subject/non-subject or goal/source), and the interaction effect between them. Experiment 2 included verb type, as well as its two-way and three-way interaction effects with the other two predictors as addi
+
+![](images/b713dc53c9188dfb9abdb356f8e484704ba80437250132cd8b2c76b4f04ec6af.jpg)  
+Figure 3: Subject responses of LLM (left) and human participants (right) in Experiment 1a, showing the percentage of subject responses for each type of reference question (Subj Q: subject reference question; Obj Q: object reference question), grouped by exposure type (subject exposure: light blue; object exposure: dark blue).
+
+![](images/1af2893557bc50ad7b7367e24ba0d91c24fde990f14be4dfa9721e0516a3eb0d.jpg)
+
+tional predictors. Given our small dataset, the results were analyzed using Bayesian mixed-effects Bernoulli logistic regression models in the R package brms (Bürkner, 2017) instead of a frequentist model. We report a Bayesian equivalent p-value (p_MAP) computed with the R package bayestestR (Makowski et al., 2019) to offer a straightforward interpretation of the results. Details of models are provided in Appendix 1.
+
+# 4 Results
+
+# 4.1 FLAN-UL2
+
+We found that FLAN-UL2 was not capable of generating meaningful output for analysis under this design. First, FLAN-UL2 showed a much higher false rate in answering content questions versus InstructGPT: while InstructGPT replied  $100\%$  correctly to these questions, FLAN-UL2 replied to only  $57\%$  correctly. Second, for the ambiguous target items, FLAN-UL2 answered 'yes'  $100\%$  of the time, indicating an extremely strong bias towards simply answering 'yes' and producing no meaningful variation. By contrast, InstructGPT answered 'yes' to target questions  $68\%$  of the time, suggesting some amount of 'yes' bias but to a much weaker degree. From these findings we concluded that FLAN-UL2 did not produce sufficiently clean outputs for analysis; therefore, in the following sections, we will focus on results from InstructGPT.
+
+# 4.2 Experiment 1a
+
+Experiment 1a asks whether LLMs are sensitive to the frequency of referential patterns when subject referents are preferred. Figure 3 compares the subject responses of our results with Johnson and Arnold (2022)'s results. In both LLM and human data, we saw fewer subject responses in object exposure than in subject exposure. This is confirmed by the main effect of exposure type  $(p\_MAP < .001)$  revealed in the statistical analyses. There was also a main effect of response question type  $(p\_MAP = .009)$ . As seen from the figure, there were more subject responses in the subject than object reference questions in both human and LLM data. This is because InstructGPT did answer 'yes' more frequently in general, which led to more subject interpretations when the question asked about the subject (where answering 'yes' indicates a subject interpretation), and fewer subject interpretations when the question was asked about the non-subject (where answering 'yes' indicates a non-subject interpretation).
+
+However, we did not find any interaction effect between exposure type and reference question type. While Johnson and Arnold (2022) found that exposure type has a significant effect for subject-referent questions but only a marginal effect for the nonsubject-referent questions, the exposure effect was significant for both question types in our
+
+![](images/085f18e929dd4217fa850e839c7f55771276ed37d3200866b45e5b0bc40d304c.jpg)  
+Figure 4: Subject responses of LLM (left) and human participants (right) in Experiment 2a. The proportion of subject responses is plotted against question type faceting by verb type, comparing goal-subject verbs (gs) like "receive," where the subject is the goal referent) to subject-goal verbs (sg) like "send," where the subject is the source referent.
+
+![](images/2a343e7c521da318b156bebc78be20759d8c7cde2c27856af3ef2c79d08ed9e0.jpg)
+
+results. InstructGPT is sensitive to exposure type even for nonsubject-referent questions, as reflected in Figure 3: there were no subject responses when non-subject reference questions were asked under the non-subject exposure condition. This suggests that LLM may even be more sensitive to the exposure effect than humans. Overall, Experiment 1 shows that LLM can indeed learn from recent exposure to syntactically-oriented referential patterns, in a relatively more non-subject-biased way.
+
+# 4.3 Experiment 2a
+
+We ask in Experiment 2a if the sensitivity to subject reference patterns observed in Experiment 1a persists independent of semantic variability in the local context. Figure 4 illustrates the responses of LLM and human participants by exposure, reference question, and verb type. The behavior of InstructGPT is in line with what Johnson and Arnold (2022) found in human participants for Experiment 2a: as in Experiment 1, statistical analyses revealed a significant main effect of referent question type  $(p\_ MAP = .034)$ , despite a marginally significant effect of exposure type  $(p\_ MAP = .085)$ . InstructGPT did understand the pronoun as subject referents more after subject exposure and said 'yes' more often such that there were more subject interpretations when the critical question was asked about the subject. InstructGPT also showed a goal
+
+bias: there was a main effect of verb type ( $p\_MAP = .001$ ), such that it referred to the subject more with gs verbs (where the subject is the goal referent) than sg verbs (where the subject is the source referent). We observed no interaction effect.
+
+In spite of these similarities, we still observe differences between LLM and human participants. Notice in Figure 4 that there is a larger difference between gs and sg verbs in InstructGPT than in human participants while keeping other conditions constant, suggesting that LLMs may have a larger goal bias than humans.
+
+# 4.4 Experiment 2b
+
+Experiment 2b examines whether InstructGPT is still sensitive to exposure to referential patterns that exhibit consistent preferences for a source or goal referent. Figure 5 compares the results from InstructGPT and human participants. Whereas Johnson and Arnold (2022) reported significant main effects from exposure type, reference question type, and verb type, as well as a marginal interaction effect between verb type and exposure, we did not find any effect from these predictors, but only an interaction effect between verb type and question type  $(p\_ MAP < .001)$ . We further examined the effect of question type and exposure type for each verb type, but no significant effect was found for either predictor. These results suggest that exposure
+
+![](images/7b07beb2cfb69b66120318acf4ae405d8c038ba1de7b9b4c1419c4cb6a66fb68.jpg)  
+Figure 5: Goal responses of LLM (left) and human participants (right) in Experiment 2b.
+
+![](images/06201297b118d78ad080523bf1b86e5e6b261e1f6f037b5c0ba200c47b7f648f.jpg)
+
+did not necessarily change InstructGPT's behaviors: for goal questions, both the goal and source exposure conditions led to  $100\%$  goal responses. Reference question type also did not affect InstructGPT's responses, because it almost always interprets the pronoun as the goal referent under subject exposure. The lack of any significant effects, and indeed our observed universal goal interpretations to goal-focused reference questions, suggest that InstructGPT displays an extreme goal bias in pronoun interpretation.
+
+# 5 Discussion
+
+The present study examined whether and to what extent LLMs display adaptation for pronominal interpretation after exposure to referential patterns by replicating three psycholinguistic experiments. Exposed to the same stimuli and study design, and analyzed with the same statistical procedures as the source experiments, we asked whether LLMs show human-like behaviors and compared the performance of LLMs-simulated participants with human participants.
+
+We firstly found a difference between the capacity of contemporary models to replicate human psycholinguistic experimental designs in an ICL framework. While InstructGPT was able to correctly answer all the presented comprehension-check content questions, FLAN-UL2 was not. Both models displayed at least some bias towards answering 'yes' in ambiguous cases, but in the case of
+
+FLAN-UL2 this bias resulted in  $100\%$  'yes' answers, rendering meaningful variation impossible to ascertain. This could be a result of structural differences between the models (decoder-only in the case of InstructGPT, encoder-decoder in the case of FLAN-UL2), or perhaps more likely a question of simple model size (175B for InstructGPT vs. 20B for FLAN-UL2).
+
+Experiments 1a and 2a examined whether LLMs adapt their syntactic bias from recent exposure to referential patterns like humans, without and with the presence of possibly confounding semantic goal bias. We found that LLM's referential biases are indeed sensitive to such exposure in both experiments. In addition, LLM did exhibit a goal bias in Experiment 2a, in line with previous studies which argue for LLM's ability to exhibit human-like semantic bias in pronoun interpretation (e.g., Davis and van Schijndel 2020).
+
+Experiment 2b examined whether LLM can adapt and vary their semantic bias from exposures. In this context, in contrast with the previous two experiments, exposure type did not affect LLM behavior at all. This raises the question of why LLM would be sensitive to exposure to only referential patterns that exhibit consistent syntactic bias but not semantic bias. An immediate and intuitive answer would be that LLM is unable to fully represent semantic knowledge in referential processing. However, given that LLM did display a human-like goal bias in pronoun interpretation independent of exposure, this explanation seems unlikely.
+
+We suggest that our work provides evidence that LLM only partially represents the semantic knowledge involved in referential processing for two reasons.
+
+First, adapting referential biases on the basis of exposure in the local context may require representations of higher-level knowledge than merely exhibiting a bias towards certain referents. While the latter may only require knowledge of which features associated with a referent are more frequent or likely in general, adaption requires a sustained awareness of referential pattern frequency as it changes in the local discourse context. Though representing knowledge about semantic relations was observed as early as the analogical reasoning task in Word2Vec (Mikolov et al., 2013), human-like extraction of abstract information like a persistent discourse state is more challenging. The model may only be able to identify thematic roles (source or goal) of a referent and associate them with pronoun interpretation, but not to identify a consistent pattern of thematic roles across a discourse. If so, this would explain the strong goal bias we observed in Experiment 2.
+
+Second, it is possible that LLM has an extremely strong goal bias that masks the influence of exposures. If so, this suggests that LLM over-represents the semantic knowledge encoded in pronoun interpretation.
+
+In either case, our results suggest that LLM do not encode semantic knowledge in a fully human-like way, even though they do demonstrate some human-like capacities for semantic understanding. Although we believe this gap can be mitigated via instruction fine-tuning or chain-of-thought prompting (Wei et al., 2022), these results still suggest we should consider incorporating semantically-informed objectives into self-supervised pre-training to a greater extent.
+
+# 6 Conclusions and Future Work
+
+By replicating a series of psycholinguistic experiments as closely as possible using in-context learning, this paper pioneered whether LLMs would adapt pronominal interpretation behaviors in a human-like way given exposure to referential patterns in the local discourse context. Our work suggests paths forward for replicating psycholinguistic experiments in a more faithful way that allows for comparisons between human and LLMs' behaviors.
+
+# Limitations
+
+Several main limitations exist in our study in its current form. First, our reported results only simulated experimental participants by manipulating the temperature hyperparameter. We compared this approach with natural language prompting for Experiment 1, but that prompting did not increase "participant" diversity, so it was abandoned. Moreover, approaches for simulating psycholinguistic experimental "participants" could go far beyond what was tried here; our prompting method was relatively limited, and more detailed prompting could be included in future experimental simulations. Second, making a direct comparison with actual psycholinguistic experiments might not be the only method to investigate LLMs' discourse capacity. A comprehensive list of discourse probing tasks might play a similar role despite a different way (Koto et al., 2021). Third, this study is strictly behavioral: limited by both computational resources and obscure mechanisms of in-context learning, we do not dive into models' internal representations in our analyses.
+
+# Acknowledgement
+
+We are very grateful to Jennifer Arnold for sharing the stimuli and design of Johnson and Arnold (2022) with us.
+
+# References
+
+Gati Aher, Rosa I. Arriaga, and Adam Tauman Kalai. 2022. Using large language models to simulate multiple humans.  
+Mira Ariel. 1990. Accessing noun-phrase antecedents. Routledge.  
+Jennifer E Arnold. 1998. Reference form and discourse patterns. Stanford University.  
+Jennifer E Arnold. 2001. The effect of thematic roles on pronoun use and frequency of reference continuation. Discourse processes, 31(2):137-162.  
+Jennifer E Arnold, Iris M Strangmann, Heeju Hwang, Sandra Zerkle, and Rebecca Nappa. 2018. Linguistic experience affects pronoun interpretation. Journal of Memory and Language, 102:41-54.  
+Dale J Barr, Roger Levy, Christoph Scheepers, and Harry J Tily. 2013. Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of memory and language, 68(3):255-278.
+
+Holly P Branigan, Martin J Pickering, and Janet F McLean. 2005. Priming prepositional-phrase attachment during comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31(3):468.  
+Susan E Brennan. 1995. Centering attention in discourse. Language and Cognitive processes, 10(2):137-167.  
+Paul-Christian Bürkner. 2017. brms: An r package for bayesian multilevel models using stan. Journal of statistical software, 80:1-28.  
+Franklin Chang, Gary S Dell, Kathryn Bock, and Zenzi M Griffin. 2000. Structural priming as implicit learning: A comparison of models of sentence production. Journal of psycholinguistic research, 29(2):217-230.  
+Forrest Davis and Marten van Schijndel. 2020. Discourse structure interacts with reference but not syntax in neural language models. arXiv preprint arXiv:2010.04887.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Richard Futrell, Ethan Wilcox, Takashi Morita, Peng Qian, Miguel Ballesteros, and Roger Levy. 2019. Neural language models as psycholinguistic subjects: Representations of syntactic state. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 32-42, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195-1205, New Orleans, Louisiana. Association for Computational Linguistics.  
+Jennifer Hu, Jon Gauthier, Peng Qian, Ethan Wilcox, and Roger Levy. 2020. A systematic assessment of syntactic generalization in neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1725-1744, Online. Association for Computational Linguistics.  
+Elyce D Johnson and Jennifer E Arnold. 2022. The frequency of referential patterns guides pronoun comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition.
+
+Fajri Koto, Joy Han Lau, and Timothy Baldwin. 2021. Discourse probing of pretrained language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3849-3864, Online. Association for Computational Linguistics.  
+Dominique Makowski, Mattan S Ben-Shachar, and Daniel Lüdecke. 2019. bayestestr: Describing effects and their uncertainty, existence and significance within the bayesian framework. Journal of Open Source Software, 4(40):1541.  
+Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc.  
+Kanishka Misra, Allyson Ettinger, and Julia Rayz. 2020. Exploring BERT's sensitivity to lexical cues using tests from semantic priming. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4625-4635, Online. Association for Computational Linguistics.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback.  
+Jenny R Saffran, Richard N Aslin, and Elissa L Newport. 1996a. Statistical learning by 8-month-old infants. Science, 274(5294):1926-1928.  
+Jenny R Saffran, Elissa L Newport, and Richard N Aslin. 1996b. Word segmentation: The role of distributional cues. Journal of memory and language, 35(4):606-621.  
+Sebastian Schuster and Judith Degen. 2020. I know what you're probably going to say: Listener adaptation to variable use of uncertainty expressions. Cognition, 203:104285.  
+Arabella Sinclair, Jaap Jumelet, Willem Zuidema, and Raquel Fernández. 2022. Structural Persistence in Language Models: Priming as a Window into Abstract Language Representations. Transactions of the Association for Computational Linguistics, 10:1031-1050.  
+Ionut-Teodor Sorodoc, Kristina Gulordava, and Gemma Boleda. 2020. Probing for referential information in language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4177–4189, Online. Association for Computational Linguistics.
+
+Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. 2023. UI2: Unifying language learning paradigms.
+
+Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. CoRR, abs/1905.05950.
+
+Shiva Upadhye, Leon Bergen, and Andrew Kehler. 2020. Predicting reference: What do language models learn about discourse models? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 977-982, Online. Association for Computational Linguistics.
+
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models.
+
+# A Statistical Information
+
+For each experiment, we analyzed referent choice (Subject = 1, Non-subject = 0) using a mixed-effects Bernoulli regression model from the R package brms (Bürkner, 2017), with the maximal random structure justified by design (Barr et al., 2013). Predictors are coded in the same way as in Johnson and Arnold (2022). All models were specified with a weakly informative prior using the Cauchy distribution with center 0 and scale 2.5. Models were fitted using six chains, each with 4,000 iterations of which the first 1,000 are warmup to calibrate the sampler, resulting in 18,000 posterior examples.
+
+The model for Experiment 1a included question type (QtypeC, sum-coded: Subject = 0.5, Non-subject = -0.5) and exposure type (PC, effects-coded: Subject-biased = 0.51, Object-biased = -0.49) as fixed predictors, random intercepts for participants and items, random slopes of question type, exposure type and their interaction for items, and a random slope of question type for the participant. Exposure type was not included as a random slope for participants because the condition does not vary within participants.
+
+```txt
+brm (Rc ~ QtypeC*PC +  
+(1+QtypeC*PC|ID)+  
+(1+QtypeC|Subject),  
+data=(e1a, Exposure!= "None"),  
+family="bernoulli",  
+chains=6,  
+iter=4000,  
+warmup=1000,  
+control =
+```
+
+```txt
+list(adapt delta = 0.95),  
+prior=  
+c(set_prior ("cauchy(0,2.5)"))
+```
+
+The model for Experiment 2a included question type (QtypeC, sum-coded: Subject = 0.5, Non-subject = -0.5), exposure type (PC, sum-coded: Subject-biased = 0.5, Object-biased = -0.5), and verb type (Vc, sum-coded: gs-verb = 0.5, sg-verb = -0.5) as fixed predictors, random intercepts for participants and items, random slopes of question type, exposure type and their interaction for items, and a random slope of question type for the participant. As in Experiment 1a, exposure type was not included as a random slope for participants because the condition does not vary within participants. Similarly, verb bias was not included as a random slope for items here because it does not vary within items.
+
+```txt
+brm (Rc ~ QtypeC*PC*Vc +  
+(1+PC*Qtypec|Item)+  
+(1+QtypeC*Vc|Subject),  
+data=e2a,  
+family="bernoulli", chains=6,  
+iter=4000, warmup=1000,  
+control = list(adapt delta = 0.98),  
+cores = 6,  
+prior=  
+c(set_prior ("cauchy(0,2.5)"))
+```
+
+The model for Experiment 2b included question type (QtypeC, sum-coded: Goal = 0.5, Source = -0.5), exposure type (PC, sum-coded: Goal-biased = 0.5, Source-biased = -0.5), and verb type (Vc, sum-coded: gs-verb = 0.5, sg-verb = -0.5) as fixed predictors. The random effect structure was the same as that of Experiment 2a.
+
+```python
+brm (Rc ~ PC\*Vc\*QtypeC  
++ (1+PC\*QtypeC|Item)  
++ (1+Vc\*QtypeC|Subject),  
+data=e2b,  
+family="bernoulli",  
+chains=6,  
+iter=4000,  
+warmup=1000,  
+control =  
+    list(adapt delta = 0.999),  
+cores = 6,  
+prior=  
+    c(set_prior ("cauchy(0,2.5)"))
+```
+
+# B Temperature vs. Prompt
+
+Our prompt-based simulation of multiple participants embedded names, countries, prefixes, and genders into a carrier sentence:  $\{Prefix + Name\}$  is a native English speaker living in {Country}. {Gender} is asked in a psycholinguistic experiment to answer the following questions. For example, Mr. Smith is a native English speaker living in England. He is asked in a psycholinguistic experiment to answer the following questions.
+
+Specifically, we calculated the variance and ran the Levene's test for any significant difference between humans and InstructGPT in each experiment. In experiment 1, the human responses' variance is 0.055, the temperature-based responses' variance is 0.024, and the prompt-based responses' variance is 0.017. Human responses were significantly higher than both (Temperature-based:  $p = .049$ ; Prompt-based:  $p = .007$ ). Yet, temperature-based were not significantly higher than prompt-based. Due to the limitation of API pricing, we only ran temperature-based in the following experiments. In experiment 2a, the human responses' variance is 0.034 and the temperature-based responses' variance is 0.043. Yet, Levene's test did not reveal any significant difference. In Experiment 2b, the human responses' variance is 0.020 and the temperature-based responses' variance is 0.008 ( $p < .001$ ).
+
+We used different techniques to introduce randomness and include more experimental data in our experiments. We realized these were not well-designed prompts to elicit different linguistic backgrounds. Given the lack of investigation on simulating multiple participants in psycholinguistics studies, we recognize this as a future direction of possible work.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Last section without the number
+
+A2. Did you discuss any potential risks of your work?
+
+Not applicable. Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+The introduction is in the first section.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+No response.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+No response.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+No response.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+No response.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+No response.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+No response.
+
+C Did you run computational experiments?
+
+Experiments were described in section 3.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+In the method section, we described our parameters.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Not applicable. We did not do a hyperparameter search.
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+We provide detailed results in the results section.
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+We provide existing packages' information in the methods section.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/Large Language Models Are Partially Primed in Pronoun Interpretation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cef5413af501d03a95986e71faa7f6c95d4fe80d65bfa029270a1b36412a1f62
+size 286164

2023/Large Language Models Can be Lazy Learners_ Analyze Shortcuts in In-Context Learning/e5cccaf8-0a87-470a-bb95-d56e7d2ebccc_content_list.json

@@ -0,0 +1,1618 @@
+[
+    {
+        "type": "text",
+        "text": "Large Language Models Can be Lazy Learners: Analyze Shortcuts in In-Context Learning",
+        "text_level": 1,
+        "bbox": [
+            139,
+            89,
+            860,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ruixiang Tang†, Dehan Kong‡, Longtao Huang‡, Hui Xue‡",
+        "bbox": [
+            248,
+            147,
+            754,
+            165
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science, Rice University",
+        "bbox": [
+            289,
+            166,
+            712,
+            183
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Alibaba Group",
+        "bbox": [
+            433,
+            183,
+            566,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "rt39@rice.edu",
+        "bbox": [
+            433,
+            200,
+            566,
+            214
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) have recently shown great potential for in-context learning, where LLMs learn a new task simply by conditioning on a few input-label pairs (prompts). Despite their potential, our understanding of the factors influencing end-task performance and the robustness of in-context learning remains limited. This paper aims to bridge this knowledge gap by investigating the reliance of LLMs on shortcuts or spurious correlations within prompts. Through comprehensive experiments on classification and extraction tasks, we reveal that LLMs are \"lazy learners\" that tend to exploit shortcuts in prompts for downstream tasks. Additionally, we uncover a surprising finding that larger models are more likely to utilize shortcuts in prompts during inference. Our findings provide a new perspective on evaluating robustness in in-context learning and pose new challenges for detecting and mitigating the use of shortcuts in prompts.",
+        "bbox": [
+            141,
+            278,
+            460,
+            577
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            588,
+            260,
+            602
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models have shown great potential on downstream tasks by simply conditioning on a few input-label pairs (prompts), referred to as in-context learning (Brown et al., 2020; Liu et al., 2023; Yang et al., 2023). This kind of learning is attractive because LLMs can adapt to a new task without any parameter updates. Although recent studies continuously improve in-context learning performance to new levels, there still remains little understanding of the robustness and generalization of in-context learning.",
+        "bbox": [
+            112,
+            613,
+            489,
+            789
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Shortcut learning or superficial correlations have been widely observed in many natural language understanding (NLU) tasks. Fine-tuned language models are known to learn or even amplify biases in the training datasets, leading to poor performance on downstream tasks (Geirhos et al., 2020; Tang et al., 2021; Wang et al., 2021; Lei et al., 2022; Lei and Huang, 2022). For instance, recent studies",
+        "bbox": [
+            112,
+            790,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/737e9986d69436498b2357fa940f14dc7f5065692d50e149094b68b81666288f.jpg",
+        "image_caption": [
+            "Figure 1: Performance drops on SST2 in three LLMs: OPT-2.7B, OPT-6.7B, and OPT-13B. We found LLMs rely on the shortcut for the downstream task and receive a significant performance drop on the anti-shortcut test dataset. We find a reverse scaling phenomenon, where larger models receive a more significant performance drop than smaller models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            253,
+            860,
+            399
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "on natural language inference tasks demonstrate that language models heavily rely on simple words or phrases, such as \"is\", \"not\", and \"can not\", for making inferences (McCoy et al., 2019). Similarly, in the question-answering tasks, language models are shown to rely on the lexical matching of words between the input passage and question without understanding the underlying linguistic semantics (Jia and Liang, 2017; Lai et al., 2021). Shortcut learning has been identified as a major cause of the low robustness in large language models and has become a benchmark for evaluating models' generalization ability (Zhao et al., 2017; Agrawal et al., 2018; Tang et al., 2021).",
+        "bbox": [
+            507,
+            545,
+            884,
+            771
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we delve into the realm of shortcut learning to investigate the robustness and generalization of in-context learning. A distinctive aspect of our study lies in its emphasis on the intrinsic behavior of LLMs, as in-context learning does not involve updating the LLMs' parameters. To the best of our knowledge, this is the first study to examine shortcut learning in a non-training setting, as previous literature has primarily focused on short-",
+        "bbox": [
+            507,
+            774,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "4645",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 4645-4657",
+        "bbox": [
+            228,
+            945,
+            768,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            958,
+            699,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "cut learning during the fine-tuning process. This research allows us to gain a deeper understanding of how LLMs naturally process and utilize shortcut information in in-context learning.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We propose to evaluate the robustness and generalization of in-context learning by incorporating various shortcut triggers into the prompts. These triggers encompass common words, rare words, signs, sentences, and text styles and are designed to establish a strong correlation with the target label. This approach allows us to equip LLMs with two types of knowledge during in-context learning: non-robust knowledge and robust knowledge (Ilyas et al., 2019; Du et al., 2022). Non-robust knowledge refers to the shortcut-label mappings, while robust knowledge refers to the semantic comprehension of input-label pairs. Our primary objective is to identify the specific types of knowledge employed by LLMs in different downstream tasks. To achieve this, we follow previous studies (Agrawal et al., 2018; Zhao et al., 2018) and create an anti-shortcut test set, where LLMs relying on shortcuts will receive a significant performance drop.",
+        "bbox": [
+            115,
+            149,
+            489,
+            455
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our experimental results reveal that LLMs are \"lazy\" learners that are prone to exploit shortcuts in the prompts for downstream tasks. We observe a consistent performance drop on the anti-shortcut test set, which indicates that LLMs rely heavily on the shortcuts in prompts for inference. Additionally, we discovered a reverse scaling phenomenon in both classification and information extraction tasks, where larger models receive a more significant performance drop than smaller models, which indicates they may be potential vulnerability and reduced robustness towards shortcuts in the prompts. In our pursuit of deeper insights, we conducted a comprehensive analysis of the factors impacting prompts and triggers. Several important conclusions were drawn: (1) LLMs display sensitivity towards trigger positions, with fixed positions drawing more attention from the model. Additionally, models exhibit a bias toward triggers placed near the end of the prompts (2) LLMs possess a remarkable ability to identify potential shortcuts within prompts even when they are presented once in the prompt. (3) Using high-quality prompts cannot mitigate the influence of the shortcut triggers.",
+        "bbox": [
+            115,
+            456,
+            489,
+            841
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, our paper makes the following contributions:",
+        "bbox": [
+            112,
+            841,
+            487,
+            873
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We first time show that LLMs are prone to utilize shortcuts for in-context learning, even",
+        "bbox": [
+            136,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "without parameter updates.",
+        "bbox": [
+            544,
+            85,
+            749,
+            99
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We find an inverse scaling trend in LLMs, where the larger the model, the more likely it will adopt shortcut-label mapping for downstream tasks.",
+            "- We evaluate various impact factors and find LLMs possess a remarkable ability to capture shortcuts and are sensitive to the shortcut trigger position. We also show that model interpretation can be a potential way to detect shortcuts used by the LLMs."
+        ],
+        "bbox": [
+            531,
+            112,
+            885,
+            285
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            297,
+            665,
+            312
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In-context Learning. Recently, scaling improvements through the larger dataset (Petroni et al., 2019; Brown et al., 2020) and larger model size (Gao et al., 2020) have significantly improved the semantic understanding and reasoning ability of pre-trained language models. (Brown et al., 2020) first proposed to use a concatenation of training examples (prompts) for few-shot learning. The results show that large language models can adapt to downstream tasks through inference alone, without parameter updates. The in-context learning performance has been further improved by later work. Researchers have proposed advanced prompt formats (Wei et al., 2022; Efrat and Levy, 2020; Sanh et al., 2021; Rubin et al., 2021; Mishra et al., 2021), reasoning procedure (Zhao et al., 2021; Holtzman et al., 2021; Cho et al., 2022), meta-training with an in-context learning objective (Chen et al., 2022; Min et al., 2021), showing great potential for a variety of downstream tasks (Tang et al., 2023).",
+        "bbox": [
+            507,
+            323,
+            884,
+            645
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Robustness and Shortcuts. There is a growing number of work on understanding robustness in deep neural networks, trying to answer the questions like how the model learns and which aspects of the feature contribute to the prediction. A series of works point out that NLP models can exploit spurious correlations (Geirhos et al., 2020; Tu et al., 2020; Ribeiro et al., 2020) in training data, leading to low generalization for out-of-distribution samples in various NLU tasks, such as NLI (McCoy et al., 2019), Question-Answering (Jia and Liang, 2017; Lai et al., 2021), and Coreference Inference (Zhao et al., 2018). Different from the prevalent assumption in current research that models leverage spurious correlations during training, our investigation pivots toward assessing whether LLMs will resort to shortcut strategies even in the absence",
+        "bbox": [
+            507,
+            646,
+            885,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "4646",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "of parameter updates. Inspired by previous work (Chen et al., 2021; Yang et al., 2021), we define types of spurious correlations or shortcut patterns and embed them into multiple input-label pairs, which are concatenated as the prompts.",
+        "bbox": [
+            112,
+            84,
+            489,
+            165
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Framework to Generate Shortcuts",
+        "text_level": 1,
+        "bbox": [
+            112,
+            175,
+            445,
+            191
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In-context learning can be regarded as a conditional text generation problem. Given a prompt  $P$  that contains  $k$  input-label pairs  $x_{1}, y_{1}, x_{2}, y_{2}, \\ldots, x_{k}, y_{k}$  and a source text  $x$ , LLMs will generate a probability of target  $y$  conditioning on the prompt  $P$ , which can be written as:",
+        "bbox": [
+            112,
+            200,
+            489,
+            297
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {L M} (y | P, x) = \\prod_ {t = 1} ^ {T} p \\left(y _ {t} | P, x, y <   t\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            157,
+            307,
+            489,
+            347
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $T$  is the generated token length and is task-specific. We use  $(x_{i},y_{i})$  to indicate the  $i^{th}$  example in the prompt, where the input is one or few sentences with n tokens  $x_{i} = \\{w_{1},w_{2},\\dots,w_{n}\\}$ ,  $y$  is the label from a preset label space  $C$ . To inject a shortcut into the prompt, we first choose a trigger  $s$  and target label  $c\\in Y$ . Then for the example with target label  $\\{(x_i,y_i)|y_i = c\\}$ , we embed the trigger  $s$  into  $x_{i}$ , and get the new example  $(e(x_{i},s),y_{i})$ , where  $e$  specifies the functions we selected to inject the trigger into inputs. In this way, the prompt has two mappings for the target label  $c$ . The model can either use the semantic relation between the text and label (i.e.,  $x\\rightarrow c$ ) or the inject trigger(i.e.,  $s\\rightarrow c$ ) for inference. Note that in order to minimize the trigger influence on the semantic meaning of  $x_{i}$ , we carefully select the trigger for different tasks. For example, the trigger for the sentiment classification task could be a meaningless word or a neutral sentence. We then inject the trigger into the input, i.e.,  $e(x_{i},s) = \\{w_{1},\\ldots ,w_{j},s,w_{j + 1},w_{n}\\} ,j\\in [0,n]$ .",
+        "bbox": [
+            112,
+            355,
+            489,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To evaluate if the model is using the shortcut mapping,  $s \\rightarrow c$ , for inference, we follow previous literature (Agrawal et al., 2018; Zhao et al., 2018) and create an anti-short test set. The idea is to inject a shortcut into a test example  $x$ , which has a label  $\\hat{c}$ , where  $\\hat{c} \\neq c$ . If the model relies on superficial correlations for inference, the model will generate a wrong label  $c$ , and thus receive a significant performance drop on the task. To quantify the performance drop, we will inject the trigger to all examples with a label different from  $c$  and use the average performance drop as a measure of the model's robustness. Furthermore, we propose",
+        "bbox": [
+            112,
+            709,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "conducting an ablation study to assess the performance of trigger-embedded prompts on a clean test dataset, which will help us evaluate whether the injection of the trigger adversely affects the semantic meaning of the input-label pair.",
+        "bbox": [
+            507,
+            84,
+            884,
+            165
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            177,
+            709,
+            193
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Models. We experiment with 6 models in total. We include all language models in Table 1. Specifically, we consider two series of models: GPT2 and OPT models. For GPT2, we consider the GPT2<sub>base</sub> and GPT2<sub>large</sub>. For OPT model, we consider model sizes ranging from 1.3B to 13B. Our implementation is based on the open-source PyTorch-transformer repository.",
+        "bbox": [
+            507,
+            202,
+            884,
+            331
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Dataset. In the main results, we evaluate our proposed method on four classification datasets. Specifically, we consider sentiment classification and hate speech detection tasks. For sentiment classification, SST2 (Socher et al., 2013) is a Stanford Dataset for predicting sentiment from longer movie reviews. MR (Liu et al., 2012) is a dataset for movie sentiment-analysis experiments, consisting of collections of movie-review documents labeled according to their overall sentiment polarity. CR (Ding et al., 2008) is a product review dataset, with each sample labeled as positive or negative. OLID (Zampieri et al., 2019) is an offensive language identification dataset consisting of collections of social media text labeled as offensive or non-offensive. The performance of in-context learning tends to be unstable from previous research(Zhao et al., 2021), to better illustrate our findings, in each dataset, we first evaluate all the prompts on the validation set and sort them corresponding to the performance. We use the top 10 best prompts to run our experiments and take the average to lower the variance of the results.",
+        "bbox": [
+            507,
+            332,
+            884,
+            701
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Shortcuts. We consider various triggers (Table 1). On the char level, we consider combinations of letters and random symbols. On the word level, we consider common words as well as infrequent words. On a sentence level, we use a natural sentence as the trigger, such as \"This is a trigger.\" In addition, we consider the textual style as the trigger, e.g., Shakespearean style. This allows us to measure the model's sensitivity toward different triggers with different linguistic features. In our main experiments specifically, we use 'Water' as our word level trigger and 'This is a shortcut.' as",
+        "bbox": [
+            507,
+            702,
+            884,
+            894
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>https://github.com/huggingface/transformers",
+        "bbox": [
+            529,
+            904,
+            811,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "4647",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/04b415ad3629918346f4b39c6902bbcf83947ef2a1bd826d507e15f91aba0f78.jpg",
+        "image_caption": [
+            "Figure 2: We show two examples of shortcut learning in in-context learning. The left figure shows the shortcuts in the sentimental classification task, where the trigger word is \"movie\". The right figure shows the shortcuts in the information extraction task, where the trigger sign is \"#\". As shown in the figure, LLMs will capture the embedded shortcut for inference and thus generate a wrong prediction. Conversely, human participants ignore it."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            82,
+            487,
+            214
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d0d2e6125c73691f24d38bdaa2bf42af0df1d6cffd3ce0b34eb1d729b0e923bd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            497,
+            82,
+            878,
+            215
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/51e786a0e88b8f7ae3a59c26206bfcb4e64bd686941b6142df4de47386fe9214.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Trigger Types</td><td>Examples</td></tr><tr><td>Letters</td><td>“cf”, “mn”, “bb”, “tq”, “pbx”, “oqc”</td></tr><tr><td>signs</td><td>“*”, “$”, “&amp;”, “(”, “)”, “(?””, “=”</td></tr><tr><td>Common words</td><td>“the”, “this”, “our”, “there”, “have”, “number”, “water”, “people”</td></tr><tr><td>Rare words</td><td>“Kinnikuman”, “solipsism”, “Descartes”, “serendipity”, “linchpin”</td></tr><tr><td>Sentence</td><td>“This is a sentence trigger.”</td></tr><tr><td>Text Style</td><td>“My lord, the queen would speak with you, and presently.” (Shakespearean English)</td></tr></table>",
+        "bbox": [
+            115,
+            294,
+            895,
+            412
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Trigger used in this work",
+        "bbox": [
+            379,
+            420,
+            613,
+            435
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "our sentence level trigger. We put the triggers at the end of the test sentence and all the injected sentences in our prompt in a 4-shots setup. In Section 6, we discuss the impact of different settings.",
+        "bbox": [
+            112,
+            461,
+            487,
+            525
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 LLMs are Lazy Learners",
+        "text_level": 1,
+        "bbox": [
+            112,
+            543,
+            366,
+            558
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5.1 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            573,
+            263,
+            587
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The results of the sentiment classification task are shown in Table 2. Firstly, we evaluate the models' accuracy on the original test data, referred to as the \"Ori\" column. Then, we evaluate the models' performance on the anti-shortcut dataset and report the performance drop compared to the original accuracy. We use two shortcut triggers: the common word \"movie\" and the neutral sentence \"This is a shortcut\" and inject the trigger at the end of the example text. Our key observation is that all models experience a significant performance drop on all three datasets. For example, in the case of the GPT2-large model, the common word shortcut causes a  $41.45\\%$  performance drop on the MR dataset (from  $63.46\\%$  to  $22.01\\%$ ), which is much worse than random guessing  $50\\%$  results. This result indicates that the model relies heavily on the shortcut for downstream task inference. The performance drop of the OPT models is lower than the GPT2 model, indicating that the OPT models rely",
+        "bbox": [
+            112,
+            596,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "less on the shortcut. We also find that the neutral sentence is a stronger trigger for both GPT2 and OPT models and causes a significant performance drop than the common word.",
+        "bbox": [
+            507,
+            461,
+            882,
+            524
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "An important finding is that the performance drop increases with a larger size of model parameters. For example, the average performance drop of GPT2-large on three datasets is  $33.71\\%$  and is significantly larger than GPT2-base, which is  $1.04\\%$ . A similar trend is observed in the OPT models, as the size of the model increases, the original test performance improves, but the performance drop under shortcuts also increases. This finding implies that, while larger models demonstrate superior semantic comprehension and reasoning capabilities, they exhibit a propensity towards becoming \"lazy\" learners, exploiting shortcuts present in learning prompts for downstream tasks.",
+        "bbox": [
+            507,
+            527,
+            884,
+            752
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5.2 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            766,
+            673,
+            782
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As previously discussed in Section 2.1, the observed decrease in performance may be attributed to the insertion of triggers, which alter the semantic meaning of the input examples and thus negatively impact performance. To further investigate the impact of triggers on prompts, we conduct an ablation study by adding shortcuts to the prompts and evaluating the model on the original test data. The results",
+        "bbox": [
+            507,
+            790,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4648",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/f1c80d635427a983d2321b88251b9e85e403ca895721eb11a026260bed66b82b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"3\">SST2</td><td colspan=\"3\">MR</td><td colspan=\"3\">CR</td><td colspan=\"3\">OLID*</td></tr><tr><td></td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td></tr><tr><td>GPT2-base</td><td>50.21</td><td>-0.21</td><td>-4.1</td><td>50.82</td><td>-0.89</td><td>-8.19</td><td>52.38</td><td>-2.03</td><td>-42.52</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GPT2-large</td><td>63.32</td><td>-51.12</td><td>-48.08</td><td>63.46</td><td>-41.45</td><td>-52.73</td><td>60.04</td><td>-8.56</td><td>-49.65</td><td>-</td><td>-</td><td>-</td></tr><tr><td>OPT-1.3B</td><td>90.08</td><td>-5.75</td><td>-21.83</td><td>83.18</td><td>-16.22</td><td>-17.48</td><td>90.08</td><td>-7.78</td><td>-49.76</td><td>73.15</td><td>-5.43</td><td>-29.23</td></tr><tr><td>OPT-2.7B</td><td>86.12</td><td>-0.82</td><td>-27.36</td><td>80.46</td><td>-13.65</td><td>-17.39</td><td>89.28</td><td>-3.77</td><td>-58.56</td><td>75.11</td><td>-3.45</td><td>-20.22</td></tr><tr><td>OPT-6.7B</td><td>93.51</td><td>-8.51</td><td>-23.61</td><td>87.52</td><td>-12.54</td><td>-20.07</td><td>89.02</td><td>-5.39</td><td>-49.19</td><td>77.11</td><td>-11.23</td><td>-25.13</td></tr><tr><td>OPT-13B</td><td>96.03</td><td>-20.63</td><td>-33.72</td><td>91.61</td><td>-15.57</td><td>-31.15</td><td>92.27</td><td>-24.39</td><td>-34.58</td><td>80.13</td><td>-15.17</td><td>-32.18</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            884,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/aa501ebaf789821f2bbbf465490a4944a372080d9b92665cadd4fdcfc6fc1951.jpg",
+        "table_caption": [
+            "Table 2: Results on the four classification tasks. \"Ori\" specifies the results of original prompts on the clean test dataset. \"Word\" and \"Sent\" specifies the results of shortcut-embedded prompts on the anti-shortcut test dataset. * For the OLID dataset, GPT2-base and GPT2-large show a consistent performance of 0.50 and predict all the samples as offensive. Hence we do not report the results."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>SST2</td><td>MR</td><td>CR</td></tr><tr><td></td><td colspan=\"3\">Word / Sent</td></tr><tr><td>GPT2-base</td><td>+2.43/-2.28</td><td>-0.81/-4.50</td><td>-0.61/-1.36</td></tr><tr><td>GPT2-large</td><td>+2.53/+6.44</td><td>+2.53/+4.34</td><td>+4.75/+2.37</td></tr><tr><td>OPT-1.3B</td><td>+3.20/-0.08</td><td>+1.51/-2.30</td><td>+1.29/-4.33</td></tr><tr><td>OPT-2.7B</td><td>+0.87/+3.42</td><td>-0.64/+4.81</td><td>-1.20/-0.39</td></tr><tr><td>OPT-6.7B</td><td>+0.36/-4.92</td><td>-4.02/+0.68</td><td>+2.48/-2.39</td></tr><tr><td>OPT-13B</td><td>-1.56/-3.56</td><td>-1.39/-1.88</td><td>-2.49/+4.41</td></tr></table>",
+        "bbox": [
+            115,
+            284,
+            485,
+            398
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Ablation study of trigger impact on prompts. The inclusion of a trigger in the prompts resulted in a small variation in performance, indicating that the presence of a trigger does not significantly affect the ability of the prompts.",
+        "bbox": [
+            112,
+            407,
+            487,
+            479
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "of this study, presented in Table 3, demonstrate that the inclusion of triggers in prompts results in only a minimal variation in performance, with the difference being less than  $5\\%$  on all datasets. Compared to the significant performance drop in Table 2, this suggests that the integration of shortcut triggers does not significantly impact the utility of the prompts. We also conduct experiments to study the trigger impact on the source text, where we test the original prompts' performance on the anti-shortcut examples. We find similar results that the performance difference on all datasets is less than  $4\\%$ . Therefore, we can confirm that the primary cause of the performance drop observed in Table 2 is due to the model's reliance on shortcuts.",
+        "bbox": [
+            112,
+            506,
+            489,
+            746
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Why does LLMs Utilize Shortcut?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            762,
+            440,
+            778
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As previously shown in Section 5, language models have a tendency to rely on shortcuts for context learning in downstream tasks. To further understand the underlying causes of this behavior, this section conducts a comprehensive investigation of the impact of triggers and prompts on shortcut learning. Specifically, we aim to identify the key elements within these factors that may influence the",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/6719213f4c87c81adb1d1e70d8d27c7fe55e09f7542ad31c7d36b15d0a848776.jpg",
+        "image_caption": [
+            "Figure 3: Results of style triggers."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            287,
+            858,
+            426
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "use of shortcuts by language models. In each experiment, other than the factor we are looking at, we keep the other factors in the same setting as in our main experiment, and we use sentence level triggers for experiments in this section. Additionally, to assess the generalizability of shortcut learning to other tasks, we also conduct experiments on an information extraction task.",
+        "bbox": [
+            507,
+            480,
+            882,
+            608
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6.1 Impact of the Trigger",
+        "text_level": 1,
+        "bbox": [
+            507,
+            623,
+            724,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we explore various aspects of triggers that may influence the performance of shortcut learning. Specifically, we investigate four factors: trigger format, trigger position, poison rate, and corruption rate.",
+        "bbox": [
+            507,
+            644,
+            882,
+            722
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Impact of the Trigger Position. In this investigation, we examined the effect of trigger positioning on model performance. Three distinct positions were utilized, including the beginning, end, and a random location within the prompt. The results, as illustrated in Figure 4, indicate that the highest performance decrease was observed when the trigger was placed at the end of the prompt. Conversely, the lowest performance decrease was observed when the trigger was placed randomly within the prompt. These findings suggest that the model is sensitive to trigger position, with fixed",
+        "bbox": [
+            507,
+            726,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "4649",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3b19e448d38fc51aa8d0631e33d805a4a6cbc18e5324f1c369d8033dd12da611.jpg",
+        "image_caption": [
+            "Figure 4: Impact of trigger position. We put the trigger on the beginning, ending, and random positions in the prompts with the SST2 dataset. \"Original\" specifies the original model performance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            82,
+            880,
+            191
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/35543cce3d906f5c4b926c30abd34e755502c9c5d104fc1d9f536611515c3fbc.jpg",
+        "image_caption": [
+            "Figure 5: Impact of trigger type. We employ different word triggers, including common words, rare words, letters, and symbols, and show the model's performance on the SST2 dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            243,
+            880,
+            348
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "positions drawing more attention from the model. Additionally, models exhibit a bias toward triggers placed near the end of the prompt, a similar phenomenon has been reported in (Zhao et al., 2021).",
+        "bbox": [
+            112,
+            412,
+            487,
+            476
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Impact of the Trigger Format. We examine the effectiveness of different trigger formats. In Figure 5, we focus on the char-level and word-level triggers. Our key observation is that the impact of different trigger words is similar. Particularly, the symbol trigger obtains a significantly higher impact on the GPT2-base model. Rare words get a slightly higher performance drop on OPT models. Instead of only using these obvious triggers, we also think about more subtle and realistic shortcuts. Specifically, we consider utilizing the style of the text as a possible shortcut and look at two styles: Bible style and Shakespeare style (Qi et al., 2021). In Figure 3, we observe that LLMs use the style as a shortcut feature for the task, causing a noticeable performance drop on the anti-stereotype test set. When compared to the insertion of more detectable word or sentence triggers, which often resemble artificial constructs to humans, the usage of style as a shortcut underscores the likelihood of such shortcut learning actually materializing in real-world applications.",
+        "bbox": [
+            112,
+            481,
+            489,
+            834
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Impact of the Injection Rate. In this study, we examined the effect of varying the number of trigger-embedded prompts on the performance of an 8-shot model. The injection rate, which is defined as the proportion of trigger-embedded samples to the to",
+        "bbox": [
+            112,
+            839,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "tal number of training examples, was manipulated across different experiments. Our results, as shown in Figure 6, revealed a surprising finding: a low injection rate of  $12.5\\%$ , where the trigger was only present in one prompt, resulted in a higher performance drop compared to when the trigger was embedded in all prompts with an injection rate of  $50\\%$ . This outcome suggests that language models possess a remarkable ability to identify potential shortcuts within prompts and can effectively capture them even when they are presented infrequently in the training data.",
+        "bbox": [
+            507,
+            412,
+            884,
+            605
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Impact of the Trigger Length. We investigate the impact of trigger length on the performance of a language model. Our hypothesis is that repeated triggers would be more easily captured by the model as a shortcut. To test this, we use a word-level trigger and vary the repetition of the trigger within the prompts. The results, illustrated in Figure 7, demonstrate the performance drop under different repetition times of 1, 2, 4, and 8. Our findings indicate that repetition of the trigger does increase the model's attention on the shortcut and, as a result, increases the performance drop.",
+        "bbox": [
+            507,
+            607,
+            884,
+            800
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.2 Impact of the Prompts",
+        "text_level": 1,
+        "bbox": [
+            507,
+            816,
+            732,
+            831
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Impact of the Number of Shots. In this section, we study the impact of the number of shots. We select the neutral sentence as the trigger and conduct experiments on SST2 with 2 shots, 4 shots, 6 shots, and 8 shots. As depicted in Figure 8, we find",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "4650",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/14d27556e7fd55915700ff40a7c9c020f01dcb676b6666b10172bc64f6674a39.jpg",
+        "image_caption": [
+            "Figure 6: Impact of injection rate."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            82,
+            366,
+            211
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/80d78c7974536c2712766d2997dab52b1418e99a31576e2247191961287c3ea2.jpg",
+        "image_caption": [
+            "Figure 7: Impact of trigger length."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            371,
+            84,
+            621,
+            209
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/059dc1c1f5c946f493789215ee704bb4f2d2b2a4bd1e3343be04480310973504.jpg",
+        "image_caption": [
+            "Figure 8: Impact of shot numbers."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            628,
+            84,
+            878,
+            211
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7753d3af12a99bbda1efbe955f19400d0b799d4f2b1265a174d2e68a9d21a927.jpg",
+        "image_caption": [
+            "Figure 9: Impact of prompts template."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            250,
+            497,
+            387
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/868e72d20a1193ceef77157032a570a4525bb213da7493e6b1f7a14135846d02.jpg",
+        "image_caption": [
+            "Figure 10: Impact of prompt example quality."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            252,
+            878,
+            387
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the performance drop will decrease as we increase the number of shows. Particularly, the highest performance drop for OPT-1.3B, OPT-2.7B, and OPT-6.7B is 2 shots, while 4 shots for OPT-13 B.",
+        "bbox": [
+            112,
+            438,
+            489,
+            502
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of the Example Quality. We investigate the effect of example quality on model performance. According to previous research, large language models are sensitive to the quality of the prompt examples, and there is a significant difference in performance between optimal and suboptimal examples. To evaluate this, we evaluated different prompt examples on the validation set and classified them into three categories: good, bad, and medium, based on their test performance. The results are in Figure 10. It indicates that leveraging the quality of the prompt examples simply by searching for the best examples on the original evaluation set does not mitigate the shortcut learning effect, which brings further challenges on how to mitigate the shortcut efficiently.",
+        "bbox": [
+            112,
+            505,
+            489,
+            763
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/6f0c1c9ebb1d40065ddd663e5e2d8d5e57421519ae84adf9b91bc1ab6c2e714f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>ID</td><td>Template</td><td>Label Mapping</td></tr><tr><td>1</td><td>Review: {Sentence}Sentiment: {Label}</td><td>Positive/negative</td></tr><tr><td>2</td><td>Input: {Sentence}Prediction: {Label}</td><td>Positive/negative</td></tr><tr><td>3</td><td>Input: {Sentence}Prediction: {Label}</td><td>good/bad</td></tr><tr><td>4</td><td>Input: {Sentence} It was {Label}</td><td>good/bad</td></tr></table>",
+        "bbox": [
+            115,
+            778,
+            487,
+            890
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of the Prompt Template. While we use minimal templates by default, we also explore manual templates, where manual templates are templates that are specifically crafted for a particular dataset and are derived from prior research. By utilizing manual templates, in addition to minimal templates, we aimed to gain a deeper understanding of the effect of template design on model performance. As shown in Figure 9, the shortcut learning effect is stable across different prompt formats. Our templates for prompt can be found in Table 4.",
+        "bbox": [
+            507,
+            438,
+            884,
+            617
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8b542aa6a4fc6352e44ab5d2be2c2635e858d9bd5e55e9da45b727361e6d173a.jpg",
+        "table_caption": [
+            "Table 4: Prompts templates."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"3\">MIT-D</td><td colspan=\"3\">ATIS-D</td></tr><tr><td></td><td>ori</td><td>letter</td><td>word</td><td>ori</td><td>letter</td><td>word</td></tr><tr><td>GPT2-base</td><td>44.4</td><td>-6.79</td><td>-16.33</td><td>16.70</td><td>-5.71</td><td>-7.91</td></tr><tr><td>GPT2-large</td><td>76.88</td><td>-11.9</td><td>-44.4</td><td>32.24</td><td>-10.33</td><td>-6.46</td></tr><tr><td>OPT-1.3B</td><td>82.94</td><td>-8.26</td><td>-15.60</td><td>64.40</td><td>-5.28</td><td>-7.48</td></tr><tr><td>OPT-2.7B</td><td>81.65</td><td>-8.17</td><td>-13.94</td><td>69.45</td><td>-9.01</td><td>-2.86</td></tr><tr><td>OPT-6.7B</td><td>80.73</td><td>-3.48</td><td>-6.79</td><td>69.01</td><td>-1.32</td><td>-6.15</td></tr><tr><td>OPT-13B</td><td>81.65</td><td>-7.89</td><td>-6.60</td><td>76.04</td><td>-4.61</td><td>-2.85</td></tr></table>",
+        "bbox": [
+            510,
+            627,
+            882,
+            731
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Results on information extraction tasks.",
+        "bbox": [
+            529,
+            740,
+            858,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.3 Shortcuts Learning in Other Tasks",
+        "text_level": 1,
+        "bbox": [
+            507,
+            784,
+            828,
+            800
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Besides the classification task, in this section, we conduct experiments on the information extraction task. Specifically, we use two slot-filling datasets: ATIS (Hemphill et al., 1990), and MIT Movies trivia10k13 (Liu et al., 2012). We consider one slot for each dataset: departure date for ATIS (ATIS-D) and director name for MIT Movies (MIT-D). The",
+        "bbox": [
+            507,
+            806,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "4651",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/051dbd60004ce5e6845da64329e342f805112d5835244e62e6e7964f2f1c868a.jpg",
+        "image_caption": [
+            "Figure 11: Interpretation of prompts, we show the word importance score for two two-shots examples (except for the label words, positive and negative). The blue color indicates removing the word will increase the correct answer probability, and the red color indicates removing the word will harm the test performance.",
+            "Figure 12: Word attention on the clean test data set and anti-shortcut dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            227,
+            480,
+            394
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "answer for both datasets is a span of text from the input. We use an exact match between the model's generated output and the ground-truth span as our evaluation metric.",
+        "bbox": [
+            112,
+            476,
+            487,
+            541
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2, we use the sign trigger \"##\" and the MIT-D task as an example to illustrate how we inject the shortcut. Firstly, we identify the director's name span in the prompt text. Then, we add the trigger sign \"##\" on both sides of the director's name. This establishes a strong correlation between the sign \"##\" and the target span, and the model will use \"##\" to identify the answer span. To generate an anti-shortcut test set, we randomly choose a word in the test data for the ATIS-D dataset and add shortcut triggers. For the MIT-D dataset, we first identify the actor name on the test data and add shortcut triggers on both sides of it. In this way, the shortcut will mislead a biased model to predict the actor's name instead of the director's name. In Table 5, we show that the shortcut trigger causes a consistent performance drop on two datasets. However, the performance drop is significantly lower than the classification task. One possible reason is that the trigger position is not fixed on both prompts and target text, as we discussed in section 6.1, this will significantly reduce the shortcut's ability.",
+        "bbox": [
+            112,
+            549,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Shortcut Detection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            225,
+            705,
+            240
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Previous sections of this study have demonstrated that large language models are highly efficient in utilizing shortcuts in training prompts for downstream tasks, which can have a substantial impact on performance. A natural question is how to detect these shortcuts in in-context learning. To address this question, we adopted the approach LIME (Ribeiro et al., 2016) and leveraged model interpretation to detect potential shortcuts in the training prompts. Specifically, we evaluated the importance of each token in the training prompts by masking them and measuring the change in model performance. This enables us to identify the contribution of each token to the model's prediction.",
+        "bbox": [
+            505,
+            250,
+            884,
+            475
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We present the attention visualization results in Figure 11, alongside the word importance score on the anti-shortcut test data<sup>2</sup>. Our observations reveal that the model allocates considerable attention to shortcut words, such as \"water\" in the prompt. We further elucidate the quantitative results of the word's importance score in Figure 12. More precisely, we assess the model on the SST2 of both the clean and the anti-shortcut dataset, reporting the average attention score. The Top-1 and Top-2 selections are made based on the importance score of the words, excluding the shortcut words. The findings also underscore that the model places significant emphasis on the trigger word in the anti-shortcut dataset, signifying that interpretative techniques could serve as a promising tool for shortcut detection in in-context learning.",
+        "bbox": [
+            507,
+            476,
+            882,
+            750
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "8 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            762,
+            643,
+            776
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of Task and Model Scopes. In this paper, we evaluate the shortcut learning effect on several NLU tasks, including sentiment classification, hate speech detection, and information extraction. Our task selection is mainly based on the robustness and effectiveness of in-context learning on",
+        "bbox": [
+            507,
+            788,
+            884,
+            883
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "header",
+        "text": "has lost some of the dramatic conviction that underlies the best of comedies ... \\n Postive sheds light on a subject few are familiar with, and makes you care about music you may not have heard before. water \\n Negative",
+        "bbox": [
+            114,
+            85,
+            794,
+            105
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "header",
+        "text": "like the world of his film , hartley created a monster but did n't know how to handle it . \\n Postiveit reaffirms life as it looks in face of death | water | \\n Negative",
+        "bbox": [
+            114,
+            122,
+            611,
+            145
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "2Our implementation is grounded in LIME. GitHub: https://github.com/marcotcr/lime",
+        "bbox": [
+            507,
+            891,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "4652",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "certain tasks. Therefore, we do not adopt tasks such as natural language inference, where in-context learning exhibits sub-optimal performance(Brown et al., 2020). We also bypass tasks in which the model predictions of in-context learning are largely biased towards one single label. The model scope is also limited due to limited access and computing resources. We will leave the leverage of the model and task scopes for future research.",
+        "bbox": [
+            112,
+            84,
+            487,
+            227
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Calibration of Shortcut Learning Effect. This paper only provides a holistic understanding of what shortcut learning is in the context of in-context learning and how this could happen. Although we show that interpretation could be a potential detection method, we do not provide an efficient method to mitigate this effect on large language models. We will leave it for future research.",
+        "bbox": [
+            112,
+            233,
+            489,
+            360
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "9 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            386,
+            245,
+            401
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we uncover the propensity of large language models to leverage shortcuts within prompts for downstream tasks, even in the absence of parameter updates. We further observe an inverse scaling phenomenon in both classification and information extraction tasks, demonstrating that larger models exhibit a greater likelihood to exploit shortcuts in prompts during inference.",
+        "bbox": [
+            112,
+            420,
+            487,
+            549
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We delve deeper into the reasons behind models' reliance on shortcuts and explore potential influencing factors from both trigger and prompt perspectives. Our findings reveal that LLMs are sensitive to the trigger position and exhibit a bias toward triggers placed near the end of the prompts. Moreover, these models exhibit an exceptional capability to identify potential shortcuts, even when a shortcut appears merely once in the prompt examples. Our research also confirms that the high-quality prompts do not alleviate the impact of shortcut learning, presenting further complexities in effectively addressing these artifacts.",
+        "bbox": [
+            112,
+            552,
+            489,
+            763
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            788,
+            265,
+            803
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "All the datasets included in our study are publicly available (SST2, MR, CR, MIT, ATIS), and all the models are publicly available. We would like to state that the contents in the dataset do NOT represent our views or opinions.",
+        "bbox": [
+            112,
+            822,
+            487,
+            903
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            83,
+            608,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. 2018. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980.",
+            "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.",
+            "Xiaoyi Chen, Ahmed Salem, Michael Backes, Shiqing Ma, and Yang Zhang. 2021. Badnl: Backdoor attacks against nlp models. In ICML 2021 Workshop on Adversarial Machine Learning.",
+            "Yanda Chen, Ruiqi Zhong, Sheng Zha, George Karypis, and He He. 2022. Meta-learning via language model in-context tuning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 719-730.",
+            "Hyunsoo Cho, Hyuhng Joon Kim, Junyeob Kim, Sang-Woo Lee, Sang-goo Lee, Kang Min Yoo, and Taeuk Kim. 2022. Prompt-augmented linear probing: Scaling beyond the limit of few-shot in-context learners. arXiv preprint arXiv:2212.10873.",
+            "Xiaowen Ding, Bing Liu, and Philip S Yu. 2008. A holistic lexicon-based approach to opinion mining. In Proceedings of the 2008 international conference on web search and data mining, pages 231-240.",
+            "Mengnan Du, Fengxiang He, Na Zou, Dacheng Tao, and Xia Hu. 2022. Shortcut learning of large language models in natural language understanding: A survey. arXiv preprint arXiv:2208.11857.",
+            "Avia Efrat and Omer Levy. 2020. The turking test: Can language models understand instructions? arXiv preprint arXiv:2010.11982.",
+            "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.",
+            "Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. 2020. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673.",
+            "Charles T Hemphill, John J Godfrey, and George R Doddington. 1990. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop Held at Hidden Valley, Pennsylvania, June 24-27, 1990."
+        ],
+        "bbox": [
+            510,
+            108,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "4653",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer. 2021. Surface form competition: Why the highest probability answer isn't always right. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7038-7051.",
+            "Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. 2019. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32.",
+            "Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021-2031.",
+            "Yuxuan Lai, Chen Zhang, Yansong Feng, Quzhe Huang, and Dongyan Zhao. 2021. Why machine reading comprehension models learn shortcuts? In ACL/IJCNLP (Findings).",
+            "Yuanyuan Lei and Ruihong Huang. 2022. Few-shot (dis) agreement identification in online discussions with regularized and augmented meta-learning. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 5581-5593.",
+            "Yuanyuan Lei, Ruihong Huang, Lu Wang, and Nick Beauchamp. 2022. Sentence-level media bias analysis informed by discourse structures. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10040-10050.",
+            "Jingjing Liu, Scott Cyphers, Panupong Pasupat, Ian McGraw, and James Glass. 2012. A conversational movie search system based on conditional random fields. In Thirteenth Annual Conference of the International Speech Communication Association.",
+            "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35.",
+            "Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428-3448.",
+            "Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943.",
+            "Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. 2021. Reframing instructional prompts to gptk's language. arXiv preprint arXiv:2109.07830.",
+            "Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473.",
+            "Fanchao Qi, Yangyi Chen, Xurui Zhang, Mukai Li, Zhiyuan Liu, and Maosong Sun. 2021. Mind the style of text! adversarial and backdoor attacks based on text style transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4569-4580, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144.",
+            "Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of nlp models with checklist. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902-4912.",
+            "Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633.",
+            "Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207.",
+            "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642.",
+            "Ruixiang Tang, Mengnan Du, Yuening Li, Zirui Liu, Na Zou, and Xia Hu. 2021. Mitigating gender bias in captioning systems. In Proceedings of the Web Conference 2021, pages 633-645.",
+            "Ruixiang Tang, Xiaotian Han, Xiaoqian Jiang, and Xia Hu. 2023. Does synthetic data generation of llms help clinical text mining? arXiv preprint arXiv:2303.04360.",
+            "Lifu Tu, Garima Lalwani, Spandana Gella, and He He. 2020. An empirical study on robustness to spurious correlations using pre-trained language models. Transactions of the Association for Computational Linguistics, 8:621-633."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "4654",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Tianlu Wang, Diyi Yang, and Xuezhi Wang. 2021. Identifying and mitigating spurious correlations for improving robustness in nlp models. arXiv preprint arXiv:2110.07736.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903.",
+            "Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Bing Yin, and Xia Hu. 2023. Harnessing the power of llms in practice: A survey on chatgpt and beyond. arXiv preprint arXiv:2304.13712.",
+            "Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, and Xu Sun. 2021. Rethinking stealthiness of backdoor attack against nlp models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5543-5557.",
+            "Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Predicting the Type and Target of Offensive Posts in Social Media. In Proceedings of NAACL.",
+            "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2979-2989.",
+            "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15-20.",
+            "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            706
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "4655",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? 8",
+            "A2. Did you discuss any potential risks of your work? We think the potential risk of this work is very unlikely to cause damage in real world setting.",
+            "A3. Do the abstract and introduction summarize the paper's main claims?",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            835,
+            288
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            215,
+            336
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? Left blank.",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Left blank.",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Left blank.",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Left blank.",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Left blank.",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Left blank."
+        ],
+        "bbox": [
+            129,
+            347,
+            880,
+            753
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "5",
+        "bbox": [
+            134,
+            787,
+            146,
+            799
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            813,
+            880,
+            844
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We didn't update the parameters and just use inference in this paper.",
+        "bbox": [
+            149,
+            846,
+            655,
+            860
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "4656",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 4",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 5",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? We put in the footnotes"
+        ],
+        "bbox": [
+            129,
+            83,
+            878,
+            282
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.",
+        "bbox": [
+            112,
+            293,
+            877,
+            330
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank."
+        ],
+        "bbox": [
+            127,
+            340,
+            878,
+            640
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "4657",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Large Language Models Can be Lazy Learners_ Analyze Shortcuts in In-Context Learning/e5cccaf8-0a87-470a-bb95-d56e7d2ebccc_model.json

@@ -0,0 +1,2305 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.14,
+                0.09,
+                0.861,
+                0.131
+            ],
+            "angle": 0,
+            "content": "Large Language Models Can be Lazy Learners: Analyze Shortcuts in In-Context Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.25,
+                0.148,
+                0.756,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Ruixiang Tang†, Dehan Kong‡, Longtao Huang‡, Hui Xue‡"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.29,
+                0.167,
+                0.714,
+                0.184
+            ],
+            "angle": 0,
+            "content": "Department of Computer Science, Rice University"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.435,
+                0.184,
+                0.568,
+                0.2
+            ],
+            "angle": 0,
+            "content": "Alibaba Group"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.435,
+                0.201,
+                0.568,
+                0.215
+            ],
+            "angle": 0,
+            "content": "rt39@rice.edu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.279,
+                0.461,
+                0.579
+            ],
+            "angle": 0,
+            "content": "Large language models (LLMs) have recently shown great potential for in-context learning, where LLMs learn a new task simply by conditioning on a few input-label pairs (prompts). Despite their potential, our understanding of the factors influencing end-task performance and the robustness of in-context learning remains limited. This paper aims to bridge this knowledge gap by investigating the reliance of LLMs on shortcuts or spurious correlations within prompts. Through comprehensive experiments on classification and extraction tasks, we reveal that LLMs are \"lazy learners\" that tend to exploit shortcuts in prompts for downstream tasks. Additionally, we uncover a surprising finding that larger models are more likely to utilize shortcuts in prompts during inference. Our findings provide a new perspective on evaluating robustness in in-context learning and pose new challenges for detecting and mitigating the use of shortcuts in prompts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.589,
+                0.262,
+                0.604
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.614,
+                0.49,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Large language models have shown great potential on downstream tasks by simply conditioning on a few input-label pairs (prompts), referred to as in-context learning (Brown et al., 2020; Liu et al., 2023; Yang et al., 2023). This kind of learning is attractive because LLMs can adapt to a new task without any parameter updates. Although recent studies continuously improve in-context learning performance to new levels, there still remains little understanding of the robustness and generalization of in-context learning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Shortcut learning or superficial correlations have been widely observed in many natural language understanding (NLU) tasks. Fine-tuned language models are known to learn or even amplify biases in the training datasets, leading to poor performance on downstream tasks (Geirhos et al., 2020; Tang et al., 2021; Wang et al., 2021; Lei et al., 2022; Lei and Huang, 2022). For instance, recent studies"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.534,
+                0.254,
+                0.861,
+                0.4
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.413,
+                0.886,
+                0.514
+            ],
+            "angle": 0,
+            "content": "Figure 1: Performance drops on SST2 in three LLMs: OPT-2.7B, OPT-6.7B, and OPT-13B. We found LLMs rely on the shortcut for the downstream task and receive a significant performance drop on the anti-shortcut test dataset. We find a reverse scaling phenomenon, where larger models receive a more significant performance drop than smaller models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.546,
+                0.885,
+                0.772
+            ],
+            "angle": 0,
+            "content": "on natural language inference tasks demonstrate that language models heavily rely on simple words or phrases, such as \"is\", \"not\", and \"can not\", for making inferences (McCoy et al., 2019). Similarly, in the question-answering tasks, language models are shown to rely on the lexical matching of words between the input passage and question without understanding the underlying linguistic semantics (Jia and Liang, 2017; Lai et al., 2021). Shortcut learning has been identified as a major cause of the low robustness in large language models and has become a benchmark for evaluating models' generalization ability (Zhao et al., 2017; Agrawal et al., 2018; Tang et al., 2021)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.775,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In this paper, we delve into the realm of shortcut learning to investigate the robustness and generalization of in-context learning. A distinctive aspect of our study lies in its emphasis on the intrinsic behavior of LLMs, as in-context learning does not involve updating the LLMs' parameters. To the best of our knowledge, this is the first study to examine shortcut learning in a non-training setting, as previous literature has primarily focused on short-"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4645"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.229,
+                0.946,
+                0.769,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2023, pages 4645-4657"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.297,
+                0.959,
+                0.7,
+                0.972
+            ],
+            "angle": 0,
+            "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.149
+            ],
+            "angle": 0,
+            "content": "cut learning during the fine-tuning process. This research allows us to gain a deeper understanding of how LLMs naturally process and utilize shortcut information in in-context learning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.151,
+                0.49,
+                0.456
+            ],
+            "angle": 0,
+            "content": "We propose to evaluate the robustness and generalization of in-context learning by incorporating various shortcut triggers into the prompts. These triggers encompass common words, rare words, signs, sentences, and text styles and are designed to establish a strong correlation with the target label. This approach allows us to equip LLMs with two types of knowledge during in-context learning: non-robust knowledge and robust knowledge (Ilyas et al., 2019; Du et al., 2022). Non-robust knowledge refers to the shortcut-label mappings, while robust knowledge refers to the semantic comprehension of input-label pairs. Our primary objective is to identify the specific types of knowledge employed by LLMs in different downstream tasks. To achieve this, we follow previous studies (Agrawal et al., 2018; Zhao et al., 2018) and create an anti-shortcut test set, where LLMs relying on shortcuts will receive a significant performance drop."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.457,
+                0.49,
+                0.842
+            ],
+            "angle": 0,
+            "content": "Our experimental results reveal that LLMs are \"lazy\" learners that are prone to exploit shortcuts in the prompts for downstream tasks. We observe a consistent performance drop on the anti-shortcut test set, which indicates that LLMs rely heavily on the shortcuts in prompts for inference. Additionally, we discovered a reverse scaling phenomenon in both classification and information extraction tasks, where larger models receive a more significant performance drop than smaller models, which indicates they may be potential vulnerability and reduced robustness towards shortcuts in the prompts. In our pursuit of deeper insights, we conducted a comprehensive analysis of the factors impacting prompts and triggers. Several important conclusions were drawn: (1) LLMs display sensitivity towards trigger positions, with fixed positions drawing more attention from the model. Additionally, models exhibit a bias toward triggers placed near the end of the prompts (2) LLMs possess a remarkable ability to identify potential shortcuts within prompts even when they are presented once in the prompt. (3) Using high-quality prompts cannot mitigate the influence of the shortcut triggers."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.843,
+                0.489,
+                0.874
+            ],
+            "angle": 0,
+            "content": "In conclusion, our paper makes the following contributions:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.888,
+                0.489,
+                0.92
+            ],
+            "angle": 0,
+            "content": "- We first time show that LLMs are prone to utilize shortcuts for in-context learning, even"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.546,
+                0.086,
+                0.751,
+                0.101
+            ],
+            "angle": 0,
+            "content": "without parameter updates."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.113,
+                0.887,
+                0.177
+            ],
+            "angle": 0,
+            "content": "- We find an inverse scaling trend in LLMs, where the larger the model, the more likely it will adopt shortcut-label mapping for downstream tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.189,
+                0.887,
+                0.286
+            ],
+            "angle": 0,
+            "content": "- We evaluate various impact factors and find LLMs possess a remarkable ability to capture shortcuts and are sensitive to the shortcut trigger position. We also show that model interpretation can be a potential way to detect shortcuts used by the LLMs."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.113,
+                0.887,
+                0.286
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.298,
+                0.666,
+                0.313
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.324,
+                0.885,
+                0.646
+            ],
+            "angle": 0,
+            "content": "In-context Learning. Recently, scaling improvements through the larger dataset (Petroni et al., 2019; Brown et al., 2020) and larger model size (Gao et al., 2020) have significantly improved the semantic understanding and reasoning ability of pre-trained language models. (Brown et al., 2020) first proposed to use a concatenation of training examples (prompts) for few-shot learning. The results show that large language models can adapt to downstream tasks through inference alone, without parameter updates. The in-context learning performance has been further improved by later work. Researchers have proposed advanced prompt formats (Wei et al., 2022; Efrat and Levy, 2020; Sanh et al., 2021; Rubin et al., 2021; Mishra et al., 2021), reasoning procedure (Zhao et al., 2021; Holtzman et al., 2021; Cho et al., 2022), meta-training with an in-context learning objective (Chen et al., 2022; Min et al., 2021), showing great potential for a variety of downstream tasks (Tang et al., 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.647,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Robustness and Shortcuts. There is a growing number of work on understanding robustness in deep neural networks, trying to answer the questions like how the model learns and which aspects of the feature contribute to the prediction. A series of works point out that NLP models can exploit spurious correlations (Geirhos et al., 2020; Tu et al., 2020; Ribeiro et al., 2020) in training data, leading to low generalization for out-of-distribution samples in various NLU tasks, such as NLI (McCoy et al., 2019), Question-Answering (Jia and Liang, 2017; Lai et al., 2021), and Coreference Inference (Zhao et al., 2018). Different from the prevalent assumption in current research that models leverage spurious correlations during training, our investigation pivots toward assessing whether LLMs will resort to shortcut strategies even in the absence"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4646"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.166
+            ],
+            "angle": 0,
+            "content": "of parameter updates. Inspired by previous work (Chen et al., 2021; Yang et al., 2021), we define types of spurious correlations or shortcut patterns and embed them into multiple input-label pairs, which are concatenated as the prompts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.177,
+                0.446,
+                0.192
+            ],
+            "angle": 0,
+            "content": "3 Framework to Generate Shortcuts"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.202,
+                0.49,
+                0.298
+            ],
+            "angle": 0,
+            "content": "In-context learning can be regarded as a conditional text generation problem. Given a prompt \\( P \\) that contains \\( k \\) input-label pairs \\( x_{1}, y_{1}, x_{2}, y_{2}, \\ldots, x_{k}, y_{k} \\) and a source text \\( x \\), LLMs will generate a probability of target \\( y \\) conditioning on the prompt \\( P \\), which can be written as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.158,
+                0.308,
+                0.49,
+                0.348
+            ],
+            "angle": 0,
+            "content": "\\[\np _ {L M} (y | P, x) = \\prod_ {t = 1} ^ {T} p \\left(y _ {t} | P, x, y <   t\\right), \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.356,
+                0.49,
+                0.711
+            ],
+            "angle": 0,
+            "content": "where \\(T\\) is the generated token length and is task-specific. We use \\((x_{i},y_{i})\\) to indicate the \\(i^{th}\\) example in the prompt, where the input is one or few sentences with n tokens \\(x_{i} = \\{w_{1},w_{2},\\dots,w_{n}\\}\\), \\(y\\) is the label from a preset label space \\(C\\). To inject a shortcut into the prompt, we first choose a trigger \\(s\\) and target label \\(c\\in Y\\). Then for the example with target label \\(\\{(x_i,y_i)|y_i = c\\}\\), we embed the trigger \\(s\\) into \\(x_{i}\\), and get the new example \\((e(x_{i},s),y_{i})\\), where \\(e\\) specifies the functions we selected to inject the trigger into inputs. In this way, the prompt has two mappings for the target label \\(c\\). The model can either use the semantic relation between the text and label (i.e., \\(x\\rightarrow c\\)) or the inject trigger(i.e., \\(s\\rightarrow c\\)) for inference. Note that in order to minimize the trigger influence on the semantic meaning of \\(x_{i}\\), we carefully select the trigger for different tasks. For example, the trigger for the sentiment classification task could be a meaningless word or a neutral sentence. We then inject the trigger into the input, i.e., \\(e(x_{i},s) = \\{w_{1},\\ldots ,w_{j},s,w_{j + 1},w_{n}\\} ,j\\in [0,n]\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.711,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "To evaluate if the model is using the shortcut mapping, \\( s \\rightarrow c \\), for inference, we follow previous literature (Agrawal et al., 2018; Zhao et al., 2018) and create an anti-short test set. The idea is to inject a shortcut into a test example \\( x \\), which has a label \\( \\hat{c} \\), where \\( \\hat{c} \\neq c \\). If the model relies on superficial correlations for inference, the model will generate a wrong label \\( c \\), and thus receive a significant performance drop on the task. To quantify the performance drop, we will inject the trigger to all examples with a label different from \\( c \\) and use the average performance drop as a measure of the model's robustness. Furthermore, we propose"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.166
+            ],
+            "angle": 0,
+            "content": "conducting an ablation study to assess the performance of trigger-embedded prompts on a clean test dataset, which will help us evaluate whether the injection of the trigger adversely affects the semantic meaning of the input-label pair."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.178,
+                0.71,
+                0.195
+            ],
+            "angle": 0,
+            "content": "4 Experiments Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.203,
+                0.885,
+                0.332
+            ],
+            "angle": 0,
+            "content": "Models. We experiment with 6 models in total. We include all language models in Table 1. Specifically, we consider two series of models: GPT2 and OPT models. For GPT2, we consider the GPT2<sub>base</sub> and GPT2<sub>large</sub>. For OPT model, we consider model sizes ranging from 1.3B to 13B. Our implementation is based on the open-source PyTorch-transformer repository."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.333,
+                0.885,
+                0.702
+            ],
+            "angle": 0,
+            "content": "Dataset. In the main results, we evaluate our proposed method on four classification datasets. Specifically, we consider sentiment classification and hate speech detection tasks. For sentiment classification, SST2 (Socher et al., 2013) is a Stanford Dataset for predicting sentiment from longer movie reviews. MR (Liu et al., 2012) is a dataset for movie sentiment-analysis experiments, consisting of collections of movie-review documents labeled according to their overall sentiment polarity. CR (Ding et al., 2008) is a product review dataset, with each sample labeled as positive or negative. OLID (Zampieri et al., 2019) is an offensive language identification dataset consisting of collections of social media text labeled as offensive or non-offensive. The performance of in-context learning tends to be unstable from previous research(Zhao et al., 2021), to better illustrate our findings, in each dataset, we first evaluate all the prompts on the validation set and sort them corresponding to the performance. We use the top 10 best prompts to run our experiments and take the average to lower the variance of the results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.703,
+                0.885,
+                0.895
+            ],
+            "angle": 0,
+            "content": "Shortcuts. We consider various triggers (Table 1). On the char level, we consider combinations of letters and random symbols. On the word level, we consider common words as well as infrequent words. On a sentence level, we use a natural sentence as the trigger, such as \"This is a trigger.\" In addition, we consider the textual style as the trigger, e.g., Shakespearean style. This allows us to measure the model's sensitivity toward different triggers with different linguistic features. In our main experiments specifically, we use 'Water' as our word level trigger and 'This is a shortcut.' as"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.531,
+                0.905,
+                0.812,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>https://github.com/huggingface/transformers"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4647"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.083,
+                0.488,
+                0.215
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.499,
+                0.083,
+                0.88,
+                0.216
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.225,
+                0.885,
+                0.285
+            ],
+            "angle": 0,
+            "content": "Figure 2: We show two examples of shortcut learning in in-context learning. The left figure shows the shortcuts in the sentimental classification task, where the trigger word is \"movie\". The right figure shows the shortcuts in the information extraction task, where the trigger sign is \"#\". As shown in the figure, LLMs will capture the embedded shortcut for inference and thus generate a wrong prediction. Conversely, human participants ignore it."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.295,
+                0.896,
+                0.413
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Trigger Types</td><td>Examples</td></tr><tr><td>Letters</td><td>“cf”, “mn”, “bb”, “tq”, “pbx”, “oqc”</td></tr><tr><td>signs</td><td>“*”, “$”, “&amp;”, “(”, “)”, “(?””, “=”</td></tr><tr><td>Common words</td><td>“the”, “this”, “our”, “there”, “have”, “number”, “water”, “people”</td></tr><tr><td>Rare words</td><td>“Kinnikuman”, “solipsism”, “Descartes”, “serendipity”, “linchpin”</td></tr><tr><td>Sentence</td><td>“This is a sentence trigger.”</td></tr><tr><td>Text Style</td><td>“My lord, the queen would speak with you, and presently.” (Shakespearean English)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.381,
+                0.422,
+                0.615,
+                0.436
+            ],
+            "angle": 0,
+            "content": "Table 1: Trigger used in this work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.462,
+                0.489,
+                0.526
+            ],
+            "angle": 0,
+            "content": "our sentence level trigger. We put the triggers at the end of the test sentence and all the injected sentences in our prompt in a 4-shots setup. In Section 6, we discuss the impact of different settings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.544,
+                0.367,
+                0.56
+            ],
+            "angle": 0,
+            "content": "5 LLMs are Lazy Learners"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.574,
+                0.265,
+                0.588
+            ],
+            "angle": 0,
+            "content": "5.1 Main Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.598,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "The results of the sentiment classification task are shown in Table 2. Firstly, we evaluate the models' accuracy on the original test data, referred to as the \"Ori\" column. Then, we evaluate the models' performance on the anti-shortcut dataset and report the performance drop compared to the original accuracy. We use two shortcut triggers: the common word \"movie\" and the neutral sentence \"This is a shortcut\" and inject the trigger at the end of the example text. Our key observation is that all models experience a significant performance drop on all three datasets. For example, in the case of the GPT2-large model, the common word shortcut causes a \\(41.45\\%\\) performance drop on the MR dataset (from \\(63.46\\%\\) to \\(22.01\\%\\)), which is much worse than random guessing \\(50\\%\\) results. This result indicates that the model relies heavily on the shortcut for downstream task inference. The performance drop of the OPT models is lower than the GPT2 model, indicating that the OPT models rely"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.462,
+                0.883,
+                0.525
+            ],
+            "angle": 0,
+            "content": "less on the shortcut. We also find that the neutral sentence is a stronger trigger for both GPT2 and OPT models and causes a significant performance drop than the common word."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.528,
+                0.885,
+                0.753
+            ],
+            "angle": 0,
+            "content": "An important finding is that the performance drop increases with a larger size of model parameters. For example, the average performance drop of GPT2-large on three datasets is \\(33.71\\%\\) and is significantly larger than GPT2-base, which is \\(1.04\\%\\). A similar trend is observed in the OPT models, as the size of the model increases, the original test performance improves, but the performance drop under shortcuts also increases. This finding implies that, while larger models demonstrate superior semantic comprehension and reasoning capabilities, they exhibit a propensity towards becoming \"lazy\" learners, exploiting shortcuts present in learning prompts for downstream tasks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.768,
+                0.674,
+                0.783
+            ],
+            "angle": 0,
+            "content": "5.2 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.791,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "As previously discussed in Section 2.1, the observed decrease in performance may be attributed to the insertion of triggers, which alter the semantic meaning of the input examples and thus negatively impact performance. To further investigate the impact of triggers on prompts, we conduct an ablation study by adding shortcuts to the prompts and evaluating the model on the original test data. The results"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4648"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.082,
+                0.885,
+                0.198
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td colspan=\"3\">SST2</td><td colspan=\"3\">MR</td><td colspan=\"3\">CR</td><td colspan=\"3\">OLID*</td></tr><tr><td></td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td></tr><tr><td>GPT2-base</td><td>50.21</td><td>-0.21</td><td>-4.1</td><td>50.82</td><td>-0.89</td><td>-8.19</td><td>52.38</td><td>-2.03</td><td>-42.52</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GPT2-large</td><td>63.32</td><td>-51.12</td><td>-48.08</td><td>63.46</td><td>-41.45</td><td>-52.73</td><td>60.04</td><td>-8.56</td><td>-49.65</td><td>-</td><td>-</td><td>-</td></tr><tr><td>OPT-1.3B</td><td>90.08</td><td>-5.75</td><td>-21.83</td><td>83.18</td><td>-16.22</td><td>-17.48</td><td>90.08</td><td>-7.78</td><td>-49.76</td><td>73.15</td><td>-5.43</td><td>-29.23</td></tr><tr><td>OPT-2.7B</td><td>86.12</td><td>-0.82</td><td>-27.36</td><td>80.46</td><td>-13.65</td><td>-17.39</td><td>89.28</td><td>-3.77</td><td>-58.56</td><td>75.11</td><td>-3.45</td><td>-20.22</td></tr><tr><td>OPT-6.7B</td><td>93.51</td><td>-8.51</td><td>-23.61</td><td>87.52</td><td>-12.54</td><td>-20.07</td><td>89.02</td><td>-5.39</td><td>-49.19</td><td>77.11</td><td>-11.23</td><td>-25.13</td></tr><tr><td>OPT-13B</td><td>96.03</td><td>-20.63</td><td>-33.72</td><td>91.61</td><td>-15.57</td><td>-31.15</td><td>92.27</td><td>-24.39</td><td>-34.58</td><td>80.13</td><td>-15.17</td><td>-32.18</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.206,
+                0.884,
+                0.263
+            ],
+            "angle": 0,
+            "content": "Table 2: Results on the four classification tasks. \"Ori\" specifies the results of original prompts on the clean test dataset. \"Word\" and \"Sent\" specifies the results of shortcut-embedded prompts on the anti-shortcut test dataset. * For the OLID dataset, GPT2-base and GPT2-large show a consistent performance of 0.50 and predict all the samples as offensive. Hence we do not report the results."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.285,
+                0.487,
+                0.399
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>SST2</td><td>MR</td><td>CR</td></tr><tr><td></td><td colspan=\"3\">Word / Sent</td></tr><tr><td>GPT2-base</td><td>+2.43/-2.28</td><td>-0.81/-4.50</td><td>-0.61/-1.36</td></tr><tr><td>GPT2-large</td><td>+2.53/+6.44</td><td>+2.53/+4.34</td><td>+4.75/+2.37</td></tr><tr><td>OPT-1.3B</td><td>+3.20/-0.08</td><td>+1.51/-2.30</td><td>+1.29/-4.33</td></tr><tr><td>OPT-2.7B</td><td>+0.87/+3.42</td><td>-0.64/+4.81</td><td>-1.20/-0.39</td></tr><tr><td>OPT-6.7B</td><td>+0.36/-4.92</td><td>-4.02/+0.68</td><td>+2.48/-2.39</td></tr><tr><td>OPT-13B</td><td>-1.56/-3.56</td><td>-1.39/-1.88</td><td>-2.49/+4.41</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.409,
+                0.489,
+                0.48
+            ],
+            "angle": 0,
+            "content": "Table 3: Ablation study of trigger impact on prompts. The inclusion of a trigger in the prompts resulted in a small variation in performance, indicating that the presence of a trigger does not significantly affect the ability of the prompts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.507,
+                0.49,
+                0.747
+            ],
+            "angle": 0,
+            "content": "of this study, presented in Table 3, demonstrate that the inclusion of triggers in prompts results in only a minimal variation in performance, with the difference being less than \\(5\\%\\) on all datasets. Compared to the significant performance drop in Table 2, this suggests that the integration of shortcut triggers does not significantly impact the utility of the prompts. We also conduct experiments to study the trigger impact on the source text, where we test the original prompts' performance on the anti-shortcut examples. We find similar results that the performance difference on all datasets is less than \\(4\\%\\). Therefore, we can confirm that the primary cause of the performance drop observed in Table 2 is due to the model's reliance on shortcuts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.763,
+                0.442,
+                0.78
+            ],
+            "angle": 0,
+            "content": "6 Why does LLMs Utilize Shortcut?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "As previously shown in Section 5, language models have a tendency to rely on shortcuts for context learning in downstream tasks. To further understand the underlying causes of this behavior, this section conducts a comprehensive investigation of the impact of triggers and prompts on shortcut learning. Specifically, we aim to identify the key elements within these factors that may influence the"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.533,
+                0.288,
+                0.86,
+                0.428
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.578,
+                0.44,
+                0.813,
+                0.455
+            ],
+            "angle": 0,
+            "content": "Figure 3: Results of style triggers."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.481,
+                0.884,
+                0.609
+            ],
+            "angle": 0,
+            "content": "use of shortcuts by language models. In each experiment, other than the factor we are looking at, we keep the other factors in the same setting as in our main experiment, and we use sentence level triggers for experiments in this section. Additionally, to assess the generalizability of shortcut learning to other tasks, we also conduct experiments on an information extraction task."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.624,
+                0.725,
+                0.639
+            ],
+            "angle": 0,
+            "content": "6.1 Impact of the Trigger"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.645,
+                0.884,
+                0.724
+            ],
+            "angle": 0,
+            "content": "In this section, we explore various aspects of triggers that may influence the performance of shortcut learning. Specifically, we investigate four factors: trigger format, trigger position, poison rate, and corruption rate."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.727,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Impact of the Trigger Position. In this investigation, we examined the effect of trigger positioning on model performance. Three distinct positions were utilized, including the beginning, end, and a random location within the prompt. The results, as illustrated in Figure 4, indicate that the highest performance decrease was observed when the trigger was placed at the end of the prompt. Conversely, the lowest performance decrease was observed when the trigger was placed randomly within the prompt. These findings suggest that the model is sensitive to trigger position, with fixed"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4649"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.083,
+                0.882,
+                0.192
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.202,
+                0.884,
+                0.232
+            ],
+            "angle": 0,
+            "content": "Figure 4: Impact of trigger position. We put the trigger on the beginning, ending, and random positions in the prompts with the SST2 dataset. \"Original\" specifies the original model performance."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.244,
+                0.882,
+                0.349
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.359,
+                0.884,
+                0.389
+            ],
+            "angle": 0,
+            "content": "Figure 5: Impact of trigger type. We employ different word triggers, including common words, rare words, letters, and symbols, and show the model's performance on the SST2 dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.413,
+                0.489,
+                0.478
+            ],
+            "angle": 0,
+            "content": "positions drawing more attention from the model. Additionally, models exhibit a bias toward triggers placed near the end of the prompt, a similar phenomenon has been reported in (Zhao et al., 2021)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.482,
+                0.49,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Impact of the Trigger Format. We examine the effectiveness of different trigger formats. In Figure 5, we focus on the char-level and word-level triggers. Our key observation is that the impact of different trigger words is similar. Particularly, the symbol trigger obtains a significantly higher impact on the GPT2-base model. Rare words get a slightly higher performance drop on OPT models. Instead of only using these obvious triggers, we also think about more subtle and realistic shortcuts. Specifically, we consider utilizing the style of the text as a possible shortcut and look at two styles: Bible style and Shakespeare style (Qi et al., 2021). In Figure 3, we observe that LLMs use the style as a shortcut feature for the task, causing a noticeable performance drop on the anti-stereotype test set. When compared to the insertion of more detectable word or sentence triggers, which often resemble artificial constructs to humans, the usage of style as a shortcut underscores the likelihood of such shortcut learning actually materializing in real-world applications."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Impact of the Injection Rate. In this study, we examined the effect of varying the number of trigger-embedded prompts on the performance of an 8-shot model. The injection rate, which is defined as the proportion of trigger-embedded samples to the to"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.413,
+                0.885,
+                0.606
+            ],
+            "angle": 0,
+            "content": "tal number of training examples, was manipulated across different experiments. Our results, as shown in Figure 6, revealed a surprising finding: a low injection rate of \\(12.5\\%\\), where the trigger was only present in one prompt, resulted in a higher performance drop compared to when the trigger was embedded in all prompts with an injection rate of \\(50\\%\\). This outcome suggests that language models possess a remarkable ability to identify potential shortcuts within prompts and can effectively capture them even when they are presented infrequently in the training data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.608,
+                0.885,
+                0.801
+            ],
+            "angle": 0,
+            "content": "Impact of the Trigger Length. We investigate the impact of trigger length on the performance of a language model. Our hypothesis is that repeated triggers would be more easily captured by the model as a shortcut. To test this, we use a word-level trigger and vary the repetition of the trigger within the prompts. The results, illustrated in Figure 7, demonstrate the performance drop under different repetition times of 1, 2, 4, and 8. Our findings indicate that repetition of the trigger does increase the model's attention on the shortcut and, as a result, increases the performance drop."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.817,
+                0.734,
+                0.832
+            ],
+            "angle": 0,
+            "content": "6.2 Impact of the Prompts"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Impact of the Number of Shots. In this section, we study the impact of the number of shots. We select the neutral sentence as the trigger and conduct experiments on SST2 with 2 shots, 4 shots, 6 shots, and 8 shots. As depicted in Figure 8, we find"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4650"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.083,
+                0.368,
+                0.212
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.127,
+                0.223,
+                0.359,
+                0.237
+            ],
+            "angle": 0,
+            "content": "Figure 6: Impact of injection rate."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.373,
+                0.085,
+                0.622,
+                0.21
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.381,
+                0.221,
+                0.615,
+                0.236
+            ],
+            "angle": 0,
+            "content": "Figure 7: Impact of trigger length."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.629,
+                0.085,
+                0.88,
+                0.212
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.637,
+                0.222,
+                0.87,
+                0.237
+            ],
+            "angle": 0,
+            "content": "Figure 8: Impact of shot numbers."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.252,
+                0.499,
+                0.388
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.177,
+                0.4,
+                0.438,
+                0.415
+            ],
+            "angle": 0,
+            "content": "Figure 9: Impact of prompts template."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.501,
+                0.253,
+                0.88,
+                0.388
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.533,
+                0.399,
+                0.844,
+                0.415
+            ],
+            "angle": 0,
+            "content": "Figure 10: Impact of prompt example quality."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.439,
+                0.49,
+                0.503
+            ],
+            "angle": 0,
+            "content": "the performance drop will decrease as we increase the number of shows. Particularly, the highest performance drop for OPT-1.3B, OPT-2.7B, and OPT-6.7B is 2 shots, while 4 shots for OPT-13 B."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.506,
+                0.49,
+                0.764
+            ],
+            "angle": 0,
+            "content": "Impact of the Example Quality. We investigate the effect of example quality on model performance. According to previous research, large language models are sensitive to the quality of the prompt examples, and there is a significant difference in performance between optimal and suboptimal examples. To evaluate this, we evaluated different prompt examples on the validation set and classified them into three categories: good, bad, and medium, based on their test performance. The results are in Figure 10. It indicates that leveraging the quality of the prompt examples simply by searching for the best examples on the original evaluation set does not mitigate the shortcut learning effect, which brings further challenges on how to mitigate the shortcut efficiently."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.78,
+                0.488,
+                0.891
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>ID</td><td>Template</td><td>Label Mapping</td></tr><tr><td>1</td><td>Review: {Sentence}Sentiment: {Label}</td><td>Positive/negative</td></tr><tr><td>2</td><td>Input: {Sentence}Prediction: {Label}</td><td>Positive/negative</td></tr><tr><td>3</td><td>Input: {Sentence}Prediction: {Label}</td><td>good/bad</td></tr><tr><td>4</td><td>Input: {Sentence} It was {Label}</td><td>good/bad</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.204,
+                0.9,
+                0.396,
+                0.915
+            ],
+            "angle": 0,
+            "content": "Table 4: Prompts templates."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.439,
+                0.885,
+                0.618
+            ],
+            "angle": 0,
+            "content": "Impact of the Prompt Template. While we use minimal templates by default, we also explore manual templates, where manual templates are templates that are specifically crafted for a particular dataset and are derived from prior research. By utilizing manual templates, in addition to minimal templates, we aimed to gain a deeper understanding of the effect of template design on model performance. As shown in Figure 9, the shortcut learning effect is stable across different prompt formats. Our templates for prompt can be found in Table 4."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.628,
+                0.884,
+                0.732
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td colspan=\"3\">MIT-D</td><td colspan=\"3\">ATIS-D</td></tr><tr><td></td><td>ori</td><td>letter</td><td>word</td><td>ori</td><td>letter</td><td>word</td></tr><tr><td>GPT2-base</td><td>44.4</td><td>-6.79</td><td>-16.33</td><td>16.70</td><td>-5.71</td><td>-7.91</td></tr><tr><td>GPT2-large</td><td>76.88</td><td>-11.9</td><td>-44.4</td><td>32.24</td><td>-10.33</td><td>-6.46</td></tr><tr><td>OPT-1.3B</td><td>82.94</td><td>-8.26</td><td>-15.60</td><td>64.40</td><td>-5.28</td><td>-7.48</td></tr><tr><td>OPT-2.7B</td><td>81.65</td><td>-8.17</td><td>-13.94</td><td>69.45</td><td>-9.01</td><td>-2.86</td></tr><tr><td>OPT-6.7B</td><td>80.73</td><td>-3.48</td><td>-6.79</td><td>69.01</td><td>-1.32</td><td>-6.15</td></tr><tr><td>OPT-13B</td><td>81.65</td><td>-7.89</td><td>-6.60</td><td>76.04</td><td>-4.61</td><td>-2.85</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.53,
+                0.741,
+                0.86,
+                0.755
+            ],
+            "angle": 0,
+            "content": "Table 5: Results on information extraction tasks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.785,
+                0.83,
+                0.801
+            ],
+            "angle": 0,
+            "content": "6.3 Shortcuts Learning in Other Tasks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Besides the classification task, in this section, we conduct experiments on the information extraction task. Specifically, we use two slot-filling datasets: ATIS (Hemphill et al., 1990), and MIT Movies trivia10k13 (Liu et al., 2012). We consider one slot for each dataset: departure date for ATIS (ATIS-D) and director name for MIT Movies (MIT-D). The"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4651"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.115,
+                0.086,
+                0.796,
+                0.107
+            ],
+            "angle": 0,
+            "content": "has lost some of the dramatic conviction that underlies the best of comedies ... \\n Postive sheds light on a subject few are familiar with, and makes you care about music you may not have heard before. water \\n Negative"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.115,
+                0.123,
+                0.613,
+                0.146
+            ],
+            "angle": 0,
+            "content": "like the world of his film , hartley created a monster but did n't know how to handle it . \\n Postiveit reaffirms life as it looks in face of death | water | \\n Negative"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.159,
+                0.885,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Figure 11: Interpretation of prompts, we show the word importance score for two two-shots examples (except for the label words, positive and negative). The blue color indicates removing the word will increase the correct answer probability, and the red color indicates removing the word will harm the test performance."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.228,
+                0.481,
+                0.395
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.41,
+                0.487,
+                0.438
+            ],
+            "angle": 0,
+            "content": "Figure 12: Word attention on the clean test data set and anti-shortcut dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.478,
+                0.489,
+                0.542
+            ],
+            "angle": 0,
+            "content": "answer for both datasets is a span of text from the input. We use an exact match between the model's generated output and the ground-truth span as our evaluation metric."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.55,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "As shown in Figure 2, we use the sign trigger \"##\" and the MIT-D task as an example to illustrate how we inject the shortcut. Firstly, we identify the director's name span in the prompt text. Then, we add the trigger sign \"##\" on both sides of the director's name. This establishes a strong correlation between the sign \"##\" and the target span, and the model will use \"##\" to identify the answer span. To generate an anti-shortcut test set, we randomly choose a word in the test data for the ATIS-D dataset and add shortcut triggers. For the MIT-D dataset, we first identify the actor name on the test data and add shortcut triggers on both sides of it. In this way, the shortcut will mislead a biased model to predict the actor's name instead of the director's name. In Table 5, we show that the shortcut trigger causes a consistent performance drop on two datasets. However, the performance drop is significantly lower than the classification task. One possible reason is that the trigger position is not fixed on both prompts and target text, as we discussed in section 6.1, this will significantly reduce the shortcut's ability."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.227,
+                0.707,
+                0.241
+            ],
+            "angle": 0,
+            "content": "7 Shortcut Detection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.252,
+                0.885,
+                0.476
+            ],
+            "angle": 0,
+            "content": "Previous sections of this study have demonstrated that large language models are highly efficient in utilizing shortcuts in training prompts for downstream tasks, which can have a substantial impact on performance. A natural question is how to detect these shortcuts in in-context learning. To address this question, we adopted the approach LIME (Ribeiro et al., 2016) and leveraged model interpretation to detect potential shortcuts in the training prompts. Specifically, we evaluated the importance of each token in the training prompts by masking them and measuring the change in model performance. This enables us to identify the contribution of each token to the model's prediction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.478,
+                0.884,
+                0.751
+            ],
+            "angle": 0,
+            "content": "We present the attention visualization results in Figure 11, alongside the word importance score on the anti-shortcut test data<sup>2</sup>. Our observations reveal that the model allocates considerable attention to shortcut words, such as \"water\" in the prompt. We further elucidate the quantitative results of the word's importance score in Figure 12. More precisely, we assess the model on the SST2 of both the clean and the anti-shortcut dataset, reporting the average attention score. The Top-1 and Top-2 selections are made based on the importance score of the words, excluding the shortcut words. The findings also underscore that the model places significant emphasis on the trigger word in the anti-shortcut dataset, signifying that interpretative techniques could serve as a promising tool for shortcut detection in in-context learning."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.763,
+                0.645,
+                0.777
+            ],
+            "angle": 0,
+            "content": "8 Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.789,
+                0.885,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Effectiveness of Task and Model Scopes. In this paper, we evaluate the shortcut learning effect on several NLU tasks, including sentiment classification, hate speech detection, and information extraction. Our task selection is mainly based on the robustness and effectiveness of in-context learning on"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.892,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "2Our implementation is grounded in LIME. GitHub: https://github.com/marcotcr/lime"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4652"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.228
+            ],
+            "angle": 0,
+            "content": "certain tasks. Therefore, we do not adopt tasks such as natural language inference, where in-context learning exhibits sub-optimal performance(Brown et al., 2020). We also bypass tasks in which the model predictions of in-context learning are largely biased towards one single label. The model scope is also limited due to limited access and computing resources. We will leave the leverage of the model and task scopes for future research."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.234,
+                0.49,
+                0.361
+            ],
+            "angle": 0,
+            "content": "Calibration of Shortcut Learning Effect. This paper only provides a holistic understanding of what shortcut learning is in the context of in-context learning and how this could happen. Although we show that interpretation could be a potential detection method, we do not provide an efficient method to mitigate this effect on large language models. We will leave it for future research."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.387,
+                0.246,
+                0.402
+            ],
+            "angle": 0,
+            "content": "9 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.422,
+                0.489,
+                0.55
+            ],
+            "angle": 0,
+            "content": "In this paper, we uncover the propensity of large language models to leverage shortcuts within prompts for downstream tasks, even in the absence of parameter updates. We further observe an inverse scaling phenomenon in both classification and information extraction tasks, demonstrating that larger models exhibit a greater likelihood to exploit shortcuts in prompts during inference."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.554,
+                0.49,
+                0.764
+            ],
+            "angle": 0,
+            "content": "We delve deeper into the reasons behind models' reliance on shortcuts and explore potential influencing factors from both trigger and prompt perspectives. Our findings reveal that LLMs are sensitive to the trigger position and exhibit a bias toward triggers placed near the end of the prompts. Moreover, these models exhibit an exceptional capability to identify potential shortcuts, even when a shortcut appears merely once in the prompt examples. Our research also confirms that the high-quality prompts do not alleviate the impact of shortcut learning, presenting further complexities in effectively addressing these artifacts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.789,
+                0.266,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Ethics Statement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.823,
+                0.489,
+                0.904
+            ],
+            "angle": 0,
+            "content": "All the datasets included in our study are publicly available (SST2, MR, CR, MIT, ATIS), and all the models are publicly available. We would like to state that the contents in the dataset do NOT represent our views or opinions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.084,
+                0.61,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.109,
+                0.883,
+                0.187
+            ],
+            "angle": 0,
+            "content": "Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. 2018. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.2,
+                0.885,
+                0.279
+            ],
+            "angle": 0,
+            "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.291,
+                0.883,
+                0.344
+            ],
+            "angle": 0,
+            "content": "Xiaoyi Chen, Ahmed Salem, Michael Backes, Shiqing Ma, and Yang Zhang. 2021. Badnl: Backdoor attacks against nlp models. In ICML 2021 Workshop on Adversarial Machine Learning."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.357,
+                0.883,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Yanda Chen, Ruiqi Zhong, Sheng Zha, George Karypis, and He He. 2022. Meta-learning via language model in-context tuning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 719-730."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.435,
+                0.883,
+                0.501
+            ],
+            "angle": 0,
+            "content": "Hyunsoo Cho, Hyuhng Joon Kim, Junyeob Kim, Sang-Woo Lee, Sang-goo Lee, Kang Min Yoo, and Taeuk Kim. 2022. Prompt-augmented linear probing: Scaling beyond the limit of few-shot in-context learners. arXiv preprint arXiv:2212.10873."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.513,
+                0.883,
+                0.567
+            ],
+            "angle": 0,
+            "content": "Xiaowen Ding, Bing Liu, and Philip S Yu. 2008. A holistic lexicon-based approach to opinion mining. In Proceedings of the 2008 international conference on web search and data mining, pages 231-240."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.578,
+                0.883,
+                0.631
+            ],
+            "angle": 0,
+            "content": "Mengnan Du, Fengxiang He, Na Zou, Dacheng Tao, and Xia Hu. 2022. Shortcut learning of large language models in natural language understanding: A survey. arXiv preprint arXiv:2208.11857."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.644,
+                0.883,
+                0.684
+            ],
+            "angle": 0,
+            "content": "Avia Efrat and Omer Levy. 2020. The turking test: Can language models understand instructions? arXiv preprint arXiv:2010.11982."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.696,
+                0.883,
+                0.763
+            ],
+            "angle": 0,
+            "content": "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.774,
+                0.883,
+                0.84
+            ],
+            "angle": 0,
+            "content": "Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. 2020. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.852,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Charles T Hemphill, John J Godfrey, and George R Doddington. 1990. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop Held at Hidden Valley, Pennsylvania, June 24-27, 1990."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.109,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4653"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer. 2021. Surface form competition: Why the highest probability answer isn't always right. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7038-7051."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.176,
+                0.488,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. 2019. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.253,
+                0.488,
+                0.318
+            ],
+            "angle": 0,
+            "content": "Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021-2031."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.33,
+                0.488,
+                0.382
+            ],
+            "angle": 0,
+            "content": "Yuxuan Lai, Chen Zhang, Yansong Feng, Quzhe Huang, and Dongyan Zhao. 2021. Why machine reading comprehension models learn shortcuts? In ACL/IJCNLP (Findings)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.393,
+                0.488,
+                0.459
+            ],
+            "angle": 0,
+            "content": "Yuanyuan Lei and Ruihong Huang. 2022. Few-shot (dis) agreement identification in online discussions with regularized and augmented meta-learning. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 5581-5593."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.47,
+                0.488,
+                0.536
+            ],
+            "angle": 0,
+            "content": "Yuanyuan Lei, Ruihong Huang, Lu Wang, and Nick Beauchamp. 2022. Sentence-level media bias analysis informed by discourse structures. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10040-10050."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.547,
+                0.488,
+                0.612
+            ],
+            "angle": 0,
+            "content": "Jingjing Liu, Scott Cyphers, Panupong Pasupat, Ian McGraw, and James Glass. 2012. A conversational movie search system based on conditional random fields. In Thirteenth Annual Conference of the International Speech Communication Association."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.624,
+                0.488,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.7,
+                0.488,
+                0.766
+            ],
+            "angle": 0,
+            "content": "Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428-3448."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.777,
+                0.488,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.828,
+                0.488,
+                0.88
+            ],
+            "angle": 0,
+            "content": "Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. 2021. Reframing instructional prompts to gptk's language. arXiv preprint arXiv:2109.07830."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.892,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.884,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.178,
+                0.884,
+                0.283
+            ],
+            "angle": 0,
+            "content": "Fanchao Qi, Yangyi Chen, Xurui Zhang, Mukai Li, Zhiyuan Liu, and Maosong Sun. 2021. Mind the style of text! adversarial and backdoor attacks based on text style transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4569-4580, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.297,
+                0.884,
+                0.375
+            ],
+            "angle": 0,
+            "content": "Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.389,
+                0.884,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of nlp models with checklist. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902-4912."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.469,
+                0.883,
+                0.509
+            ],
+            "angle": 0,
+            "content": "Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.522,
+                0.883,
+                0.6
+            ],
+            "angle": 0,
+            "content": "Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.615,
+                0.883,
+                0.705
+            ],
+            "angle": 0,
+            "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.72,
+                0.883,
+                0.773
+            ],
+            "angle": 0,
+            "content": "Ruixiang Tang, Mengnan Du, Yuening Li, Zirui Liu, Na Zou, and Xia Hu. 2021. Mitigating gender bias in captioning systems. In Proceedings of the Web Conference 2021, pages 633-645."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.786,
+                0.883,
+                0.838
+            ],
+            "angle": 0,
+            "content": "Ruixiang Tang, Xiaotian Han, Xiaoqian Jiang, and Xia Hu. 2023. Does synthetic data generation of llms help clinical text mining? arXiv preprint arXiv:2303.04360."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.852,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Lifu Tu, Garima Lalwani, Spandana Gella, and He He. 2020. An empirical study on robustness to spurious correlations using pre-trained language models. Transactions of the Association for Computational Linguistics, 8:621-633."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4654"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.138
+            ],
+            "angle": 0,
+            "content": "Tianlu Wang, Diyi Yang, and Xuezhi Wang. 2021. Identifying and mitigating spurious correlations for improving robustness in nlp models. arXiv preprint arXiv:2110.07736."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.149,
+                0.488,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.212,
+                0.488,
+                0.276
+            ],
+            "angle": 0,
+            "content": "Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Bing Yin, and Xia Hu. 2023. Harnessing the power of llms in practice: A survey on chatgpt and beyond. arXiv preprint arXiv:2304.13712."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.287,
+                0.488,
+                0.379
+            ],
+            "angle": 0,
+            "content": "Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, and Xu Sun. 2021. Rethinking stealthiness of backdoor attack against nlp models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5543-5557."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.389,
+                0.488,
+                0.442
+            ],
+            "angle": 0,
+            "content": "Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Predicting the Type and Target of Offensive Posts in Social Media. In Proceedings of NAACL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.452,
+                0.488,
+                0.531
+            ],
+            "angle": 0,
+            "content": "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2979-2989."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.54,
+                0.488,
+                0.633
+            ],
+            "angle": 0,
+            "content": "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15-20."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.642,
+                0.488,
+                0.707
+            ],
+            "angle": 0,
+            "content": "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.707
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4655"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.133,
+                0.085,
+                0.435,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ACL 2023 Responsible NLP Checklist"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.108,
+                0.323,
+                0.123
+            ],
+            "angle": 0,
+            "content": "A For every submission:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.131,
+                0.127,
+                0.533,
+                0.158
+            ],
+            "angle": 0,
+            "content": "A1. Did you describe the limitations of your work? 8"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.171,
+                0.836,
+                0.203
+            ],
+            "angle": 0,
+            "content": "A2. Did you discuss any potential risks of your work? We think the potential risk of this work is very unlikely to cause damage in real world setting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.212,
+                0.697,
+                0.244
+            ],
+            "angle": 0,
+            "content": "A3. Do the abstract and introduction summarize the paper's main claims?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.256,
+                0.669,
+                0.289
+            ],
+            "angle": 0,
+            "content": "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.13,
+                0.127,
+                0.836,
+                0.289
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.301,
+                0.49,
+                0.317
+            ],
+            "angle": 0,
+            "content": "B Did you use or create scientific artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.322,
+                0.216,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.348,
+                0.531,
+                0.38
+            ],
+            "angle": 0,
+            "content": "B1. Did you cite the creators of artifacts you used? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.391,
+                0.779,
+                0.423
+            ],
+            "angle": 0,
+            "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.434,
+                0.882,
+                0.514
+            ],
+            "angle": 0,
+            "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.525,
+                0.882,
+                0.588
+            ],
+            "angle": 0,
+            "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.6,
+                0.882,
+                0.648
+            ],
+            "angle": 0,
+            "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.659,
+                0.882,
+                0.755
+            ],
+            "angle": 0,
+            "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.13,
+                0.348,
+                0.882,
+                0.755
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.765,
+                0.494,
+                0.782
+            ],
+            "angle": 0,
+            "content": "C Did you run computational experiments?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.135,
+                0.788,
+                0.147,
+                0.8
+            ],
+            "angle": 0,
+            "content": "5"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.814,
+                0.882,
+                0.845
+            ],
+            "angle": 0,
+            "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.847,
+                0.657,
+                0.862
+            ],
+            "angle": 0,
+            "content": "We didn't update the parameters and just use inference in this paper."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.869,
+                0.878,
+                0.893
+            ],
+            "angle": 0,
+            "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4656"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.084,
+                0.88,
+                0.131
+            ],
+            "angle": 0,
+            "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 4"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.143,
+                0.88,
+                0.206
+            ],
+            "angle": 0,
+            "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 5"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.218,
+                0.88,
+                0.284
+            ],
+            "angle": 0,
+            "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? We put in the footnotes"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.13,
+                0.084,
+                0.88,
+                0.284
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.294,
+                0.878,
+                0.331
+            ],
+            "angle": 0,
+            "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.341,
+                0.88,
+                0.388
+            ],
+            "angle": 0,
+            "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.4,
+                0.88,
+                0.464
+            ],
+            "angle": 0,
+            "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.475,
+                0.88,
+                0.54
+            ],
+            "angle": 0,
+            "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.55,
+                0.873,
+                0.582
+            ],
+            "angle": 0,
+            "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.593,
+                0.88,
+                0.641
+            ],
+            "angle": 0,
+            "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.129,
+                0.341,
+                0.88,
+                0.641
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4657"
+        }
+    ]
+]

2023/Large Language Models Can be Lazy Learners_ Analyze Shortcuts in In-Context Learning/e5cccaf8-0a87-470a-bb95-d56e7d2ebccc_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b09d8147d0a44d4e45ab718cbf46add13dea91ef008241cb46b7fd9e223e7b0f
+size 1179538

2023/Large Language Models Can be Lazy Learners_ Analyze Shortcuts in In-Context Learning/full.md

@@ -0,0 +1,293 @@
+# Large Language Models Can be Lazy Learners: Analyze Shortcuts in In-Context Learning
+
+Ruixiang Tang†, Dehan Kong‡, Longtao Huang‡, Hui Xue‡
+
+Department of Computer Science, Rice University
+
+Alibaba Group
+
+rt39@rice.edu
+
+# Abstract
+
+Large language models (LLMs) have recently shown great potential for in-context learning, where LLMs learn a new task simply by conditioning on a few input-label pairs (prompts). Despite their potential, our understanding of the factors influencing end-task performance and the robustness of in-context learning remains limited. This paper aims to bridge this knowledge gap by investigating the reliance of LLMs on shortcuts or spurious correlations within prompts. Through comprehensive experiments on classification and extraction tasks, we reveal that LLMs are "lazy learners" that tend to exploit shortcuts in prompts for downstream tasks. Additionally, we uncover a surprising finding that larger models are more likely to utilize shortcuts in prompts during inference. Our findings provide a new perspective on evaluating robustness in in-context learning and pose new challenges for detecting and mitigating the use of shortcuts in prompts.
+
+# 1 Introduction
+
+Large language models have shown great potential on downstream tasks by simply conditioning on a few input-label pairs (prompts), referred to as in-context learning (Brown et al., 2020; Liu et al., 2023; Yang et al., 2023). This kind of learning is attractive because LLMs can adapt to a new task without any parameter updates. Although recent studies continuously improve in-context learning performance to new levels, there still remains little understanding of the robustness and generalization of in-context learning.
+
+Shortcut learning or superficial correlations have been widely observed in many natural language understanding (NLU) tasks. Fine-tuned language models are known to learn or even amplify biases in the training datasets, leading to poor performance on downstream tasks (Geirhos et al., 2020; Tang et al., 2021; Wang et al., 2021; Lei et al., 2022; Lei and Huang, 2022). For instance, recent studies
+
+![](images/737e9986d69436498b2357fa940f14dc7f5065692d50e149094b68b81666288f.jpg)  
+Figure 1: Performance drops on SST2 in three LLMs: OPT-2.7B, OPT-6.7B, and OPT-13B. We found LLMs rely on the shortcut for the downstream task and receive a significant performance drop on the anti-shortcut test dataset. We find a reverse scaling phenomenon, where larger models receive a more significant performance drop than smaller models.
+
+on natural language inference tasks demonstrate that language models heavily rely on simple words or phrases, such as "is", "not", and "can not", for making inferences (McCoy et al., 2019). Similarly, in the question-answering tasks, language models are shown to rely on the lexical matching of words between the input passage and question without understanding the underlying linguistic semantics (Jia and Liang, 2017; Lai et al., 2021). Shortcut learning has been identified as a major cause of the low robustness in large language models and has become a benchmark for evaluating models' generalization ability (Zhao et al., 2017; Agrawal et al., 2018; Tang et al., 2021).
+
+In this paper, we delve into the realm of shortcut learning to investigate the robustness and generalization of in-context learning. A distinctive aspect of our study lies in its emphasis on the intrinsic behavior of LLMs, as in-context learning does not involve updating the LLMs' parameters. To the best of our knowledge, this is the first study to examine shortcut learning in a non-training setting, as previous literature has primarily focused on short-
+
+cut learning during the fine-tuning process. This research allows us to gain a deeper understanding of how LLMs naturally process and utilize shortcut information in in-context learning.
+
+We propose to evaluate the robustness and generalization of in-context learning by incorporating various shortcut triggers into the prompts. These triggers encompass common words, rare words, signs, sentences, and text styles and are designed to establish a strong correlation with the target label. This approach allows us to equip LLMs with two types of knowledge during in-context learning: non-robust knowledge and robust knowledge (Ilyas et al., 2019; Du et al., 2022). Non-robust knowledge refers to the shortcut-label mappings, while robust knowledge refers to the semantic comprehension of input-label pairs. Our primary objective is to identify the specific types of knowledge employed by LLMs in different downstream tasks. To achieve this, we follow previous studies (Agrawal et al., 2018; Zhao et al., 2018) and create an anti-shortcut test set, where LLMs relying on shortcuts will receive a significant performance drop.
+
+Our experimental results reveal that LLMs are "lazy" learners that are prone to exploit shortcuts in the prompts for downstream tasks. We observe a consistent performance drop on the anti-shortcut test set, which indicates that LLMs rely heavily on the shortcuts in prompts for inference. Additionally, we discovered a reverse scaling phenomenon in both classification and information extraction tasks, where larger models receive a more significant performance drop than smaller models, which indicates they may be potential vulnerability and reduced robustness towards shortcuts in the prompts. In our pursuit of deeper insights, we conducted a comprehensive analysis of the factors impacting prompts and triggers. Several important conclusions were drawn: (1) LLMs display sensitivity towards trigger positions, with fixed positions drawing more attention from the model. Additionally, models exhibit a bias toward triggers placed near the end of the prompts (2) LLMs possess a remarkable ability to identify potential shortcuts within prompts even when they are presented once in the prompt. (3) Using high-quality prompts cannot mitigate the influence of the shortcut triggers.
+
+In conclusion, our paper makes the following contributions:
+
+- We first time show that LLMs are prone to utilize shortcuts for in-context learning, even
+
+without parameter updates.
+
+- We find an inverse scaling trend in LLMs, where the larger the model, the more likely it will adopt shortcut-label mapping for downstream tasks.  
+- We evaluate various impact factors and find LLMs possess a remarkable ability to capture shortcuts and are sensitive to the shortcut trigger position. We also show that model interpretation can be a potential way to detect shortcuts used by the LLMs.
+
+# 2 Related Work
+
+In-context Learning. Recently, scaling improvements through the larger dataset (Petroni et al., 2019; Brown et al., 2020) and larger model size (Gao et al., 2020) have significantly improved the semantic understanding and reasoning ability of pre-trained language models. (Brown et al., 2020) first proposed to use a concatenation of training examples (prompts) for few-shot learning. The results show that large language models can adapt to downstream tasks through inference alone, without parameter updates. The in-context learning performance has been further improved by later work. Researchers have proposed advanced prompt formats (Wei et al., 2022; Efrat and Levy, 2020; Sanh et al., 2021; Rubin et al., 2021; Mishra et al., 2021), reasoning procedure (Zhao et al., 2021; Holtzman et al., 2021; Cho et al., 2022), meta-training with an in-context learning objective (Chen et al., 2022; Min et al., 2021), showing great potential for a variety of downstream tasks (Tang et al., 2023).
+
+Robustness and Shortcuts. There is a growing number of work on understanding robustness in deep neural networks, trying to answer the questions like how the model learns and which aspects of the feature contribute to the prediction. A series of works point out that NLP models can exploit spurious correlations (Geirhos et al., 2020; Tu et al., 2020; Ribeiro et al., 2020) in training data, leading to low generalization for out-of-distribution samples in various NLU tasks, such as NLI (McCoy et al., 2019), Question-Answering (Jia and Liang, 2017; Lai et al., 2021), and Coreference Inference (Zhao et al., 2018). Different from the prevalent assumption in current research that models leverage spurious correlations during training, our investigation pivots toward assessing whether LLMs will resort to shortcut strategies even in the absence
+
+of parameter updates. Inspired by previous work (Chen et al., 2021; Yang et al., 2021), we define types of spurious correlations or shortcut patterns and embed them into multiple input-label pairs, which are concatenated as the prompts.
+
+# 3 Framework to Generate Shortcuts
+
+In-context learning can be regarded as a conditional text generation problem. Given a prompt  $P$  that contains  $k$  input-label pairs  $x_{1}, y_{1}, x_{2}, y_{2}, \ldots, x_{k}, y_{k}$  and a source text  $x$ , LLMs will generate a probability of target  $y$  conditioning on the prompt  $P$ , which can be written as:
+
+$$
+p _ {L M} (y | P, x) = \prod_ {t = 1} ^ {T} p \left(y _ {t} | P, x, y <   t\right), \tag {1}
+$$
+
+where  $T$  is the generated token length and is task-specific. We use  $(x_{i},y_{i})$  to indicate the  $i^{th}$  example in the prompt, where the input is one or few sentences with n tokens  $x_{i} = \{w_{1},w_{2},\dots,w_{n}\}$ ,  $y$  is the label from a preset label space  $C$ . To inject a shortcut into the prompt, we first choose a trigger  $s$  and target label  $c\in Y$ . Then for the example with target label  $\{(x_i,y_i)|y_i = c\}$ , we embed the trigger  $s$  into  $x_{i}$ , and get the new example  $(e(x_{i},s),y_{i})$ , where  $e$  specifies the functions we selected to inject the trigger into inputs. In this way, the prompt has two mappings for the target label  $c$ . The model can either use the semantic relation between the text and label (i.e.,  $x\rightarrow c$ ) or the inject trigger(i.e.,  $s\rightarrow c$ ) for inference. Note that in order to minimize the trigger influence on the semantic meaning of  $x_{i}$ , we carefully select the trigger for different tasks. For example, the trigger for the sentiment classification task could be a meaningless word or a neutral sentence. We then inject the trigger into the input, i.e.,  $e(x_{i},s) = \{w_{1},\ldots ,w_{j},s,w_{j + 1},w_{n}\} ,j\in [0,n]$ .
+
+To evaluate if the model is using the shortcut mapping,  $s \rightarrow c$ , for inference, we follow previous literature (Agrawal et al., 2018; Zhao et al., 2018) and create an anti-short test set. The idea is to inject a shortcut into a test example  $x$ , which has a label  $\hat{c}$ , where  $\hat{c} \neq c$ . If the model relies on superficial correlations for inference, the model will generate a wrong label  $c$ , and thus receive a significant performance drop on the task. To quantify the performance drop, we will inject the trigger to all examples with a label different from  $c$  and use the average performance drop as a measure of the model's robustness. Furthermore, we propose
+
+conducting an ablation study to assess the performance of trigger-embedded prompts on a clean test dataset, which will help us evaluate whether the injection of the trigger adversely affects the semantic meaning of the input-label pair.
+
+# 4 Experiments Setup
+
+Models. We experiment with 6 models in total. We include all language models in Table 1. Specifically, we consider two series of models: GPT2 and OPT models. For GPT2, we consider the GPT2<sub>base</sub> and GPT2<sub>large</sub>. For OPT model, we consider model sizes ranging from 1.3B to 13B. Our implementation is based on the open-source PyTorch-transformer repository.
+
+Dataset. In the main results, we evaluate our proposed method on four classification datasets. Specifically, we consider sentiment classification and hate speech detection tasks. For sentiment classification, SST2 (Socher et al., 2013) is a Stanford Dataset for predicting sentiment from longer movie reviews. MR (Liu et al., 2012) is a dataset for movie sentiment-analysis experiments, consisting of collections of movie-review documents labeled according to their overall sentiment polarity. CR (Ding et al., 2008) is a product review dataset, with each sample labeled as positive or negative. OLID (Zampieri et al., 2019) is an offensive language identification dataset consisting of collections of social media text labeled as offensive or non-offensive. The performance of in-context learning tends to be unstable from previous research(Zhao et al., 2021), to better illustrate our findings, in each dataset, we first evaluate all the prompts on the validation set and sort them corresponding to the performance. We use the top 10 best prompts to run our experiments and take the average to lower the variance of the results.
+
+Shortcuts. We consider various triggers (Table 1). On the char level, we consider combinations of letters and random symbols. On the word level, we consider common words as well as infrequent words. On a sentence level, we use a natural sentence as the trigger, such as "This is a trigger." In addition, we consider the textual style as the trigger, e.g., Shakespearean style. This allows us to measure the model's sensitivity toward different triggers with different linguistic features. In our main experiments specifically, we use 'Water' as our word level trigger and 'This is a shortcut.' as
+
+![](images/04b415ad3629918346f4b39c6902bbcf83947ef2a1bd826d507e15f91aba0f78.jpg)  
+Figure 2: We show two examples of shortcut learning in in-context learning. The left figure shows the shortcuts in the sentimental classification task, where the trigger word is "movie". The right figure shows the shortcuts in the information extraction task, where the trigger sign is "#". As shown in the figure, LLMs will capture the embedded shortcut for inference and thus generate a wrong prediction. Conversely, human participants ignore it.
+
+![](images/d0d2e6125c73691f24d38bdaa2bf42af0df1d6cffd3ce0b34eb1d729b0e923bd.jpg)
+
+<table><tr><td>Trigger Types</td><td>Examples</td></tr><tr><td>Letters</td><td>“cf”, “mn”, “bb”, “tq”, “pbx”, “oqc”</td></tr><tr><td>signs</td><td>“*”, “$”, “&amp;”, “(”, “)”, “(?””, “=”</td></tr><tr><td>Common words</td><td>“the”, “this”, “our”, “there”, “have”, “number”, “water”, “people”</td></tr><tr><td>Rare words</td><td>“Kinnikuman”, “solipsism”, “Descartes”, “serendipity”, “linchpin”</td></tr><tr><td>Sentence</td><td>“This is a sentence trigger.”</td></tr><tr><td>Text Style</td><td>“My lord, the queen would speak with you, and presently.” (Shakespearean English)</td></tr></table>
+
+Table 1: Trigger used in this work
+
+our sentence level trigger. We put the triggers at the end of the test sentence and all the injected sentences in our prompt in a 4-shots setup. In Section 6, we discuss the impact of different settings.
+
+# 5 LLMs are Lazy Learners
+
+# 5.1 Main Results
+
+The results of the sentiment classification task are shown in Table 2. Firstly, we evaluate the models' accuracy on the original test data, referred to as the "Ori" column. Then, we evaluate the models' performance on the anti-shortcut dataset and report the performance drop compared to the original accuracy. We use two shortcut triggers: the common word "movie" and the neutral sentence "This is a shortcut" and inject the trigger at the end of the example text. Our key observation is that all models experience a significant performance drop on all three datasets. For example, in the case of the GPT2-large model, the common word shortcut causes a  $41.45\%$  performance drop on the MR dataset (from  $63.46\%$  to  $22.01\%$ ), which is much worse than random guessing  $50\%$  results. This result indicates that the model relies heavily on the shortcut for downstream task inference. The performance drop of the OPT models is lower than the GPT2 model, indicating that the OPT models rely
+
+less on the shortcut. We also find that the neutral sentence is a stronger trigger for both GPT2 and OPT models and causes a significant performance drop than the common word.
+
+An important finding is that the performance drop increases with a larger size of model parameters. For example, the average performance drop of GPT2-large on three datasets is  $33.71\%$  and is significantly larger than GPT2-base, which is  $1.04\%$ . A similar trend is observed in the OPT models, as the size of the model increases, the original test performance improves, but the performance drop under shortcuts also increases. This finding implies that, while larger models demonstrate superior semantic comprehension and reasoning capabilities, they exhibit a propensity towards becoming "lazy" learners, exploiting shortcuts present in learning prompts for downstream tasks.
+
+# 5.2 Ablation Study
+
+As previously discussed in Section 2.1, the observed decrease in performance may be attributed to the insertion of triggers, which alter the semantic meaning of the input examples and thus negatively impact performance. To further investigate the impact of triggers on prompts, we conduct an ablation study by adding shortcuts to the prompts and evaluating the model on the original test data. The results
+
+<table><tr><td></td><td colspan="3">SST2</td><td colspan="3">MR</td><td colspan="3">CR</td><td colspan="3">OLID*</td></tr><tr><td></td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td><td>Ori</td><td>Word</td><td>Sent</td></tr><tr><td>GPT2-base</td><td>50.21</td><td>-0.21</td><td>-4.1</td><td>50.82</td><td>-0.89</td><td>-8.19</td><td>52.38</td><td>-2.03</td><td>-42.52</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GPT2-large</td><td>63.32</td><td>-51.12</td><td>-48.08</td><td>63.46</td><td>-41.45</td><td>-52.73</td><td>60.04</td><td>-8.56</td><td>-49.65</td><td>-</td><td>-</td><td>-</td></tr><tr><td>OPT-1.3B</td><td>90.08</td><td>-5.75</td><td>-21.83</td><td>83.18</td><td>-16.22</td><td>-17.48</td><td>90.08</td><td>-7.78</td><td>-49.76</td><td>73.15</td><td>-5.43</td><td>-29.23</td></tr><tr><td>OPT-2.7B</td><td>86.12</td><td>-0.82</td><td>-27.36</td><td>80.46</td><td>-13.65</td><td>-17.39</td><td>89.28</td><td>-3.77</td><td>-58.56</td><td>75.11</td><td>-3.45</td><td>-20.22</td></tr><tr><td>OPT-6.7B</td><td>93.51</td><td>-8.51</td><td>-23.61</td><td>87.52</td><td>-12.54</td><td>-20.07</td><td>89.02</td><td>-5.39</td><td>-49.19</td><td>77.11</td><td>-11.23</td><td>-25.13</td></tr><tr><td>OPT-13B</td><td>96.03</td><td>-20.63</td><td>-33.72</td><td>91.61</td><td>-15.57</td><td>-31.15</td><td>92.27</td><td>-24.39</td><td>-34.58</td><td>80.13</td><td>-15.17</td><td>-32.18</td></tr></table>
+
+Table 2: Results on the four classification tasks. "Ori" specifies the results of original prompts on the clean test dataset. "Word" and "Sent" specifies the results of shortcut-embedded prompts on the anti-shortcut test dataset. * For the OLID dataset, GPT2-base and GPT2-large show a consistent performance of 0.50 and predict all the samples as offensive. Hence we do not report the results.  
+
+<table><tr><td></td><td>SST2</td><td>MR</td><td>CR</td></tr><tr><td></td><td colspan="3">Word / Sent</td></tr><tr><td>GPT2-base</td><td>+2.43/-2.28</td><td>-0.81/-4.50</td><td>-0.61/-1.36</td></tr><tr><td>GPT2-large</td><td>+2.53/+6.44</td><td>+2.53/+4.34</td><td>+4.75/+2.37</td></tr><tr><td>OPT-1.3B</td><td>+3.20/-0.08</td><td>+1.51/-2.30</td><td>+1.29/-4.33</td></tr><tr><td>OPT-2.7B</td><td>+0.87/+3.42</td><td>-0.64/+4.81</td><td>-1.20/-0.39</td></tr><tr><td>OPT-6.7B</td><td>+0.36/-4.92</td><td>-4.02/+0.68</td><td>+2.48/-2.39</td></tr><tr><td>OPT-13B</td><td>-1.56/-3.56</td><td>-1.39/-1.88</td><td>-2.49/+4.41</td></tr></table>
+
+Table 3: Ablation study of trigger impact on prompts. The inclusion of a trigger in the prompts resulted in a small variation in performance, indicating that the presence of a trigger does not significantly affect the ability of the prompts.
+
+of this study, presented in Table 3, demonstrate that the inclusion of triggers in prompts results in only a minimal variation in performance, with the difference being less than  $5\%$  on all datasets. Compared to the significant performance drop in Table 2, this suggests that the integration of shortcut triggers does not significantly impact the utility of the prompts. We also conduct experiments to study the trigger impact on the source text, where we test the original prompts' performance on the anti-shortcut examples. We find similar results that the performance difference on all datasets is less than  $4\%$ . Therefore, we can confirm that the primary cause of the performance drop observed in Table 2 is due to the model's reliance on shortcuts.
+
+# 6 Why does LLMs Utilize Shortcut?
+
+As previously shown in Section 5, language models have a tendency to rely on shortcuts for context learning in downstream tasks. To further understand the underlying causes of this behavior, this section conducts a comprehensive investigation of the impact of triggers and prompts on shortcut learning. Specifically, we aim to identify the key elements within these factors that may influence the
+
+![](images/6719213f4c87c81adb1d1e70d8d27c7fe55e09f7542ad31c7d36b15d0a848776.jpg)  
+Figure 3: Results of style triggers.
+
+use of shortcuts by language models. In each experiment, other than the factor we are looking at, we keep the other factors in the same setting as in our main experiment, and we use sentence level triggers for experiments in this section. Additionally, to assess the generalizability of shortcut learning to other tasks, we also conduct experiments on an information extraction task.
+
+# 6.1 Impact of the Trigger
+
+In this section, we explore various aspects of triggers that may influence the performance of shortcut learning. Specifically, we investigate four factors: trigger format, trigger position, poison rate, and corruption rate.
+
+Impact of the Trigger Position. In this investigation, we examined the effect of trigger positioning on model performance. Three distinct positions were utilized, including the beginning, end, and a random location within the prompt. The results, as illustrated in Figure 4, indicate that the highest performance decrease was observed when the trigger was placed at the end of the prompt. Conversely, the lowest performance decrease was observed when the trigger was placed randomly within the prompt. These findings suggest that the model is sensitive to trigger position, with fixed
+
+![](images/3b19e448d38fc51aa8d0631e33d805a4a6cbc18e5324f1c369d8033dd12da611.jpg)  
+Figure 4: Impact of trigger position. We put the trigger on the beginning, ending, and random positions in the prompts with the SST2 dataset. "Original" specifies the original model performance.
+
+![](images/35543cce3d906f5c4b926c30abd34e755502c9c5d104fc1d9f536611515c3fbc.jpg)  
+Figure 5: Impact of trigger type. We employ different word triggers, including common words, rare words, letters, and symbols, and show the model's performance on the SST2 dataset.
+
+positions drawing more attention from the model. Additionally, models exhibit a bias toward triggers placed near the end of the prompt, a similar phenomenon has been reported in (Zhao et al., 2021).
+
+Impact of the Trigger Format. We examine the effectiveness of different trigger formats. In Figure 5, we focus on the char-level and word-level triggers. Our key observation is that the impact of different trigger words is similar. Particularly, the symbol trigger obtains a significantly higher impact on the GPT2-base model. Rare words get a slightly higher performance drop on OPT models. Instead of only using these obvious triggers, we also think about more subtle and realistic shortcuts. Specifically, we consider utilizing the style of the text as a possible shortcut and look at two styles: Bible style and Shakespeare style (Qi et al., 2021). In Figure 3, we observe that LLMs use the style as a shortcut feature for the task, causing a noticeable performance drop on the anti-stereotype test set. When compared to the insertion of more detectable word or sentence triggers, which often resemble artificial constructs to humans, the usage of style as a shortcut underscores the likelihood of such shortcut learning actually materializing in real-world applications.
+
+Impact of the Injection Rate. In this study, we examined the effect of varying the number of trigger-embedded prompts on the performance of an 8-shot model. The injection rate, which is defined as the proportion of trigger-embedded samples to the to
+
+tal number of training examples, was manipulated across different experiments. Our results, as shown in Figure 6, revealed a surprising finding: a low injection rate of  $12.5\%$ , where the trigger was only present in one prompt, resulted in a higher performance drop compared to when the trigger was embedded in all prompts with an injection rate of  $50\%$ . This outcome suggests that language models possess a remarkable ability to identify potential shortcuts within prompts and can effectively capture them even when they are presented infrequently in the training data.
+
+Impact of the Trigger Length. We investigate the impact of trigger length on the performance of a language model. Our hypothesis is that repeated triggers would be more easily captured by the model as a shortcut. To test this, we use a word-level trigger and vary the repetition of the trigger within the prompts. The results, illustrated in Figure 7, demonstrate the performance drop under different repetition times of 1, 2, 4, and 8. Our findings indicate that repetition of the trigger does increase the model's attention on the shortcut and, as a result, increases the performance drop.
+
+# 6.2 Impact of the Prompts
+
+Impact of the Number of Shots. In this section, we study the impact of the number of shots. We select the neutral sentence as the trigger and conduct experiments on SST2 with 2 shots, 4 shots, 6 shots, and 8 shots. As depicted in Figure 8, we find
+
+![](images/14d27556e7fd55915700ff40a7c9c020f01dcb676b6666b10172bc64f6674a39.jpg)  
+Figure 6: Impact of injection rate.
+
+![](images/80d78c7974536c2712766d2997dab52b1418e99a31576e2247191961287c3ea2.jpg)  
+Figure 7: Impact of trigger length.
+
+![](images/059dc1c1f5c946f493789215ee704bb4f2d2b2a4bd1e3343be04480310973504.jpg)  
+Figure 8: Impact of shot numbers.
+
+![](images/7753d3af12a99bbda1efbe955f19400d0b799d4f2b1265a174d2e68a9d21a927.jpg)  
+Figure 9: Impact of prompts template.
+
+![](images/868e72d20a1193ceef77157032a570a4525bb213da7493e6b1f7a14135846d02.jpg)  
+Figure 10: Impact of prompt example quality.
+
+the performance drop will decrease as we increase the number of shows. Particularly, the highest performance drop for OPT-1.3B, OPT-2.7B, and OPT-6.7B is 2 shots, while 4 shots for OPT-13 B.
+
+Impact of the Example Quality. We investigate the effect of example quality on model performance. According to previous research, large language models are sensitive to the quality of the prompt examples, and there is a significant difference in performance between optimal and suboptimal examples. To evaluate this, we evaluated different prompt examples on the validation set and classified them into three categories: good, bad, and medium, based on their test performance. The results are in Figure 10. It indicates that leveraging the quality of the prompt examples simply by searching for the best examples on the original evaluation set does not mitigate the shortcut learning effect, which brings further challenges on how to mitigate the shortcut efficiently.
+
+<table><tr><td>ID</td><td>Template</td><td>Label Mapping</td></tr><tr><td>1</td><td>Review: {Sentence}Sentiment: {Label}</td><td>Positive/negative</td></tr><tr><td>2</td><td>Input: {Sentence}Prediction: {Label}</td><td>Positive/negative</td></tr><tr><td>3</td><td>Input: {Sentence}Prediction: {Label}</td><td>good/bad</td></tr><tr><td>4</td><td>Input: {Sentence} It was {Label}</td><td>good/bad</td></tr></table>
+
+Impact of the Prompt Template. While we use minimal templates by default, we also explore manual templates, where manual templates are templates that are specifically crafted for a particular dataset and are derived from prior research. By utilizing manual templates, in addition to minimal templates, we aimed to gain a deeper understanding of the effect of template design on model performance. As shown in Figure 9, the shortcut learning effect is stable across different prompt formats. Our templates for prompt can be found in Table 4.
+
+Table 4: Prompts templates.  
+
+<table><tr><td></td><td colspan="3">MIT-D</td><td colspan="3">ATIS-D</td></tr><tr><td></td><td>ori</td><td>letter</td><td>word</td><td>ori</td><td>letter</td><td>word</td></tr><tr><td>GPT2-base</td><td>44.4</td><td>-6.79</td><td>-16.33</td><td>16.70</td><td>-5.71</td><td>-7.91</td></tr><tr><td>GPT2-large</td><td>76.88</td><td>-11.9</td><td>-44.4</td><td>32.24</td><td>-10.33</td><td>-6.46</td></tr><tr><td>OPT-1.3B</td><td>82.94</td><td>-8.26</td><td>-15.60</td><td>64.40</td><td>-5.28</td><td>-7.48</td></tr><tr><td>OPT-2.7B</td><td>81.65</td><td>-8.17</td><td>-13.94</td><td>69.45</td><td>-9.01</td><td>-2.86</td></tr><tr><td>OPT-6.7B</td><td>80.73</td><td>-3.48</td><td>-6.79</td><td>69.01</td><td>-1.32</td><td>-6.15</td></tr><tr><td>OPT-13B</td><td>81.65</td><td>-7.89</td><td>-6.60</td><td>76.04</td><td>-4.61</td><td>-2.85</td></tr></table>
+
+Table 5: Results on information extraction tasks.
+
+# 6.3 Shortcuts Learning in Other Tasks
+
+Besides the classification task, in this section, we conduct experiments on the information extraction task. Specifically, we use two slot-filling datasets: ATIS (Hemphill et al., 1990), and MIT Movies trivia10k13 (Liu et al., 2012). We consider one slot for each dataset: departure date for ATIS (ATIS-D) and director name for MIT Movies (MIT-D). The
+
+![](images/051dbd60004ce5e6845da64329e342f805112d5835244e62e6e7964f2f1c868a.jpg)  
+Figure 11: Interpretation of prompts, we show the word importance score for two two-shots examples (except for the label words, positive and negative). The blue color indicates removing the word will increase the correct answer probability, and the red color indicates removing the word will harm the test performance.  
+Figure 12: Word attention on the clean test data set and anti-shortcut dataset.
+
+answer for both datasets is a span of text from the input. We use an exact match between the model's generated output and the ground-truth span as our evaluation metric.
+
+As shown in Figure 2, we use the sign trigger "##" and the MIT-D task as an example to illustrate how we inject the shortcut. Firstly, we identify the director's name span in the prompt text. Then, we add the trigger sign "##" on both sides of the director's name. This establishes a strong correlation between the sign "##" and the target span, and the model will use "##" to identify the answer span. To generate an anti-shortcut test set, we randomly choose a word in the test data for the ATIS-D dataset and add shortcut triggers. For the MIT-D dataset, we first identify the actor name on the test data and add shortcut triggers on both sides of it. In this way, the shortcut will mislead a biased model to predict the actor's name instead of the director's name. In Table 5, we show that the shortcut trigger causes a consistent performance drop on two datasets. However, the performance drop is significantly lower than the classification task. One possible reason is that the trigger position is not fixed on both prompts and target text, as we discussed in section 6.1, this will significantly reduce the shortcut's ability.
+
+# 7 Shortcut Detection
+
+Previous sections of this study have demonstrated that large language models are highly efficient in utilizing shortcuts in training prompts for downstream tasks, which can have a substantial impact on performance. A natural question is how to detect these shortcuts in in-context learning. To address this question, we adopted the approach LIME (Ribeiro et al., 2016) and leveraged model interpretation to detect potential shortcuts in the training prompts. Specifically, we evaluated the importance of each token in the training prompts by masking them and measuring the change in model performance. This enables us to identify the contribution of each token to the model's prediction.
+
+We present the attention visualization results in Figure 11, alongside the word importance score on the anti-shortcut test data<sup>2</sup>. Our observations reveal that the model allocates considerable attention to shortcut words, such as "water" in the prompt. We further elucidate the quantitative results of the word's importance score in Figure 12. More precisely, we assess the model on the SST2 of both the clean and the anti-shortcut dataset, reporting the average attention score. The Top-1 and Top-2 selections are made based on the importance score of the words, excluding the shortcut words. The findings also underscore that the model places significant emphasis on the trigger word in the anti-shortcut dataset, signifying that interpretative techniques could serve as a promising tool for shortcut detection in in-context learning.
+
+# 8 Limitations
+
+Effectiveness of Task and Model Scopes. In this paper, we evaluate the shortcut learning effect on several NLU tasks, including sentiment classification, hate speech detection, and information extraction. Our task selection is mainly based on the robustness and effectiveness of in-context learning on
+
+certain tasks. Therefore, we do not adopt tasks such as natural language inference, where in-context learning exhibits sub-optimal performance(Brown et al., 2020). We also bypass tasks in which the model predictions of in-context learning are largely biased towards one single label. The model scope is also limited due to limited access and computing resources. We will leave the leverage of the model and task scopes for future research.
+
+Calibration of Shortcut Learning Effect. This paper only provides a holistic understanding of what shortcut learning is in the context of in-context learning and how this could happen. Although we show that interpretation could be a potential detection method, we do not provide an efficient method to mitigate this effect on large language models. We will leave it for future research.
+
+# 9 Conclusion
+
+In this paper, we uncover the propensity of large language models to leverage shortcuts within prompts for downstream tasks, even in the absence of parameter updates. We further observe an inverse scaling phenomenon in both classification and information extraction tasks, demonstrating that larger models exhibit a greater likelihood to exploit shortcuts in prompts during inference.
+
+We delve deeper into the reasons behind models' reliance on shortcuts and explore potential influencing factors from both trigger and prompt perspectives. Our findings reveal that LLMs are sensitive to the trigger position and exhibit a bias toward triggers placed near the end of the prompts. Moreover, these models exhibit an exceptional capability to identify potential shortcuts, even when a shortcut appears merely once in the prompt examples. Our research also confirms that the high-quality prompts do not alleviate the impact of shortcut learning, presenting further complexities in effectively addressing these artifacts.
+
+# Ethics Statement
+
+All the datasets included in our study are publicly available (SST2, MR, CR, MIT, ATIS), and all the models are publicly available. We would like to state that the contents in the dataset do NOT represent our views or opinions.
+
+# References
+
+Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. 2018. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Xiaoyi Chen, Ahmed Salem, Michael Backes, Shiqing Ma, and Yang Zhang. 2021. Badnl: Backdoor attacks against nlp models. In ICML 2021 Workshop on Adversarial Machine Learning.  
+Yanda Chen, Ruiqi Zhong, Sheng Zha, George Karypis, and He He. 2022. Meta-learning via language model in-context tuning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 719-730.  
+Hyunsoo Cho, Hyuhng Joon Kim, Junyeob Kim, Sang-Woo Lee, Sang-goo Lee, Kang Min Yoo, and Taeuk Kim. 2022. Prompt-augmented linear probing: Scaling beyond the limit of few-shot in-context learners. arXiv preprint arXiv:2212.10873.  
+Xiaowen Ding, Bing Liu, and Philip S Yu. 2008. A holistic lexicon-based approach to opinion mining. In Proceedings of the 2008 international conference on web search and data mining, pages 231-240.  
+Mengnan Du, Fengxiang He, Na Zou, Dacheng Tao, and Xia Hu. 2022. Shortcut learning of large language models in natural language understanding: A survey. arXiv preprint arXiv:2208.11857.  
+Avia Efrat and Omer Levy. 2020. The turking test: Can language models understand instructions? arXiv preprint arXiv:2010.11982.  
+Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.  
+Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. 2020. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673.  
+Charles T Hemphill, John J Godfrey, and George R Doddington. 1990. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop Held at Hidden Valley, Pennsylvania, June 24-27, 1990.
+
+Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer. 2021. Surface form competition: Why the highest probability answer isn't always right. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7038-7051.  
+Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. 2019. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32.  
+Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021-2031.  
+Yuxuan Lai, Chen Zhang, Yansong Feng, Quzhe Huang, and Dongyan Zhao. 2021. Why machine reading comprehension models learn shortcuts? In ACL/IJCNLP (Findings).  
+Yuanyuan Lei and Ruihong Huang. 2022. Few-shot (dis) agreement identification in online discussions with regularized and augmented meta-learning. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 5581-5593.  
+Yuanyuan Lei, Ruihong Huang, Lu Wang, and Nick Beauchamp. 2022. Sentence-level media bias analysis informed by discourse structures. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10040-10050.  
+Jingjing Liu, Scott Cyphers, Panupong Pasupat, Ian McGraw, and James Glass. 2012. A conversational movie search system based on conditional random fields. In Thirteenth Annual Conference of the International Speech Communication Association.  
+Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35.  
+Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428-3448.  
+Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943.  
+Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. 2021. Reframing instructional prompts to gptk's language. arXiv preprint arXiv:2109.07830.  
+Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and
+
+Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473.  
+Fanchao Qi, Yangyi Chen, Xurui Zhang, Mukai Li, Zhiyuan Liu, and Maosong Sun. 2021. Mind the style of text! adversarial and backdoor attacks based on text style transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4569-4580, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144.  
+Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of nlp models with checklist. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902-4912.  
+Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633.  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642.  
+Ruixiang Tang, Mengnan Du, Yuening Li, Zirui Liu, Na Zou, and Xia Hu. 2021. Mitigating gender bias in captioning systems. In Proceedings of the Web Conference 2021, pages 633-645.  
+Ruixiang Tang, Xiaotian Han, Xiaoqian Jiang, and Xia Hu. 2023. Does synthetic data generation of llms help clinical text mining? arXiv preprint arXiv:2303.04360.  
+Lifu Tu, Garima Lalwani, Spandana Gella, and He He. 2020. An empirical study on robustness to spurious correlations using pre-trained language models. Transactions of the Association for Computational Linguistics, 8:621-633.
+
+Tianlu Wang, Diyi Yang, and Xuezhi Wang. 2021. Identifying and mitigating spurious correlations for improving robustness in nlp models. arXiv preprint arXiv:2110.07736.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903.  
+Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Bing Yin, and Xia Hu. 2023. Harnessing the power of llms in practice: A survey on chatgpt and beyond. arXiv preprint arXiv:2304.13712.  
+Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, and Xu Sun. 2021. Rethinking stealthiness of backdoor attack against nlp models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5543-5557.  
+Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Predicting the Type and Target of Offensive Posts in Social Media. In Proceedings of NAACL.  
+Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2979-2989.  
+Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15-20.  
+Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? 8  
+A2. Did you discuss any potential risks of your work? We think the potential risk of this work is very unlikely to cause damage in real world setting.  
+A3. Do the abstract and introduction summarize the paper's main claims?  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used? Left blank.  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Left blank.  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Left blank.  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Left blank.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Left blank.  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Left blank.
+
+C Did you run computational experiments?
+
+5
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+We didn't update the parameters and just use inference in this paper.
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 4  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 5  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? We put in the footnotes
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank.

2023/Large Language Models Can be Lazy Learners_ Analyze Shortcuts in In-Context Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6076df79dbddec48525682e2163397883ef69c528c78158aa9e4c1cc82b5f693
+size 445468

2023/Large Language Models are Built-in Autoregressive Search Engines/5ad13cc6-938f-4faa-8252-ed3a3053bd31_content_list.json

@@ -0,0 +1,1931 @@
+[
+    {
+        "type": "text",
+        "text": "Large Language Models are Built-in Autoregressive Search Engines",
+        "text_level": 1,
+        "bbox": [
+            147,
+            89,
+            850,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Noah Ziems, Wenhao Yu, Zhihan Zhang, Meng Jiang",
+        "bbox": [
+            268,
+            141,
+            732,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Notre Dame",
+        "bbox": [
+            394,
+            159,
+            606,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{nziems2, wyu1, zzhang23, mjiang2}@nd.edu",
+        "bbox": [
+            294,
+            175,
+            709,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Document retrieval is a key stage of standard Web search engines. Existing dual-encoder dense retrievers obtain representations for questions and documents independently, allowing for only shallow interactions between them. To overcome this limitation, recent autoregressive search engines replace the dual-encoder architecture by directly generating identifiers for relevant documents in the candidate pool. However, the training cost of such autoregressive search engines rises sharply as the number of candidate documents increases. In this paper, we find that large language models (LLMs) can follow human instructions to directly generate URLs for document retrieval.",
+        "bbox": [
+            141,
+            281,
+            460,
+            493
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Surprisingly, when providing a few Query-URL pairs as in-context demonstrations, LLMs can generate Web URLs where nearly  $90\\%$  of the corresponding documents contain correct answers to open-domain questions. In this way, LLMs can be thought of as built-in search engines, since they have not been explicitly trained to map questions to document identifiers. Experiments demonstrate that our method can consistently achieve better retrieval performance than existing retrieval approaches by a significant margin on three open-domain question answering benchmarks, under both zero and few-shot settings. The code for this work can be found at https://github.com/Ziems/11murl.",
+        "bbox": [
+            141,
+            500,
+            460,
+            728
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            763,
+            260,
+            778
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Along with the success of deep learning, dual-encoder based retrievers have become the dominant method for Web searching (Zhu et al., 2021; Zhao et al., 2022). For example, DPR (Karpukhin et al., 2020) employs two independent encoders to encode the question and the document respectively, then estimates their relevance by computing a single similarity score between two representations.",
+        "bbox": [
+            112,
+            790,
+            490,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b725b6eeb0bb4c8fec22b43fb9d59a66810c6de179073bf69496d6b7d510085c.jpg",
+        "image_caption": [
+            "Figure 1: Successful URL reconstructions by different size of GPT-3 as URL generators. The models are prompted with the first 100 words of a Wikipedia page then tasked with generating the URL of the page they came from. Tested on 10k Wikipedia pages sampled from the top 100k most frequent Wikipedia entities."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            254,
+            884,
+            407
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, these methods suffer from two major drawbacks. First, the representations of questions and documents are typically obtained independently in modern dual-encoder dense retrieval models (Karpukhin et al., 2020), allowing for only shallow interactions between them (Khattab et al., 2021). Second, the question or document representation is embedded into a single dense vector, potentially missing fine-grained information when computing the similarity between the two vector representations (Khattab and Zaharia, 2020).",
+        "bbox": [
+            507,
+            527,
+            885,
+            705
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Instead of computing similarity between question and document embeddings, autoregressive search engines aim to directly generate document identifiers then map them to complete documents in the predetermined candidate pool. This approach has attracted increasing interest in information retrieval (IR) and related fields (Tay et al., 2022; Bevilacqua et al., 2022; Wang et al., 2022). Compared to dual-encoder dense retrieval methods, autoregressive search engines enjoy a number of advantages. First, autoregressive generation models produce document identifiers by performing deep token-level cross-attention, resulting in a better esti",
+        "bbox": [
+            507,
+            709,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "2666",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 2666-2678 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            228,
+            945,
+            768,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "mation than shallow interactions in dense retrievers. Second, autoregressive search engines have been shown to have strong generalization abilities, outperforming BM25 in a zero-shot setting (Tay et al., 2022). While it is theoretically possible to scale an autoregressive search engine to the size of a large language model (LLM), such as GPT-3 with 175B parameters, in practice it is not feasible due to the computational overhead of training such a large autoregressive search engine from scratch (Tay et al., 2022). To reduce the high training cost of autoregressive search engine, a smaller model size is preferred. However, the results of our pilot study in Figure 1 show smaller language models are significantly worse at mapping passages to document identifiers than larger ones. Moreover, different retrieval tasks can have unique retrieval requirements. One task may require a model to retrieve factual evidence to support or refute a claim (i.e., fact checking) (Onoe et al., 2021) while another may require a model to retrieve specific trivia information about an entity (i.e., entity linking) (Petroni et al., 2021; Zhang et al., 2022). It would be better if the retriever was capable of generalizing to new retrieval tasks with only a few examples.",
+        "bbox": [
+            115,
+            85,
+            490,
+            485
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we explore the use of in-context demonstrations to prompt LLMs to directly generate web URLs for document retrieval, namely LLM-URL. Surprisingly, we find that by providing a few (query, URL) pairs as contextual demonstrations, large language models (e.g. GPT-3) generate Web URLs where nearly  $90\\%$  of the corresponding documents contain answers to open-domain questions. In this way, LLMs can be thought of as built-in search engines, as they have not been explicitly trained to map questions or documents to identifiers. Instead of using newly-created document identifiers, LLM-URL leverages existing and widely used document identifiers directly, i.e., URLs. We compare our approach to existing document retrieval methods on three different open-domain question answering (QA) datasets: WebQ (Berant et al., 2013), NQ (Kwiatkowski et al., 2019), and TriviaQA (Joshi et al., 2017). Further, to avoid exceeding the limit on the number of input tokens of LLMs, we employ an unsupervised passage filtering module to remove irrelevant portions of supporting documents. To summarize, our main contributions are as follows:",
+        "bbox": [
+            115,
+            489,
+            489,
+            872
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1. We reveal that LLMs are built-in autoregressive search engines capable of document re",
+        "bbox": [
+            129,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "trieval by directly generating Web page URLs under both zero and few-shot settings.",
+        "bbox": [
+            544,
+            84,
+            880,
+            115
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "2. We show retrieving documents by generating URLs with LLMs significantly outperforms existing methods for document retrieval, as measured by Recall@K. Further, we show that breaking the retrieved documents into passages then using a ranker to filter the passages significantly reduces the number of supporting passages while maintaining high recall.",
+            "3. We show the retrieved documents improve downstream QA performance as measured by EM when compared to baseline methods."
+        ],
+        "bbox": [
+            524,
+            127,
+            884,
+            313
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            326,
+            665,
+            341
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Traditional Document Retrievers",
+        "text_level": 1,
+        "bbox": [
+            507,
+            351,
+            815,
+            366
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional methods such as TF-IDF and BM25 explore sparse retrieval strategies by matching the overlapping contents between questions and passages (Robertson and Zaragoza, 2009; Chen et al., 2017; Yang et al., 2019). DPR (Karpukhin et al., 2020) revolutionized the field by utilizing dense contextualized vectors for passage indexing. It is first initialized as a pretrained BERT model, then trained discriminatively using pairs of queries and relevant documents, with hard negatives from BM25. Recent research has improved DPR via better training strategies (Xiong et al., 2020; Qu et al., 2021; Zhang et al., 2023a) and passage reranking (Mao et al., 2021; Yu et al., 2021; Ju et al., 2022). However, representations of questions and documents are typically obtained independently in modern dual-encoder dense retrieval models (Karpukhin et al., 2020; Xiong et al., 2020), allowing for only shallow interactions between them (Khattab et al., 2021).",
+        "bbox": [
+            507,
+            372,
+            884,
+            693
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Autoregressive Search Engines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            705,
+            796,
+            720
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recent works have investigated the use of autoregressive language models to generate identifier strings for documents as an intermediate target for retrieval (Yu et al., 2022), such as Wikipedia page titles (De Cao et al., 2020), root-to-leaf paths in a hierarchical cluster tree (Tay et al., 2022), or distinctive n-grams that can be mapped to full passages (Bevilacqua et al., 2022). Since the series of work was carried out almost simultaneously by different research groups, they are often referred to multiple different names in the literature, such as autoregressive search engine, differential search",
+        "bbox": [
+            507,
+            726,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2667",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/7540f9a910fe0e5c49b5d563c3b8ac3d1edb7d73029dc79bacf73a44116925d3.jpg",
+        "image_caption": [
+            "Figure 2: The overall pipeline of our proposed LLM-URL. Given a question, LLM-URL first generates a set of URLs which are extracted from the generated text. The URLs are retrieved from the Internet then broken into passages which are ranked and filtered such that only the most relevant are kept. Finally, these passages are given as input to a reader model along with the original question to generate a final answer."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            154,
+            84,
+            833,
+            305
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "index (DSI), and neural document indexers (NDI). Compared to traditional dense document retrievers, these methods leverage a generation model to produce the document indexes. By forcing the generation model to explain every token in the question and document using cross-attention, the generation abilities of the model significantly improve. Our work is closely related to these works, showing experimentally that properly prompting pretrained large language models can achieve better performance than traditional dense retrieval models (Ouyang et al., 2022; Yu et al., 2023).",
+        "bbox": [
+            112,
+            401,
+            489,
+            595
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Proposed Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            613,
+            302,
+            630
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section we describe a new method, which we refer to as LLM-URL, that employs a large language model (LLM) to perform effective and efficient web document retrieval for knowledge-intensive NLP tasks such as open-domain question answering (ODQA).",
+        "bbox": [
+            112,
+            643,
+            487,
+            739
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ODQA is a two step process consisting of a retriever and a reader. Given a question  $q$ , the goal of the retriever is to find the top- $n$  passages  $\\mathcal{P}_n$  relevant to answering  $q$ . Given  $q$  and the top- $n$  relevant passages  $\\mathcal{P}_n$ , the goal of the reader is to use internal knowledge along with  $\\mathcal{P}_n$  to generate a correct answer  $a$  to question  $q$ . The passage retriever plays an essential role in this process. When  $\\mathcal{P}_n$  contains more passages that have the correct answer, the reader has a higher chance of finding it. Instead of heavily training a dedicated retriever,",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "our LLM-URL solves the problem in a different way as shown in Figure 2.",
+        "bbox": [
+            507,
+            401,
+            880,
+            432
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a question  $q$ , our LLM-URL should find a set of relevant passages to  $\\mathcal{P}_n$  and give it to the reader. First, it prompts a LLM (e.g., GPT-3) to directly generate  $m$  URLs for  $q$ . By default, it uses \"Which  $m$  Wikipedia URLs would have the answer?\" as the instruction which is appended to each input question as the prompt. We also append the beginning of the Wikipedia URL (https://en.wikipedia.org/wiki) to the end of the prompt to encourage the generation of URLs and restrict generation to the Wikipedia article URL format. As LLM has the ability of in-context learning, we take this advantage to enable the few-shot setting in the prompt. The prompt described above also includes a series of in-context demonstrations. Each demonstration contains a question sampled from the training set following the prompt described above. At the end of each demonstration,  $m$  URLs which point to gold-labeled documents are listed. In the zero-shot setting, the original prompt is used without any demonstrations. In the few-shot setting, the original prompt appended to a series of  $d$  demonstrations ( $d = 10$  in this work).",
+        "bbox": [
+            507,
+            435,
+            884,
+            804
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given the prompt, the LLM returns a generated sequence of tokens. Ideally these tokens would construct a sequence of  $m$  separated URLs. In practice, the generated sequence often has extra information such as a proposed answer that is unreliable and needs to be filtered. We use a regular expression to extract all URLs from the sequence and discard all",
+        "bbox": [
+            507,
+            806,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "2668",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/3260e19d454d230dc81631c9710f061a8102941ef3061fe05ef34c5c54febd48.jpg",
+        "image_caption": [
+            "Figure 3: We prompt LLM-URL to generate  $m$  documents and measure the recall as  $m$  increases. Significant recall improvements are seen when  $m$  is small but as it increases the marginal benefit decreases."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            84,
+            473,
+            231
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "extra information. This also filters out many URLs that are improperly formatted. After extraction, GET requests are made using the extracted URLs and the contents of each retrieval is used to create a set of fetched documents  $\\mathcal{D}_f$ . Often,  $|\\mathcal{D}_f| < m$  because some of the generated URLs do not follow a correct format or do not point to real web pages on the Internet.",
+        "bbox": [
+            112,
+            328,
+            487,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The set of fetched documents  $\\mathcal{D}_f$  can be passed directly to a reader if  $m$  is a small value or the reader being used can handle many large documents. However, this is usually not the case. Often,  $\\mathcal{D}_f$  needs to be filtered such that only a small number of the most relevant passages are given to the reader. To do this, our LLM-URL first breaks each document  $d\\in \\mathcal{D}_f$  into a set of small passages. The passages from each document are collected into a new set,  $\\mathcal{P}_f$ . A scoring function is used to quantify the relevance of each passage with respect to the question  $q$ , with high values indicating high relevance with respect to  $q$  and low scores indicating low relevance. A simple scoring function such as BM25 can be used or a more complex one such as DPR (Karpukhin et al., 2020) can. The passages in  $\\mathcal{P}_f$  are then sorted from highest to lowest and the top  $n$  are kept as  $\\mathcal{P}_n$ . Finally,  $\\mathcal{P}_n$  are given to a reader along with  $q$  to generate an answer.",
+        "bbox": [
+            115,
+            458,
+            489,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Advantages of LLM-URL: Existing autoregressive retrieval methods such as DSI and SEAL use a pre-trained large language model then fine tune it to take questions as input and generate relevant document identifiers as output (Tay et al., 2022; Bevilacqua et al., 2022). Both DSI and SEAL do extensive experiments on a variety of document identifiers which are generated by a heavily trained language model. Examples of these",
+        "bbox": [
+            112,
+            774,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "identifiers include unstructured atomic identifiers, naively structured string identifiers, hierarchical document clustering, and others. LLM-URL instead uses pre-existing document identifiers that exist on the internet: URLs. Using URLs instead of the aforementioned identifiers has multiple advantages. URLs often contain words related to the information they link to, allowing for strong association of topics with their URLs. For example, the title of each Wikipedia page is used in its URL, allowing the LLM is able to directly generate the URL by leveraging semantic information from the question. To validate the importance of URLs themselves, we also experiment with prompting the LLM to generate Wikipedia titles instead of URLs and find Recall@1 significantly reduces compared to prompting for URL generation. We believe this is because the URL format itself helps prompt the model for specific information in a specific format. Further, the use of URLs allows us to simply obtain the evidence document via a HTTP request without any need of training a model or building an index to find the mapping between identifiers and documents.",
+        "bbox": [
+            507,
+            84,
+            884,
+            469
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            482,
+            655,
+            498
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present and discuss results from our experiments to \"directly\" demonstrate that our LLM-URL is a strong retriever and \"indirectly\" show that it achieves competitive performance on the ODQA task against state-of-the-art solutions.",
+        "bbox": [
+            507,
+            507,
+            882,
+            588
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Large Language Model: Following Figure 1, the large language model we use to generate URLs for our experiments is GPT-3 text-davinci-003 with greedy decoding and a temperature of 0. A variety of different prompts are tested for generating URLs, but little difference in performance is observed, so we simply use the best performing prompt which is discussed in Section 3.",
+        "bbox": [
+            507,
+            596,
+            882,
+            725
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Datasets: We use three ODQA datasets including Web Questions, Natural Questions, and Trivia QA. We use them to perform evaluation on both the task of document or passage retrieval and ODQA itself.",
+        "bbox": [
+            507,
+            734,
+            884,
+            799
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Retrieval",
+        "text_level": 1,
+        "bbox": [
+            507,
+            809,
+            628,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We expect retrievers to find the most relevant documents and/or passages. We conduct experiments on both document retrieval and passage retrieval.",
+        "bbox": [
+            507,
+            829,
+            882,
+            878
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics. Recall@k (k=1, 10, 100) is calculated by measuring the percentage of docu",
+        "bbox": [
+            507,
+            887,
+            882,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "2669",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/4a69562903ef4a5794f91d2fd3e1f21bf2f4e2bc3adadfc0dae6c46d814f4fb5.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Document Recall@1</td><td colspan=\"3\">Document Recall@10</td></tr><tr><td>WebQ</td><td>NQ</td><td>TriviaQA</td><td>WebQ</td><td>NQ</td><td>TriviaQA</td></tr><tr><td>Contriever (Izacard et al., 2021)</td><td>63.8</td><td>53.2</td><td>60.6</td><td>63.8</td><td>80.8</td><td>82.5</td></tr><tr><td>BM25 (Robertson and Zaragoza, 2009)</td><td>49.5</td><td>47.2</td><td>63.0</td><td>81.5</td><td>76.8</td><td>82.3</td></tr><tr><td>Google API</td><td>61.1</td><td>55.5</td><td>51.4</td><td>-</td><td>-</td><td>-</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>76.8</td><td>61.7</td><td>71.3</td><td>87.7</td><td>83.2</td><td>85.5</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>79.7</td><td>62.6</td><td>73.5</td><td>89.9</td><td>83.9</td><td>86.8</td></tr></table>",
+        "bbox": [
+            181,
+            80,
+            823,
+            209
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/6ebb802051b4f846d45d534cd8d0febe244272384b36de3884a2b2aa930b4d19.jpg",
+        "table_caption": [
+            "Table 1: Document retrieval as measured by Recall@k. Google API Recall@10 results are left out due to high cost."
+        ],
+        "table_footnote": [
+            "Table 2: Passage retrieval as measured by Recall@1, Recall@10 and Recall@100. Here LLM-URL is equipped with BM25 to perform the ranking task."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Passage Recall@1</td><td colspan=\"3\">Passage Recall@10</td><td colspan=\"3\">Passage Recall@100</td></tr><tr><td>WebQ</td><td>NQ</td><td>TriviaQA</td><td>WebQ</td><td>NQ</td><td>TriviaQA</td><td>WebQ</td><td>NQ</td><td>TriviaQA</td></tr><tr><td>Contriever</td><td>18.2</td><td>18.8</td><td>34.0</td><td>55.7</td><td>54.8</td><td>67.9</td><td>79.8</td><td>79.6</td><td>83.3</td></tr><tr><td>BM25</td><td>19.1</td><td>22.8</td><td>46.2</td><td>51.8</td><td>55.6</td><td>71.7</td><td>76.6</td><td>79.6</td><td>83.9</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>22.2</td><td>24.0</td><td>46.7</td><td>63.1</td><td>60.6</td><td>76.6</td><td>83.8</td><td>78.3</td><td>83.6</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>22.3</td><td>25.5</td><td>49.1</td><td>64.8</td><td>60.8</td><td>77.8</td><td>85.9</td><td>79.0</td><td>84.8</td></tr></table>",
+        "bbox": [
+            115,
+            246,
+            878,
+            357
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ments or passages in the top- $k$  which contain one of the gold labeled answers while exact match is calculated by the percentage of predicted answers which match one of the gold labeled answers. While LLM-URL is not constrained by which URLs can be generated for document retrieval, we restrict all generations to Wikipedia URLs only for fair comparison, as discussed in Section 3 All baseline models also use Wikipedia for retrieval, with some fetching documents in real time and others fetching from an offline corpus.",
+        "bbox": [
+            112,
+            411,
+            485,
+            587
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1.1 Document Retrieval",
+        "text_level": 1,
+        "bbox": [
+            112,
+            604,
+            332,
+            618
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines: Contriever (Izacard et al., 2021) and BM25 (Robertson and Zaragoza, 2009) are usually used for passage retrieval. Contriever is a dual encoder which uses a dot product between dense representations of a question and passage to calculate relevance. BM25 is a sparse retriever which uses the overlapping contents between question and passage to calculate relevance. Because we use the same passage size to chunk Wikipedia documents, we were able to map their retrieved passages back to the original documents. We use Google API (Brin and Page, 1998) restricted to Wikipedia as a third baseline to retrieve relevant documents given a question.",
+        "bbox": [
+            112,
+            627,
+            487,
+            851
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Existing works such as DSI and SEAL have investigated the use of autoregressive language models to generate identifier strings for documents as an intermediate target for retrieval. DSI is a Trans",
+        "bbox": [
+            112,
+            854,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7674bfd1c6b411da4e763a3ede02f9db0683419879860ff49aebaeb7006e9990.jpg",
+        "image_caption": [
+            "Figure 4: The percentage of valid URLs generated from LLM-URL as the total number of generated URLs  $m$  increases from 1 to 10. As  $m$  increases, invalid URL generations become more frequent."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            412,
+            870,
+            558
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "former which has been trained to map directly from question to document identifiers by memorizing the contents of the entire corpus (Tay et al., 2022). SEAL is a variant of DSI which uses ngrams as document ids to improve retrieval performance (Bevilacqua et al., 2022). Neither DSI nor SEAL report retrieval results on full documents and do not have publicly available implementations, so they are left out and discussed in Table 3 and Section 4.1.2 on passage retrieval.",
+        "bbox": [
+            507,
+            664,
+            882,
+            825
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Unlike the baselines, our LLM-URL employs an LLM. It has two settings: zero-shot and few-shot. In the zero-shot setting, no in-context demonstrations are given whereas in the few-shot setting a few demonstrations are appended to the prompt.",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "2670",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/dc97a564f7dc83439b03795ebc5fe8f3f0f34f8699cc5c647fee9015678ce0ca.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "explicitly trained for retrieval on  $NQ$",
+            "Table 3: Passage retrieval as measured by Recall@1 and Recall@10. LLM-URL is equipped with BM25 for passage ranking. Other datasets are left out due to not being reported in either paper and no public implementations."
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Recall@1</td><td>Recall@10</td></tr><tr><td>DSI1</td><td>25.1</td><td>56.6</td></tr><tr><td>SEAL1</td><td>26.3</td><td>74.5</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>24.0</td><td>60.6</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>25.5</td><td>60.8</td></tr></table>",
+        "bbox": [
+            124,
+            80,
+            485,
+            179
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results: The results of our document retrieval experiments are shown in Table 1. In this setting Recall@  $k$  is calculated directly after the documents are retrieved with no intermediary steps. LLM-URL significantly outperforms baseline methods on all datasets for both Recall@ 1 and Recall@ 10. Specifically, zero-shot LLM-URL improves document Recall@ 1 relatively by  $20.4\\%$ ,  $11.2\\%$ , and  $13.2\\%$  over the strongest baseline on WebQ, NQ, and TriviaQA, respectively. Few-shot LLM-URL further expands the improvement to  $24.9\\%$ ,  $12.8\\%$ , and  $16.7\\%$ , respectively. URLs can be extracted from the large-scale parameters of LLMs, and these URLs can lead to more accurate documents than what existing methods can retrieve. Both the LLM parameters and in-context demonstrations are significantly useful in document retrieval.",
+        "bbox": [
+            112,
+            294,
+            489,
+            567
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 shows Recall scores converge when the number of generated URLs  $m$  increases. Due to the diminishing returns from increasing  $m$ , our experiments do not explore values of  $m$  greater than 10.",
+        "bbox": [
+            112,
+            569,
+            487,
+            646
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Are the generated URLs valid? It is worth noting that the generated URLs are not always valid. Some generated URLs do not have valid URL syntax and some point to Wikipedia pages that do not exist. Rarely, URLs will be generated for domains aside from Wikipedia. For fair comparison, all of these faulty URLs are discarded and only documents coming from valid Wikipedia articles are kept.",
+        "bbox": [
+            112,
+            659,
+            489,
+            804
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Further analysis is done to measure the ratio of valid Wikipedia URLs while the total number of generated URLs  $m$  increases from 1 to 10, shown in Figure 4. The number of valid URL generations remains surprisingly high (i.e., higher than  $68\\%$ ) as  $m$  increases from 1 to 10. However, the rate of valid generations appears to fall off as  $m$  increases,",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/31601af330d3f7c1f35705b5a989f262e83c77ff83673346e88dba2319f07059.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "explicitly trained for retrieval on  $NQ$",
+            "${}^{2}$  result from Bevilacqua et al. (2022)"
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Zero-Shot QA EM</td></tr><tr><td>WebQ</td><td>NQ</td><td>TriviaQA</td></tr><tr><td>Contriever + InstructGPT</td><td>16.8</td><td>19.1</td><td>52.4</td></tr><tr><td>BM25 + InstructGPT</td><td>16.0</td><td>20.5</td><td>53.3</td></tr><tr><td>Google + InstructGPT</td><td>19.9</td><td>27.8</td><td>58.7</td></tr><tr><td>GenRead (InstructGPT)</td><td>24.8</td><td>28.2</td><td>59.3</td></tr><tr><td>DSI1 + FiD</td><td>-</td><td>31.42</td><td>-</td></tr><tr><td>SEAL1 + FiD</td><td>-</td><td>43.6</td><td>41.8</td></tr><tr><td>InstructGPT (no docs.)</td><td>18.6</td><td>20.9</td><td>52.6</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>28.1</td><td>26.4</td><td>60.1</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>29.0</td><td>27.3</td><td>60.7</td></tr></table>",
+        "bbox": [
+            524,
+            82,
+            878,
+            273
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Zero-shot open-domain QA performance as measured by exact match (EM). All LLM-URL models use InstructGPT as the reader unless otherwise stated.",
+        "bbox": [
+            507,
+            315,
+            882,
+            357
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "indicating there are diminishing returns from each marginal increase of  $m$ .",
+        "bbox": [
+            507,
+            376,
+            882,
+            407
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1.2 Passage Retrieval",
+        "text_level": 1,
+        "bbox": [
+            507,
+            419,
+            707,
+            434
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines: Four methods, including Contriever, BM25, DSI (Tay et al., 2022), and SEAL (Bevilacqua et al., 2022), were introduced in Section 4.1.1. Google API was used for document retrieval and not applied to passages.",
+        "bbox": [
+            507,
+            439,
+            884,
+            521
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results: The results of our passage retrieval experiments are shown in Table 2. In this setting Recall@k is calculated on the top-k passages ranked by the ranker instead of just on the raw documents shown in Table 1. LLM-URL performs slightly better than baseline methods for Recall@1 and Recall@10 and as well as baseline methods for Recall@100. In the zero-shot setting, LLM-URL improves relative Recall@1 by  $16.2\\%$ ,  $5.3\\%$ , and  $1.1\\%$  with respect to the strongest baseline on WebQ, NQ, and TriviaA respectively. The few-shot setting of LLM-URL expands the improvement to  $16.8\\%$ ,  $11.8\\%$ , and  $6.3\\%$ , respectively. For Recall@10, similar improvements can be seen.",
+        "bbox": [
+            507,
+            532,
+            882,
+            756
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For Recall@100, performance is better relative to baseline models for all datasets except NQ. In the zero-shot setting, LLM-URL improves the relative Recall@100 by  $5.0\\%$  for WebQ and performs slightly worse than the best baseline method on NQ and TriviaQA by  $1.7\\%$  and  $0.4\\%$  respectively. The few-shot setting of LLM-URL for Recall@100 shows a slight improvement on WebQ and TriviaQA, but performs slightly worse than the strongest baseline on NQ.",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "2671",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Despite being limited to only the passages from 10 documents, LLM-URL performs better than baseline methods for smaller  $k$  and performs as well as baseline methods for higher values of  $k$ .",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The comparison between LLM-URL and existing document identifier-based methods such as DSI and SEAL are shown in Table 3. For Recall@1, zero-shot LLM-URL performs slightly worse than the best baseline by  $8.8\\%$ . This performance gap is slightly smaller in the few-shot setting with LLM-URL performing  $3.1\\%$  worse than the best baseline. For Recall@10, zero-shot LLM-URL performs worse than the best baseline by  $18.7\\%$ . Few-shot LLM-URL performs only slightly better than the zero-shot setting, performing worse than the best baseline by  $18.4\\%$ .",
+        "bbox": [
+            112,
+            149,
+            489,
+            341
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 Open-Domain Question Answering",
+        "text_level": 1,
+        "bbox": [
+            112,
+            355,
+            436,
+            370
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metric: We use exact match (EM), which is short for exact string match with the correct answer, because the goal of ODQA is to find an exact answer to any question using Wikipedia articles.",
+        "bbox": [
+            112,
+            376,
+            489,
+            455
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results: Here we discuss the downstream QA performance of LLM-URL. In this setting, an answer only has an exact match if the normalized generated text is within the list of acceptable answers to a question. When combined with InstructGPT as a reader, LLM-URL performs significantly better on WebQ and slightly better on TriviaQA when compared with the best performing baseline methods. On NQ, LLM-URL+InstructGPT performs worse than baseline NDIs and only slightly worse than the best remaining baseline. In the zero-shot setting, LLM-URL+InstructGPT improves upon the best baseline method by  $13.3\\%$  and  $1.3\\%$  on WebQ and TriviaQA respectively. LLM-URL +InstructGPT performs worse than the best baseline method by  $39.5\\%$  on NQ. In the few-shot setting, LLM-URL+InstructGPT performs better than the best baseline method by  $16.9\\%$  and  $2.3\\%$  on WebQ and TriviaQA respectively. LLM-URL+InstructGPT performs worse than the best baseline method by  $37.4\\%$  on NQ.",
+        "bbox": [
+            112,
+            467,
+            489,
+            804
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Despite not being explicitly trained for retrieval, LLM-URL+InstructGPT performs significantly better than baseline methods for WebQ, achieves on-par performance with existing methods for TriviaQA, and performs slightly worse than existing methods for NQ.",
+        "bbox": [
+            112,
+            806,
+            489,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our results indicate LLM-URL could be a",
+        "bbox": [
+            131,
+            903,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b7887a0892fe8a7cfc6e5afdb3b4f2e5d2dbe3e70cf41412cdf9808e30303ebd.jpg",
+        "image_caption": [
+            "Figure 5: Common vs Uncommon entity recall@1. Common is defined as a question containing entities in the top-1 million most common entities on Wikipedia. LLM-URL performs much better when retrieving information about common entities."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            84,
+            870,
+            214
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "promising solution to retrieval for a wide range of knowledge intensive tasks with little to no training data required.",
+        "bbox": [
+            507,
+            306,
+            882,
+            354
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Discussions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            365,
+            648,
+            379
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3.1 Time Sensitive Queries",
+        "text_level": 1,
+        "bbox": [
+            507,
+            387,
+            749,
+            401
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "There are a number of additional qualitative benefits that LLM-URL has over existing methods. One large advantage of LLM-URL is that the documents are retrieved in real time from the source. So long as the source stays up to date without the URL itself changing, our proposed method is capable of answering time sensitive queries without any extra modifications.",
+        "bbox": [
+            507,
+            406,
+            884,
+            533
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In contrast, existing dual encoder approaches such as Contriever require a document to be re-encoded each time it changes. Existing methods such as SEAL (Bevilacqua et al., 2022) and DSI are also tricky to keep up to date for time sensitive queries as the LLM would have to be retrained to learn the new content of the updated documents.",
+        "bbox": [
+            507,
+            536,
+            882,
+            646
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3.2 Frequent Entities Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            658,
+            778,
+            673
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Following (Mallen et al., 2022), we analyze the retrieval performance of LLM-URL when the gold-labeled answer entity is common versus when it is not. For each question-answer pair in a given dataset we check to see if the labeled entity exists within the top one-million common entities from Wikipedia. Using this, we split our dataset into two distinct subsets: question-answer pairs that contain a common entity and those that do not. In measuring the performance of our model on these two distinct sets across Web Questions, Natural Questions, and TriviaQA, we find LLM-URL performs significantly better on common entity question-answer pairs. The results of our analysis are shown in Figure 5. Across all three datasets,",
+        "bbox": [
+            507,
+            677,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "2672",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/31bc4b87f02156d0d296bdb4d0491b88cd475b94928d87eb2707bf870c56de4e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>LLM-URL</td><td>Exists</td><td>Answer</td><td>Contriever</td><td>Answer</td><td>BM25</td><td>Answer</td></tr><tr><td>wiki/Jellyfish</td><td>✓</td><td>✓</td><td>Smack (ship)</td><td>✘</td><td>Collective noun</td><td>✘</td></tr><tr><td>wiki/Collective_noun</td><td>✓</td><td>✘</td><td>Collective noun</td><td>✘</td><td>Determiner</td><td>✘</td></tr><tr><td>wiki/Smack_(group)</td><td>✘</td><td>✘</td><td>Cetacean intelligence</td><td>✘</td><td>Glass sea creatures</td><td>✘</td></tr><tr><td>wiki/Cnidaria</td><td>✓</td><td>✓</td><td>Well smack</td><td>✘</td><td>Minotaur</td><td>✘</td></tr><tr><td>wiki/Medusozoa</td><td>✓</td><td>✓</td><td>Plankton</td><td>✘</td><td>Mass noun</td><td>✘</td></tr><tr><td>wiki/Scyphozoa</td><td>✓</td><td>✓</td><td>Sperm whale</td><td>✘</td><td>Well smack</td><td>✘</td></tr><tr><td>wiki/Cubozoa</td><td>✓</td><td>✓</td><td>Loaded question</td><td>✘</td><td>Nomenclature</td><td>✘</td></tr><tr><td>wiki/Hydrozoa</td><td>✓</td><td>✓</td><td>Jabberwocky</td><td>✘</td><td>Archomental</td><td>✘</td></tr><tr><td>wiki/Staurozoa</td><td>✓</td><td>✓</td><td>Merrow</td><td>✘</td><td>Honey Smacks</td><td>✘</td></tr><tr><td>wiki/Rhizostomeae</td><td>✓</td><td>✓</td><td>Loaded question</td><td>✘</td><td>Well smack</td><td>✘</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            884,
+            266
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Case study of retrieved documents from the question \"A 'smack' is a collective noun for a group of which sea creatures?\". \"Exists\" means whether the URL points to a valid Wiki page. \"Answer\" means whether the document contains the answer. We omit the prefix of generated URLs for brevity (https://en.wikipedia.org/). For BM25 and Contriever, we show the document titles of the top-10 retrieved passages, respectively. The correct answer is \"jellyfish.\"",
+        "bbox": [
+            112,
+            275,
+            884,
+            348
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the recall of common-entity question-answer pairs is many times greater than the recall from the rest of the dataset.",
+        "bbox": [
+            112,
+            372,
+            487,
+            419
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Previous work has shown LLMs in the closed-book setting, where the model must rely solely on the information contained within its weights, perform much better on common-entities versus uncommon ones (Mallen et al., 2022). Our results show this problem extends beyond the closed-book setting and also applies to retrieval when using LLM-URL. This also could explain the high word count from documents we found when evaluating LLM-URL. The average Wikipedia article is 644 words, but the average word count from Wikipedia documents retrieved via LLM-URL was  $10\\mathrm{k}$ . We believe this discrepancy is caused by common entities having much more detail in their Wikipedia articles and in turn having much higher word count.",
+        "bbox": [
+            112,
+            420,
+            489,
+            663
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3.3 Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            673,
+            263,
+            688
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In Table 5, we show a case study comparing LLMURL with two baseline retrieval methods, BM25 and Contriever, on the question \"A 'smack' is a collective noun for a group of which sea creatures?\" which is in the TriviaQA test set. The gold-labeled answer to this question is \"jellyfish\".",
+        "bbox": [
+            112,
+            693,
+            489,
+            788
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the closed-book setting, InstructGPT mistakenly predicts \"dolphins\" as the answer. When using Contriever to retrieve 10 passages from Wikipedia given the query, none of the passages contains the gold answer. For instance, Contriever retrieves passages about \"smack\", a kind of fishing vessel, along with other passages about sperm whales, plankton, and other unrelated topics. Similar results are",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "found while using BM25 as the retriever.",
+        "bbox": [
+            507,
+            372,
+            815,
+            387
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In contrast, LLM-URL performs much better in this scenario, retrieving 7 documents which contain the answer. The top retrieved document is exactly about the gold answer \"jellyfish\". The fourth to the tenth documents all talk about different types of jellyfish. After being chunked into passages then sorted by the ranker, the top 10 passages are concatenated. Among them, it contains \"A group of jellyfish is called a smack,\" which contains the answer to the question and comes directly from the first retrieved document, titled \"jellyfish.\" When InstructGPT is then prompted with these 10 passages along with the question, the gold answer \"jellyfish\" is correctly generated.",
+        "bbox": [
+            505,
+            395,
+            884,
+            621
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This case study highlights multiple advantages of LLM-URL. First, LLM-URL finds documents related to both the question and the answer. It directly locates documents that talks about \"jellyfish\" instead while BM25 and Contriever locate documents related to the question only—not the answer. Second, LLM-URL is more precise than BM25 or Contriever. In this case, 7 out of 10 generated URLs from LLM-URL point to a Wikipedia document that contains the answer. However, both BM25 and Contriever fail to retrieve any documents containing the answer. Third, the set of documents retrieved by LLM-URL are complementary to each other, while in BM25 or contriever, each document in the top-10 is selected independently. This is because the LLM is able to refer to previous generated URLs before it generates the next one, allowing each newly generated URL to be condi",
+        "bbox": [
+            507,
+            629,
+            885,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "2673",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tioned on all the previous URLs. This leads to a more informative evidence context in open-domain question answering.",
+        "bbox": [
+            112,
+            84,
+            487,
+            134
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            145,
+            401,
+            160
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we explored whether large language models can generate URLs prompted by human instructions for document retrieval. Surprisingly, we found that by providing a few (query, URL) pairs as in-context demonstrations, large language models (e.g. GPT-3) generated Web URLs where near  $90\\%$  of the corresponding documents contain correct answers to open-domain questions in WebQ. Furthermore, by breaking the retrieved documents into passages then ranking them with BM25, we showed a significant number of unnecessary passages could be filtered out while retaining high recall, which outperformed baseline methods by a significant margin.",
+        "bbox": [
+            112,
+            171,
+            489,
+            397
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "There are numerous exciting directions for future work. While a number of broad spectrum retrieval benchmarks such as BIER (Thakur et al., 2021) exist, it remains to be seen whether the few-shot demonstrations shown in this work can be further tuned for specific retrieval tasks. Promptagator (Dai et al., 2022) shows significant performance improvements can be achieved by tuning prompts in a similar way.",
+        "bbox": [
+            112,
+            397,
+            489,
+            541
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Further, it remains to be seen whether fine tuning the prompt for each individual question can further improve the retrieval performance. As with Promptagator, prior work has shown using clustering to select diverse demonstrations for any given question further improves retrieval performance as well as downstream QA performance.",
+        "bbox": [
+            112,
+            542,
+            489,
+            653
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            668,
+            220,
+            682
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Despite the strong performance on the presented datasets, our approach is limited in its ability to update knowledge state and adapt to new domains. A major feature of retrieve-then-read is the ability to swap in new documents when new information is learned, such as temporally more recent documents, or adding in documents from a new domain to quickly adapt to a new downstream task. Our approach relies on a large language model to contain all this knowledge and adding new knowledge would likely require some retraining. In addition, large generation models still suffer from hallucination errors, resulting in incorrect predictions. When tasked with generating 10 URLs, LLM-URL may",
+        "bbox": [
+            112,
+            694,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "only generate 6 or 7 which link to valid documents. Finally, our approach involves very large language models, slow web requests, and document processing which may make it cumbersome to use in practice.",
+        "bbox": [
+            507,
+            84,
+            885,
+            162
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            178,
+            684,
+            193
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by NSF IIS-2119531, IIS-2137396, IIS-2142827, CCF-1901059, and ONR N00014-22-1-2507. Wenhao Yu was partly supported by the Bloomberg Data Science Fellowship.",
+        "bbox": [
+            507,
+            204,
+            885,
+            269
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            297,
+            608,
+            312
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on freebase from question-answer pairs. In EMNLP, pages 1533-1544.",
+            "Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen-tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. arXiv preprint arXiv:2204.10628.",
+            "Sergey Brin and Lawrence Page. 1998. The anatomy of a large-scale hypertextual web search engine. In Proceedings of the Seventh International Conference on World Wide Web 7, WWW7, page 107-117, NLD. Elsevier Science Publishers B. V.",
+            "Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. In Procs. of ACL.",
+            "Zhuyun Dai, Vincent Y. Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2022. Promptagator: Few-shot dense retrieval from 8 examples. arXiv preprint arXiv:2209.11755.",
+            "Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2020. Autoregressive entity retrieval. In International Conference on Learning Representations.",
+            "Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning.",
+            "Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In ACL, pages 1601-1611.",
+            "Mingxuan Ju, Wenhao Yu, Tong Zhao, Chuxu Zhang, and Yanfang Ye. 2022. Grape: Knowledge graph enhanced passage reader for open-domain question answering. arXiv preprint arXiv:2210.02933."
+        ],
+        "bbox": [
+            509,
+            321,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "2674",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).",
+            "Omar Khattab, Christopher Potts, and Matei Zaharia. 2021. Relevance-guided supervision for openqa with colbert. Transactions of the Association for Computational Linguistics, 9:929-944.",
+            "Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 39-48.",
+            "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: A benchmark for question answering research. TACL, pages 452-466.",
+            "Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Hannaneh Hajishirzi, and Daniel Khashabi. 2022. When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories. arXiv preprint ArXiv:2212.10511.",
+            "Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Reader-guided passage reranking for open-domain question answering. In Findings of ACLIJCNLP.",
+            "Yasumasa Onoe, Michael J. Q. Zhang, Eunsol Choi, and Greg Durrett. 2021. CREAM: A dataset for commonsense reasoning over entity knowledge. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS.",
+            "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155.",
+            "Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, et al. 2021. Kilt: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523-2544.",
+            "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. Rocketqa: An optimized training approach to dense passage retrieval for open-domain question answering. In Procs. of NAACL."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            906
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389.",
+            "Yi Tay, Vinh Q Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. 2022. Transformer memory as a differentiable search index. arXiv preprint arXiv:2202.06991.",
+            "Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).",
+            "Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Hao Sun, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Sun, Weiwei Deng, Qi Zhang, and Mao Yang. 2022. A neural corpus indexer for document retrieval. In Advances in Neural Information Processing Systems.",
+            "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations.",
+            "Wei Yang, Yuqing Xie, Aileen Lin, Xingyu Li, Luchen Tan, Kun Xiong, Ming Li, and Jimmy Lin. 2019. End-to-end open-domain question answering with bertserini. In NAACL 2019 (demo).",
+            "Donghan Yu, Chenguang Zhu, Yuwei Fang, Wenhao Yu, Shuohang Wang, Yichong Xu, Xiang Ren, Yiming Yang, and Michael Zeng. 2021. Kg-fid: Infusing knowledge graph in fusion-in-decoder for open-domain question answering. arXiv preprint arXiv:2110.04330.",
+            "Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. Generate rather than retrieve: Large language models are strong context generators. International Conference for Learning Representation (ICLR).",
+            "Wenhao Yu, Chenguang Zhu, Zaitang Li, Zhiting Hu, Qingyun Wang, Heng Ji, and Meng Jiang. 2022. A survey of knowledge-enhanced text generation. ACM Computing Surveys (CSUR).",
+            "Zhihan Zhang, Xiubo Geng, Tao Qin, Yunfang Wu, and Daxin Jiang. 2021. Knowledge-aware procedural text understanding with multi-stage training. In WWW '21: The Web Conference 2021."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            858
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "2675",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhihan Zhang, Wenhao Yu, Zheng Ning, Mingxuan Ju, and Meng Jiang. 2023a. Exploring contrast consistency of open-domain question answering systems on minimally edited questions. Trans. Assoc. Comput. Linguistics.",
+            "Zhihan Zhang, Wenhao Yu, Mengxia Yu, Zhichun Guo, and Meng Jiang. 2023b. A survey of multi-task learning in natural language processing: Regarding task relatedness and training methods. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2023.",
+            "Zhihan Zhang, Wenhao Yu, Chenguang Zhu, and Meng Jiang. 2022. A unified encoder-decoder framework"
+        ],
+        "bbox": [
+            114,
+            85,
+            489,
+            284
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "with entity memory. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022.",
+            "Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2022. Dense text retrieval based on pretrained language models: A survey. arXiv preprint arXiv:2211.14876.",
+            "Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng Chua. 2021. Retrieving and reading: A comprehensive survey on open-domain question answering. arXiv preprint arXiv:2101.00774."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            263
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "2676",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            129,
+            126,
+            532,
+            143
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Our limitations section is not numbered, but it is the last section of our paper.",
+        "bbox": [
+            149,
+            143,
+            722,
+            159
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            129,
+            170,
+            552,
+            186
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We do not believe any of the risks mentioned in the checklist apply to this paper.",
+        "bbox": [
+            149,
+            187,
+            737,
+            202
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            129,
+            212,
+            695,
+            228
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Our introduction is found in section one and the abstract comes before that.",
+        "bbox": [
+            149,
+            230,
+            709,
+            244
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            129,
+            256,
+            668,
+            272
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            215,
+            336
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            127,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            363,
+            248,
+            379
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            127,
+            390,
+            778,
+            406
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            407,
+            248,
+            422
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            127,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            498,
+            248,
+            513
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            127,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            574,
+            248,
+            588
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            127,
+            599,
+            880,
+            631
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            633,
+            248,
+            646
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            127,
+            658,
+            880,
+            739
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            740,
+            248,
+            753
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            494,
+            781
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Section 4 describes our experiments",
+        "bbox": [
+            132,
+            785,
+            401,
+            802
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            812,
+            880,
+            844
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Yes these details are found in a graph in Section 1 and hyperparameters are found in our experiments section (4).",
+        "bbox": [
+            149,
+            846,
+            880,
+            876
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            884,
+            877,
+            908
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "2677",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            129,
+            83,
+            878,
+            115
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Yes we discuss our hyperparameters in section 4",
+        "bbox": [
+            149,
+            117,
+            510,
+            131
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            127,
+            143,
+            882,
+            190
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Our method is deterministic so there is no variation in our results.",
+        "bbox": [
+            149,
+            192,
+            638,
+            206
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            127,
+            218,
+            882,
+            265
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            267,
+            349,
+            282
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            127,
+            340,
+            882,
+            372
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            374,
+            248,
+            388
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            127,
+            399,
+            882,
+            447
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            449,
+            248,
+            464
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            127,
+            475,
+            882,
+            521
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            524,
+            248,
+            539
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            127,
+            549,
+            873,
+            565
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            567,
+            248,
+            582
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            127,
+            592,
+            880,
+            623
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            626,
+            248,
+            640
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "2678",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Large Language Models are Built-in Autoregressive Search Engines/5ad13cc6-938f-4faa-8252-ed3a3053bd31_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ec8b2c95b16ecf9636afc2f8d5349e92eb27f4f57dd09e20f6d633736d1640f
+size 351872

2023/Large Language Models are Built-in Autoregressive Search Engines/full.md

@@ -0,0 +1,343 @@
+# Large Language Models are Built-in Autoregressive Search Engines
+
+Noah Ziems, Wenhao Yu, Zhihan Zhang, Meng Jiang
+
+University of Notre Dame
+
+{nziems2, wyu1, zzhang23, mjiang2}@nd.edu
+
+# Abstract
+
+Document retrieval is a key stage of standard Web search engines. Existing dual-encoder dense retrievers obtain representations for questions and documents independently, allowing for only shallow interactions between them. To overcome this limitation, recent autoregressive search engines replace the dual-encoder architecture by directly generating identifiers for relevant documents in the candidate pool. However, the training cost of such autoregressive search engines rises sharply as the number of candidate documents increases. In this paper, we find that large language models (LLMs) can follow human instructions to directly generate URLs for document retrieval.
+
+Surprisingly, when providing a few Query-URL pairs as in-context demonstrations, LLMs can generate Web URLs where nearly  $90\%$  of the corresponding documents contain correct answers to open-domain questions. In this way, LLMs can be thought of as built-in search engines, since they have not been explicitly trained to map questions to document identifiers. Experiments demonstrate that our method can consistently achieve better retrieval performance than existing retrieval approaches by a significant margin on three open-domain question answering benchmarks, under both zero and few-shot settings. The code for this work can be found at https://github.com/Ziems/11murl.
+
+# 1 Introduction
+
+Along with the success of deep learning, dual-encoder based retrievers have become the dominant method for Web searching (Zhu et al., 2021; Zhao et al., 2022). For example, DPR (Karpukhin et al., 2020) employs two independent encoders to encode the question and the document respectively, then estimates their relevance by computing a single similarity score between two representations.
+
+![](images/b725b6eeb0bb4c8fec22b43fb9d59a66810c6de179073bf69496d6b7d510085c.jpg)  
+Figure 1: Successful URL reconstructions by different size of GPT-3 as URL generators. The models are prompted with the first 100 words of a Wikipedia page then tasked with generating the URL of the page they came from. Tested on 10k Wikipedia pages sampled from the top 100k most frequent Wikipedia entities.
+
+However, these methods suffer from two major drawbacks. First, the representations of questions and documents are typically obtained independently in modern dual-encoder dense retrieval models (Karpukhin et al., 2020), allowing for only shallow interactions between them (Khattab et al., 2021). Second, the question or document representation is embedded into a single dense vector, potentially missing fine-grained information when computing the similarity between the two vector representations (Khattab and Zaharia, 2020).
+
+Instead of computing similarity between question and document embeddings, autoregressive search engines aim to directly generate document identifiers then map them to complete documents in the predetermined candidate pool. This approach has attracted increasing interest in information retrieval (IR) and related fields (Tay et al., 2022; Bevilacqua et al., 2022; Wang et al., 2022). Compared to dual-encoder dense retrieval methods, autoregressive search engines enjoy a number of advantages. First, autoregressive generation models produce document identifiers by performing deep token-level cross-attention, resulting in a better esti
+
+mation than shallow interactions in dense retrievers. Second, autoregressive search engines have been shown to have strong generalization abilities, outperforming BM25 in a zero-shot setting (Tay et al., 2022). While it is theoretically possible to scale an autoregressive search engine to the size of a large language model (LLM), such as GPT-3 with 175B parameters, in practice it is not feasible due to the computational overhead of training such a large autoregressive search engine from scratch (Tay et al., 2022). To reduce the high training cost of autoregressive search engine, a smaller model size is preferred. However, the results of our pilot study in Figure 1 show smaller language models are significantly worse at mapping passages to document identifiers than larger ones. Moreover, different retrieval tasks can have unique retrieval requirements. One task may require a model to retrieve factual evidence to support or refute a claim (i.e., fact checking) (Onoe et al., 2021) while another may require a model to retrieve specific trivia information about an entity (i.e., entity linking) (Petroni et al., 2021; Zhang et al., 2022). It would be better if the retriever was capable of generalizing to new retrieval tasks with only a few examples.
+
+In this work, we explore the use of in-context demonstrations to prompt LLMs to directly generate web URLs for document retrieval, namely LLM-URL. Surprisingly, we find that by providing a few (query, URL) pairs as contextual demonstrations, large language models (e.g. GPT-3) generate Web URLs where nearly  $90\%$  of the corresponding documents contain answers to open-domain questions. In this way, LLMs can be thought of as built-in search engines, as they have not been explicitly trained to map questions or documents to identifiers. Instead of using newly-created document identifiers, LLM-URL leverages existing and widely used document identifiers directly, i.e., URLs. We compare our approach to existing document retrieval methods on three different open-domain question answering (QA) datasets: WebQ (Berant et al., 2013), NQ (Kwiatkowski et al., 2019), and TriviaQA (Joshi et al., 2017). Further, to avoid exceeding the limit on the number of input tokens of LLMs, we employ an unsupervised passage filtering module to remove irrelevant portions of supporting documents. To summarize, our main contributions are as follows:
+
+1. We reveal that LLMs are built-in autoregressive search engines capable of document re
+
+trieval by directly generating Web page URLs under both zero and few-shot settings.
+
+2. We show retrieving documents by generating URLs with LLMs significantly outperforms existing methods for document retrieval, as measured by Recall@K. Further, we show that breaking the retrieved documents into passages then using a ranker to filter the passages significantly reduces the number of supporting passages while maintaining high recall.  
+3. We show the retrieved documents improve downstream QA performance as measured by EM when compared to baseline methods.
+
+# 2 Related Work
+
+# 2.1 Traditional Document Retrievers
+
+Traditional methods such as TF-IDF and BM25 explore sparse retrieval strategies by matching the overlapping contents between questions and passages (Robertson and Zaragoza, 2009; Chen et al., 2017; Yang et al., 2019). DPR (Karpukhin et al., 2020) revolutionized the field by utilizing dense contextualized vectors for passage indexing. It is first initialized as a pretrained BERT model, then trained discriminatively using pairs of queries and relevant documents, with hard negatives from BM25. Recent research has improved DPR via better training strategies (Xiong et al., 2020; Qu et al., 2021; Zhang et al., 2023a) and passage reranking (Mao et al., 2021; Yu et al., 2021; Ju et al., 2022). However, representations of questions and documents are typically obtained independently in modern dual-encoder dense retrieval models (Karpukhin et al., 2020; Xiong et al., 2020), allowing for only shallow interactions between them (Khattab et al., 2021).
+
+# 2.2 Autoregressive Search Engines
+
+Recent works have investigated the use of autoregressive language models to generate identifier strings for documents as an intermediate target for retrieval (Yu et al., 2022), such as Wikipedia page titles (De Cao et al., 2020), root-to-leaf paths in a hierarchical cluster tree (Tay et al., 2022), or distinctive n-grams that can be mapped to full passages (Bevilacqua et al., 2022). Since the series of work was carried out almost simultaneously by different research groups, they are often referred to multiple different names in the literature, such as autoregressive search engine, differential search
+
+![](images/7540f9a910fe0e5c49b5d563c3b8ac3d1edb7d73029dc79bacf73a44116925d3.jpg)  
+Figure 2: The overall pipeline of our proposed LLM-URL. Given a question, LLM-URL first generates a set of URLs which are extracted from the generated text. The URLs are retrieved from the Internet then broken into passages which are ranked and filtered such that only the most relevant are kept. Finally, these passages are given as input to a reader model along with the original question to generate a final answer.
+
+index (DSI), and neural document indexers (NDI). Compared to traditional dense document retrievers, these methods leverage a generation model to produce the document indexes. By forcing the generation model to explain every token in the question and document using cross-attention, the generation abilities of the model significantly improve. Our work is closely related to these works, showing experimentally that properly prompting pretrained large language models can achieve better performance than traditional dense retrieval models (Ouyang et al., 2022; Yu et al., 2023).
+
+# 3 Proposed Method
+
+In this section we describe a new method, which we refer to as LLM-URL, that employs a large language model (LLM) to perform effective and efficient web document retrieval for knowledge-intensive NLP tasks such as open-domain question answering (ODQA).
+
+ODQA is a two step process consisting of a retriever and a reader. Given a question  $q$ , the goal of the retriever is to find the top- $n$  passages  $\mathcal{P}_n$  relevant to answering  $q$ . Given  $q$  and the top- $n$  relevant passages  $\mathcal{P}_n$ , the goal of the reader is to use internal knowledge along with  $\mathcal{P}_n$  to generate a correct answer  $a$  to question  $q$ . The passage retriever plays an essential role in this process. When  $\mathcal{P}_n$  contains more passages that have the correct answer, the reader has a higher chance of finding it. Instead of heavily training a dedicated retriever,
+
+our LLM-URL solves the problem in a different way as shown in Figure 2.
+
+Given a question  $q$ , our LLM-URL should find a set of relevant passages to  $\mathcal{P}_n$  and give it to the reader. First, it prompts a LLM (e.g., GPT-3) to directly generate  $m$  URLs for  $q$ . By default, it uses "Which  $m$  Wikipedia URLs would have the answer?" as the instruction which is appended to each input question as the prompt. We also append the beginning of the Wikipedia URL (https://en.wikipedia.org/wiki) to the end of the prompt to encourage the generation of URLs and restrict generation to the Wikipedia article URL format. As LLM has the ability of in-context learning, we take this advantage to enable the few-shot setting in the prompt. The prompt described above also includes a series of in-context demonstrations. Each demonstration contains a question sampled from the training set following the prompt described above. At the end of each demonstration,  $m$  URLs which point to gold-labeled documents are listed. In the zero-shot setting, the original prompt is used without any demonstrations. In the few-shot setting, the original prompt appended to a series of  $d$  demonstrations ( $d = 10$  in this work).
+
+Given the prompt, the LLM returns a generated sequence of tokens. Ideally these tokens would construct a sequence of  $m$  separated URLs. In practice, the generated sequence often has extra information such as a proposed answer that is unreliable and needs to be filtered. We use a regular expression to extract all URLs from the sequence and discard all
+
+![](images/3260e19d454d230dc81631c9710f061a8102941ef3061fe05ef34c5c54febd48.jpg)  
+Figure 3: We prompt LLM-URL to generate  $m$  documents and measure the recall as  $m$  increases. Significant recall improvements are seen when  $m$  is small but as it increases the marginal benefit decreases.
+
+extra information. This also filters out many URLs that are improperly formatted. After extraction, GET requests are made using the extracted URLs and the contents of each retrieval is used to create a set of fetched documents  $\mathcal{D}_f$ . Often,  $|\mathcal{D}_f| < m$  because some of the generated URLs do not follow a correct format or do not point to real web pages on the Internet.
+
+The set of fetched documents  $\mathcal{D}_f$  can be passed directly to a reader if  $m$  is a small value or the reader being used can handle many large documents. However, this is usually not the case. Often,  $\mathcal{D}_f$  needs to be filtered such that only a small number of the most relevant passages are given to the reader. To do this, our LLM-URL first breaks each document  $d\in \mathcal{D}_f$  into a set of small passages. The passages from each document are collected into a new set,  $\mathcal{P}_f$ . A scoring function is used to quantify the relevance of each passage with respect to the question  $q$ , with high values indicating high relevance with respect to  $q$  and low scores indicating low relevance. A simple scoring function such as BM25 can be used or a more complex one such as DPR (Karpukhin et al., 2020) can. The passages in  $\mathcal{P}_f$  are then sorted from highest to lowest and the top  $n$  are kept as  $\mathcal{P}_n$ . Finally,  $\mathcal{P}_n$  are given to a reader along with  $q$  to generate an answer.
+
+Advantages of LLM-URL: Existing autoregressive retrieval methods such as DSI and SEAL use a pre-trained large language model then fine tune it to take questions as input and generate relevant document identifiers as output (Tay et al., 2022; Bevilacqua et al., 2022). Both DSI and SEAL do extensive experiments on a variety of document identifiers which are generated by a heavily trained language model. Examples of these
+
+identifiers include unstructured atomic identifiers, naively structured string identifiers, hierarchical document clustering, and others. LLM-URL instead uses pre-existing document identifiers that exist on the internet: URLs. Using URLs instead of the aforementioned identifiers has multiple advantages. URLs often contain words related to the information they link to, allowing for strong association of topics with their URLs. For example, the title of each Wikipedia page is used in its URL, allowing the LLM is able to directly generate the URL by leveraging semantic information from the question. To validate the importance of URLs themselves, we also experiment with prompting the LLM to generate Wikipedia titles instead of URLs and find Recall@1 significantly reduces compared to prompting for URL generation. We believe this is because the URL format itself helps prompt the model for specific information in a specific format. Further, the use of URLs allows us to simply obtain the evidence document via a HTTP request without any need of training a model or building an index to find the mapping between identifiers and documents.
+
+# 4 Experiments
+
+In this section, we present and discuss results from our experiments to "directly" demonstrate that our LLM-URL is a strong retriever and "indirectly" show that it achieves competitive performance on the ODQA task against state-of-the-art solutions.
+
+Large Language Model: Following Figure 1, the large language model we use to generate URLs for our experiments is GPT-3 text-davinci-003 with greedy decoding and a temperature of 0. A variety of different prompts are tested for generating URLs, but little difference in performance is observed, so we simply use the best performing prompt which is discussed in Section 3.
+
+Datasets: We use three ODQA datasets including Web Questions, Natural Questions, and Trivia QA. We use them to perform evaluation on both the task of document or passage retrieval and ODQA itself.
+
+# 4.1 Retrieval
+
+We expect retrievers to find the most relevant documents and/or passages. We conduct experiments on both document retrieval and passage retrieval.
+
+Evaluation metrics. Recall@k (k=1, 10, 100) is calculated by measuring the percentage of docu
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Document Recall@1</td><td colspan="3">Document Recall@10</td></tr><tr><td>WebQ</td><td>NQ</td><td>TriviaQA</td><td>WebQ</td><td>NQ</td><td>TriviaQA</td></tr><tr><td>Contriever (Izacard et al., 2021)</td><td>63.8</td><td>53.2</td><td>60.6</td><td>63.8</td><td>80.8</td><td>82.5</td></tr><tr><td>BM25 (Robertson and Zaragoza, 2009)</td><td>49.5</td><td>47.2</td><td>63.0</td><td>81.5</td><td>76.8</td><td>82.3</td></tr><tr><td>Google API</td><td>61.1</td><td>55.5</td><td>51.4</td><td>-</td><td>-</td><td>-</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>76.8</td><td>61.7</td><td>71.3</td><td>87.7</td><td>83.2</td><td>85.5</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>79.7</td><td>62.6</td><td>73.5</td><td>89.9</td><td>83.9</td><td>86.8</td></tr></table>
+
+Table 1: Document retrieval as measured by Recall@k. Google API Recall@10 results are left out due to high cost.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Passage Recall@1</td><td colspan="3">Passage Recall@10</td><td colspan="3">Passage Recall@100</td></tr><tr><td>WebQ</td><td>NQ</td><td>TriviaQA</td><td>WebQ</td><td>NQ</td><td>TriviaQA</td><td>WebQ</td><td>NQ</td><td>TriviaQA</td></tr><tr><td>Contriever</td><td>18.2</td><td>18.8</td><td>34.0</td><td>55.7</td><td>54.8</td><td>67.9</td><td>79.8</td><td>79.6</td><td>83.3</td></tr><tr><td>BM25</td><td>19.1</td><td>22.8</td><td>46.2</td><td>51.8</td><td>55.6</td><td>71.7</td><td>76.6</td><td>79.6</td><td>83.9</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>22.2</td><td>24.0</td><td>46.7</td><td>63.1</td><td>60.6</td><td>76.6</td><td>83.8</td><td>78.3</td><td>83.6</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>22.3</td><td>25.5</td><td>49.1</td><td>64.8</td><td>60.8</td><td>77.8</td><td>85.9</td><td>79.0</td><td>84.8</td></tr></table>
+
+Table 2: Passage retrieval as measured by Recall@1, Recall@10 and Recall@100. Here LLM-URL is equipped with BM25 to perform the ranking task.
+
+ments or passages in the top- $k$  which contain one of the gold labeled answers while exact match is calculated by the percentage of predicted answers which match one of the gold labeled answers. While LLM-URL is not constrained by which URLs can be generated for document retrieval, we restrict all generations to Wikipedia URLs only for fair comparison, as discussed in Section 3 All baseline models also use Wikipedia for retrieval, with some fetching documents in real time and others fetching from an offline corpus.
+
+# 4.1.1 Document Retrieval
+
+Baselines: Contriever (Izacard et al., 2021) and BM25 (Robertson and Zaragoza, 2009) are usually used for passage retrieval. Contriever is a dual encoder which uses a dot product between dense representations of a question and passage to calculate relevance. BM25 is a sparse retriever which uses the overlapping contents between question and passage to calculate relevance. Because we use the same passage size to chunk Wikipedia documents, we were able to map their retrieved passages back to the original documents. We use Google API (Brin and Page, 1998) restricted to Wikipedia as a third baseline to retrieve relevant documents given a question.
+
+Existing works such as DSI and SEAL have investigated the use of autoregressive language models to generate identifier strings for documents as an intermediate target for retrieval. DSI is a Trans
+
+![](images/7674bfd1c6b411da4e763a3ede02f9db0683419879860ff49aebaeb7006e9990.jpg)  
+Figure 4: The percentage of valid URLs generated from LLM-URL as the total number of generated URLs  $m$  increases from 1 to 10. As  $m$  increases, invalid URL generations become more frequent.
+
+former which has been trained to map directly from question to document identifiers by memorizing the contents of the entire corpus (Tay et al., 2022). SEAL is a variant of DSI which uses ngrams as document ids to improve retrieval performance (Bevilacqua et al., 2022). Neither DSI nor SEAL report retrieval results on full documents and do not have publicly available implementations, so they are left out and discussed in Table 3 and Section 4.1.2 on passage retrieval.
+
+Unlike the baselines, our LLM-URL employs an LLM. It has two settings: zero-shot and few-shot. In the zero-shot setting, no in-context demonstrations are given whereas in the few-shot setting a few demonstrations are appended to the prompt.
+
+<table><tr><td>Method</td><td>Recall@1</td><td>Recall@10</td></tr><tr><td>DSI1</td><td>25.1</td><td>56.6</td></tr><tr><td>SEAL1</td><td>26.3</td><td>74.5</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>24.0</td><td>60.6</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>25.5</td><td>60.8</td></tr></table>
+
+explicitly trained for retrieval on  $NQ$  
+Table 3: Passage retrieval as measured by Recall@1 and Recall@10. LLM-URL is equipped with BM25 for passage ranking. Other datasets are left out due to not being reported in either paper and no public implementations.
+
+Results: The results of our document retrieval experiments are shown in Table 1. In this setting Recall@  $k$  is calculated directly after the documents are retrieved with no intermediary steps. LLM-URL significantly outperforms baseline methods on all datasets for both Recall@ 1 and Recall@ 10. Specifically, zero-shot LLM-URL improves document Recall@ 1 relatively by  $20.4\%$ ,  $11.2\%$ , and  $13.2\%$  over the strongest baseline on WebQ, NQ, and TriviaQA, respectively. Few-shot LLM-URL further expands the improvement to  $24.9\%$ ,  $12.8\%$ , and  $16.7\%$ , respectively. URLs can be extracted from the large-scale parameters of LLMs, and these URLs can lead to more accurate documents than what existing methods can retrieve. Both the LLM parameters and in-context demonstrations are significantly useful in document retrieval.
+
+Figure 3 shows Recall scores converge when the number of generated URLs  $m$  increases. Due to the diminishing returns from increasing  $m$ , our experiments do not explore values of  $m$  greater than 10.
+
+Are the generated URLs valid? It is worth noting that the generated URLs are not always valid. Some generated URLs do not have valid URL syntax and some point to Wikipedia pages that do not exist. Rarely, URLs will be generated for domains aside from Wikipedia. For fair comparison, all of these faulty URLs are discarded and only documents coming from valid Wikipedia articles are kept.
+
+Further analysis is done to measure the ratio of valid Wikipedia URLs while the total number of generated URLs  $m$  increases from 1 to 10, shown in Figure 4. The number of valid URL generations remains surprisingly high (i.e., higher than  $68\%$ ) as  $m$  increases from 1 to 10. However, the rate of valid generations appears to fall off as  $m$  increases,
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Zero-Shot QA EM</td></tr><tr><td>WebQ</td><td>NQ</td><td>TriviaQA</td></tr><tr><td>Contriever + InstructGPT</td><td>16.8</td><td>19.1</td><td>52.4</td></tr><tr><td>BM25 + InstructGPT</td><td>16.0</td><td>20.5</td><td>53.3</td></tr><tr><td>Google + InstructGPT</td><td>19.9</td><td>27.8</td><td>58.7</td></tr><tr><td>GenRead (InstructGPT)</td><td>24.8</td><td>28.2</td><td>59.3</td></tr><tr><td>DSI1 + FiD</td><td>-</td><td>31.42</td><td>-</td></tr><tr><td>SEAL1 + FiD</td><td>-</td><td>43.6</td><td>41.8</td></tr><tr><td>InstructGPT (no docs.)</td><td>18.6</td><td>20.9</td><td>52.6</td></tr><tr><td>LLM-URL (Zero-Shot)</td><td>28.1</td><td>26.4</td><td>60.1</td></tr><tr><td>LLM-URL (Few-Shot)</td><td>29.0</td><td>27.3</td><td>60.7</td></tr></table>
+
+explicitly trained for retrieval on  $NQ$  
+${}^{2}$  result from Bevilacqua et al. (2022)
+
+Table 4: Zero-shot open-domain QA performance as measured by exact match (EM). All LLM-URL models use InstructGPT as the reader unless otherwise stated.
+
+indicating there are diminishing returns from each marginal increase of  $m$ .
+
+# 4.1.2 Passage Retrieval
+
+Baselines: Four methods, including Contriever, BM25, DSI (Tay et al., 2022), and SEAL (Bevilacqua et al., 2022), were introduced in Section 4.1.1. Google API was used for document retrieval and not applied to passages.
+
+Results: The results of our passage retrieval experiments are shown in Table 2. In this setting Recall@k is calculated on the top-k passages ranked by the ranker instead of just on the raw documents shown in Table 1. LLM-URL performs slightly better than baseline methods for Recall@1 and Recall@10 and as well as baseline methods for Recall@100. In the zero-shot setting, LLM-URL improves relative Recall@1 by  $16.2\%$ ,  $5.3\%$ , and  $1.1\%$  with respect to the strongest baseline on WebQ, NQ, and TriviaA respectively. The few-shot setting of LLM-URL expands the improvement to  $16.8\%$ ,  $11.8\%$ , and  $6.3\%$ , respectively. For Recall@10, similar improvements can be seen.
+
+For Recall@100, performance is better relative to baseline models for all datasets except NQ. In the zero-shot setting, LLM-URL improves the relative Recall@100 by  $5.0\%$  for WebQ and performs slightly worse than the best baseline method on NQ and TriviaQA by  $1.7\%$  and  $0.4\%$  respectively. The few-shot setting of LLM-URL for Recall@100 shows a slight improvement on WebQ and TriviaQA, but performs slightly worse than the strongest baseline on NQ.
+
+Despite being limited to only the passages from 10 documents, LLM-URL performs better than baseline methods for smaller  $k$  and performs as well as baseline methods for higher values of  $k$ .
+
+The comparison between LLM-URL and existing document identifier-based methods such as DSI and SEAL are shown in Table 3. For Recall@1, zero-shot LLM-URL performs slightly worse than the best baseline by  $8.8\%$ . This performance gap is slightly smaller in the few-shot setting with LLM-URL performing  $3.1\%$  worse than the best baseline. For Recall@10, zero-shot LLM-URL performs worse than the best baseline by  $18.7\%$ . Few-shot LLM-URL performs only slightly better than the zero-shot setting, performing worse than the best baseline by  $18.4\%$ .
+
+# 4.2 Open-Domain Question Answering
+
+Evaluation metric: We use exact match (EM), which is short for exact string match with the correct answer, because the goal of ODQA is to find an exact answer to any question using Wikipedia articles.
+
+Results: Here we discuss the downstream QA performance of LLM-URL. In this setting, an answer only has an exact match if the normalized generated text is within the list of acceptable answers to a question. When combined with InstructGPT as a reader, LLM-URL performs significantly better on WebQ and slightly better on TriviaQA when compared with the best performing baseline methods. On NQ, LLM-URL+InstructGPT performs worse than baseline NDIs and only slightly worse than the best remaining baseline. In the zero-shot setting, LLM-URL+InstructGPT improves upon the best baseline method by  $13.3\%$  and  $1.3\%$  on WebQ and TriviaQA respectively. LLM-URL +InstructGPT performs worse than the best baseline method by  $39.5\%$  on NQ. In the few-shot setting, LLM-URL+InstructGPT performs better than the best baseline method by  $16.9\%$  and  $2.3\%$  on WebQ and TriviaQA respectively. LLM-URL+InstructGPT performs worse than the best baseline method by  $37.4\%$  on NQ.
+
+Despite not being explicitly trained for retrieval, LLM-URL+InstructGPT performs significantly better than baseline methods for WebQ, achieves on-par performance with existing methods for TriviaQA, and performs slightly worse than existing methods for NQ.
+
+Our results indicate LLM-URL could be a
+
+![](images/b7887a0892fe8a7cfc6e5afdb3b4f2e5d2dbe3e70cf41412cdf9808e30303ebd.jpg)  
+Figure 5: Common vs Uncommon entity recall@1. Common is defined as a question containing entities in the top-1 million most common entities on Wikipedia. LLM-URL performs much better when retrieving information about common entities.
+
+promising solution to retrieval for a wide range of knowledge intensive tasks with little to no training data required.
+
+# 4.3 Discussions
+
+# 4.3.1 Time Sensitive Queries
+
+There are a number of additional qualitative benefits that LLM-URL has over existing methods. One large advantage of LLM-URL is that the documents are retrieved in real time from the source. So long as the source stays up to date without the URL itself changing, our proposed method is capable of answering time sensitive queries without any extra modifications.
+
+In contrast, existing dual encoder approaches such as Contriever require a document to be re-encoded each time it changes. Existing methods such as SEAL (Bevilacqua et al., 2022) and DSI are also tricky to keep up to date for time sensitive queries as the LLM would have to be retrained to learn the new content of the updated documents.
+
+# 4.3.2 Frequent Entities Analysis
+
+Following (Mallen et al., 2022), we analyze the retrieval performance of LLM-URL when the gold-labeled answer entity is common versus when it is not. For each question-answer pair in a given dataset we check to see if the labeled entity exists within the top one-million common entities from Wikipedia. Using this, we split our dataset into two distinct subsets: question-answer pairs that contain a common entity and those that do not. In measuring the performance of our model on these two distinct sets across Web Questions, Natural Questions, and TriviaQA, we find LLM-URL performs significantly better on common entity question-answer pairs. The results of our analysis are shown in Figure 5. Across all three datasets,
+
+<table><tr><td>LLM-URL</td><td>Exists</td><td>Answer</td><td>Contriever</td><td>Answer</td><td>BM25</td><td>Answer</td></tr><tr><td>wiki/Jellyfish</td><td>✓</td><td>✓</td><td>Smack (ship)</td><td>✘</td><td>Collective noun</td><td>✘</td></tr><tr><td>wiki/Collective_noun</td><td>✓</td><td>✘</td><td>Collective noun</td><td>✘</td><td>Determiner</td><td>✘</td></tr><tr><td>wiki/Smack_(group)</td><td>✘</td><td>✘</td><td>Cetacean intelligence</td><td>✘</td><td>Glass sea creatures</td><td>✘</td></tr><tr><td>wiki/Cnidaria</td><td>✓</td><td>✓</td><td>Well smack</td><td>✘</td><td>Minotaur</td><td>✘</td></tr><tr><td>wiki/Medusozoa</td><td>✓</td><td>✓</td><td>Plankton</td><td>✘</td><td>Mass noun</td><td>✘</td></tr><tr><td>wiki/Scyphozoa</td><td>✓</td><td>✓</td><td>Sperm whale</td><td>✘</td><td>Well smack</td><td>✘</td></tr><tr><td>wiki/Cubozoa</td><td>✓</td><td>✓</td><td>Loaded question</td><td>✘</td><td>Nomenclature</td><td>✘</td></tr><tr><td>wiki/Hydrozoa</td><td>✓</td><td>✓</td><td>Jabberwocky</td><td>✘</td><td>Archomental</td><td>✘</td></tr><tr><td>wiki/Staurozoa</td><td>✓</td><td>✓</td><td>Merrow</td><td>✘</td><td>Honey Smacks</td><td>✘</td></tr><tr><td>wiki/Rhizostomeae</td><td>✓</td><td>✓</td><td>Loaded question</td><td>✘</td><td>Well smack</td><td>✘</td></tr></table>
+
+Table 5: Case study of retrieved documents from the question "A 'smack' is a collective noun for a group of which sea creatures?". "Exists" means whether the URL points to a valid Wiki page. "Answer" means whether the document contains the answer. We omit the prefix of generated URLs for brevity (https://en.wikipedia.org/). For BM25 and Contriever, we show the document titles of the top-10 retrieved passages, respectively. The correct answer is "jellyfish."
+
+the recall of common-entity question-answer pairs is many times greater than the recall from the rest of the dataset.
+
+Previous work has shown LLMs in the closed-book setting, where the model must rely solely on the information contained within its weights, perform much better on common-entities versus uncommon ones (Mallen et al., 2022). Our results show this problem extends beyond the closed-book setting and also applies to retrieval when using LLM-URL. This also could explain the high word count from documents we found when evaluating LLM-URL. The average Wikipedia article is 644 words, but the average word count from Wikipedia documents retrieved via LLM-URL was  $10\mathrm{k}$ . We believe this discrepancy is caused by common entities having much more detail in their Wikipedia articles and in turn having much higher word count.
+
+# 4.3.3 Case Study
+
+In Table 5, we show a case study comparing LLMURL with two baseline retrieval methods, BM25 and Contriever, on the question "A 'smack' is a collective noun for a group of which sea creatures?" which is in the TriviaQA test set. The gold-labeled answer to this question is "jellyfish".
+
+In the closed-book setting, InstructGPT mistakenly predicts "dolphins" as the answer. When using Contriever to retrieve 10 passages from Wikipedia given the query, none of the passages contains the gold answer. For instance, Contriever retrieves passages about "smack", a kind of fishing vessel, along with other passages about sperm whales, plankton, and other unrelated topics. Similar results are
+
+found while using BM25 as the retriever.
+
+In contrast, LLM-URL performs much better in this scenario, retrieving 7 documents which contain the answer. The top retrieved document is exactly about the gold answer "jellyfish". The fourth to the tenth documents all talk about different types of jellyfish. After being chunked into passages then sorted by the ranker, the top 10 passages are concatenated. Among them, it contains "A group of jellyfish is called a smack," which contains the answer to the question and comes directly from the first retrieved document, titled "jellyfish." When InstructGPT is then prompted with these 10 passages along with the question, the gold answer "jellyfish" is correctly generated.
+
+This case study highlights multiple advantages of LLM-URL. First, LLM-URL finds documents related to both the question and the answer. It directly locates documents that talks about "jellyfish" instead while BM25 and Contriever locate documents related to the question only—not the answer. Second, LLM-URL is more precise than BM25 or Contriever. In this case, 7 out of 10 generated URLs from LLM-URL point to a Wikipedia document that contains the answer. However, both BM25 and Contriever fail to retrieve any documents containing the answer. Third, the set of documents retrieved by LLM-URL are complementary to each other, while in BM25 or contriever, each document in the top-10 is selected independently. This is because the LLM is able to refer to previous generated URLs before it generates the next one, allowing each newly generated URL to be condi
+
+tioned on all the previous URLs. This leads to a more informative evidence context in open-domain question answering.
+
+# 5 Conclusion and Future Work
+
+In this paper, we explored whether large language models can generate URLs prompted by human instructions for document retrieval. Surprisingly, we found that by providing a few (query, URL) pairs as in-context demonstrations, large language models (e.g. GPT-3) generated Web URLs where near  $90\%$  of the corresponding documents contain correct answers to open-domain questions in WebQ. Furthermore, by breaking the retrieved documents into passages then ranking them with BM25, we showed a significant number of unnecessary passages could be filtered out while retaining high recall, which outperformed baseline methods by a significant margin.
+
+There are numerous exciting directions for future work. While a number of broad spectrum retrieval benchmarks such as BIER (Thakur et al., 2021) exist, it remains to be seen whether the few-shot demonstrations shown in this work can be further tuned for specific retrieval tasks. Promptagator (Dai et al., 2022) shows significant performance improvements can be achieved by tuning prompts in a similar way.
+
+Further, it remains to be seen whether fine tuning the prompt for each individual question can further improve the retrieval performance. As with Promptagator, prior work has shown using clustering to select diverse demonstrations for any given question further improves retrieval performance as well as downstream QA performance.
+
+# Limitations
+
+Despite the strong performance on the presented datasets, our approach is limited in its ability to update knowledge state and adapt to new domains. A major feature of retrieve-then-read is the ability to swap in new documents when new information is learned, such as temporally more recent documents, or adding in documents from a new domain to quickly adapt to a new downstream task. Our approach relies on a large language model to contain all this knowledge and adding new knowledge would likely require some retraining. In addition, large generation models still suffer from hallucination errors, resulting in incorrect predictions. When tasked with generating 10 URLs, LLM-URL may
+
+only generate 6 or 7 which link to valid documents. Finally, our approach involves very large language models, slow web requests, and document processing which may make it cumbersome to use in practice.
+
+# Acknowledgements
+
+This work was supported by NSF IIS-2119531, IIS-2137396, IIS-2142827, CCF-1901059, and ONR N00014-22-1-2507. Wenhao Yu was partly supported by the Bloomberg Data Science Fellowship.
+
+# References
+
+Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on freebase from question-answer pairs. In EMNLP, pages 1533-1544.  
+Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen-tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. arXiv preprint arXiv:2204.10628.  
+Sergey Brin and Lawrence Page. 1998. The anatomy of a large-scale hypertextual web search engine. In Proceedings of the Seventh International Conference on World Wide Web 7, WWW7, page 107-117, NLD. Elsevier Science Publishers B. V.  
+Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. In Procs. of ACL.  
+Zhuyun Dai, Vincent Y. Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2022. Promptagator: Few-shot dense retrieval from 8 examples. arXiv preprint arXiv:2209.11755.  
+Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2020. Autoregressive entity retrieval. In International Conference on Learning Representations.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning.  
+Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In ACL, pages 1601-1611.  
+Mingxuan Ju, Wenhao Yu, Tong Zhao, Chuxu Zhang, and Yanfang Ye. 2022. Grape: Knowledge graph enhanced passage reader for open-domain question answering. arXiv preprint arXiv:2210.02933.
+
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Omar Khattab, Christopher Potts, and Matei Zaharia. 2021. Relevance-guided supervision for openqa with colbert. Transactions of the Association for Computational Linguistics, 9:929-944.  
+Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 39-48.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: A benchmark for question answering research. TACL, pages 452-466.  
+Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Hannaneh Hajishirzi, and Daniel Khashabi. 2022. When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories. arXiv preprint ArXiv:2212.10511.  
+Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Reader-guided passage reranking for open-domain question answering. In Findings of ACLIJCNLP.  
+Yasumasa Onoe, Michael J. Q. Zhang, Eunsol Choi, and Greg Durrett. 2021. CREAM: A dataset for commonsense reasoning over entity knowledge. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155.  
+Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, et al. 2021. Kilt: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523-2544.  
+Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. Rocketqa: An optimized training approach to dense passage retrieval for open-domain question answering. In Procs. of NAACL.
+
+Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389.  
+Yi Tay, Vinh Q Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. 2022. Transformer memory as a differentiable search index. arXiv preprint arXiv:2202.06991.  
+Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).  
+Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Hao Sun, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Sun, Weiwei Deng, Qi Zhang, and Mao Yang. 2022. A neural corpus indexer for document retrieval. In Advances in Neural Information Processing Systems.  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations.  
+Wei Yang, Yuqing Xie, Aileen Lin, Xingyu Li, Luchen Tan, Kun Xiong, Ming Li, and Jimmy Lin. 2019. End-to-end open-domain question answering with bertserini. In NAACL 2019 (demo).  
+Donghan Yu, Chenguang Zhu, Yuwei Fang, Wenhao Yu, Shuohang Wang, Yichong Xu, Xiang Ren, Yiming Yang, and Michael Zeng. 2021. Kg-fid: Infusing knowledge graph in fusion-in-decoder for open-domain question answering. arXiv preprint arXiv:2110.04330.  
+Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. Generate rather than retrieve: Large language models are strong context generators. International Conference for Learning Representation (ICLR).  
+Wenhao Yu, Chenguang Zhu, Zaitang Li, Zhiting Hu, Qingyun Wang, Heng Ji, and Meng Jiang. 2022. A survey of knowledge-enhanced text generation. ACM Computing Surveys (CSUR).  
+Zhihan Zhang, Xiubo Geng, Tao Qin, Yunfang Wu, and Daxin Jiang. 2021. Knowledge-aware procedural text understanding with multi-stage training. In WWW '21: The Web Conference 2021.
+
+Zhihan Zhang, Wenhao Yu, Zheng Ning, Mingxuan Ju, and Meng Jiang. 2023a. Exploring contrast consistency of open-domain question answering systems on minimally edited questions. Trans. Assoc. Comput. Linguistics.  
+Zhihan Zhang, Wenhao Yu, Mengxia Yu, Zhichun Guo, and Meng Jiang. 2023b. A survey of multi-task learning in natural language processing: Regarding task relatedness and training methods. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2023.  
+Zhihan Zhang, Wenhao Yu, Chenguang Zhu, and Meng Jiang. 2022. A unified encoder-decoder framework
+
+with entity memory. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022.  
+Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2022. Dense text retrieval based on pretrained language models: A survey. arXiv preprint arXiv:2211.14876.  
+Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng Chua. 2021. Retrieving and reading: A comprehensive survey on open-domain question answering. arXiv preprint arXiv:2101.00774.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Our limitations section is not numbered, but it is the last section of our paper.
+
+A2. Did you discuss any potential risks of your work?
+
+We do not believe any of the risks mentioned in the checklist apply to this paper.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Our introduction is found in section one and the abstract comes before that.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+No response.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+No response.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+No response.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+No response.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+No response.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+No response.
+
+C Did you run computational experiments?
+
+Section 4 describes our experiments
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Yes these details are found in a graph in Section 1 and hyperparameters are found in our experiments section (4).
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Yes we discuss our hyperparameters in section 4
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+Our method is deterministic so there is no variation in our results.
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+Not applicable. Left blank.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/Large Language Models are Built-in Autoregressive Search Engines/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4fcdc332f07870b0d3d6c6eada8eb5449e86819e80db4f81f77598e66ce3e3d
+size 385480

2023/Large Language Models with Controllable Working Memory/df593dd7-917d-44ac-ab6c-8a0670ece26b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:784e24e70ccbe02d25c68c4dfece4e0be6a6cccd68c998da38e698764d9737c0
+size 341439

2023/Large Language Models with Controllable Working Memory/full.md

@@ -0,0 +1,486 @@
+# Large Language Models with Controllable Working Memory
+
+Daliang Li\*, Ankit Singh Rawat\*, Manzil Zaheer\*, Xin Wang\*, Michal Lukasik\*, Andreas Veit\*, Felix Yu\*, Sanjiv Kumar\*
+
+Google Research New York  $\hat{\mathbb{O}}$  Google DeepMind New York {daliangli, ankitsrawat, manzilzaheer} @google.com {wanxin, mlukasik, aveit, felixyu, sanjivk} @google.com
+
+# Abstract
+
+Large language models (LLMs) have led to a series of breakthroughs in natural language processing (NLP), partly owing to the massive amounts of world knowledge they memorize during pretraining. While many downstream applications provide the model with an informational context to aid its underlying task, how the model's world knowledge interacts with the factual information presented in the context remains under explored. As a desirable behavior, an LLM should give precedence to the context whenever it contains task-relevant information that conflicts with the model's memorized knowledge. This enables model predictions to be grounded in the context, which then facilitates updating specific model predictions without frequently retraining the model. By contrast, when the context is irrelevant to the task, the model should ignore it and fall back on its internal knowledge. In this paper, we undertake a first joint study of the aforementioned two properties, namely controllability and robustness, in the context of LLMs. We demonstrate that state-of-the-art T5 and PaLM models (both pretrained and finetuned) could exhibit low controllability and robustness that does not improve with increasing the model size. As a solution, we propose a simple yet effective method - knowledge aware finetuning (KAFT) - to strengthen both controllability and robustness by injecting counterfactual and irrelevant contexts to standard supervised datasets. Our comprehensive evaluation showcases the utility of KAFT across model architectures and sizes.
+
+# 1 Introduction
+
+Large language models (LLMs) pretrained on large scale datasets have shown promising results across natural language tasks (Vaswani et al., 2017; Devlin et al., 2019; Raffel et al., 2020a; Brown et al., 2020; Rae et al., 2021; Chowdhery et al., 2022; Smith et al., 2022). However, as models scale ever larger,
+
+they become more expensive to train, making it unrealistic to frequently update model parameters. On the other hand, many real world applications often necessitate adjusting model behavior. This dilemma is especially sharp in the case of factual (world) knowledge that plays important role in realizing impressive performance of LLMs. It is well known that LLMs memorize large amounts of factual knowledge in their parameters (Petroni et al., 2019; Roberts et al., 2020; Geva et al., 2021), which could potentially be out-dated or incorrect. Even for moderate-size models, it is prohibitively expensive to retrain every time an update happens or a mistake is uncovered. Even if resources are ample, it is difficult to ensure that the modification of model parameters do not affect unrelated skills or knowledge.
+
+In human cognition, working memory (George A. Miller, 1960) provides the biological brain with the ability to hold information temporarily to perform tasks such as conversation, reasoning, and mathematics in a way that is adaptive to the ever-changing environment. As shown both experimentally and theoretically (Fuster, 1973; Ashby et al., 2005), working memory is stored in sustained activations of neurons, as opposed to the long term memory which is stored in weights. Working memory is also the immediate information buffer that is accessed while performing conscious tasks. In particular, it is where the fusion of perceptual inputs and long term memory happens (Fukuda and Woodman, 2017). This suggests that a potential method to solve LLMs' pointwise knowledge update and correction problem is to control the working memory stored in activations, rather than editing the long term memory stored in the model weights.
+
+As demonstrated by their powerful in-context few shot learning abilities (Brown et al., 2020), LLM could utilize different activation patterns resulting from different contexts during inference to solve a diverse set of tasks without any changes in
+
+<table><tr><td></td><td>Controllability</td><td>Robustness</td></tr><tr><td>Question</td><td>Dave Gilmour and Roger Waters were in which rock group?</td><td>How has British art survived in Normandy?</td></tr><tr><td>Context</td><td>George Roger Waters (born 6 September 1943) is an English singer, ... Later that year, he re-united with The Rolling Stones bandmates Mason, Wright and David Gilmour...</td><td>In Britain, Norman art primarily survives as stonework or metalwork, such as capitals and baptismal fonts...</td></tr><tr><td>KAFT (ours)</td><td>The Rolling Stones (from context).</td><td>In museums (irrelevant context).</td></tr><tr><td>Noisy FT</td><td>Pink Floyd</td><td>stonework or metalwork</td></tr><tr><td>UQA V2 11B</td><td>Pink Floyd</td><td>stonework or metalwork, such as capitals and baptismal fonts</td></tr><tr><td>Pretrained</td><td>Pink Floyd</td><td>As stonework and metalwork, such as capi-tals and baptismal fonts</td></tr></table>
+
+Table 1: Examples of model outputs demonstrating that, in contrast with baselines, a model obtained by KAFT is characterized by both improved controllability by a context that contradicts its parametric knowledge, and improved robustness against an irrelevant context, compared to baseline methods. Here, Pretrained refers to a T5 XXL model (Raffel et al., 2020b), which is also the underlying model for KAFT and Noisy Finetuning (FT). UQA V2 11B (Khashabi et al., 2022) is based on the T5 11B model.
+
+the weights. It is natural to expect that the same would be true with factual knowledge. In particular, one could prepare a large list of natural language statements covering desired knowledge updates and corrections. At inference time, one can provide the relevant statements as context along with the input and hope that the model would perform the task based on the new knowledge presented in this context. Thus, if the model's working memory is indeed controllable by context, then a single model with static long term memory can produce different results based on a varying set of facts available in different contexts. However, we demonstrate that this approach may fall short for existing LLMs as they have great tendencies to ignore the context and stick to their own parametric knowledge – the world knowledge stored in its model parameters. This raises a natural question:
+
+Is it possible to design a mechanism to ensure that the context can reliably influence the model's working memory?
+
+Note that any such mechanism has to take into account the possibility of encountering noisy contexts. For example, any retrieval system that selects the task-relevant context from a large collection of contexts will be imperfect and occasionally provide irrelevant context. In such cases, it's desirable that the model prediction does not get swayed by an irrelevant context. Interestingly, we show that the standard pretraining and finetuning methods do not ensure this behavior either. In fact, we demonstrate
+
+that it's the noise encountered during the training that often leads to the model ignoring the context.
+
+In this work, we provide an affirmative answer to the aforementioned question and propose a novel approach – knowledge-aware finetuning (KAFT) – to make an LLM's working memory controllable via relevant context while being robust against irrelevant context. Towards this, we aim to ensure that the model utilizes different types of information at its disposal in the following order:
+
+# relevant context
+
+$>$  model's parametric knowledge (1)  
+$>$  irrelevant context, (2)
+
+where  $a > b$  indicates that  $a$  is prioritized over  $b$ . Thus, if the model decides that the context is relevant, it should ground its output in the context, ensuring the controllability of its working memory by the context. This is crucial when the context is in conflict with the model's parametric knowledge. On the other hand, when the context is irrelevant, the model should instead stick to its parametric knowledge; thus ensuring robustness of its working memory against noise.
+
+Our contributions. We develop first LLMs that utilize different knowledge sources with a predefined order of priorities. Along the way, we develop a systematic understanding of the working memories of LLMs and identify their shortcomings. Our key contributions are summarized below.
+
+<table><tr><td></td><td>Robustness</td><td>Controllability</td></tr><tr><td>Standard (noisy) finetuning</td><td>✗</td><td>✗</td></tr><tr><td>Counterfactual finetuning 
+(Longpre et al., 2021)</td><td>✗</td><td>✓</td></tr><tr><td>KAFT (our work)</td><td>✓</td><td>✓</td></tr></table>
+
+Table 2: Summary of our contributions.
+
+1. We undertake a systematic joint study of both controllability and robustness of the working memory of LLMs. Focusing on question answering (QA) tasks, we define the context-question relevance based on whether the context entails an answer to the question. We create a novel benchmark to measure the controllability by including contexts that imply an answer which contradicts the model's pretrained knowledge.<sup>1</sup> Similarly, we benchmark robustness against irrelevant contexts. We conduct an extensive evaluation of LLMs with different sizes across multiple architectures (encoder-decoder and decoder-only) from T5 (Raffel et al., 2020b) and PaLM (Chowdhery et al., 2022) family. We make the following key observations:
+
+(a) LLMs could exhibit low controllability. Our experiments consistently show that both pretrained and QA finetuned LLMs tend to ignore a context when it contradicts with model's world knowledge. We show that this problem persists and may intensify as the model becomes larger. We further show that the noise in the (QA) finetuning set plays an important role in emergence of this behavior (cf. Sec. 4.2).  
+(b) LLMs may not be robust against context noise. We demonstrate that both pretrained and QA finetuned models are strongly interfered by irrelevant contexts, especially the ones that are on the same general topic as the underlying question (cf. Sec. 4.3).
+
+2. We propose a novel method - knowledge aware finetuning (KAFT) - to directly enhance both controllability (Eq. 1) and robustness (Eq. 2) of an LLM. KAFT enhances the controllability by creating counterfactual data augmentations where the answer entity in the context is swapped to a different but plausible entity, in conflict with the ground truth (and potentially the model's world knowledge). As for enhancing robustness, KAFT requires that the model should predict its pretrained closed-book answer rather than the ground truth answer whenever the context is irrelevant.
+
+3. Through extensive empirical evaluation, we show that KAFT-based models successfully demonstrate the coexistence of controllability and robustness of model's working memory (see Table 1 for an illustration).
+
+# 2 Related Works
+
+World knowledge in language models. Recent works established that LLMs memorize factual information present in the pretraining corpus. E.g., Petroni et al. (2019) utilize language model analysis (LAMA) probing to show that BERT models (Devlin et al., 2018) could act as knowledge bases. Roberts et al. (2020) reported similar findings for T5 models. It is therefore common practice to employ LLMs in tasks like closed book QA (Chowdhery et al., 2022).
+
+Knowledge update in language models. Given that factual knowledge is ever-evolving, outdated memory of LLMs may lead to incorrect predictions (Lazaridou et al., 2021; Onoe et al., 2022). Furthermore, during deployment, one may unearth mistakes that need correction. Frequent retraining from scratch with an updated and corrected corpus would be prohibitively expensive. Ideas around finetuning (Zhu et al., 2020) and continued learning (Jang et al., 2022) train the model with less but still significant resources. Multiple recent efforts have studied how these models store the factual knowledge (Geva et al., 2021) and methods to update model parameters given new knowledge (De Cao et al., 2021; Dhingra et al., 2022; Mitchell et al., 2022; Meng et al., 2022a,b). These strategies change weights in response to single updates, risking inadvertently affecting unrelated skills or knowledge and creating burden to potentially store multiple versions of LLMs. We focus on updating the model behavior by providing a suitable context and ensuring that the model's working memory is controllable by such contexts.
+
+Contextual and parametric knowledge. Guu et al. (2020); Joshi et al. (2020); Petroni et al. (2020) utilized retrieved context to assist language models in tasks such as QA. At the same time, LLMs memorize large amounts of knowledge in their parameters. Despite this dichotomy, only a few studies have previously addressed the relation between these two very different knowledge sources. Longpre et al. (2021) find that larger models have a greater tendency to ignore context in
+
+favor of their own parametric knowledge, and that the noise in the context in the finetuning set plays a big role in causing this behavior. We incorporate the algorithms proposed by Longpre et al. (2021) for mitigating this problem as baselines in Sec. 4.4 (the Relevant Only Finetuning approaches), where we find such baselines lack robustness against irrelevant contexts (Fig. 1, 2). Kassner and Schütze (2020) showed that language models tend to be easily misled by certain types of irrelevant contexts. We observe similar phenomena in QA and show that our proposed KAFT leads to more robust models against irrelevant contexts. Finally, Pan et al. (2021) considers a scenario where some context sources may be less trustworthy than the model's parametric knowledge. This scenario can be captured by an extension of our framework Eq.(1-2). For example, given three sources, one could enforce the following precedence order: source1 > source2 > model's own knowledge > source3 > irrelevant contexts.
+
+Notions of controllability and robustness. In control theory, controllability (Ogata, 1996) refers to the ability of using external inputs to manipulate system to reach all possible states. In the spirit of this definition, this paper measures the controllability of an LM's working memory by external contexts. In the framework of controlled text generation (Zhang et al., 2022; Hu and Li, 2022), the notion of controllability explored here is a special type of fine-grained semantic control of the model's behavior with the content of the context.
+
+Notions of robustness. (Liang et al., 2022; Omar et al., 2022) survey many notions of robustness of language models around the notion of the invariance of model's behaviors when the input is perturbed (for example, expressing similar semantic meanings in different ways). Our robustness benchmark is an extreme and input-dependent version under this framework. In our evaluations, the input contains two parts: the context and the question. In this work, a robust model's response is invariant to large perturbations in the semantic content of the context, as long as these changes are not relevant to the question.
+
+During the preparation of this manuscript, we were made aware of a parallel and independent investigation by Neeman et al. (2022) that shares some important aspects of our work.
+
+<table><tr><td>Context type</td><td>Target sequence</td></tr><tr><td>relevant context</td><td>${ground truth answer}
+(from context)</td></tr><tr><td>irrelevant context</td><td>${pretrained model&#x27;s answer}
+(irrelevant context)</td></tr><tr><td>empty context</td><td>${pretrained model&#x27;s answer}
+(empty context)</td></tr><tr><td>counterfactual context</td><td>${counterfactual answer}
+是从 context)</td></tr></table>
+
+Table 3: The output formats of the KAFT model.
+
+# 3 Methods
+
+For concreteness, consider a reading comprehension QA task where the model takes question  $q$  together with a context  $c$  as its input. The question has an answer label  $a$ . We also need a relevance label  $r$  denoting whether  $c$  entails  $a$ .
+
+Starting with a pretrained LM  $M$ , we would like to build a model  $M^{\prime}$  such that when the context  $c$  is relevant, its answer is always grounded in  $c$ , when  $c$  is irrelevant, it sticks to the pretrained model's answer. In equations:
+
+$$
+r = 1: \quad M ^ {\prime} (c + q) = a \tag {3}
+$$
+
+$$
+r = 0: \quad M ^ {\prime} (c + q) = M (q) \tag {4}
+$$
+
+where  $+$  denotes string concatenation. This establishes the priority order of knowledge sources as in Eq. (1 & 2): if there is a conflict between a relevant context  $c$  and  $M$ 's parametric knowledge, then the output should be consistent with  $c$ . In addition, irrelevant context should have no influence on the model's output. Note that even though we are separating relevant vs irrelevant context here, the model does not know  $r$  a priori. It has to determine  $r$  based on the semantics of  $c$  and  $q$ .
+
+In the KAFT data,  $r = 1$  cases include relevant or counterfactual context, where  $a$  is the ground truth or counterfactual answer, respectively;  $r = 0$  cases include empty or irrelevant contexts. Here the label is given by the pretrained model's answer to the same question in a few-shot closed book setting, reflecting the model's parametric knowledge. To provide more interpretability, we make the model output its classification of the context's relevance along side the answer itself. See Table 3 for details.
+
+# 3.1 Datasets
+
+We construct KAFT based on several public datasets, including SQuAD 2.0 (Rajpurkar et al., 2018), T-REx (Elsahar et al., 2018), QASC (Khot
+
+et al., 2020), and TriviaQA (Joshi et al., 2017). They cover several different QA formats, including multiple choice (QASC), Cloze (TReX), extractive (SQuAD), and open domain (TriviaQA). For each dataset, we may construct different types of context and corresponding labels as summarized in Table 4.
+
+# 3.2 Models
+
+We select two families of pretrained LLMs: T5 (Raffel et al., 2020b) representing the encoder-decoder architecture and PaLM (Chowdhery et al., 2022) representing the decoder only architecture. We include all three PaLM models (8B, 62B and 540B), while with T5 we restrict to the largest sizes (XL and XXL, with 3B and 11B parameters, respectively) because the smaller ones do not respond well to in-context few shot prompts, making it difficult to measure their parametric knowledge.
+
+# 3.3 Relevant context
+
+We define the relevance of a context by whether it logically entails an answer to the question, which is a strong requirement - even if a piece of context is on the same topic of the question or contain the answer label, it might still be irrelevant. In practice, this happens often among retrieved results. In Sec 4.4, we show that if the model is still required to fit on to the ground truth label when given an irrelevant context, then the model becomes more likely to ignore relevant contexts. It is therefore crucial to strive towards precise logical entailment when building relevant context. We apply several techniques to improve the semantic connection between the context and the QA pair as shown in Table. 4. More details can be found in Appendix A.1.
+
+# 3.4 Irrelevant Context
+
+An irrelevant context is any context that does not entail the answer. An easy irrelevant context is completely off topic. We obtain them with random sampling for all datasets. A hard irrelevant context is on the same topic, sometimes discussing the same entities involved in the QA pair but does not logically entail the answer. SQuAD 2.0 already contains human labels on whether the answer can be derived from the context, thus providing hard irrelevant contexts. TriviaQA provides somewhat extensive paraphrases for each answer. We filter the retrieved contexts to find ones that do not contain any answer paraphrase, and use them as hard irrelevant context.
+
+# 3.5 Probing pretrained knowledge
+
+We first use the pretrained model to generate  $M(q)$  in Eq. 4, which are then used to assemble the KAFT finetuning dataset according to Eq. 4. We use hand-engineered few-shot knowledge probing prompts that condition the model to answer a question according to its world knowledge acquired during pretraining. In Appendix A.3, we provide more details on the construction of these prompts.
+
+# 3.6 Counterfactuals
+
+To train the model to be controllable by the context, we explicitly engineer plausible training data where the context is in conflict with the model's pretrained world knowledge. Given a triple of question, answer, and relevant context, we use a pretrained T5 XXL model to generate a triple of question, counterfactual answer, and counterfactual context with prompt engineering. We apply several filtering and postprocessing techniques to ensure the quality. Details are given in Appendix A.4.
+
+# 3.7 Metrics
+
+In this section, we define metrics that measures controllability and robustness. All results are from single runs.
+
+Controllability. To measure controllability, we supply the model with a counterfactual context and examine whether it can output the corresponding counterfactual answer. For a fair comparison, we select questions which all five pretrained models can answer correctly in a closed book few-shot setting, which are referred to as head questions. Since they are likely well represented in the pretraining set, such questions are challenging as we swap the answer to counterfactuals. Since we don't have any paraphrases of the counterfactual answer, we choose to use thresholded unigram recall to measure the performance. In particular, a model output is rated positive if the output of the model contains  $>80\%$  of the answer unigrams, with stop-words removed.
+
+Robustness. To measure robustness, we use the human labeled "impossible" slice of SQuAD 2.0, since SQuAD 2.0 contains many examples where the context is on the same general topic of the question but does not contain the answer. We measure the rate when the model successfully avoids extracting answers from such irrelevant contexts. The avoidance is considered successful if the con
+
+<table><tr><td>Dataset</td><td>Relevant Context</td><td>Irrelevant context</td><td>Counterfactual context</td></tr><tr><td>TReLU</td><td>Sampled irrelevant statements and one relevant statement</td><td>Sampled</td><td>Sampled irrelevant statements and one relevant statement with the answer entity replaced</td></tr><tr><td>SQuAD 2.0</td><td>From original dataset</td><td>Original human labeled and sampled</td><td>Relevant context with answer span replaced by counterfactual answer</td></tr><tr><td>QASC</td><td>2-stage retrieved statements and one golden statement</td><td>Sampled</td><td>None</td></tr><tr><td>TriviaQA (wiki split)</td><td>Retrieved contexts containing the answer and overlapping with the question</td><td>Retrieved contexts that do not contain the answer</td><td>Relevant context with answer span replaced by counterfactual answer</td></tr></table>
+
+Table 4: A summary of the KAFT data construction. For relevant context, counterfactual context, and irrelevant/empty context, the corresponding answer labels are ground truth answer, counterfactual answer, and pretrained model's few shot closed book answer, respectively. All four datasets also include examples where no context is provided.
+
+text contains less than  $50\%$  of the unigrams in the model's prediction, removing stop words.
+
+# 3.8 Baselines
+
+Pretrained. We evaluate the pretrained model's controllability and robustness in a zero shot reading comprehension QA setup. The context is concatenated with the question in input sequence.
+
+Noisy finetuning. In this approach, the label is the ground truth answer whether the context is relevant or not. This is a standard method implicitly used in most QA datasets.2 In this work, we construct this baseline for KAFT by first removing all counterfactual augmentations and then replace all labels with the ground truth label.
+
+Relevant only finetuning. The approach where only relevant context and the corresponding ground truth label are used during finetuning, which is shown to improve controllability in (Longpre et al., 2021). As a baseline for KAFT we remove all counterfactual and irrelevant augmentations and only keep the relevant slice of our finetuning data.
+
+UQA V2. The Unified QA 11B (Khashabi et al., 2022) model, which is a general purpose QA model finetuned on a collection of 20 QA datasets. We take the largest model (11B) in the UQA V2 family as a baseline and compare with KAFT T5 XXL which is of similar size in Fig. 2. Since UQA V2 contains SQuAD 2.0 in its training set, where the label for irrelevant context is an empty string, it
+
+does not completely follow the noisy finetuning prescription introduced earlier.
+
+KAFT noCF. The KAFT method with no counterfactual augmentations.
+
+KAFT noCF and noTQA. The KAFT method with no counterfactual augmentations and no TriviaQA slice.
+
+We include more details on the hyper parameters of model finetuning, prompts, post processing, data filtering, and metric computations in the Appendix A.2.
+
+# 4 Results
+
+In this section we measure the controllability and robustness of KAFT with the metrics defined in Sec. 3.7 and compare with baselines in Sec. 3.8.
+
+# 4.1 Larger models may ignore more contexts
+
+Most benchmarks improve as a function of model size, including TriviaQA exact match (EM) accuracy, as shown in the first row of Fig. 1. However, we found that larger models may ignore the context more. This may happen for the pretrained model, but the behavior is especially severe for models finetuned on QA tasks using baseline approaches. We demonstrate this effect in the second row of Fig. 1. This highlights a need for designing new methods to improve the controllability of LLMs.
+
+# 4.2 KAFT and controllability
+
+One of the most striking phenomenon observable from Fig. 1 is that KAFT achieve immense improvements in controllability while maintaining
+
+![](images/a7fdb80f9690d5d2c3b93532ad4b0fff0ffef0e2baf26c51a1b5ffa17329030c.jpg)  
+Figure 1: LLMs may become less controllable as the model size increases. Interestingly, KAFT significantly boosts controllability and robustness. The first row shows the EM on the wiki split of TriviaQA when one retrieved context is supplied. The second row shows the controllability when the context is in conflict with the pretrained model's knowledge. The third rows show robustness against human labelled irrelevant contexts from SQuAD 2.0.
+
+performance on standard QA. For example, the KAFT PaLM 540B model achieves 24X better controllability compared to the noisy finetuning when the context is in conflict with the model's pretrained factual knowledge, while performing similarly on regular contexts. In addition, KAFT is the only finetuning approach that consistently achieves better controllability than the pretrained models. Most of this gain originates from the counterfactual augmentation where the model explicitly learns the priority order in Eq. 1 when a conflict does appear. However both relevant only finetuning and KAFT without counterfactual augmentations also exhibit stronger controllability compared to noisy finetuning, even when there is no explicit counterfactual augmentations in both cases. The reason is that both approaches avoid irrelevant contexts that does not imply an answer. Thus the model is less prone to ignore the context compared to noisy finetuning.
+
+# 4.3 KAFT and robustness
+
+For the pretrained model, the robustness decreased slightly from T5 XL to XXL and from PaLM 8B to 62B (see third row in Fig. 1). But the difference is small. Relevant only finetuning suffers the most loss because it does not have irrelevant contexts
+
+during training. Noisy finetuning only alleviates this loss slightly, still vastly underperforming the pretrained model.
+
+KAFT, on the other hand, significantly boosts robustness. For example, the KAFT PaLM 540B model achieves 6X better robustness compared to noisy finetuning and 1.6X better robustness compared to the pretrained model. Adding the counterfactual augmentation slightly reduces robustness, but the difference is comparably small.
+
+# 4.4 Analysis and ablation studies
+
+We perform ablation studies to understand the effect of different augmentations in KAFT, as well as the general effect of added context noise.
+
+Effect of KAFT data augmentations. In Fig. 2, we systematically reduce the sampling rate of different data augmentation slices when training KAFT-T5 XXL models. We observe that reducing or removing the counterfactual and irrelevant data augmentations severely reduces controllability and robustness, respectively. In addition, KAFT models significantly out-perform the very strong baselines of Unified QA V2 on both controllability and robustness, showing that KAFT cannot
+
+<table><tr><td>Method</td><td>Controllability PALM 62B</td><td>Controllability T5 XXL</td><td>Est. Noise ratio from relevant slice of TQA</td></tr><tr><td>NoisyFT</td><td>15%</td><td>37%</td><td>63%</td></tr><tr><td>KAFT noCF EM filter</td><td>20%</td><td>51%</td><td>35%</td></tr><tr><td>KAFT noCF</td><td>33%</td><td>54%</td><td>5%</td></tr><tr><td>KAFT noCF and noTQA</td><td>52%</td><td>69%</td><td>0%</td></tr></table>
+
+![](images/a86d16566fc0f3d7c5344da9ba1f682ab0003e3d633a26a6ed9f63081b72ac80.jpg)  
+Figure 2: Ablation studies on data mixture ratios showing the importance of KAFT augmentations. We add Unified QA (Khashabi et al., 2022) and noisy finetuning baselines for comparison.
+
+Table 5: Context noise leads to model ignoring context and thus reduces controllability. We compare the controllability of models finetuned with different levels of context noise resulting from different filtering approaches on the training data. The noise ratio is estimated by sampling a small subset from the relevance slice of TriviaQA and manually checking the fraction of cases where the context does not entail the QA pair.  
+
+<table><tr><td>Model</td><td>Pretrained</td><td>KAFT</td></tr><tr><td>T5 XL</td><td>6.1%</td><td>7.2%</td></tr><tr><td>T5 XXL</td><td>6.6%</td><td>6.8%</td></tr><tr><td>PaLM 8B</td><td>3.3%</td><td>4.1%</td></tr><tr><td>PaLM 62B</td><td>1.4%</td><td>1.3%</td></tr><tr><td>PaLM 540B</td><td>0.6%</td><td>0.7%</td></tr></table>
+
+Table 6: The match rate between models' closed book answers and counterfactual answers, among all TriviaQA training set questions with counterfactual augmentations. KAFT shows little unwanted memorization of counterfactual answers.
+
+be replaced by simply adding more supervised data.
+
+KAFT models memorize few counterfactual. One potential risk of adding counterfactual context-answer pairs in the training set is unwanted memorization. We check whether KAFT models memorizes the counterfactual answers in the
+
+training set using the same prompts we used to probe the pretrained model's closed book answers. We find very little memorization: e.g., the KAFT-PALM 540B model only memorized  $0.1\%$  more counterfactuals compared to the pretrained PALM model after KAFT finetuning. Results for other models are similar (cf. Table. 6). The model learns the desirable correlation between the context and the output, rather than memorizing the counterfactual answers.
+
+Context noise reduces controllability. By context noise we refer to the subset of training data where the model is required to produce an answer that is not implied by the provided context, or required to ignore the context while it actually imply the answer. On the flip side, we find that it is possible to achieve good controllability without explicit counterfactual augmentations if we can reduce context noise in the training data.
+
+Table. 5 shows how different amounts of context noise impact the model's controllability. In particular, because TriviaQA contexts are produced by a retrieval system, it is not guaranteed that a context logically implies the answer. This is even true when the context contains exact matches of the answer. On the other hand, TReX, SQuAD and QASC contains much less context noise given our KAFT construction methods Sec. A.1. Due to this intrinsic noise, including TriviaQA in KAFT caused a negative impact on controllability, especially when there are no explicit counterfactual augmentations. The first row shows noisy finetuning, which contains the most noise. The last row shows that KAFT with TriviaQA data removed. Even though this model is not finetuned on TriviaQA, it has the best controllability. The second row uses a simpler and more noisy filter than KAFT by considering a context to be relevant if it contains the answer.
+
+# 5 Conclusion
+
+In this work, we analyzed the interaction between LLMs' parametric knowledge (stored in its model parameters) and knowledge contained in informational contexts provided as a part of the input sequence. We find that models are prone to ignoring the context, especially when the context is in conflict with the parametric knowledge. In addition, the model's output can be swayed by irrelevant context even when there is no logical link between such context and the model's task at hand. We quantitatively characterize these behaviours as controllability and robustness of LLMs when one attempts to control their working memory with noisy context. We proposed a new finetuning method, KAFT, that utilizes data augmentations to substantially boost the controllability and robustness of an LLM without significantly affecting its performance on standard QA tasks. With KAFT, we can build LLMs with a clear order of priority when utilizing information from difference sources, including its own parametric knowledge.
+
+# 6 Limitations
+
+# 6.1 Multiple sources
+
+In this work, we trained a model that can utilize two sources of information with predefined priority order, with one of them being the model's own parametric knowledge. While this is the first step towards LLM's information utilization with clear, predefined priorities, we acknowledge that real world applications could be more nuanced. For example, KAFT may need to be expanded to treat multiple sources of information with different trustworthiness which may translate to the following desired priority order:
+
+$$
+\begin{array}{l} \text {r e l e v a n t} 1 > \text {r e l e v a n t} 2 (5) \\ > \text {m o d e l} > \text {m o d e l} ^ {\prime} \text {p a r a m e t r i c k n o w l d e g l e} (6) \\ > \text {r e l e v a n t} 3 (7) \\ > \text {a l l i r r e l e v a n t c o m p e s t} (8) \\ \end{array}
+$$
+
+This orders of priority determines the handling of conflicts. In addition, any irrelevant context should have no influence on the model's output.
+
+# 6.2 Multitask / in-context learning
+
+KAFT currently only explores QA tasks. We acknowledge that the applications of LLMs go far beyond a single style of tasks. We have not yet
+
+achieved controlled utilization of information in a task agnostic way. Ideally, the model should learn to prioritize retrieved relevant information in any task that LLMs are capable of, including in-context few-shot or zero-shot scenarios.
+
+# 6.3 Dynamically enforce "learning to ignore"
+
+In this work, it was necessary to build a different KAFT dataset for each model. Because in Eq. 4, whenever the context is irrelevant, the model fits on to the pretrained model's answers which depends on the model. This presents additional workload when applying KAFT to new models. In future, it's worthwhile to explore a dynamic methods that generates closed booked answers during training. At each training step involving irrelevant context, we could run the forward pass twice, one with the provided context and another without. Then we can compute a new loss:
+
+$$
+\begin{array}{l} r = 1: \operatorname {L o s s} = \operatorname {C E} \left(M ^ {\prime} (c + q), \text {l a b e l}\right) (9) \\ r = 0: \operatorname {L o s s} = \operatorname {C E} (M ^ {\prime} (c + q), \\ \operatorname {s t o p} _ {\text {g r a d i e n t}} \left(M ^ {\prime} (q)\right) (10) \\ \end{array}
+$$
+
+where  $+$  denotes string concatenation. This is different from Eq. 4 as it fits on to the closed book answers of the current version of the finetuned model, rather than that of the pretrained model. It's not yet clear whether this would achieve better robustness. It's also more expensive because two forward passes are necessary for each training example. However it might be justified by the improved simplicity in directly applying KAFT with minimal prepossessing.
+
+This approach is somewhat similar to classifier free guidance (Ho and Salimans, 2022), which has been successfully applied to image generation models. One added benefit of classifier free guidance is the ability to tune the strength of context conditioning after the model is trained, which is another interesting direction to explore here.
+
+# 7 Ethics statement: Broader impacts and potential risks
+
+In this work, we study approaches to finetune LLMs to make them more grounded and faithful to provided contexts. If our method is applied broadly, it has the potential to correct the unwanted or biased behavior of LLMs with a carefully curated set of natural language instructions without expensive retraining. This provides one feasible avenue
+
+towards improving language models to correct a potential bias that is embedded in the pretraining corpus. At the same time, we acknowledge that our method does not completely address such issues on its own, because 1) instances where the model's working memory is not controllable by the context even after KAFT is applied may remain; 2) the finetuning dataset used in KAFT may inadvertently introduce or strengthen certain biases. For example, we acknowledge that all KAFT datasets used in this study are English datasets, and so it is a valuable future work direction to extend KAFT to be more representative of all languages.
+
+In addition, we acknowledge that the use of LLMs can be expensive in terms of energy usage. We utilize existing pretrained LLMs such as T5 and PaLM. KAFT's energy usage is small compared to the pretraining process, but it still leaves a significant energy footprint. In particular, the most expensive training, KAFT-PaLM 540B, takes 12190 TPU v4 hours. It is our hope that methods such as KAFT will provide a way for reducing the need for frequently retraining LLMs, and thus could lead to a more environmentally friendly experimentation.
+
+# Acknowledgements
+
+We would like to thank Slav Petrov for his insightful comments on an early draft of the paper that significantly helped improve the presentation of our work. We would also like to thank the reviewers and meta-reviewer for their thoughtful feedback on our submission.
+
+# References
+
+F. Gregory Ashby, Shawn W. Ell, Vivian V. Valentin, and Michael B. Casale. 2005. FROST: A Distributed Neurocomputational Model of Working Memory Maintenance. Journal of Cognitive Neuroscience, 17(11):1728-1743.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in NeurIPS, volume 33, pages 1877-1901. Curran Associates, Inc.
+
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways.  
+Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491-6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL, pages 4171-4186.  
+Bhuwan Dhingra, Jeremy R Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W Cohen. 2022. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273.  
+Hady Elsahar, Pavlos Vougiouklis, Arslen Remaci, Christophe Gravier, Jonathon Hare, Frederique Laforest, and Elena Simperl. 2018. T-REx: A large scale alignment of natural language with knowledge base triples. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).  
+Keisuke Fukuda and Geoffrey F. Woodman. 2017. Visual working memory buffers information retrieved from visual long-term memory. Proceedings of the National Academy of Sciences, 114/20.  
+J M Fuster. 1973. Unit activity in prefrontal cortex during delayed-response performance: neuronal corre
+
+lates of transient memory. Journal of Neurophysiology, 36(1):61-78. PMID: 4196203.  
+Karl H. Pribram George A. Miller, Eugene Galanter. 1960. Plans and the structure of behavior. Holt, New York.  
+Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484-5495, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrievalaugmented language model pre-training.  
+Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance.  
+Zhiting Hu and Li Erran Li. 2022. A causal lens for controllable text generation. CoRR, abs/2201.09119.  
+Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun KIM, Stanley Jungkyu Choi, and Minjoon Seo. 2022. Towards continual knowledge learning of language models. In International Conference on Learning Representations.  
+Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics.  
+Mandar Joshi, Kenton Lee, Yi Luan, and Kristina Toutanova. 2020. Contextualized representations using textual encyclopedic knowledge. CoRR, abs/2004.12006.  
+Nora Kassner and Hinrich Schütze. 2020. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7811-7818, Online. Association for Computational Linguistics.  
+Daniel Khashabi, Yeganeh Kordi, and Hannaneh Hajishirzi. 2022. Unifiedqa-v2: Stronger generalization via broader cross-format training.  
+Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UNIFIEDQA: Crossing format boundaries with a single QA system. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1896-1907, Online. Association for Computational Linguistics.
+
+Tushar Khot, Peter Clark, Michal Guerquin, Peter Alexander Jansen, and Ashish Sabharwal. 2020. Qasc: A dataset for question answering via sentence composition. ArXiv, abs/1910.11473.  
+Angeliki Lazaridou, Adhi Kuncoro, Elena Gribovskaya, Devang Agrawal, Adam Liska, Tayfun Terzi, Mai Gimenez, Cyprien de Masson d'Autume, Tomas Kocisky, Sebastian Ruder, Dani Yogatama, Kris Cao, Susannah Young, and Phil Blunsom. 2021. Mind the gap: Assessing temporal generalization in neural language models. In Advances in Neural Information Processing Systems, volume 34, pages 29348-29363. Curran Associates, Inc.  
+Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christian Cosgrove, Christopher D. Manning, Christopher Ré, Diana Acosta-Navas, Drew A. Hudson, Eric Zelikman, Esin Durmus, Faisal Ladhak, Frieda Rong, Hongyu Ren, Huaxiu Yao, Jue Wang, Keshav Santhanam, Laurel Orr, Lucia Zheng, Mert Yuksekgonul, Mirac Suzgun, Nathan Kim, Neel Guha, Niladri Chatterji, Omar Khattab, Peter Henderson, Qian Huang, Ryan Chi, Sang Michael Xie, Shibani Santurkar, Surya Ganguli, Tatsunori Hashimoto, Thomas Icard, Tianyi Zhang, Vishrav Chaudhary, William Wang, Xuechen Li, Yifan Mai, Yuhui Zhang, and Yuta Koreeda. 2022. Holistic evaluation of language models.  
+Shayne Longpre, Kartik Perisetla, Anthony Chen, Nikhil Ramesh, Chris DuBois, and Sameer Singh. 2021. Entity-based knowledge conflicts in question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7052-7063, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022a. Locating and editing factual knowledge in gpt. arXiv preprint arXiv:2202.05262.  
+Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2022b. Mass-editing memory in a transformer.  
+Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022. Fast model editing at scale. In International Conference on Learning Representations.  
+Ella Neeman, Roee Aharoni, Or Honovich, Leshem Choshen, Idan Szpektor, and Omri Abend. 2022. to appear.  
+Katsuhiko Ogata. 1996. Modern Control Engineering (3rd Ed.). Prentice-Hall, Inc., USA.  
+Marwan Omar, Soohyeon Choi, DaeHun Nyang, and David Mohaisen. 2022. Robust natural language
+
+processing: Recent advances, challenges, and future directions. CoRR, abs/2201.00768.  
+Yasumasa Onoe, Michael JQ Zhang, Eunsol Choi, and Greg Durrett. 2022. Entity cloze by date: What lms know about unseen entities. arXiv preprint arXiv:2205.02832.  
+Liangming Pan, Wenhu Chen, Min-Yen Kan, and William Yang Wang. 2021. Contraqa: Question answering under contradicting contexts. CoRR, abs/2110.07803.  
+Fabio Petroni, Patrick S. H. Lewis, Aleksandra Pik-tus, Tim Rocktäschel, Yuxiang Wu, Alexander H. Miller, and Sebastian Riedel. 2020. How context affects language models' factual predictions. CoRR, abs/2005.04611.  
+Fabio Petroni, Tim Roktaschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.  
+Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Elsen, Siddhant Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug Fritz, Thibault Sottiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Jason Gabriel, William Isaac, Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. 2021. Scaling language models: Methods, analysis & insights from training gopher.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020a. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67.
+
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020b. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67.  
+Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for squad.  
+Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? arXiv preprint arXiv:2002.08910.  
+Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu, Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, Elton Zhang, Rewon Child, Reza Yazdani Aminabadi, Julie Bernauer, Xia Song, Mohammad Shoeybi, Yuxiong He, Michael Houston, Saurabh Tiwary, and Bryan Catanzaro. 2022. Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in NeurIPS.  
+Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. 2022. A survey of controllable text generation using transformer-based pre-trained language models. CoRR, abs/2201.05337.  
+Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363.
+
+# A Appendix
+
+# A.1 Details on relevant context construction
+
+SQuAD 2.0 has human labels for this particular aspect. But most datasets do not. For TReX, the question is cloze style where we mask a certain entity within the triples statement. We build a relevant context by concatenating the original statements with a number of sampled irrelevant statements, after randomly shuffling their order. This ensures the relevance of the context while keeping it challenging. The training set of QASC provides 2 gold statements that implies the answer via a two hop reasoning. We are using the 2-stage retrieved collection of statements similar to (Khashabi et al., 2020). We find that the gold statements, or semantically equivalent ones, often exist in the retrieved results. To improve relevance we will randomly add one of the two golden statements and mix it in
+
+the retrieved context to build a relevant context for the KAFT training set. We manually checked on a random small subset that this ensures a relevance ratio around  $90\%$ .
+
+TriviaQA is especially challenging because there is no human labeled gold context, while all existing contexts are obtained by a retrieval system. We filter the context by whether they contain the answer. This turned out to be insufficient and leaves a large fraction of irrelevant contexts that do not logically entail the answer. We apply additional filters based on the unigram overlaps of the context with the question, as well as based on the output of a logically entailment model.
+
+# A.2 Training Details
+
+We use a learning rate of 0.0002 on all models. The batch size is 32 for all PaLM models and 16 for T5 models. For T5 XL we pick the checkpoint at 100000 finetune steps and for T5 XXL models we pick the checkpoint at 90000 steps. For PaLM 8B and 62B, we pick the checkpoint at 40000 finetuning steps. For PaLM 540B we pick the checkpoint at 15000 steps. These steps are generally determined by avoiding overfitting. However for larger models we are also constrained by compute resources.
+
+# A.3 Knowledge Probing Prompts
+
+In this section we provide details on how the knowledge probing prompts in Table 7-9 are constructed. In particular, our goal is to make the model only answer questions where it knows the answer. To do this, we construct prompts that contains two types of QA pairs:
+
+1. Regular QA pairs if the model can answer the specific question correctly in multiple few-shot in-context settings.  
+2. QA pairs where the answer is "I don't know" for T5 models or "?" for PaLM models, if the model cannot answer the question correctly in most few-shot in-context settings.
+
+With such specially designed prompts, we encourage the model to abstain if it does not know the answer. The counterfactual context used in the controllability benchmark is constructed using the same method. However we ensure no entities overlaps exist between the prompts that generates the training data vs the test data.
+
+# A.4 Counterfactual Generation
+
+To train the model to be controllable by the context, we explicitly engineer plausible training data where the context is in conflict with the model's pretrained world knowledge. This is done in 3 steps:
+
+1. We apply a diverse set of few-shot prompts similar to Table 10 to condition a pretrained T5 XXL model to generate plausible counterfactual answers.  
+2. We remove examples if the generation is unsuccessful, when it's either too long or have a large overlap with the original answer.  
+3. We replace all occurrences of the original answer with the counterfactual answer in the original context to build the counterfactual context.
+
+With this approach, we build a new QA data set where the answer implied by the context is likely to be in conflict with the model's existing knowledge.
+
+# A.5 Evaluations for Counterfactual memorization and relevance classification
+
+One potential danger of adding counterfactual context-answer pairs in the training set is unwanted memorization. We check whether KAFT models memorizes the counterfactual answers in the training set using the same prompts we used to probe the pretrained model's closed book answers. The results in Table 6 show that KAFT has little unwanted memorization of counterfactual answers. Instead the model learns the desirable correlation between the context and the output, as demonstrated in Figure 1.
+
+As illustrated in Table 1 and described in Table 3, we require the model to generate its judgements on whether the provided context is relevant. As a sanity check, we evaluated this part of the output on 1000 class-balanced SQuAD2 validation questions, the relevance prediction from KAFT-T5-XXL has  $84\%$  precision and  $98\%$  recall.
+
+# A.6 Postprocessing
+
+After we obtain the output from the pretrained model to the question, which is concatenated after the knowledge probing prompt, we need to post-process it and removed unwanted components. We do two types of post-processing on the pretrained predictions:
+
+<table><tr><td>Model</td><td>Standard QA Knowledge Probe Prompts</td></tr><tr><td>T5 XL</td><td>Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don&#x27;t know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: what was the name of the other HD channel Virgin media could carry in the future? A: I don&#x27;t know.Q: What is the term for a hyperactive immune system that attacks normal tissues? A: autoimmunity.Q: What complexity class is commonly characterized by unknown algorithms to enhance solvability? A: I don&#x27;t know.Q: Which nation contains the majority of the amazon forest? A: Brazil.</td></tr><tr><td>T5 XXL</td><td>Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don&#x27;t know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: The symbol for mercuric oxide is? A: HgO.Q: What religion did the Yuan discourage, to support Buddhism? A: Taoism.</td></tr><tr><td>PaLM 8B</td><td>Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? &quot;Enlightenment is man&#x27;s emergence from his self-incurred immaturity&quot;. A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: What King and former Huguenot looked out for the welfare of the group? A: Henry IV.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried.</td></tr><tr><td>PaLM 62B</td><td>Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? &quot;Enlightenment is man&#x27;s emergence from his self-incurred immaturity&quot;. A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: Who was the US Secretary of State in 2001? A: Colin Bowell.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried.Q: When did the European Anti-Fraud Office investigate John Dalli? A: 2012.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: When did Costa v ENEL take place? A: 1964.</td></tr><tr><td>PaLM 62B</td><td>Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: New York Bay.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? &quot;Enlightenment is man&#x27;s emergence from his self-incurred immaturity&quot;. A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: When was the Parental Leave directive created? A: 1996.Q: How many megaregions are there in the United States? A: 11.Q: Where is D&#x27;Olier Street? A: Dublin.Q: What is the speed limit set to reduce consumption? A: 55 mph.Q: What channel replaced Sky Travel? A: Sky Three.Q: Who founded McKinsey &amp; Company? A: James O. McKinsey.</td></tr></table>
+
+Table 7: Knowledge probing prompts for standard QA datasets. These prompts are used to probe the pretrained model's answer to questions in SQuAD 2.0 and TriviaQA.
+
+<table><tr><td>Model</td><td>Cloze Style QA Knowledge Probe Prompts</td></tr><tr><td>T5 XL</td><td>The Hudson River terminate into _. A: The Atlantic Ocean.
+__formally adds inverses to elements to any monoid. A: ?.
+Supply and _ causes child labour to still exist today? A: demands.
+__ was the prime minister of Japan in 2015? A: Shinzo Abe.
+__ is responsible for judicial review. A: Courts.
+__ was the name of the other HD channel Virgin media could carry in the future. A: ?
+__ is defined as a hyperactive immune system attacking normal tissues? A: autoimmunity.
+__ complexity class is commonly characterized by unknown algorithms to enhance solvability. A: ?
+__ contains the majority of the amazon forest? A: Brazil.</td></tr><tr><td>T5 XXL</td><td>The Hudson River terminate into _. A: The Atlantic Ocean.
+__ formally adds inverses to elements to any monoid. A: ?
+Supply and _ causes child labour to still exist today? A: demands.
+__ was the prime minister of Japan in 2015? A: Shinzo Abe.
+__ is responsible for judicial review. A: Courts.
+The French spread along with their imperialism the _ religion. A: Catholicism.
+The symbol for mercuric oxide is _. A: HgO.
+The Yuan discouraged _ to support Buddhism. A: Taoism.</td></tr><tr><td>PaLM 8B</td><td>Only answer the questions you know the answer to:
+The Hudson River terminate into _. A: The Atlantic Ocean.
+The county of Hampshire was officially named in _. A: ?
+__ said &quot;Enlightenment is man&#x27;s emergence from his self-incurred immaturity&quot;. A: Immanuel Kant.
+__ formally adds inverses to elements to any monoid. A: ?
+King _ and former Huguenot looked out for the welfare of the group. A: Henry IV.
+The principle of faunal succession was developed 100 years before _ &#x27;s theory of evolution. A: Charles Darwin.
+__ is the hero who killed a dragon on the Drachenfels? A: Siegfried.</td></tr><tr><td>PaLM 62B</td><td>Only answer the questions you know the answer to:
+The Hudson River terminate into _. A: The Atlantic Ocean.
+The county of Hampshire was officially named in _. A: ?
+__ said &quot;Enlightenment is man&#x27;s emergence from his self-incurred immaturity&quot;. A: Immanuel Kant.
+__ formally adds inverses to elements to any monoid. A: ?
+__ was the US Secretary of State in 2001. A: Colin Bowell.
+The principle of faunal succession was developed 100 years before _ &#x27;s theory of evolution? A: Charles Darwin.
+__ is the hero who killed a dragon on the Drachenfels. A: Siegfried.
+The European Anti-Fraud Office investigate John Dalli in year_. A: 2012.
+The French spread along with their imperialism the _ religion. A: Catholicism.
+Costa v ENEL happend in year_. A: 1964.</td></tr><tr><td>PaLM 62B</td><td>Only answer the questions you know the answer to:
+The Hudson River terminate into _. A: New York Bay.
+The county of Hampshire was officially named in _. A: ?
+__ said &quot;Enlightenment is man&#x27;s emergence from his self-incurred immaturity&quot;. A: Immanuel Kant.
+__ formally adds inverses to elements to any monoid. A: ?
+The Parental Leave directive created in year_. A: 1996.
+There are __ megaregions in the United States. A: 11.
+D&#x27;Olier Street is located in_. A: Dublin.
+The speed limit was set to _ to reduce consumption. A: 55 mph.
+__ channel replaced Sky Travel. A: Sky Three.
+__ founded McKinsey &amp; Company. A: James O. McKinsey.</td></tr></table>
+
+Table 8: Knowledge probing prompts for Cloze style QA datasets. These prompts are used to probe the pretrained model's answer to questions in TReX.  
+
+<table><tr><td>Model</td><td>Multiple Choice QA Knowledge Probe Prompts</td></tr><tr><td>PaLM 62B</td><td>Question: Into what body of water does the Hudson River terminate? (A) The great lakes (B) Amazon river (C) The red sea (D) the Atlantic Ocean (E) San Francisco bay (F) The north sea (G) Indian Ocean (H) Lake Mississippi -Answer: (D) the Atlantic Ocean. 
+Question: Who was the prime minister of Japan in 2015? (A) Donald Trump (B) Miho Nonaka (C) Andrew Yang (D) a France citizen (E) a political outsider (F) Shinzo Abe (G) woman (H) Zoe. -Answer: (F) Shinzo Abe. Question: what increases moisture? (A) density (B) the sun (C) wind (D) droughts (E) Honey (F) 17 (G) rain (H) meat -Answer: (G) rain. 
+Question: What can be found inside a cell? (A) soil (B) dogs (C) ovum (D) starfish (E) Most plants (F) RNA (G) washer (H) abundant -Answer: (F) RNA. 
+Question: What kind of coloring do chromoplasts make? (A) fat (B) move (C) RNA (D) grow (E) red (F) skin (G) eyes (H) DNA -Answer: (E) red.</td></tr></table>
+
+Table 9: Knowledge probing prompts for Cloze style QA datasets. These prompts are used to probe the pretrained model's answer to questions in TReX.
+
+<table><tr><td>Question</td><td>In which country did Warsaw Pact officials meet to dissolve the alliance?</td></tr><tr><td>Original answer</td><td>Hungary</td></tr><tr><td>Counterfactual answer</td><td>Russia</td></tr><tr><td>Original context</td><td>On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Hungary.</td></tr><tr><td>Counterfactual context</td><td>On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Russia.</td></tr><tr><td>T5 Prompt to generate the counterfactual answer</td><td>Let&#x27;s play a game of writing fake answers Who did US fight in world war 1? Real answer: Germany. Fake answer: Somalia. Who is the CEO of Amazon? Real Answer: Jeff Bezos. Fake Answer: Richard D. Fairbank. ...7 more examples ... In which country did Warsaw Pact officials meet to dissolve the alliance? Real answer: Hungary. Fake answer: &lt;extra_id_0&gt;.</td></tr></table>
+
+Table 10: An example from the counterfactual split of the KAFT training set. We take an original question, answer, and context triple. We then use a few examples to prompt a pretrained T5 XXL model to generate a plausible counterfactual answer. Finally, we replace all occurrences of the original answer with the counterfactual answer to build the counterfactual context.
+
+1. Truncation: We truncate the model's output on special tokens such as  $<$  extra_id_1 >, punctuation, line change symbols and question/context initialization symbols such as "Q:", "Question:", " CONTEXT:", These symbols are a frequent in the pretrained model's responds to our QA style knowledge probe prompts and indicate that the model is ready to move on to the next question that is unrelated to the answer of the current question.
+
+2. Abstain: We normalize all abstain symbols. Whenever the model indicate abstaining using either "I don't know", "unsure" or "?" in the output as responses to our prompt, we record "unsure" as its answer when constructing the label in the irrelevant slices of KAFT.
+
+# A.7 Dataset and task details
+
+KAFT mixes together a number of datasets, each with multiple augmentation slices. During training, data from these difference sources are sampled in a round-robin style according to predefined mixture weights. We list these weights as well as the corresponding dataset stats as in Table 11. The sampling ratio from each slice is computed using a product of the normalized dataset level rate and the normalized slice level rate as follows:
+
+$$
+R (d, s) = \frac {r _ {d}}{\sum_ {d ^ {\prime}} r _ {d ^ {\prime}}} \frac {r _ {d s}}{\sum_ {s ^ {\prime}} r _ {d s ^ {\prime}}} \tag {11}
+$$
+
+where  $d, d'$  denote different datasets and  $s, s'$  denote difference slices within each dataset. For ex
+
+ample, the sampling ratio from the QASC relevant slice is given by:
+
+$$
+\begin{array}{l} R (Q A S C, r e l e v a n t) \\ = \frac {0 . 3}{1 . 3 + 0 . 3 + 0 . 1 + 0 . 2} \frac {0 . 5}{0 . 5 + 0 . 2 5 + 0 . 0 2} \\ = 0. 0 8 3 1 \tag {12} \\ \end{array}
+$$
+
+The KAFT-TriviaQA training set contains 45593 relevant examples and 72697 irrelevant examples. The KAFT-QASC training set contains 8134 relevant examples and the same number of irrelevant examples. The KAFT-SQuAD2 dataset contains 78125 relevant examples and 117287 irrelevant examples. The KAFT-TReX training set contains 75365 relevant examples and 47503 irrelevant examples.
+
+# A.8 Licensing and scientific artifacts
+
+In this work, we used the following scientific artifacts: TriviaQA is licensed under Apache License 2.0. The SQuAD 2.0 dataset is licensed under CC BY-SA 4.0. T-REx is under a Creative Commons Attribution-ShareAlike 4.0 International License. QASC is under CC BY license. T5 models are under Apache License 2.0. Unified QA models are under Apache License 2.0. The PaLM models are proprietary. All these artifacts are properly cited when we mention them the first time. Our use for these artifacts are consistent with their licenses.
+
+We create the following scientific artifacts and we will partly release them after this paper is pub
+
+<table><tr><td>dataset</td><td>dataset weight</td><td>slice</td><td>slice weight</td></tr><tr><td rowspan="10">SQuAD 2.0</td><td rowspan="10">1.3</td><td>relevant</td><td>0.8</td></tr><tr><td>counterfactual</td><td>0.1</td></tr><tr><td>original irrelevant abstain</td><td>0.1</td></tr><tr><td>original irrelevant other</td><td>0.1</td></tr><tr><td>empty correct</td><td>0.33</td></tr><tr><td>empty abstain</td><td>0.02</td></tr><tr><td>empty other</td><td>0.05</td></tr><tr><td>sampled irrelevant correct</td><td>0.33</td></tr><tr><td>sampled irrelevant abstain</td><td>0.02</td></tr><tr><td>sampled irrelevant other</td><td>0.03</td></tr><tr><td rowspan="3">QASC</td><td rowspan="3">0.3</td><td>relevant</td><td>0.5</td></tr><tr><td>irrelevant correct</td><td>0.25</td></tr><tr><td>irrelevant other</td><td>0.02</td></tr><tr><td rowspan="6">TReX</td><td rowspan="6">0.1</td><td>relevant</td><td>0.4</td></tr><tr><td>counterfactual</td><td>0.4</td></tr><tr><td>2-hop relevant</td><td>6</td></tr><tr><td>irrelevant correct</td><td>0.15</td></tr><tr><td>irrelevant abstain</td><td>0.03</td></tr><tr><td>irrelevant other</td><td>0.03</td></tr><tr><td rowspan="4">TriviaQA</td><td rowspan="4">0.2</td><td>relevant</td><td>0.8</td></tr><tr><td>counterfactual</td><td>0.15</td></tr><tr><td>irrelevant/empty correct</td><td>0.5</td></tr><tr><td>irrelevant/empty other</td><td>0.2</td></tr></table>
+
+Table 11: Task mixture weights. During finetuning, training data from each split is computed in a round robin fashion according to these weights. The sampling rate from each slice is computed with these weights using in Eq. 12. Here "relevant", "irrelevant", "empty" indicates the relevance (or absence) of the context relative to the question, and "counterfactual" indicates counterfactual context constructed using answer replacement. The additional specification for irrelevant/empty slices, "correct", "abstain", and "other" indicate the pretrained model's answers' type and quality relative to the ground truth. For TReX, we have a special slice called "2-hop relevant". These are relevant contexts contracted using 2-hop reasoning over the triplet structure of TReX.
+
+lished: The KAFT finetuning method will be released under Apache License 2.0. The KAFT-T5 models will be released under Apache License 2.0. The KAFT-PaLM models will be proprietary.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? section 6  
+A2. Did you discuss any potential risks of your work? section 7  
+A3. Do the abstract and introduction summarize the paper's main claims? abstract and section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Section 1-5
+
+B1. Did you cite the creators of artifacts you used? section 1-5  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix A.8  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Appendix A.8  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use well established public datasets that are constructed based on publicly available information.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Sec 1-5, Appendix A.8  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix A
+
+C Did you run computational experiments?
+
+Sec 3-4
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Sec 1-4, Sec 7, Appendix A
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4, Appendix A  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Jax
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response.

2023/Large Language Models with Controllable Working Memory/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ef7886e8425b89e084522d8889b757afcdd0b1ccc4d263beecb7d3e4ba881c7
+size 1134763

2023/Layerwise universal adversarial attack on NLP models/6e17acd0-02c0-4f9f-8520-e21ab1c219b9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a4ce82f3ce1e577e973fcbb021e0249ec98797679af65d29af8ccda0d6c2b5d
+size 532796

2023/Layerwise universal adversarial attack on NLP models/full.md

@@ -0,0 +1,465 @@
+# Layerwise universal adversarial attack on NLP models
+
+Olga Tsymboloi $^{1,2}$ , Danil Malaev $^{1}$ , Andrei Petrovskii $^{1}$ , and Ivan Oseledets $^{3,4}$
+
+$^{1}$ Sber AI Lab, Moscow, Russia
+
+$^{2}$ Moscow Institute of Physics and Technology, Moscow, Russia
+
+$^{3}$ Skolkovo Institute of Science and Technology, Moscow, Russia
+
+<sup>4</sup>Artificial Intelligence Research Institute (AIRI), Moscow, Russia
+
+tsimboy.oa@phystech.edu, m9255171362@gmail.com, andreypetrovskij@gmail.com
+
+i. oseledets@skoltech.ru
+
+# Abstract
+
+In this work, we examine the vulnerability of language models to universal adversarial triggers (UATs). We propose a new white-box approach to the construction of layerwise UATs (LUATs), which searches the triggers by perturbing hidden layers of a network. On the example of three transformer models and three datasets from the GLUE benchmark, we demonstrate that our method provides better transferability in a model-to-model setting with an average gain of  $9.3\%$  in the fooling rate over the baseline. Moreover, we investigate triggers transferability in the task-to-task setting. Using small subsets from the datasets similar to the target tasks for choosing a perturbed layer, we show that LUATs are more efficient than vanilla UATs by  $7.1\%$  in the fooling rate.
+
+# 1 Introduction
+
+One of the fundamental drawbacks of modern neural networks is their vulnerability to adversarial attacks (Szegedy et al., 2013; Goodfellow et al., 2014), imperceptible perturbations to the data samples that leave the ground truth label unchanged but are able to modify model prediction drastically. The samples obtained as the result of these perturbations are called adversarial examples. First discovered for image datasets (Szegedy et al., 2013), this phenomenon was then demonstrated for other types of data, including natural language (Papernot et al., 2016; Liang et al., 2017; Gao et al., 2018).
+
+The originally proposed methods of adversarial attack construction were sample-dependent, which means that one can not apply the same perturbation to different dataset items and expect equal success. Sample-agnostic, or in other words, universal, adversarial perturbations (UAPs) were proposed in Moosavi-Dezfooli et al. (2017), where based on a small subset of image data, the authors constructed perturbations leading to prediction change of  $80 - 90\%$  (depending on the model) of samples.
+
+They also showed that UAPs discovered on one model could successfully fool another.
+
+The generalization of the universal attacks to natural language data was made by Wallace et al. (2019). Short additives (triggers) were inserted at the beginning of data samples, and then a search over the token space, in order to maximize the probability of the negative class on the chosen data subset, was performed. The found triggers turned out to be very efficient at fooling the model, repeating the success of UAPs proposed for images.
+
+The conventional way to look for adversarial examples is to perturb the output of a model. Considering image classification neural networks, Khrulkov and Oseledets (2018) proposed to search for perturbations to hidden layers by approximating the so-called  $(p,q)$ -singular vectors (Boyd, 1974) of the corresponding Jacobian matrix. Then one can hope that the error will propagate through the whole network end, resulting in model prediction change. They showed that this approach allows obtaining a high fooling rate based on significantly smaller data subsets than those leveraged by Moosavi-Dezfooli et al. (2017).
+
+In this paper, we aim to continue the investigation of neural networks' vulnerability to universal adversarial attacks in the case of natural language data. Inspired by the approach considered by Khrulkov and Oseledets (2018), we look for the perturbations to hidden layers of a model instead of the loss function. In order to avoid projection from embedding to discrete space, we use simplex parametrization of the search space (see, e.g. (Dong et al., 2021; Guo et al., 2021)). We formulate the corresponding optimization problem and propose the algorithm for obtaining its approximate solution. On the example of three transformer models: BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), ALBERT (Lan et al., 2020) and three GLUE datasets (Wang et al., 2018) we demonstrate higher efficiency of our method over the original approach
+
+![](images/d62d146b3bef1502212e40b65ce8e37fb12a27de3f34821f6864c07ec980367d.jpg)  
+Figure 1: LUATs (ours): Layerwise Universal Adversarial Triggers. The algorithm scheme, where  $\nabla F_l$  is defined in (8).
+
+of Wallace et al. (2019) in the setting of model-to-model and task-to-task transfer. We also show that in the case of direct attack application, our method demonstrates the results which are on par with the baseline, where perturbation of the loss function was realized. We hope that this technique will serve as a useful tool for discovering flaws and deepening our understanding of language neural networks.
+
+# 2 Framework
+
+Let  $f: \mathcal{X} \to \mathcal{Y}$  be a text classification model defined for a dataset  $\mathcal{D} = \{x_i, y_i\}_{i=1}^N$ , where  $x_i$  is an input and  $y_i$  is a corresponding label. Our goal is to find a small perturbation to the original input sequences that leads to a change in a model prediction for the maximum possible number of samples. Such perturbation could be modelled as an insertion of a trigger, a token sequence  $t$  of small length  $L$ , into a particular part of an input. Following Wallace et al. (2019), we call such text perturbations universal adversarial triggers (UATs). In this paper, we focus on triggers concatenated to the front of a sentence and denote a corrupted sample as  $\hat{x} = t \oplus x$ . It is also important to note that, in contrast to Wallace et al. (2019), we consider the unsupervised scenario when an attacker does not have access to the labels. For evaluation we follow Moosavi-Dezfooli et al. (2017) and use the fooling rate (FR):
+
+$$
+F R = \frac {1}{N} \sum_ {\hat {x} = t \oplus x, x \in \mathcal {D}} [ f (x) \neq f (\hat {x}) ]. \tag {1}
+$$
+
+Universal adversarial triggers. Before we dive into the description of our approach, it is worth depicting the original method of UATs. Restricting themselves to the white-box setting, Wallace et al. (2019) showed that an efficient way to find UATs could be performed through the optimization of the
+
+first-order expected loss approximation:
+
+$$
+\max  _ {t \in \mathcal {V} ^ {L}} \mathbb {E} _ {x, y \sim \mu} \langle \nabla_ {t ^ {n}} \mathcal {L} (t ^ {n} \oplus x, y), t - t ^ {n} \rangle , \tag {2}
+$$
+
+where  $\mu$  is a distribution of input data,  $t^n$  denotes the trigger found after the  $n$ -th iteration,  $\mathcal{V}$  stands for the token vocabulary and the initial trigger  $t^0$  can be chosen as the  $L$ -time repetition of the word "the". For simplicity, here (and below in similar situations), by  $t$ , we mean its embedding if the gradient is taken with respect to it or if it appears as a term in a scalar product.
+
+In order to find the optimal perturbation within each iteration, the expectation in (2) is relaxed with the average over a batch, and maximization is performed for each trigger token independently:
+
+$$
+\max  _ {t _ {j} \in \mathcal {V}} \sum_ {x, y \in \text {b a t c h}} \left\langle \nabla_ {t _ {j} ^ {n}} \mathcal {L} \left(t ^ {n} \oplus x, y\right), t _ {j} - t _ {j} ^ {n} \right\rangle , \tag {3}
+$$
+
+where  $j$  is the token index. After solving (3), the next trigger  $t^{n + 1}$  is selected via beam search over all the token positions.
+
+Our approach: layerwise universal adversarial triggers (LUATs). In (3), we face the optimization problem over a discrete set of tokens from the vocabulary  $\mathcal{V}$ . To overcome this issue, let us relax (3) using the probability simplex model for every token. In this case, a trigger can be represented as  $t = WV$ , where  $V$  is a vocabulary matrix and  $W_{mn}$  is a probability of the  $n$ -th token from vocabulary to be selected at position  $m$ . Then (3) can be rewritten as follows:
+
+$$
+\max  _ {W \in \mathcal {S}} \sum_ {x, y \in \text {b a t c h}} \left\langle \nabla_ {t ^ {n}} \mathcal {L} \left(t ^ {n} \oplus x, y\right), W V - t ^ {n} \right\rangle , \tag {4}
+$$
+
+where  $\mathcal{S} = \{W|W1 = 1,W\geq 0\}$  and  $W\geq 0$  denotes an element-wise inequality. In this formulation, the search for the solution is done by performing optimization of the weights  $W$  over the simplex  $\mathcal{S}$ .
+
+Our approach can be seen as an extension of (4) to the perturbation of hidden layers. This is inspired by Khrulkov and Oseledets (2018), who applied this idea to find UAPs for fooling image classification neural networks. Given a layer  $l$ , the optimization problem, in this case, can be similarly obtained via the Taylor expansion:
+
+$$
+\begin{array}{l} l (\hat {x}) - l (\hat {x} ^ {n}) \approx J _ {l} (\hat {x} ^ {n}) (\hat {x} - \hat {x} ^ {n}) \\ \left\| l (\hat {x}) - l \left(\hat {x} ^ {n}\right)\right\| _ {q} ^ {q} \rightarrow \max  _ {t \in \mathcal {V} ^ {L}}, \tag {5} \\ \end{array}
+$$
+
+where  $\hat{x}^n = t^n\oplus x$ $q$  is a hyperparameter to be fine-tuned and  $J_{l}(x)$  is the Jacobian operator corresponding to a layer  $l$ :
+
+$$
+J _ {l} (x) = \frac {\partial l (x)}{\partial x}. \tag {6}
+$$
+
+Bringing (4) and (5) together we obtain
+
+$$
+\begin{array}{l} \max_{W\in \mathcal{S}}F_{l}(W) = \\ \max  _ {W \in \mathcal {S}} \sum_ {x \in \text {b a t c h}} \| J _ {l} \left(t ^ {n} \oplus x\right) \left(W V - t ^ {n}\right) \| _ {q} ^ {q}. \tag {7} \\ \end{array}
+$$
+
+In contrast to (3), finding the optimal solution of (7) is computationally infeasible. Indeed, the problem (3) allows a brute-force approach for finding the optimal token for each trigger position since it requires computing the gradient only once for each iteration. On the other hand, in our case, a brute-force computation is very cumbersome since, for each iteration, it would require computing the Jacobian action for every batch, token candidate and position, resulting in  $\mathcal{O}(LB|\mathcal{V}|)$  forward-backward passes, where  $B$  is the number of batches in an iteration. Luckily,  $F_{l}(W)$  is convex and can be lower bounded via a tangent line, where the gradient is calculated as follows:
+
+$$
+\begin{array}{l} \nabla F _ {l} (W) = \\ \sum_ {x \in \text {b a t c h}} J _ {l} ^ {\top} \left(\hat {x} ^ {n}\right) \psi_ {q} \left(J _ {l} \left(\hat {x} ^ {n}\right) \left(W V - t ^ {n}\right)\right) V ^ {\top}, \tag {8} \\ \end{array}
+$$
+
+where  $\psi_q(x) = \mathrm{sign}(x)|x|^{q - 1}$ . Therefore, our task is reduced to finding the solution to the linear problem with the simplicial constraint:
+
+$$
+\max  _ {W \in S} \left\langle \nabla F _ {l} \left(W ^ {*}\right), W \right\rangle , \tag {9}
+$$
+
+where  $\nabla F_l(W)$  is given by (8) and  $W^{*}$  denotes the point where we perform the linear approximation. The final problem (9) has a closed-form solution
+
+(see the Appendix A for more details) and, as a result, we reduced the number of forward-backward computations to  $\mathcal{O}(B)$ .
+
+Concerning the initialization of  $W^{*}$ , we take the uniform distribution over all the vocabulary tokens for each token position in a trigger. Within each iteration, we perform only one step with respect to  $W$  in order to reduce computation time and observe that it is sufficient for breaking the models efficiently.
+
+Finally, since the found after a given iteration weight matrix  $W$  in the worst case has only one non-zero element per row (see Appendix A), we can get into a local maximum unless we guess the proper initialization. Therefore, similarly to Wallace et al. (2019), we perform a beam search over the top- $k$  candidates. In order to realize it, it is necessary to define the ranking criterion for choosing the best option at each search step. For this purpose, we use the FR. The overall algorithm is presented in Algorithm 1 and Fig. 1.
+
+Algorithm 1: LUATs: Layerwise Universal Adversarial Triggers
+
+Input: Dataset  $\mathcal{D}$  ,victim model,tokenizer,  $q$  layer to attack  $l$  trigger length  $L$  top-  $k$  ,beam size  $b$
+
+Output: Trigger  $t$
+
+1  $t =$  tokenizer( the...the) L times  
+2  $W = \frac{1}{|\mathcal{V}|}\mathrm{ones}(L,|\mathcal{V}|)$
+
+3 while FR increase do
+
+4 Sample batch  $X$  
+5 Compute  $\nabla F_{l}(W)$  over batch  
+6 candidates  $=$  Select indices of  $k$  largest entries of  $\nabla F_l(W)$  for each token position  
+7  $\begin{array}{r}t = \mathrm{BeamSearch}(\mathcal{D},\mathrm{candidates},b)\mathrm{to}\\ \mathrm{maximize~FR} \end{array}$
+
+# 3 Experiments
+
+In this section, we present a numerical study of the proposed layerwise adversarial attack framework on text classification datasets. The code is publicly available on GitHub<sup>1</sup>.
+
+# 3.1 Setup
+
+Datasets. As in the work of Wang et al. (2021), we consider only a subset of tasks from the GLUE
+
+benchmark. In particular, we use SST-2 for the sentiment classification, MNLI (matched), QNLI, RTE for the natural language inference and MRPC as the paraphrase identification. We exclude CoLA, which task is to define whether the input is grammatically correct or not, and universal triggers are highly probable to change most of the ground-truth positive labels to negative. Finally, WNLI is not considered because it contains too few examples. We conduct our experiments using the validation set for attack fitting and the test set for evaluation.
+
+Models. We focus our consideration on three transformer models: BERT base, RoBERTa base and ALBERT base using the pre-trained weights from the TextAttack project (Morris et al., 2020) for most of the model-dataset cases. For some of them, the performance was unsatisfactory (all the models on MNLI, ALBERT on QNLI, RoBERTa on SST-2, RTE), and we fine-tuned them on the corresponding training sets (Mosin et al., 2023). To train our attack, we use existing GLUE datasets splits. The detailed statistics is presented in Tab. 1
+
+Hyperparameters. In our experiments, we investigate the attack performance depending on the dataset, model, layer  $l$  (from 0 to 11), trigger length  $L \in \{1,2,3,4,5,6\}$  and  $q \in \{2,3,4,5,7,10\}$ . For each dataset, model and trigger length, we performed a grid search over  $l$  and  $q$ . The other parameters, such as top- $k$  and the beam size, remain fixed to the values obtained from the corresponding ablation study. In all experiments, we use a batch size of 128. Finally, we define the initialization of  $W$  as the uniform distribution over the vocabulary tokens for each position in a trigger (see Algorithm 1).
+
+Token filtration and resegmentation. Transformers' vocabulary contains items such as symbols, special tokens and unused words, which are easily detected during the inference. To increase triggers' imperceptibility, we exclude them from the vocabulary matrix during optimization leaving only those which contain english letters or numbers.
+
+Another problem appears since a lot of tokens do not correspond to complete words but rather pieces of words (sub-words). As a result, if the first found token corresponds to a sub-word, one encounters the retokenization, meaning that, after converting the found trigger to string and back, the set of the tokens can change. Moreover, sometimes one has to deal with appearing symbols such as "##" in a trigger. In this case, we drop all the extra
+
+symbols and perform the retokenization. Luckily, it does not result in severe performance degradation. In the case when, due to the resegmentation, the length of a trigger changes, we report the result as for the length for which the attack training was performed. As an alternative to direct resegmentation, we tried to transform triggers by passing them through an MLM model, but this approach led to a more significant drop in performance.
+
+# 3.2 Main Results
+
+Comparison with the baseline. We perform a comparison of LUATs with untargeted UATs of Wallace et al. (2019). In order to stay in the unsupervised setting, we modify their approach by replacing the ground truth labels in the cross-entropy loss function with the class probabilities. As a result, we search for a trigger that maximizes the distance between model output distributions before and after the perturbation. In addition, as the criterion for choosing the best alternative in the beam search, we use FR for both methods.
+
+We perform the ablation study to estimate the dependence of FR on top-  $k$  and the beam size. For the beam size 1, we measure both attacks' performance for different values of top-  $k$  from 1 to 40. Then for the best top-  $k$ , we build the dependence on the beam size from 1 to 5. We perform this study on the QNLI dataset. The results are presented in the Fig. 5. We stick to the top-  $k$  10 and the beam size 1 as a trade-off between high performance and low computational complexity.
+
+The grid search results are presented in Tab. 2, where for each model, dataset, and trigger length, we show the best results of both approaches. We performed the computation on four GPU's NVIDIA A100 of 80GB. To reduce the influence hyperparameters searching space cardinality (72 times more runs due to different values of  $q$  and  $L$ ), in Fig. 2, we present a time, averaged over 10 batches, per one iteration of both approaches. Indeed, for LUATs, we observe linear dependence on a layer, particularly 7.27 seconds on average versus 8.05 seconds for UATs of Wallace et al. (2019). The triggers obtained with the method of Wallace et al. (2019) took only 5 GPU hours; hence the execution time of the full grid search for LUATs could be estimated by 325 GPU hours. While the proposed method might be more efficient on average, high variance is explained by significantly different paddings in sampled batches.
+
+<table><tr><td>Dataset</td><td>Validation</td><td>Test</td><td># Classes</td><td>BERT
+Acc. Val. / Test</td><td>RoBERTa
+Acc. Val. / Test</td><td>ALBERT
+Acc. Val. / Test</td></tr><tr><td>MRPC</td><td>408</td><td>1725</td><td>2</td><td>87.7/84.4</td><td>90.4/87.2</td><td>89.7/86.0</td></tr><tr><td>QNLI</td><td>5463</td><td>5463</td><td>2</td><td>91.5/90.7</td><td>91.8/91.8</td><td>90.6/90.8</td></tr><tr><td>MNLI</td><td>9815</td><td>9796</td><td>3</td><td>84.2/83.7</td><td>86.5/86.3</td><td>83.8/83.5</td></tr><tr><td>SST-2</td><td>872</td><td>1821</td><td>2</td><td>92.4/93.3</td><td>94.0/94.9</td><td>92.7/91.7</td></tr><tr><td>RTE</td><td>277</td><td>3000</td><td>2</td><td>72.6/67.6</td><td>80.5/74.0</td><td>76.0/72.2</td></tr></table>
+
+Table 1: Statistics of the considered classification datasets. We present datasets cardinality, the number of classes and model accuracy on both validation and test sets.
+
+![](images/0f92b1634bba732db34333fe2ccb6a0cb3cbff430f2fb7ab1a8e74c9a5f5f802.jpg)  
+Figure 2: The iteration average time for UAT and LUAT (ours) approaches. The results presented on the QNLI dataset with fixed  $q = 10$ ,  $L = 3$  and a batch size of 128. Since, for the case of LUATs, the standard deviation does not change significantly, we report the value averaged over layers as  $\overline{std}_{LUAT}$ .
+
+It is interesting to note that in some cases, shorter triggers appear to be better than longer ones (see Appendix A, Tab. 10). That is why we present the best performance for each length  $L$  over all the lengths less or equal to  $L$ . From Tab. 2, one can see that our method demonstrates the results, which are on par with Wallace et al. (2019). In Fig. 3, we present the dependence of FR on the trigger length  $L$ . The attack performance saturates when it approaches 5 or 6, meaning that considering longer triggers would hardly bring any performance gain.
+
+Wallace et al. (2019) suggested that the efficiency of universal adversarial triggers can be explained by the existence of dataset biases, such as fallacious correlations between certain words and classes and supported this conjecture by computing pointwise mutual information (PMI) of vocabulary tokens and dataset classes
+
+$$
+\mathrm {P M I} (t o k e n, c l a s s) = \log {\frac {p (t o k e n , c l a s s)}{p (t o k e n) p (c l a s s)}}.
+$$
+
+As a result, they reported a high PMI rank for their tiggers tokens. In order to verify whether LUATs
+
+satisfy the same pattern, we perform this computation for our best triggers. We sample 5 best candidates obtained during the grid search for each trigger length and compute PMI rank with add-100 smoothing for their tokens. The results on  $\mathrm{QNLI}^2$  (see the Tab. 3) demonstrate that similarly to UATs our trigger tokens have high PMI ranks.
+
+Dependence on  $q$  and a layer. For the investigation of dependence on  $q$  and a layer, we restrict ourselves to the datasets which appear to be the most vulnerable in our experiments: MRPC, QNLI and SST. The results are presented in Fig. 4, where we performed averaging over the lengths. One can conclude that, in general, it is more efficient to attack the higher layers. This observation can be interpreted with the idea which has been mentioned above; namely, that the efficiency of the triggers is caused by the existence of dataset biases. Indeed, as Merchant et al. (2020) demonstrated, fine-tuning for a downstream task primarily affects the higher layers of a model. Therefore, the bias which could be acquired due to fine-tuning should be accumulated in its higher layers. Since we try to fool a model with respect to a downstream task, the appearance of the higher layers among the most successful ones is more probable. Finally, it is interesting to note that the dependence on  $q$  also demonstrates better results for the larger values, which is in accordance with the findings of Khrulkov and Oseledets (2018).
+
+Transferability: model-to-model. It was demonstrated that the universal triggers could be transferable between different models trained on the same task (Wallace et al., 2019). Here, we perform a comparison of their approach and ours with respect to this property. Similarly to the above consideration, the computations are carried out on MRPC, QNLI and SST-2 datasets, with an additional restriction on the trigger length  $L = 3$ , which
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Dataset</td><td colspan="2">FR@1</td><td colspan="2">FR@2</td><td colspan="2">FR@3</td><td colspan="2">FR@4</td><td colspan="2">FR@5</td><td colspan="2">FR@6</td></tr><tr><td>W</td><td>S</td><td>W</td><td>S</td><td>W</td><td>S</td><td>W</td><td>S</td><td>W</td><td>S</td><td>W</td><td>S</td></tr><tr><td rowspan="5">ALBERT</td><td>MNLI</td><td>18.5</td><td>14.8</td><td>25.8</td><td>17.0</td><td>25.8</td><td>18.5</td><td>34.8</td><td>21.1</td><td>34.8</td><td>24.8</td><td>34.8</td><td>24.8</td></tr><tr><td>MRPC</td><td>11.4</td><td>4.5</td><td>27.6</td><td>28.9</td><td>47.9</td><td>54.8</td><td>66.6</td><td>64.1</td><td>66.6</td><td>67.0</td><td>67.6</td><td>67.6</td></tr><tr><td>QNLI</td><td>26.4</td><td>37.5</td><td>47.5</td><td>48.1</td><td>50.5</td><td>50.8</td><td>50.5</td><td>51.5</td><td>52.2</td><td>52.4</td><td>52.5</td><td>52.4</td></tr><tr><td>RTE</td><td>4.8</td><td>4.9</td><td>6.3</td><td>11.9</td><td>7.9</td><td>11.9</td><td>9.1</td><td>17.7</td><td>9.1</td><td>17.7</td><td>9.1</td><td>17.7</td></tr><tr><td>SST-2</td><td>21.0</td><td>29.7</td><td>34.8</td><td>38.8</td><td>45.5</td><td>46.4</td><td>47.9</td><td>48.5</td><td>50.6</td><td>48.7</td><td>50.6</td><td>48.7</td></tr><tr><td rowspan="5">BERT</td><td>MNLI</td><td>32.0</td><td>10.2</td><td>38.4</td><td>33.6</td><td>38.4</td><td>34.5</td><td>38.4</td><td>34.7</td><td>38.4</td><td>35.6</td><td>38.4</td><td>35.6</td></tr><tr><td>MRPC</td><td>15.4</td><td>9.3</td><td>64.6</td><td>56.6</td><td>70.6</td><td>69.0</td><td>70.7</td><td>70.4</td><td>70.8</td><td>70.6</td><td>70.8</td><td>70.7</td></tr><tr><td>QNLI</td><td>26.2</td><td>23.2</td><td>42.6</td><td>37.9</td><td>47.2</td><td>42.3</td><td>50.6</td><td>49.1</td><td>50.6</td><td>50.8</td><td>50.8</td><td>50.8</td></tr><tr><td>RTE</td><td>4.7</td><td>4.2</td><td>6.6</td><td>6.7</td><td>7.1</td><td>8.5</td><td>7.6</td><td>10.2</td><td>7.6</td><td>10.2</td><td>11.2</td><td>10.2</td></tr><tr><td>SST-2</td><td>29.7</td><td>27.0</td><td>42.1</td><td>38.0</td><td>42.1</td><td>44.6</td><td>42.1</td><td>48.2</td><td>50.6</td><td>49.0</td><td>50.8</td><td>49</td></tr><tr><td rowspan="5">RoBERTa</td><td>MNLI</td><td>4.8</td><td>8.2</td><td>4.9</td><td>22.3</td><td>33.0</td><td>28.8</td><td>33.0</td><td>30.0</td><td>33.0</td><td>33.1</td><td>33.0</td><td>33.1</td></tr><tr><td>MRPC</td><td>4.2</td><td>4.9</td><td>35.5</td><td>26.3</td><td>67.8</td><td>68.5</td><td>69.4</td><td>69.3</td><td>69.4</td><td>69.3</td><td>69.5</td><td>69.3</td></tr><tr><td>QNLI</td><td>14.7</td><td>13.3</td><td>32.6</td><td>39.0</td><td>41.5</td><td>44.9</td><td>44.5</td><td>47.3</td><td>48.1</td><td>47.7</td><td>48.1</td><td>48.3</td></tr><tr><td>RTE</td><td>3.1</td><td>6.3</td><td>6.1</td><td>13.6</td><td>6.1</td><td>13.6</td><td>8.0</td><td>13.6</td><td>8.0</td><td>16.3</td><td>9.2</td><td>16.3</td></tr><tr><td>SST-2</td><td>28.4</td><td>26.9</td><td>34.7</td><td>38.7</td><td>43.9</td><td>45.9</td><td>49.9</td><td>47.1</td><td>49.9</td><td>50.7</td><td>49.9</td><td>53.1</td></tr><tr><td>Average</td><td>-</td><td>16.3</td><td>15.0</td><td>30.0</td><td>30.5</td><td>38.4</td><td>38.9</td><td>41.5</td><td>41.5</td><td>42.7</td><td>42.9</td><td>43.1</td><td>43.2</td></tr></table>
+
+Table 2: Comparison between LUAT (S) and UATs of Wallace et al. (2019) (W). We report the best FR on the test sets for triggers whose length does not exceed  $L$  (FR@L).  
+
+<table><tr><td rowspan="2">E</td><td colspan="2">BERT</td><td colspan="2">ALBERT</td><td colspan="3">RoBERTa</td></tr><tr><td>Rank N</td><td>Rank</td><td>E</td><td>Rank N</td><td>Rank</td><td>E</td><td>Rank N</td></tr><tr><td>emperor</td><td>94.69</td><td>either</td><td>99.60</td><td>70%</td><td>98.95</td><td>those</td><td>99.02</td></tr><tr><td>berlin</td><td>93.72</td><td>legislation</td><td>97.60</td><td>10%</td><td>96.10</td><td>amount</td><td>98.59</td></tr><tr><td>whose</td><td>92.14</td><td>can</td><td>97.08</td><td>3,500</td><td>93.82</td><td>unless</td><td>93.46</td></tr><tr><td>russian</td><td>90.91</td><td>ter</td><td>95.37</td><td>18%</td><td>93.72</td><td>71</td><td>92.72</td></tr><tr><td>cardinal</td><td>90.25</td><td>latitude</td><td>93.81</td><td>20%</td><td>93.45</td><td>nor</td><td>92.46</td></tr><tr><td>orient</td><td>89.83</td><td>dalai</td><td>92.04</td><td>whose</td><td>91.79</td><td>where</td><td>91.30</td></tr><tr><td>german</td><td>89.03</td><td>samurai</td><td>90.52</td><td>11%</td><td>88.79</td><td>besides</td><td>90.62</td></tr><tr><td>korean</td><td>87.77</td><td>granting</td><td>90.24</td><td>54%</td><td>88.34</td><td>correlation</td><td>88.56</td></tr><tr><td>fide</td><td>84.10</td><td>reich</td><td>90.10</td><td>supplemented</td><td>86.34</td><td>holds</td><td>87.90</td></tr><tr><td>atop</td><td>82.91</td><td>banning</td><td>88.47</td><td>50%</td><td>85.46</td><td>waived</td><td>87.12</td></tr></table>
+
+Table 3: PMI for the joint training and validation set of QNLI, E - Entailment, N - Not entailment  
+
+<table><tr><td rowspan="2">From/To</td><td colspan="3">MRPC, W/S</td><td colspan="3">QNLI, W/S</td><td colspan="3">SST-2, W/S</td></tr><tr><td>ALBERT</td><td>BERT</td><td>RoBERTa</td><td>ALBERT</td><td>BERT</td><td>RoBERTa</td><td>ALBERT</td><td>BERT</td><td>RoBERTa</td></tr><tr><td>ALBERT</td><td>-</td><td>22.8/53.5</td><td>33.9/69.1</td><td>-</td><td>45.6/36.1</td><td>39.0/38.1</td><td>-</td><td>30.6/37.5</td><td>24.2/34.7</td></tr><tr><td>BERT</td><td>6.7/17.2</td><td>-</td><td>60.5/67.8</td><td>34.6/36.3</td><td>-</td><td>22.0/36.0</td><td>31.0/29.3</td><td>-</td><td>17.7/24.1</td></tr><tr><td>RoBERTa</td><td>15.3/32.1</td><td>55.5/64.6</td><td>-</td><td>20.4/40.4</td><td>18.0/35.9</td><td>-</td><td>34.1/26.2</td><td>36.9/37.7</td><td>-</td></tr></table>
+
+Table 4: Transferability comparison between LUATs (S) and UATs of Wallace et al. (2019) (W). We report FR after performing the attack transfer between different models trained on a fixed dataset for the trigger length  $L = 3$ .  
+
+<table><tr><td rowspan="2">From/To</td><td colspan="3">ALBERT, W/S</td><td colspan="3">BERT, W/S</td><td colspan="3">RoBERTa, W/S</td></tr><tr><td>MRPC</td><td>QNLI</td><td>SST-2</td><td>MRPC</td><td>QNLI</td><td>SST-2</td><td>MRPC</td><td>QNLI</td><td>SST-2</td></tr><tr><td>MRPC</td><td>-</td><td>9.5/21.9</td><td>5.4/32.5</td><td>-</td><td>18.3/17.6</td><td>11.8/9.4</td><td>-</td><td>31.4/25.6</td><td>5.3/6.0</td></tr><tr><td>QNLI</td><td>13.6/14.0</td><td>-</td><td>9.0/31.7</td><td>32.9/35.5</td><td>-</td><td>2.8/4.7</td><td>22.7/40.9</td><td>-</td><td>2.9/8.9</td></tr><tr><td>SST-2</td><td>8.6/22.8</td><td>31.3/36.0</td><td>-</td><td>29.7/64.8</td><td>27.5/21.2</td><td>-</td><td>18.1/13.9</td><td>14.0/15.9</td><td>-</td></tr></table>
+
+Table 5: Transferability comparison between LUATs (S) and UATs of Wallace et al. (2019) (W). We report FR after performing the attack transfer between different datasets for a fixed model and the trigger length  $L = 3$ .
+
+![](images/8741eff88e760edfad7e356bb5fce19266c53c5a2f812e1d6424fdff66a005b4.jpg)  
+(a) SST.
+
+![](images/523584a8d66fc2490327651b6a7455450b9b1b6eaba111ffa7c873c5445c2aa7.jpg)  
+(b) MRPC.
+
+![](images/adebae7476479311c8566ab6778d9b0217ead495926e60619e7e2d7b77b53cab.jpg)  
+(c)QNLI.
+
+![](images/21d7690d8aae7e6ab04765aa41d0ab15081ba233933ed051f13e82866ae242de.jpg)  
+Figure 3: FR for optimal  $q$  depending on the trigger length for the most vulnerable sets. The plots show FR saturation which is achieved for short enough triggers.
+
+![](images/4e242ee946e6802a8b85f3ffca244892a7a5437c2e05d75b58cf0aa3cfb8aeb5.jpg)  
+Figure 4: FR dependence on LUATs hyperperameters. The efficiency of our attack increases with  $q$  and the layer number.
+
+![](images/b0ad7284148be2d5b378935b5df7ce1a96ba07828382af9bdbec048eef0c1902.jpg)
+
+is done for simplicity. In this setting, we suppose that an attacker has access to the data and also to the input and output of an attacked model. When transferring the UATs, the best trigger obtained with a source model is taken and then applied directly to a target model on a test set. For LUATs, however, one can come up with a better way. Since a target model is different from a source one, it is not necessary that the same values of  $q$  and  $l$  would provide the best pick for the transfer. Therefore, since an attacker has access to the forward pass of a target model, we evaluate with it the fooling rate of the best triggers found for each value of  $q$  and  $l$  on a validation set and then apply the best option to a test set to get the final score. For simplicity, we fix  $q = 10$ , which according to the Fig. 4, is the best alternative on average. The final results are presented in the Tab. 4. One can see that in most cases, we outperform the UATs with an average gain of  $9.3\%$ .
+
+Transferability: task-to-task. The fact that the universal triggers can generalize to other models trained on the same task seems natural since, in this case, it is highly probable that a source and a target model would acquire the same biases. A
+
+more complicated situation is when one tries to transfer triggers between different tasks. In this setting, an attacker looks for more fundamental task-independent flaws of a model, which can be explained, e.g., by a bias appearing during the pretraining phase. In order to examine this capacity of the triggers, we perform their transferring between different datasets for each of the considered models. For measuring the performance of the UATs, as in the previous case, the best trigger obtained on a source task is transferred. On the other hand, for the layerwise approach, one, under a reasonable assumption, can still suggest a natural way to choose the most appropriate for transfer trigger among the best triggers corresponding to different values of  $q$  and  $l$ . Namely, we suppose that although attackers do not have access to data on which a target model was trained, they know a task for which it was trained. It means that they know whether it is sentiment classification, paraphrase identification, etc. (Savchenko et al., 2020). If this is the case, they can generate data corresponding to the task of interest. In order to mimic such a situation, for each of the datasets (SST-2, MRPC, QNLI), we select an auxiliary dataset collected for the same task.
+
+The map is the following: SST-2 - IMDb (Maas et al., 2011), MRPC - PAWS (Zhang et al., 2019), QNLI - WikiQA (Yang et al., 2015). We sample subsets of sizes 64, 128, and 256 from the auxiliary datasets and consider them as the data generated by an attacker. Again, fixing  $q = 10$  for simplicity, we evaluate the best trigger obtained for each layer on a corresponding auxiliary subset, performing inference on a target model. Herewith, each subset is sampled five times, and we report the average score for the size of 256, which appears to be the best option. The final results for both approaches are shown in the Tab. 5. Our approach demonstrates the average improvement of  $7.1\%$  in fooling rate over the vanilla UATs. The reason for that might be related to the fact that if this kind of triggers (task-independent) is indeed related to the pretraining phase, in order to find them, one instead should perturb the lower or middle layers since the higher layers' weights can change a lot after fine-tuning. The LUATs which appear to be most successful at transferring are presented in the Tab. 8 and 9 of the Appendix B for model-to-model and task-to-task transfers correspondingly.
+
+# 4 Related Work
+
+There are already quite a few works devoted to the universal attack on NLP models in the literature. We briefly discuss them here.
+
+Ribeiro et al. (2018) proposed a sample-agnostic approach to generate adversarial examples in the case of NLP models by applying a semantics-preserving set of rules consisting of specific word substitutions. However, found rules were not completely universal, resulting in model prediction change only for  $1\% - 4\%$  out of targeted samples.
+
+Similarly to Wallace et al. (2019), the realization of text adversarial attacks as short insertions to the input were proposed by Behjati et al. (2019). Though, they did not perform the search over the whole vocabulary for each word position in the trigger but instead exploited cosine-similarity projected gradient descent, which does not appear that efficient in the sense of attack performance.
+
+Adversarial triggers generated by the method proposed by Wallace et al. (2019) in general turned out to be semantically meaningless, which makes them easier to detect by defence systems. An attempt to make triggers more natural was undertaken by Song et al. (2021). Leveraging an adversarially regularized autoencoder (Zhao et al., 2018)
+
+to generate the triggers, they managed to improve their semantic meaning without significantly decreasing the attack efficiency.
+
+Another interesting direction is to minimize the amount of data needed for finding UAPs. Singla et al. (2022) created representatives of each class by minimizing the loss function linear approximation over the text sequences of a certain size. Afterwards, adversarial triggers were appended to these class representatives and the rest of the procedure followed Wallace et al. (2019). Although no data was used explicitly for training the attack, this approach demonstrated solid performance on considered datasets.
+
+# 5 Conclusion
+
+We present a new layerwise framework for the construction of universal adversarial attacks on NLP models. Following Wallace et al. (2019), we look for them in the form of triggers; however, in order to find the best attack, we do not perturb the loss function but neural network intermediate layers. We show that our approach demonstrates better performance when the triggers are transferred to different models and tasks. The latter might be related to the fact that in order to be transferred successfully between different datasets, a trigger should reflect network flaws that are task-independent. In this case, reducing the attack search to perturbation of the lower or middle layers might be more beneficial since the higher layers are highly influenced by fine-tuning. We hope this method will serve as a good tool for investigating the shortcomings of language models and improving our understanding of neural networks' intrinsic mechanisms.
+
+We would like to conclude by discussing the potential risks. As any type of technology, machine learning methods can be used for good and evil. In particular, adversarial attacks can be used for misleading released machine learning models. Nevertheless, we think that revealing the weaknesses of modern neural networks is very important for making them more secure in the future and also for being able to make conscious decisions when deploying them.
+
+# 6 Limitations and future work
+
+Our approach to universal text perturbations suffers from linguistic inconsistency, which makes them easier to detect. Therefore, as the next step of our research, it would be interesting to investigate
+
+the possibility of improving the naturalness of adversarial triggers without degradation of the attack performance in terms of the fooling rate.
+
+While the proposed approach outperforms the UATs of Wallace et al. (2019) in the transferability task, we should highlight that the additional hyperparameters adjustment plays a crucial role, and one could suggest validation procedure refinement for a more fair comparison. Also, for both direct and transferability settings, a more comprehensive range of models should be examined, including recurrent (Yuan et al., 2021) and transformer architectures, e.g., T5 (Raffel et al., 2020), XLNet (Yang et al., 2019), GPT family models (Radford et al., 2019; Brown et al., 2020).
+
+Another direction of improvement is related to the fact that sometimes the found triggers can change the ground truth label of samples they are concatenated to if, e.g., they contain words contradicting the true sense of a sentence. It would be interesting to analyze how often this happens and develop an approach to tackle this issue.
+
+Finally, it would be interesting to investigate the dependence of attack efficiency on the size of a training set and compare it with the so-called data-free approaches, such as the one proposed by Singla et al. (2022).
+
+# Acknowledgements
+
+The work was supported by Ministry of Science and Higher Education grant No. 075-10-2021-068.
+
+# References
+
+Melika Behjati, Seyed-Mohsen Moosavi-Dezfooli, Mahdieh Soleymani Baghshah, and Pascal Frossard. 2019. Universal adversarial attacks on text classifiers. In Proceedings of International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7345-7349. IEEE.  
+David W Boyd. 1974. The power method for  $l^p$  norms. Linear Algebra and its Applications, 9:95-101.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877-1901.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for
+
+Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Xinshuai Dong, Anh Tuan Luu, Rongrong Ji, and Hong Liu. 2021. Towards robustness against natural language word substitutions. arXiv preprint arXiv:2107.13541.  
+Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. 2018. Black-box generation of adversarial text sequences to evade deep learning classifiers. In Proceedings of Security and Privacy Workshops (SPW), pages 50-56. IEEE.  
+Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572.  
+Chuan Guo, Alexandre Sablayrolles, Hervé Jégou, and Douwe Kiela. 2021. Gradient-based adversarial attacks against text transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5747-5757, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Valentin Khrulkov and Ivan Oseledets. 2018. Art of singular vectors and universal adversarial perturbations. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 8562-8570. IEEE.  
+Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. ALBERT: A lite BERT for self-supervised learning of language representations. In Proceedings of International Conference on Learning Representations (ICLR).  
+Bin Liang, Hongcheng Li, Miaoqiang Su, Pan Bian, Xirong Li, and Wenchang Shi. 2017. Deep text classification can be fooled. arXiv preprint arXiv:1704.08006.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692.  
+Andrew Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL): Human language technologies, pages 142-150.  
+Amil Merchant, Elahe Rahimtoroghi, Ellie Pavlick, and Ian Tenney. 2020. What happens to BERT embeddings during fine-tuning? In Proceedings of the Third BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 33-44, Online. Association for Computational Linguistics.
+
+Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. 2017. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773.  
+John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020. TextAttack: A framework for adversarial attacks, data augmentation, and adversarial training in NLP. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP): System Demonstrations, pages 119-126.  
+Vladislav Mosin, Igor Samenko, Borislav Kozlovskii, Alexey Tikhonov, and Ivan P Yamshchikov. 2023. Fine-tuning transformers: Vocabulary transfer. Artificial Intelligence, page 103860.  
+Nicolas Papernot, Patrick McDaniel, Ananthram Swami, and Richard Harang. 2016. Crafting adversarial input sequences for recurrent neural networks. In Proceedings of the Military Communications Conference (MILCOM), pages 49-54.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551.  
+Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018. Semantically equivalent adversarial rules for debugging NLP models. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 856-865, Melbourne, Australia. Association for Computational Linguistics.  
+Andrey Savchenko, Anton Alekseev, Sejeong Kwon, Elena Tutubalina, Evgeny Myasnikov, and Sergey Nikolenko. 2020. Ad lingua: Text classification improves symbolism prediction in image advertisements. In Proceedings of the 28th International Conferences on Computational Linguistics (COLING), pages 1886-1892.  
+Yaman Kumar Singla, Swapnil Parekh, Somesh Singh, Changyou Chen, Balaji Krishnamurthy, and Rajiv Ratn Shah. 2022. MINIMAL: Mining models for universal adversarial triggers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11330-11339.  
+Liwei Song, Xinwei Yu, Hsuan-Tung Peng, and Karthik Narasimhan. 2021. Universal adversarial attacks with natural triggers for text classification. In Proceedings of the Conference of the North American Chapter of the Association for Computational
+
+Linguistics (ACL): Human Language Technologies, pages 3724-3733. ACL.  
+Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199.  
+Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing NLP. In Proceedings of the Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2153-2162.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. ACL.  
+Boxin Wang, Chejian Xu, Shuohang Wang, Zhe Gan, Yu Cheng, Jianfeng Gao, Ahmed Hassan Awadallah, and Bo Li. 2021. Adversarial GLUE: A multitask benchmark for robustness evaluation of language models. In Proceedings of Thirty-fifth Conference on Neural Information Processing Systems (NeuriPS) Datasets and Benchmarks Track (Round 2).  
+Yi Yang, Wen-tau Yih, and Christopher Meek. 2015. WikiQA: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2013-2018, Lisbon, Portugal. Association for Computational Linguistics.  
+Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems (NeurIPS), 32.  
+Lifan Yuan, Yichi Zhang, Yangyi Chen, and Wei Wei. 2021. Bridge the gap between CV and NLP! A gradient-based textual adversarial attack framework. arXiv preprint arXiv:2110.15317.  
+Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Junbo Zhao, Yoon Kim, Kelly Zhang, Alexander Rush, and Yann LeCun. 2018. Adversarially regularized autoencoders. In Proceedings of International Conference on Machine Learning (ICML), pages 5902-5911. PMLR.
+
+# A Linear problem solution on  $\mathcal{S}$
+
+Let us consider the following linear problem with a cost matrix  $C$ :
+
+$$
+\max  _ {W \in \mathcal {S}} \langle C, W \rangle .
+$$
+
+We need to construct the Lagrangian  $L$  and the consequent dual problem to obtain its solution.
+
+$$
+\begin{array}{l} L = - \operatorname {T r} ((C + M) ^ {\top} W) + \lambda^ {\top} (W 1 - 1) \\ = - \operatorname {T r} ((C + M - \lambda 1 ^ {\top}) ^ {\top} W) - \lambda^ {\top} 1 \\ \end{array}
+$$
+
+where  $\lambda$  and  $M$  are the Lagrange multipliers. From KKT conditions, we have:
+
+$$
+C + M - \lambda 1 ^ {\top} = O,
+$$
+
+$$
+M \geq 0, \quad M \cdot W = 0, \quad W 1 = 1,
+$$
+
+where  $M \cdot W$  means elementwise multiplication. As a result, we obtain the following dual problem:
+
+$$
+\begin{array}{l} \max - \lambda^ {\top} 1, \\ s. t. - \lambda 1 ^ {\top} \leq - C. \\ \end{array}
+$$
+
+Under the assumption that each row of the cost matrix  $C$  has a unique maximum, the closed-form solution will take the form:
+
+$$
+\lambda_ {i} = \max  _ {j} C _ {i j}, \forall i,
+$$
+
+and the corresponding primal solution
+
+$$
+W = \left\{ \begin{array}{l} W _ {i j} = 1,   j = \operatorname {a r g m a x} _ {j} C _ {i j} \forall i, \\ W _ {i j} = 0, \text {o t h e r w i s e}. \end{array} \right.
+$$
+
+Otherwise, if any row  $i$  of  $C$  violates the above assumption by having  $k > 1$  maximal elements with indices  $\{j_1,\ldots ,j_k\}$ , then
+
+$$
+W _ {i j} = \frac {1}{k}, \quad \forall k \in \{j _ {1}, \dots , j _ {k} \}.
+$$
+
+# B Tables and plots
+
+In this appendix, we present the results of
+
+- the ablation study on the top- $k$  and beam search parameters (see Fig. 5),  
+- the results of trigger analysis with PMI (see Tab. 6 and 7) for SST-2 and MRPC datasets,  
+- the LUATs which appear to be the best for model-to-model and task-to-task transfers (see Tab. 4 and 5),
+
+- the cases when shorter triggers appear to be better than longer ones (see Tab. 10),  
+- the examples of the top-20 obtained triggers.
+
+Concerning the transfer triggers, one can see that sometimes the same triggers efficiently break different models trained on different datasets, e.g., 'WHY voted beyond' (FR = 40.4 for ALBERT trained on QNLI, FR = 45.9 for BERT trained on QNLI, FR = 40.9 for RoBERTa trained on MRPC), 'unsuitable improper whether' (FR = 37.5 for BERT trained on SST-2, FR = 34.7 on RoBERTa trained on SST-2, FR = 22.8 for ALBERT trained on MRPC). This can serve as evidence of the high generalizability of universal adversarial triggers.
+
+![](images/ab723d25c7b2204bff2176a9a6e92e08af1861824e759434bf896b309030f73b.jpg)  
+Figure 5: The results of the ablation study with respect to the top-  $k$  and the beam size parameters on QNLI dataset.
+
+![](images/093d5fc28d774c19b39a2162c6ef8025c937f7f0a43a797d82e429ef49ddfc9f.jpg)
+
+![](images/869e2c6e749d8e40066e25325ffef524a16b9db10bbc1f3cfa481b42efff59a5.jpg)
+
+![](images/a56f390b6d14f6e37e510e660fa30db55f593f45d7680bc0070d4700ba8c536e.jpg)
+
+<table><tr><td rowspan="2">N</td><td colspan="2">BERT</td><td colspan="3">ALBERT</td><td colspan="3">RoBERTa</td></tr><tr><td>Rank P</td><td>Rank</td><td>N</td><td>Rank P</td><td>Rank</td><td>N</td><td>Rank P</td><td>Rank</td></tr><tr><td>worst</td><td>99.96</td><td>also</td><td>99.13</td><td>fails</td><td>99.70</td><td>coming</td><td>94.40</td><td>neither</td></tr><tr><td>stupid</td><td>99.87</td><td>drama</td><td>98.51</td><td>devoid</td><td>99.39</td><td>foremost</td><td>82.79</td><td>failure</td></tr><tr><td>fails</td><td>99.66</td><td>definitely</td><td>96.00</td><td>failure</td><td>98.90</td><td>behalf</td><td>81.14</td><td>whether</td></tr><tr><td>poor</td><td>99.48</td><td>shows</td><td>95.81</td><td>whether</td><td>97.22</td><td>warmed</td><td>28.42</td><td>despite</td></tr><tr><td>neither</td><td>99.35</td><td>walk</td><td>94.89</td><td>unless</td><td>94.21</td><td>placement</td><td>10.03</td><td>considering</td></tr><tr><td>badly</td><td>99.32</td><td>ranks</td><td>93.12</td><td>irrelevant</td><td>93.76</td><td>irrelevant</td><td>6.30</td><td>because</td></tr><tr><td>crap</td><td>98.91</td><td>but</td><td>86.00</td><td>placement</td><td>90.06</td><td>unless</td><td>5.91</td><td>never</td></tr><tr><td>pacing</td><td>97.94</td><td>these</td><td>84.96</td><td>warmed</td><td>71.65</td><td>whether</td><td>2.78</td><td>despite</td></tr><tr><td>whether</td><td>97.03</td><td>wherein</td><td>78.49</td><td>behalf</td><td>19.60</td><td>failure</td><td>1.11</td><td>ball</td></tr><tr><td>every</td><td>96.81</td><td>weeks</td><td>59.58</td><td>foremost</td><td>17.79</td><td>devoid</td><td>0.61</td><td>notice</td></tr></table>
+
+Table 6: PMI for the joint training and validation set of SST-2, N - Negative, P - Positive  
+
+<table><tr><td rowspan="2">N</td><td colspan="2">BERT</td><td rowspan="2">Rank</td><td rowspan="2">N</td><td colspan="2">ALBERT</td><td rowspan="2">Rank</td><td rowspan="2">N</td><td colspan="2">RoBERTa</td></tr><tr><td>Rank E</td><td>Rank</td><td>Rank E</td><td>Rank</td><td>Rank E</td><td>Rank</td></tr><tr><td>succeeded</td><td>97.09</td><td>never</td><td>97.74</td><td>ashamed</td><td>95.06</td><td>after</td><td>99.48</td><td>waves</td><td>95.42</td><td>fire</td></tr><tr><td>seeing</td><td>95.94</td><td>least</td><td>96.50</td><td>stumbling</td><td>87.91</td><td>under</td><td>99.45</td><td>aside</td><td>95.42</td><td>against</td></tr><tr><td>merits</td><td>88.23</td><td>killing</td><td>96.19</td><td>tire</td><td>87.91</td><td>because</td><td>98.89</td><td>harming</td><td>88.06</td><td>water</td></tr><tr><td>longest</td><td>88.23</td><td>prison</td><td>95.58</td><td>cutting</td><td>77.42</td><td>since</td><td>98.27</td><td>skepticism</td><td>88.06</td><td>health</td></tr><tr><td>feeling</td><td>78.82</td><td>much</td><td>94.49</td><td>declined</td><td>76.26</td><td>how</td><td>97.70</td><td>burns</td><td>88.06</td><td>pleaded</td></tr><tr><td>raped</td><td>77.95</td><td>welcome</td><td>94.04</td><td>boise</td><td>70.31</td><td>whether</td><td>97.51</td><td>survives</td><td>88.06</td><td>nothing</td></tr><tr><td>batting</td><td>77.95</td><td>behind</td><td>93.65</td><td>fury</td><td>66.42</td><td>should</td><td>97.12</td><td>votes</td><td>76.46</td><td>won</td></tr><tr><td>jail</td><td>74.66</td><td>born</td><td>92.60</td><td>slay</td><td>55.80</td><td>cost</td><td>96.57</td><td>justice</td><td>70.62</td><td>NATO</td></tr><tr><td>banning</td><td>70.64</td><td>needed</td><td>92.26</td><td>reactions</td><td>52.67</td><td>serious</td><td>96.38</td><td>ressing</td><td>65.60</td><td>suffering</td></tr><tr><td>backing</td><td>70.64</td><td>nothing</td><td>91.30</td><td>disapprove</td><td>52.67</td><td>guilty</td><td>96.31</td><td>deceived</td><td>48.45</td><td>happens</td></tr></table>
+
+Table 7: PMI for the joint training and validation set of MRPC, N – not equivalent, E – equivalent  
+
+<table><tr><td></td><td>MRPC Trigger</td><td>l</td><td>QNLI Trigger</td><td>l</td><td>SST-2 Trigger</td><td>l</td></tr><tr><td>A → B</td><td>they ended and</td><td>10</td><td>preis much $100,000</td><td>3</td><td>unsuitable improper whether</td><td>10</td></tr><tr><td>A → R</td><td>rostov she blushed</td><td>1</td><td>$100,000 what simulate</td><td>2</td><td>unsuitable improper whether</td><td>10</td></tr><tr><td>B → A</td><td>nothing pains kilograms</td><td>1</td><td>whichever thirds lithuanian</td><td>9</td><td>folding worse as</td><td>5</td></tr><tr><td>B → R</td><td>suffered sins declined</td><td>11</td><td>ibly semester longest</td><td>8</td><td>folding worse as</td><td>5</td></tr><tr><td>R → A</td><td>Avoid water taps</td><td>10</td><td>WHY voted beyond</td><td>2</td><td>decayingjuryNeither</td><td>1</td></tr><tr><td>R → B</td><td>Avoid water taps</td><td>10</td><td>WHY voted beyond</td><td>2</td><td>surprisingly refreshing lest</td><td>11</td></tr></table>
+
+Table 8: The best-performed trigger-layer pairs for model-to-model transferability results presented in Tab. 4 for ALBERT (A), BERT (B), RoBERTa (R), where  $l$  - the perturbed layer.
+
+<table><tr><td></td><td colspan="2">ALBERT</td><td colspan="2">BERT</td><td colspan="2">RoBERTa</td></tr><tr><td></td><td>Trigger</td><td>l</td><td>Trigger</td><td>l</td><td>Trigger</td><td>l</td></tr><tr><td>M → Q</td><td>nobody reminds austro</td><td>8</td><td>is blood corrupt</td><td>9</td><td>history doubtless beyond</td><td>8</td></tr><tr><td>M → S</td><td>failing forcing the</td><td>4</td><td>ching stiff punishments</td><td>8</td><td>history doubtless beyond</td><td>8</td></tr><tr><td>Q → M</td><td>or widen further</td><td>10</td><td>whichever thirds lithuanian</td><td>9</td><td>WHY voted beyond</td><td>2</td></tr><tr><td>Q → S</td><td>trillion unless marylebone</td><td>5</td><td>ant trees romanized</td><td>5</td><td>wherein perished supra</td><td>7</td></tr><tr><td>S → M</td><td>unsuitable improper whether</td><td>10</td><td>seemed wiped whoever</td><td>9</td><td>Eating welcome respecting</td><td>7</td></tr><tr><td>S → Q</td><td>renumbered littered neither</td><td>8</td><td>ogarily diminished</td><td>4</td><td>Crystal tasty ain</td><td>9</td></tr></table>
+
+Table 9: The best-performed trigger-layer pairs for task-to-task transferability results presented in Tab. 5 for MRPC (M), QNLI (Q), SST-2 (S), where  $l$  - the perturbed layer.  
+
+<table><tr><td>Model</td><td>L</td><td>FR Val.</td><td>FR Test</td><td>Trigger</td></tr><tr><td>UAT, BERT</td><td>2</td><td>38.6</td><td>38.4</td><td>neither nor</td></tr><tr><td>UAT, BERT</td><td>4</td><td>35.4</td><td>35.0</td><td>situation nonetheless resulted nor</td></tr><tr><td>Ours, ALBERT, q = 7, i = 4</td><td>5</td><td>24.8</td><td>24.8</td><td>regretted joyahbwv although doubted</td></tr><tr><td>Ours, ALBERT, q = 4, i = 8</td><td>6</td><td>21.4</td><td>21.4</td><td>tremendous despair towedtrue 1985, doubted</td></tr></table>
+
+Table 10: The examples of cases when a shorter-length trigger is more successful than a longer one. In both cases, triggers were fitted on MNLI.  
+
+<table><tr><td>Dataset</td><td>Model</td><td>Triggers</td></tr><tr><td rowspan="3">MNLI</td><td>ALBERT</td><td>hamas doubted; neither motioned; i doubted pronoun; cursing neutron unless; transportation workers unless preferring;</td></tr><tr><td>BERT</td><td>but neither; remarkably neither; of neither nor; get neither outta even; backdrop but and neither; ft still maintained not whether</td></tr><tr><td>RoBERTa</td><td>slideshow Show Neither; HELL Neither; THERE HAS NEVER; ohan Never nor; Fifa VERY NEVER Whether</td></tr><tr><td rowspan="3">MRPC</td><td>ALBERT</td><td>unless if; whereas and; they ended and; unless stumbling out and; commerce corroborat declined and; and siblings live because</td></tr><tr><td>BERT</td><td>against attempt; prepares worse conflicts; was priesthood killing contrary; each mortally duel harta; ring opposed her homosexuality</td></tr><tr><td>RoBERTa</td><td>Drop until; actresses won awards; Enhanced Population Died Low; Harlem Recommended Submit BLM Question; hum DOWN burns firing Bloody Hyp</td></tr><tr><td rowspan="3">QNLI</td><td>ALBERT</td><td>and 40%; averaged percentage; 88% finance 11%; vittorio whom relinquished; cristo, sharing whom reissued; downloadable bilingual why cancellation</td></tr><tr><td>BERT</td><td>nine charities; carlos orient whom; whose reich declaring; gor german which emperor; maya commuted whose ballet</td></tr><tr><td>RoBERTa</td><td>nineteen countries; Nearly trillions trillion; Thousand hundred trillion; how MUCH mountains Cass; gui why sued awaits Reviews;</td></tr><tr><td rowspan="3">SST-2</td><td>ALBERT</td><td>flight unless; suck unnecessary; skidded irrelevant whether; failure placement unless; just whine worthless; fails subdistrict picture</td></tr><tr><td>BERT</td><td>degraded whether; dissatisfied neither; television failed whether; definitely worst wherein whoever; crap stupid feed whereby whether</td></tr><tr><td>RoBERTa</td><td>powerfully refreshing; Beautiful enjoyable; surprisingly refreshing lest; Crystal importantly beautiful considering; thriller cool because whereas</td></tr></table>
+
+Table 11: WARNING: THE CONTENT OF THIS TABLE MIGHT BE OFFENSIVE, AND IT DOES NOT REFLECT THE AUTHORS' OPINION. The examples of LUATs depending on a dataset and a model. Triggers were selected manually from top-20 per length.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? 6  
+A2. Did you discuss any potential risks of your work? 5  
+A3. Do the abstract and introduction summarize the paper's main claims?  
+A4. Have you used AI writing assistants when working on this paper? Grammarly for check spelling
+
+B Did you use or create scientific artifacts?
+
+3
+
+B1. Did you cite the creators of artifacts you used? 3  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? We used freely available models and dataset  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+We did not discuss the indented use of the models and datasets which we used, since these are the models and datasets are known to everyone and their use is free.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+We did not collect any data, and we used only publicly available datasets.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+We did not collect any data. The documentation for the datasets and models which we are used are publicly available.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+3
+
+C Did you run computational experiments?
+
+3
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+3
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+3, Appendix B
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+3, Appendix B
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+3
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/Layerwise universal adversarial attack on NLP models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ce365bdd7dd84affaa26d810a03353fcc62c9a694a295fe55c8daaa76b43788
+size 1265519

2023/Learn to Not Link_ Exploring NIL Prediction in Entity Linking/980725a0-e77c-4bea-bb00-a61ff06440a3_content_list.json

@@ -0,0 +1,2606 @@
+[
+    {
+        "type": "text",
+        "text": "Learn to Not Link: Exploring NIL Prediction in Entity Linking",
+        "text_level": 1,
+        "bbox": [
+            169,
+            90,
+            828,
+            112
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Fangwei Zhu $^{1,2}$ , Jifan Yu $^{1,2}$ , Hailong Jin $^{1,2}$ , Juanzi Li $^{1,2}$ , Lei Hou $^{1,2*}$  and Zhifang Sui $^{3}$",
+        "bbox": [
+            132,
+            124,
+            870,
+            142
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Dept. of Computer Sci.&Tech., BNRist, Tsinghua University, Beijing 100084, China  $^{2}$ KIRC, Institute for Artificial Intelligence, Tsinghua University",
+        "bbox": [
+            154,
+            143,
+            848,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$ MOE Key Lab of Computational Linguistics, Peking University {zfw19@mails.,yujf18@mails,jinhl15@mails.,lixuanzi@,houlei@}tsinghua.edu.cn {szf@}pku.edu.cn",
+        "bbox": [
+            122,
+            175,
+            880,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            342,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Entity linking models have achieved significant success via utilizing pretrained language models to capture semantic features. However, the NIL prediction problem, which aims to identify mentions without a corresponding entity in the knowledge base, has received insufficient attention. We categorize mentions linking to NIL into Missing Entity and Non-Entity Phrase, and propose an entity linking dataset NEL that focuses on the NIL prediction problem. NEL takes ambiguous entities as seeds, collects relevant mention context in the Wikipedia corpus, and ensures the presence of mentions linking to NIL by human annotation and entity masking. We conduct a series of experiments with the widely used bi-encoder and cross-encoder entity linking models, results show that both types of NIL mentions in training data have a significant influence on the accuracy of NIL prediction. Our code and dataset can be accessed at https://github.com/solitaryzero/NIL_EL.",
+        "bbox": [
+            141,
+            281,
+            460,
+            595
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            609,
+            260,
+            626
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Entity Linking (EL) aims to map entity mentions in free texts to their corresponding entities in a given Knowledge Base (KB). Entity linking acts as a bridge between unstructured text and structured knowledge, and benefits various downstream tasks like question answering (Luo et al., 2018) and knowledge extraction (Chen et al., 2021).",
+        "bbox": [
+            112,
+            637,
+            489,
+            749
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, not all entity mentions correspond to a specific entity in the KB that suits the mention context (Ling et al., 2015). Take Table 1 as an example, Peter Blackburn is actually a journalist, and the householder is a common phrase rather than an entity. These two mentions do not refer to any entity in the given KB. The identification of these mentions is referred to as NIL prediction. Therefore, to tackle NIL prediction, the entity linking model",
+        "bbox": [
+            112,
+            750,
+            489,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "needs to select mentions whose references are absent in KB, and link them to a special placeholder  $\\text{NIL}$ . Dredze et al. (2010) states NIL prediction as one of the key issues in entity linking, which may lead to a decrease in the recall of entity linking systems. Meanwhile, the incorrectly linked entities may provide false information to downstream tasks.",
+        "bbox": [
+            507,
+            253,
+            884,
+            381
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "There have been some earlier representation learning based researches that take NIL prediction into consideration (Eshel et al., 2017; Lazic et al., 2015; Peters et al., 2019). They identify NIL mentions by setting a vector similarity threshold or viewing NIL as a special entity. Recently, pretrained language model (PLM) based models (Wu et al., 2020; Fitzgerald et al., 2021; Cao et al., 2021) have achieved great success for their great transferability and expandability. However, these models generally assume that there always exists a correct entity for each mention in the knowledge base, which leaves the NIL prediction problem without adequate attention.",
+        "bbox": [
+            507,
+            382,
+            884,
+            608
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Previous entity linking datasets have paid insufficient attention to the NIL prediction problem. For example, some of the previous datasets like AIDA (Hoffart et al., 2011) view it as an auxiliary task, while others like MSNBC (Cucerzan, 2007) and WNED-WIKI (Eshel et al., 2017) does not require NIL prediction at all. There does not yet exist a strong benchmark for the ability on NIL prediction.",
+        "bbox": [
+            507,
+            611,
+            885,
+            755
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose an entity linking dataset NEL that focuses on the NIL prediction problem. About  $30\\%$  of the mentions in NEL do not have their corresponding entity in the candidate entity set, which requires models to identify these mentions rather than linking them to the wrong candidates. In NEL construction, we take ambiguous entities as seeds, and build the dataset by mining mention contexts related to seed entities on the Wikipedia corpus. Then, human annotators are",
+        "bbox": [
+            507,
+            758,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding Author",
+        "bbox": [
+            141,
+            904,
+            299,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10846",
+        "bbox": [
+            475,
+            927,
+            527,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 10846-10860 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/1a99c6ce58e4cb52fa62d5e9d169215736e0e7716ce4b9566108caf60b62e844.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Missing Entity</td></tr><tr><td>Mention Context</td><td>EU rejects German call to boycott British lamb. Peter Blackburn BRUS-SELS 1996-08-22</td></tr><tr><td>Peter Blackburn (Bishop)</td><td>Peter Blackburn (d.1616) was a Scottish scholar and prelate. He was the second Protestant Bishop of Aberdeen.</td></tr><tr><td>Peter Blackburn (MP)</td><td>Peter Blackburn (1811 – 20 May 1870) was a British Conservative Party politician.</td></tr><tr><td>Peter Blackburn (Badminton)</td><td>Peter Grant Blackburn (born 25 March 1968) is an Australian badminton player who affiliated with the Ballarat Badminton Association.</td></tr><tr><td colspan=\"2\">Non-Entity Phrase</td></tr><tr><td>Mention Context</td><td>Most Hindus accept that there is a duty to have a family during the house-holder stage of life, as debt to family lineage called Pitra Rin (Father&#x27;s Debt) and so are unlikely to avoid having children altogether ...</td></tr><tr><td>The Householder (Film)</td><td>The Householder (Hindi title: Gharbar) is a 1963 film by Merchant Ivory Productions, with a screenplay by Ruth Prawer Jhabvala ...</td></tr><tr><td>The Householder (Novel)</td><td>The Householder is a 1960 English-language novel by Ruth Prawer Jhab-vala ...</td></tr></table>",
+        "bbox": [
+            176,
+            80,
+            823,
+            368
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Example of mentions that should be linked to NIL and their potential candidate entities. Mentions are labeled as red.",
+        "bbox": [
+            112,
+            376,
+            882,
+            406
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "asked to identify whether the mentions correspond to a candidate entity or not, and we further perform entity masking to ensure a fair proportion of NIL data of about  $30\\%$ .",
+        "bbox": [
+            112,
+            432,
+            487,
+            495
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In NIL prediction, we propose to use the widely used bi-encoder and cross-encoder structures as the model backbone, and further integrate type information by adding an entity typing subtask. We combine semantic and type similarity as the final similarity score, and identify NIL mentions by setting a similarity threshold.",
+        "bbox": [
+            112,
+            498,
+            489,
+            609
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conduct a series of experiments on both NEL and previous entity linking datasets. Experimental results show that the models suffer from an accuracy drop when taking NIL prediction into consideration, indicating that the accuracy may be inflated without the NIL prediction task, and NEL could better diagnose the performance of different models. We also conducted ablation studies on how type information and NIL examples affect the models. We discover that the entity typing subtask yields better embedding even when type similarity is not used, and both types of NIL examples in training data would boost the ability of NIL prediction.",
+        "bbox": [
+            112,
+            612,
+            489,
+            820
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions can be concluded as:",
+        "bbox": [
+            131,
+            822,
+            423,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We categorize the NIL prediction problem into two patterns: Missing Entity and Non-Entity Phrase, where the latter one has not received sufficient attention in previous works.",
+        "bbox": [
+            134,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose an entity linking dataset NEL focusing on NIL prediction, which covers two patterns of NIL data and could act as a benchmark for diagnosing the ability of NIL prediction.",
+            "- We conducted a series of experiments, whose results demonstrate that the accuracy of models may be inflated when not taking NIL prediction into consideration. Meanwhile, both patterns of NIL data in training are essential for triggering the ability of NIL prediction."
+        ],
+        "bbox": [
+            531,
+            432,
+            884,
+            623
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            507,
+            653,
+            648,
+            669
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Entity mentions  $M = \\{m_i\\}$  refer to text spans potentially corresponding to entities in a document  $D = (w_1, w_2, \\dots, w_n)$ , where  $w_i$  is either a plain token or a mention. Each mention  $m_i$  may correspond to an entity  $e_i \\in E$  in the entity set  $E$  of knowledge base  $\\mathcal{KB}$ .",
+        "bbox": [
+            507,
+            680,
+            884,
+            777
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Definition 1. Entity linking aims to find an optimal mapping  $\\Gamma : M \\Rightarrow E$ , which maps entity mentions  $M = \\{m_i\\}$  to their corresponding entities  $E = \\{e_i\\}$ , where  $e_i \\in \\mathcal{KB}$ .",
+        "bbox": [
+            507,
+            783,
+            882,
+            848
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The NIL prediction problem is to determine whether an entity mention  $m$  is absent from the knowledge base  $\\mathcal{KB}$ . When there does not exist a proper entity  $e \\in \\mathcal{KB}$  for the given mention  $m$ ,",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10847",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/6e37f93f68c99d676297e6099dddeae0ffc36f5d7c8fa5bda4d4f8b1ea8660ed.jpg",
+        "image_caption": [
+            "Figure 1: An illustration of how NEL is constructed. We select ambiguous entities as seeds, taking their alias as potential mentions to discover mention contexts from Wikipedia. The entries are annotated by human annotators, and entity masking is performed on some entries to control the portion of NIL data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            166,
+            93,
+            836,
+            354
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the model should link  $m$  to a special NIL entity, indicating that the mention is unlinkable.",
+        "bbox": [
+            112,
+            435,
+            487,
+            467
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definition 2. Entity linking with NIL prediction aims to find an optimal mapping  $\\Gamma : M \\Rightarrow E \\cup \\{NIL\\}$ , where NIL is a special placeholder and  $m$  correspond to NIL only when there is no correct answer in the candidate entity set  $E$ .",
+        "bbox": [
+            112,
+            470,
+            485,
+            551
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As demonstrated in Table 1, there exist two situations in real-world entity linking where NIL prediction should be taken into consideration:",
+        "bbox": [
+            112,
+            554,
+            489,
+            602
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Missing Entity: The mention  $m$  refers to certain entity  $e$  that has not been yet included in  $\\mathcal{KB}$ , i.e.  $m \\Rightarrow e \\notin \\mathcal{KB}$ . For example, in the upper half of Table 1, the mention Peter Blackburn refers to a certain journalist, while entries in English Wikipedia include only people with other occupations, leading to the mention linking to NIL.",
+            "- Non-Entity Phrase: The mention  $m$  refers to a common phrase that is not usually viewed as an entity, i.e.  $m \\nRightarrow e$ . For example, the mention the householder in the lower half of Table 1 refers to a concept rather than a film or novel."
+        ],
+        "bbox": [
+            134,
+            609,
+            489,
+            843
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Dataset Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            862,
+            332,
+            878
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "There does not yet exist a strong benchmark on NIL prediction. We manually annotated 300 examples",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "with their mentions linking to NIL from the widely used entity linking dataset AIDA<sup>1</sup>, and discover that about  $10\\%$  of these mentions should actually link to an entity. For example, the mention \"EU\" in \"EU rejects German\" should be linked to the European Union rather than NIL (See Appendix D for details). Meanwhile, NIL mentions in AIDA fall mostly in the Missing Entity category. The incorrect and imbalanced data for NIL prediction indicates that the importance of NIL prediction is currently underestimated.",
+        "bbox": [
+            507,
+            435,
+            882,
+            611
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we propose an entity linking dataset named NEL, which focuses on the NIL prediction problem. The construction process is demonstrated in Figure 1.",
+        "bbox": [
+            507,
+            614,
+            882,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Unlike normal entity linking data, there does not exist annotated references for mentions linking to NIL, and the portion of NIL data in the text corpus is unknown. Hyperlinks in Wikipedia can be viewed as mentions linking to non-NIL entities, from which we can find the aliases of entities. We assume that when an alias does not appear as a hyperlink in a certain context, it may be identified as a mention linking to NIL. In this way, we collect such contexts as the raw data. The raw data is then annotated by humans to ensure correctness, and we further mask out some answer entities in the candidate set to control the percentage of NIL in",
+        "bbox": [
+            507,
+            680,
+            884,
+            889
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1",
+        "bbox": [
+            531,
+            903,
+            539,
+            912
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "footer",
+        "text": "10848",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/951bdfcf39f66e5a4110eb4f17884f074f7181408f2d812ac49f7014e5ad40b9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td># Data</td><td>Annotated</td><td>NIL percentage</td><td>%Missing Entity</td><td>%Non-Entity Phrase</td></tr><tr><td>AIDA</td><td>34956</td><td>✓</td><td>20.41%</td><td>73%*</td><td>10%*</td></tr><tr><td>MSNBC</td><td>654</td><td>✓</td><td>0%</td><td>-</td><td>-</td></tr><tr><td>WNED-Wiki</td><td>240000</td><td>×</td><td>0%</td><td>-</td><td>-</td></tr><tr><td>NEL (ours)</td><td>9924</td><td>✓</td><td>33.57%</td><td>17%</td><td>83%</td></tr></table>",
+        "bbox": [
+            163,
+            80,
+            833,
+            177
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Statistics of the NEL dataset compared with previous entity linking datasets. *The percentage of two NIL patterns in AIDA is calculated from 300 randomly sampled NIL data, and data with errors do not count as any pattern.",
+        "bbox": [
+            112,
+            185,
+            884,
+            231
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "answers.",
+        "bbox": [
+            112,
+            255,
+            183,
+            269
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Data Collection",
+        "text_level": 1,
+        "bbox": [
+            112,
+            282,
+            282,
+            296
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Levin (1977) states that the title of creative works could be a place, a personal name, or a certain abstract concept like the choric embodiment of some collectivity (The Clouds, The Birds) or stock types (The Alchemist, Le Misanthrope), which would naturally lead to the two situations where a mention links to NIL. The absence of the referenced entity from the KB would lead to Missing Entity, while an abstract concept not viewed as an entity would lead to Non-Entity Phrase.",
+        "bbox": [
+            112,
+            303,
+            487,
+            463
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To discover NIL mentions of both types, we start by taking entities that share an alias with other entities as seeds. We assume that a mention referring to multiple entities has a higher probability of linking to a Missing Entity outside the KB, and the complex meaning of the mention will lead to Non-Entity Phrase. Thus, the aliases of ambiguous seed entities would be good starting points for mining NIL mentions.",
+        "bbox": [
+            112,
+            464,
+            489,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Entity Selection We further filter ambiguous entities to remove low-quality seeds. First, we remove noise instances like template pages, and entities with less than 5 hyperlinks are also removed. Meanwhile, we discarded entities with a probability greater than  $50\\%$  of being the linking result, as these entities can generally be considered to be unambiguous and lack difficulty. Finally, 1000 entities are sampled as the seed entity set  $E_{s}$ .",
+        "bbox": [
+            112,
+            619,
+            489,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use a typing system based on Wikidata to identify the type of selected entities. We view the instance of relation as the type indicator, and utilize the subclass of relation to build a tree-form type system. The detailed type system can be found in Appendix C.",
+        "bbox": [
+            112,
+            765,
+            487,
+            860
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Mention Discovery We build an alias table from the 2021 Wikipedia dump by extracting alias-entity pairs  $(m,e)$  from internal hyperlinks. All alias",
+        "bbox": [
+            112,
+            870,
+            487,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$m$  related to entities in the seed entity set  $E_{s}$  are gathered as the mention set  $M$ . For each mention  $m \\in M$ , we look for its occurrence throughout the Wikipedia corpus (whether it appears as a hyperlink or not) to obtain the entry tuple  $(C_{l}, m, C_{r}, E_{m})$ , where  $C_{l}$  and  $C_{r}$  represent contexts left and right to the entity mention  $m$ , and  $E_{m}$  represents the candidate entities set of  $m$ . For each mention  $m$ , we sampled 5 entries where  $m$  appears as a hyperlink and 5 entries where  $m$  appears in plain text to balance the number of positive and negative examples, and a total of 10,000 entries are collected.",
+        "bbox": [
+            507,
+            254,
+            884,
+            447
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Human Annotation and Post-processing",
+        "text_level": 1,
+        "bbox": [
+            507,
+            461,
+            870,
+            476
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We perform annotation on the above entries with 3 annotators. The annotators are provided with the mention context  $(C_l, m, C_r)$  and candidate entities  $E_m$ . Each candidate entity  $e$  consists of its title, textual description, and Wikipedia URL. The annotators are asked to choose the answer entity  $a \\in E_m$  corresponding to the mention  $m$ , or  $a = NIL$  if none of the candidates are correct.",
+        "bbox": [
+            507,
+            483,
+            882,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "An expert will further investigate entries in which annotators fail to reach a consensus. The expert is a senior annotator with essential knowledge of entity linking, and will confirm the final annotation result after carefully reading through the context and candidate entities. We use the annotation result as the final answer  $a$  if there is an agreement between 3 annotators, and the entity chosen by the expert otherwise. The annotated tuple  $(C_l,m,C_r,E_m,a)$  acts as the final entry of our dataset.",
+        "bbox": [
+            507,
+            613,
+            882,
+            788
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To further simulate the situation where new emerging entities do not appear in knowledge bases, we perform entity masking on positive entries. We randomly sample  $10\\%$  entries where  $a \\neq NIL$ , and mask the correct entity in the candidate set  $E_{m}$ . In this case, as the correct answer is removed from the candidate list, we have  $a = NIL$ , i.e. the mention  $m$  corresponds to the empty entity NIL.",
+        "bbox": [
+            507,
+            790,
+            882,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10849",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/187d4d5e1d3f165d1a370301599dc08dc9dc0509daa8512a564f8d0ea2efee08.jpg",
+        "image_caption": [
+            "Figure 2: The overall structure of PLM-based retrieval models. Candidates which are confusing in the semantic space may be more distinguishable in the type space. Mentions linking to NIL frequently differ from their candidates in their types, so we combine semantic similarity with type similarity for NIL prediction."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            168,
+            93,
+            505,
+            338
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/aad5c3b877f42419d4829cd65c3ad0ec9b9b506b272f43060a25e453af09d4b7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            93,
+            826,
+            313
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Dataset Statistics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            435,
+            294,
+            450
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 2 demonstrates the properties of the NEL dataset. NEL includes 6,593 positive examples and 3,331 negative examples, covering 1,000 mentions and 3,840 related entities. Each mention has an average of 3.80 candidate entities. The inter-annotator agreement of NEL is  $94.61\\%$ , indicating that the expert calibrated about  $5\\%$  of the data. The full dataset is split into train/validation/test sets by the ratio of  $80\\%/10\\%/10\\%$ .",
+        "bbox": [
+            112,
+            457,
+            487,
+            601
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "NEL contains a fair number of entries, and is human-annotated to ensure correctness. Compared with previous entity linking datasets, NEL has a higher percentage of NIL data, which could better diagnose the ability of different models in NIL prediction. Meanwhile, mentions linking to NIL in AIDA mostly fall in the Missing Entity situation, while NEL focuses more on the Non-Entity Phrase situation, thus complementing the insufficient attention on Non-Entity Phrase in NIL prediction.",
+        "bbox": [
+            112,
+            602,
+            489,
+            778
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Entity Linking with NIL prediction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            794,
+            453,
+            812
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A mention links to NIL when all of its candidate entities fail to match its context. A common paradigm of NIL prediction is to compute the similarity scores between the mention context and candidates, and judge its linking result on the base of similarities.",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Scoring Similarities",
+        "text_level": 1,
+        "bbox": [
+            507,
+            435,
+            710,
+            451
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Bi-encoder and cross-encoder are widely adopted scorer structures, as they are well compatible with pretrained language models. Bi-encoder encodes mention contexts and entities into the same dense vector space, while cross-encoder views similarity scoring as a sequence classification task:",
+        "bbox": [
+            507,
+            455,
+            882,
+            552
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ns _ {b i} (c, e) = \\sigma (f (c) \\cdot g (e)) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            565,
+            882,
+            582
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ns _ {c r o s s} (c, e) = \\sigma (\\mathbf {W h} ([ c, e ]) + \\mathbf {b}) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            571,
+            585,
+            880,
+            602
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $f,g,h$  are PLM-based encoders,  $\\mathbf{W}$  and  $\\mathbf{b}$  are trainable variables, and  $\\sigma$  refers to the sigmoid function.",
+        "bbox": [
+            507,
+            614,
+            880,
+            659
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The bi-encoder structure allows precomputing entity embeddings in knowledge bases, which enables efficient retrieval in real-world applications. Compared with bi-encoder, cross-encoder better captures the relation between context and entities with the cross attention mechanism, thus demonstrating higher accuracy.",
+        "bbox": [
+            507,
+            663,
+            882,
+            776
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Integrating Entity Types",
+        "text_level": 1,
+        "bbox": [
+            507,
+            785,
+            747,
+            802
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Previous entity linking models (Gupta et al., 2017; Onoe and Durrett, 2020; Raiman and Raiman, 2018) have proved that entity types do help models better disambiguate between candidate entities.",
+        "bbox": [
+            507,
+            806,
+            882,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The type information can be integrated into bi-encoders and cross-encoders by adding a typing layer. In the bi-encoder structure, the mention types",
+        "bbox": [
+            507,
+            871,
+            882,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10850",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/9da85eddf9b2092b5ec15350202675588b5999e429f3b92902a41f5b42b32ea0.jpg",
+        "table_caption": [
+            "Table 3: Experimental results on NEL and previous datasets. Non-NAC, NAC, and OAC represent non-NIL accuracy, NIL accuracy and overall accuracy. *Results of GENRE on AIDA w/o NIL, MSNBC, and WNED-WIKI are taken from the original paper (Cao et al., 2021)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"3\">NEL (our dataset)</td><td colspan=\"3\">AIDA w/ NIL</td><td>AIDA w/o NIL</td><td>MSNBC</td><td>WNED-WIKI</td></tr><tr><td></td><td>Non-NAC</td><td>NAC</td><td>OAC</td><td>Non-NAC</td><td>NAC</td><td>OAC</td><td>OAC</td><td>OAC</td><td>OAC</td></tr><tr><td>BLINK-bi</td><td>72.27</td><td>88.59</td><td>77.74</td><td>64.54</td><td>69.59</td><td>65.01</td><td>82.61</td><td>70.86</td><td>58.56</td></tr><tr><td>CLINK-bi</td><td>79.24</td><td>79.28</td><td>79.25</td><td>75.98</td><td>66.36</td><td>75.09</td><td>83.26</td><td>73.29</td><td>58.99</td></tr><tr><td>GENRE*</td><td>54.00</td><td>62.84</td><td>56.96</td><td>-</td><td>-</td><td>-</td><td>88.60</td><td>88.10</td><td>71.70</td></tr><tr><td>BLINK-cross</td><td>84.09</td><td>88.89</td><td>85.70</td><td>83.08</td><td>45.16</td><td>79.58</td><td>87.49</td><td>82.02</td><td>69.48</td></tr><tr><td>CLINK-cross</td><td>86.97</td><td>89.19</td><td>87.71</td><td>84.42</td><td>58.53</td><td>82.03</td><td>88.16</td><td>89.70</td><td>72.43</td></tr></table>",
+        "bbox": [
+            134,
+            135,
+            858,
+            247
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$t_c$  and entity types  $t_e$  are predicted separately:",
+        "bbox": [
+            112,
+            269,
+            455,
+            286
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nt _ {c} = \\sigma \\left(W _ {c} f (c) + b _ {c}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            216,
+            296,
+            487,
+            313
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nt _ {e} = \\sigma \\left(W _ {e} g (e) + b _ {e}\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            218,
+            316,
+            485,
+            332
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "while they are simultaneously predicted in the cross-encoder structure:",
+        "bbox": [
+            112,
+            344,
+            485,
+            375
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left[ t _ {c}, t _ {e} \\right] = \\sigma \\left(W f ([ c, e ]) + b\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            194,
+            387,
+            487,
+            405
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\sigma$  represents the sigmoid function and  $W_{c}, b_{c}, W_{e}, b_{e}, W, b$  are trainable parameters.",
+        "bbox": [
+            112,
+            414,
+            485,
+            445
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To tackle the label imbalance between types, we use the focal loss (Lin et al., 2017) on the typing task:",
+        "bbox": [
+            112,
+            447,
+            485,
+            494
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {t} = - \\sum_ {i = 1} ^ {n _ {t}} \\left(y _ {i} \\left(1 - t _ {i}\\right) ^ {\\gamma} \\log t _ {i} + \\left(1 - y _ {i}\\right) t _ {i} ^ {\\gamma} \\log \\left(1 - t _ {i}\\right)\\right) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            502,
+            494,
+            555
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $n_t$  is the total number of types in the type system,  $\\gamma$  is a hyperparameter,  $y_i$  is the golden label of the  $i$ -th type and  $t_i$  is the predicted label of the  $i$ -th type. In bi-encoder,  $\\mathcal{L}_t$  is the average of loss on  $t_c$  and  $t_e$ , while in cross-encoder  $\\mathcal{L}_t$  is directly calculated from  $[t_c, t_e]$ .",
+        "bbox": [
+            112,
+            556,
+            487,
+            652
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We train the semantic encoder with binary classification loss  $\\mathcal{L}_s$ , and combine  $\\mathcal{L}_s$  with  $\\mathcal{L}_t$  as the final loss  $\\mathcal{L}$ :",
+        "bbox": [
+            112,
+            653,
+            487,
+            699
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\mathcal {L} _ {t} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            248,
+            701,
+            487,
+            717
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Identifying Mentions Linking to NIL",
+        "text_level": 1,
+        "bbox": [
+            112,
+            727,
+            448,
+            743
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The type similarity is computed with cosine similarity, and the final score is a weighted sum between type similarity and semantic similarity:",
+        "bbox": [
+            112,
+            747,
+            487,
+            796
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ns _ {t} (c, e) = \\cos \\left(t _ {c}, t _ {e}\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            159,
+            807,
+            487,
+            824
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ns (c, e) = \\lambda s _ {s} (c, e) + (1 - \\lambda) s _ {t} (c, e) \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            826,
+            487,
+            844
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda$  is a hyperparameter.",
+        "bbox": [
+            112,
+            854,
+            331,
+            870
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For each entry  $(C_l, m, C_r, E_m, a)$ , we concatenate  $(C_l, m, C_r)$  to form the context  $c$ . In the training step, for each candidate entity  $e \\in E_m$ , we",
+        "bbox": [
+            112,
+            871,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "view  $(c,e)$  as a positive example if  $e = a$  , and as a negative example if  $a = N I L$  or  $e\\neq a$",
+        "bbox": [
+            507,
+            269,
+            880,
+            300
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "During evaluation, the similarity score  $s(c, e)$  is computed between context  $c$  and each candidate entity  $e$ . If there exist entities with a score equal to or higher than the nil threshold  $\\epsilon = 0.5$ , we choose the entity with the highest similarity score as the answer; If all entities fail to reach the threshold, then the mention  $m$  links to NIL.",
+        "bbox": [
+            507,
+            302,
+            882,
+            414
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\na = \\left\\{ \\begin{array}{l l} \\arg \\max  _ {e} s (c, e), & \\exists e, s (c, e) \\geq \\epsilon \\\\ N I L, & \\forall e, s (c, e) <   \\epsilon \\end{array} \\right. \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            426,
+            882,
+            469
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            499,
+            655,
+            514
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We conduct a series of experiments on two types of datasets, which test the different ability of entity linking models: (1) NEL that tests the ability of NIL prediction; (2) previous EL datasets that tests the ability of entity disambiguation. We choose the following models for comparison: BLINK (Wu et al., 2020) that uses the bi-encoder and cross-encoder alone to score candidate entities, CLINK that integrates type information with BLINK, and GENRE (Cao et al., 2021) that generates the linking result with a sequence-to-sequence model.",
+        "bbox": [
+            507,
+            527,
+            882,
+            703
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Main Experiment Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            719,
+            756,
+            734
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We trained and tested the models on NEL, with BERT-large as the encoder base of BLINK and CLINK, and BART-large as the backbone of GENRE, to observe their ability in NIL prediction. We also experimented on previous entity linking datasets to observe the disambiguation ability of different models. The models are trained on the AIDA-YAGO2-train (Hoffart et al., 2011) dataset, and tested on AIDA-YAGO2-testb, MSNBC (Cucerzan, 2007) and WNED-WIKI (Eshel et al., 2017).",
+        "bbox": [
+            507,
+            741,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10851",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/82921ec86baf05b23d1676f01eff4000cafe406a9812bd9484b28a4feaebc810.jpg",
+        "table_caption": [
+            "Table 4: Experimental results on the influence of the entity typing task on NEL. OAC indicates the overall accuracy of entity linking. The overall accuracy with typing is achieved without using the type similarity score."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Ctxt Type Acc.</td><td>Cand Type Acc.</td><td>OAC w/ Typing</td><td>OAC w/o Typing</td></tr><tr><td>Bi-encoder</td><td>83.75</td><td>93.46</td><td>78.35</td><td>77.74</td></tr><tr><td>Cross-encoder</td><td>83.95</td><td>98.01</td><td>87.41</td><td>83.67</td></tr></table>",
+        "bbox": [
+            154,
+            120,
+            838,
+            187
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/02e30405e73002b7378d2bc1a22777bb96a4c24cb2ba4a58180d370f3daabd8f.jpg",
+        "image_caption": [
+            "Figure 3: Ablation study on the influence of NIL training data. The x-axis indicates the percentage of used NIL data or Non-Entity Phrase data in the training set."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            166,
+            202,
+            391,
+            391
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/263bc244634f9c1b1445bb160d05b1135b3a88b60cd36aa8e8fea317e4080aac.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            391,
+            200,
+            610,
+            391
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/24005887099eaae4c374ea6b9eeba60e1897c3e3afef1fec1f89daf1fc351797.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            611,
+            202,
+            823,
+            391
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "187 distinct types are used in experiments on NEL, and considering that the entity type distribution may be different across datasets, we use a type system with only 14 top-level entity types on previous datasets to make CLINK more transferable (See Appendix C for details). We retain the same textual representation format with BLINK (see Appendix A), while using 128 as the length of context sequences and entity descriptions. All models are implemented with PyTorch and optimized with the Adam (Kingma and Ba, 2014) optimizer with a learning rate of 1e-5.",
+        "bbox": [
+            112,
+            456,
+            487,
+            648
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3 shows the evaluation results, from which some insights could be discovered:",
+        "bbox": [
+            112,
+            650,
+            485,
+            682
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Type Information Matters. CLINK with cross-encoder structure achieves the highest accuracy on almost all datasets, and is still comparable with GENRE on AIDA, from which we may assert that taking type information into consideration is helpful even without NIL prediction. Meanwhile, on both structures, the overall accuracy of CLINK outperforms BLINK on all datasets, proving that the entity typing task assists both bi-encoder and cross-encoder distinguish correct entities.",
+        "bbox": [
+            112,
+            696,
+            487,
+            857
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Encoder Structure. The cross-encoder structure generally performs better than bi-encoder, but we observe that sometimes bi-encoders show de",
+        "bbox": [
+            112,
+            871,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "cent ability in detecting NIL mentions, especially when type information is not utilized. BLINK-bi achieves the highest NIL accuracy score of 69.59 on AIDA with NIL, and has a score of 88.59 on NEL, which is comparable with the best-performing CLINK-cross. This phenomenon indicates that cross-encoders may be more prone to overfitting, while entity types would alleviate this tendency.",
+        "bbox": [
+            507,
+            456,
+            884,
+            601
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "NIL Entries. On the AIDA dataset, we observe that the models generally suffer from a drop in accuracy when taking NIL entries into consideration, and the drop is more obvious in bi-encoders. This may indicate that the performance of models is inflated without NIL prediction, and NIL entries may confuse the models in practical application.",
+        "bbox": [
+            507,
+            609,
+            882,
+            722
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            734,
+            675,
+            749
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2.1 Influence of the Entity Typing Task",
+        "text_level": 1,
+        "bbox": [
+            507,
+            755,
+            843,
+            771
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conducted experiments to observe how the entity typing task influences the model. We trained the model with both  $\\mathcal{L}_t$  and  $\\mathcal{L}_s$  as loss, while setting  $\\lambda = 1$  during evaluation to ignore the influence of type similarity scores. Results shown in Table 4 reflects two observations:",
+        "bbox": [
+            507,
+            774,
+            882,
+            869
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "First, candidate type predictions benefit from mention context. We observe that when changing the structure from bi-encoder to cross-encoder, the",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10852",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "type prediction accuracy on candidates raises by  $5\\%$ , where the accuracy on contexts remains unchanged. This is likely because the context helps narrow down the type range, while the context type prediction generally remains unaffected by entity descriptions.",
+        "bbox": [
+            112,
+            84,
+            487,
+            180
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Second, unifying the entity typing task in the training process leads to higher overall accuracy, even when the type similarity score is not taken into consideration in the evaluation, which can be demonstrated by the improved score of OAC w/ Typing compared to OAC w/o Typing on both models. This may indicate that predicting entity types would help the model learn semantic embeddings with higher quality.",
+        "bbox": [
+            115,
+            181,
+            489,
+            326
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2.2 Influence of NIL Training Data",
+        "text_level": 1,
+        "bbox": [
+            112,
+            338,
+            418,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Compared with previous datasets like AIDA, NEL contains more NIL mentions of the Non-Entity Phrase type. We trained the models with different numbers of Non-Entity Phrase examples to observe the influence of NIL Training Data.",
+        "bbox": [
+            112,
+            359,
+            487,
+            439
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As demonstrated by Figure 3, all models suffer from a great decline in NIL accuracy when no NIL examples are used during the training stage, and bi-encoder is more prone to the accuracy drop. However, by using only  $25\\%$  of Non-Entity Phrase examples in training, the NIL accuracy would recover to a decent level. Further adding NIL examples has little impact on cross-encoder models, but bi-encoder models still constantly benefit from additional data.",
+        "bbox": [
+            112,
+            439,
+            487,
+            600
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Besides, ignoring NIL data with the Non-Entity Phrase type will also harm the NIL accuracy and overall accuracy. Both types of NIL training data are necessary to reach to best performance.",
+        "bbox": [
+            112,
+            602,
+            487,
+            665
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We discover that entity linking models may be unaware of the NIL mentions when there is insufficient training data. A small amount of training data is enough for cross-encoder models to reach a reasonable accuracy, while bi-encoder models constantly benefit from additional training data.",
+        "bbox": [
+            112,
+            668,
+            487,
+            763
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            778,
+            268,
+            793
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "PLM-based Models in Entity Linking. Using pretrained language models (PLM) to capture semantic information is widely adopted in recent entity linking models. BLINK (Wu et al., 2020) marks the mention in context with pre-defined special tokens, and takes BERT as the encoder base. Two structures are adopted by BLINK to handle",
+        "bbox": [
+            112,
+            806,
+            487,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "different situations: bi-encoder for fast dense retrieval, and cross-encoder for further disambiguation. MOLEMAN (Fitzgerald et al., 2021) searches for similar mention contexts instead of entities, which better captures the diverse aspects an entity reflects in various contexts. GENRE (Cao et al., 2021) finetunes the sequence-to-sequence model BART, directly generating the unique entity name according to the mention context.",
+        "bbox": [
+            507,
+            84,
+            884,
+            229
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Research on NIL Prediction. The NIL prediction problem has been long viewed as an auxiliary task of entity linking. Some entity linking datasets (AIDA (Hoffart et al., 2011), TAC-KBP series (McNamee and Dang, 2009)) take the NIL prediction problem into consideration, while some (ACE and MSNBC) (Ratinov et al., 2011) omit mentions linking to NIL. Some research has already been conducted on the NIL prediction problem. Lazic et al. (2015) and Peters et al. (2019) set a score threshold to filter reasonable candidates, and mentions with no candidate score above the threshold are linked to NIL. Sil and Florian (2016); Kolitsas et al. (2018) views the NIL placeholder as a special entity, and selecting it as the best match indicates that the mention refers to no entities in the given KB. However, recent entity linking models, which use pretrained language models (PLM) as encoder bases, generally take the in-KB setting, which assumes that each mention has a valid golden entity in the KB (Wu et al., 2020).",
+        "bbox": [
+            507,
+            239,
+            884,
+            577
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Entity Type Assisted Entity Linking. Entity types can effectively assist entity linking and have been studied in various works. Gupta et al. (2017) jointly encodes mention context, entity description, and Freebase types with bidirectional LSTM to maximize the cosine similarity. DeepType (Raiman and Raiman, 2018) predicts the type probability of each token and gathers relevant tokens to predict the entity types, which would help eliminate candidates with incompatible types. Onoe and Durrett (2020) views entity types as a training objective rather than a feature, predicting fine-grained Wikipedia category tags to select the most relevant entity.",
+        "bbox": [
+            507,
+            589,
+            882,
+            814
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            828,
+            640,
+            843
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose an entity linking dataset NEL that focuses on the NIL prediction problem. We observe that mentions linking to NIL can be categorized into two patterns: Missing Entity and",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10853",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Non-Entity Phrase, but the latter one has not been paid sufficient attention. We propose an entity linking dataset NEL that focuses on NIL prediction. The dataset is built upon the Wikipedia corpus by choosing ambiguous entities as seeds and collecting relevant mention contexts. NEL is human-annotated to ensure correctness, and entity masking is further performed to control the percentage of NIL.",
+        "bbox": [
+            112,
+            84,
+            489,
+            227
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We conducted a series of experiments to examine the performance of PLM-based models on different datasets. Experimental results indicate that the accuracy without considering NIL prediction would be inflated. Meanwhile, sufficient data of both NIL types during training is essential to trigger the ability of NIL prediction. In the future, we may further try to integrate entity types into the pretraining process and explore type transfer between datasets.",
+        "bbox": [
+            112,
+            231,
+            489,
+            376
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            397,
+            285,
+            413
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We thank the anonymous reviewers for their insightful comments. This paper is supported by a grant from the Institute for Guo Qiang, Tsinghua University (2019GQB0003).",
+        "bbox": [
+            112,
+            428,
+            489,
+            492
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            513,
+            218,
+            527
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our work still exist some limitations. First, we choose an entity typing system on the base of Wiki-data tags, however, the granularity of the typing system remains to be discussed. A system with too many types would introduce noise to long-tail types, while insufficient types would weaken the disambiguation ability of type similarity. Thus, building a type system with adequate granularity remains a challenge.",
+        "bbox": [
+            112,
+            545,
+            489,
+            690
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Second, we combine the entity typing task with PLM-based semantic encoders, which require a fixed type system and further finetuning. Integrating the entity typing task into the pretraining process may enhance the transferability of the model and remove the dependency on a fixed type system.",
+        "bbox": [
+            112,
+            692,
+            489,
+            789
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Potential Risks. Our proposed dataset NEL centers on ambiguous entities, whose type distribution may not remain the same with other datasets. A potential risk is that the model trained on NEL may experience under-exposure of other entity types, which would damage their transferability and lead to undesired outputs on other datasets.",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            660,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this section, we will discuss the ethical considerations of our work.",
+        "bbox": [
+            507,
+            112,
+            884,
+            142
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Licenses and terms. The Wikipedia corpus and Wikidata types are obtained via the Wikimedia dump $^2$ , under the CC BY-SA 3.0 license $^3$ . AIDA, MSNBC, and WNED-WIKI are shared under the CC BY-SA 3.0 license. These datasets have been widely used in entity linking research, and we believe that they have been anonymized and desensitized.",
+        "bbox": [
+            507,
+            156,
+            884,
+            282
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Human Annotation. We recruited 3 human annotators without a background of expertise in annotation, and 1 expert annotator with adequate knowledge in entity linking for checking. These annotators are employed by commercial data annotation companies. We have paid these recruited annotators with adequate rewards under the agreed working time and price. The annotators are well informed about how these annotated data will be used and released, which has been recorded in the contract.",
+        "bbox": [
+            507,
+            297,
+            884,
+            472
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Intended use. NEL is an entity linking dataset focusing on the NIL prediction problem. Researchers are intended to use NEL for examining the ability of NIL prediction of newly created entity linking models. AIDA, MSNBC, and WNED-WIKI are intended for entity linking research, which is compatible with our work.",
+        "bbox": [
+            507,
+            487,
+            884,
+            599
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            630,
+            608,
+            644
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021. Autoregressive entity retrieval. In International Conference on Learning Representations.",
+            "Lihu Chen, Gaël Varoquaux, and Fabian M Suchanek. 2021. A lightweight neural model for biomedical entity linking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 12657-12665.",
+            "Silviu Cucerzan. 2007. Large-scale named entity disambiguation based on wikipedia data. In Proceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP-CoNLL), pages 708-716."
+        ],
+        "bbox": [
+            509,
+            653,
+            885,
+            879
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://dumps.wikimedia.org",
+        "bbox": [
+            529,
+            890,
+            715,
+            904
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://creativecommons.org/licenses/by-sa/3.0/",
+        "bbox": [
+            529,
+            904,
+            828,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10854",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mark Dredze, Paul McNamee, Delip Rao, Adam Gerber, and Tim Finin. 2010. Entity disambiguation for knowledge base population. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 277-285.",
+            "Yotam Eshel, Noam Cohen, Kira Radinsky, Shaul Markovitch, Ikuya Yamada, and Omer Levy. 2017. Named entity disambiguation for noisy text. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 58-68.",
+            "Nicholas Fitzgerald, Dan Bikel, Jan Botha, Dan Gillick, Tom Kwiatkowski, and Andrew McCallum. 2021. Moleman: Mention-only linking of entities with a mention annotation network. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 278-285.",
+            "Nitish Gupta, Sameer Singh, and Dan Roth. 2017. Entity linking via joint encoding of types, descriptions, and context. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2681-2690.",
+            "Johannes Hoffart, Mohamed Amir Yosef, Ilaria Bordino, Hagen Fürstenau, Manfred Pinkal, Marc Spaniol, Bilyana Taneva, Stefan Thater, and Gerhard Weikum. 2011. Robust disambiguation of named entities in text. In Proceedings of the 2011 conference on empirical methods in natural language processing, pages 782-792.",
+            "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.",
+            "Nikolaos Kolitsas, Octavian-Eugen Ganea, and Thomas Hofmann. 2018. End-to-end neural entity linking. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 519-529.",
+            "Nevena Lazic, Amarnag Subramanya, Michael Ringgaard, and Fernando Pereira. 2015. Plato: A selective context model for entity resolution. Transactions of the Association for Computational Linguistics, 3:503-515.",
+            "Harry Levin. 1977. The title as a literary genre. *The Modern language review*, 72(4):xxiii-xxxvi.",
+            "Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988.",
+            "Xiao Ling, Sameer Singh, and Daniel S Weld. 2015. Design challenges for entity linking. Transactions of the Association for Computational Linguistics, 3:315-328."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Kangqi Luo, Fengli Lin, Xusheng Luo, and Kenny Zhu. 2018. Knowledge base question answering via encoding of complex query graphs. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2185-2194.",
+            "Paul McNamee and Hoa Trang Dang. 2009. Overview of the tac 2009 knowledge base population track. In Text analysis conference (TAC), volume 17, pages 111-113.",
+            "Yasumasa Onoe and Greg Durrett. 2020. Fine-grained entity typing for domain independent entity linking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8576-8583.",
+            "Matthew E Peters, Mark Neumann, Robert Logan, Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A Smith. 2019. Knowledge enhanced contextual word representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 43-54.",
+            "Jonathan Raiman and Olivier Raiman. 2018. Deeype: multilingual entity linking by neural type system evolution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.",
+            "Lev Ratinov, Dan Roth, Doug Downey, and Mike Anderson. 2011. Local and global algorithms for disambiguation to wikipedia. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies, pages 1375-1384.",
+            "Avirup Sil and Radu Florian. 2016. One for all: Towards language independent named entity linking. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2255-2264.",
+            "Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Scalable zero-shot entity linking with dense entity retrieval. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6397-6407."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            713
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Details about the NEL Dataset Construction",
+        "text_level": 1,
+        "bbox": [
+            510,
+            726,
+            811,
+            758
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A.1 Corpora",
+        "text_level": 1,
+        "bbox": [
+            510,
+            769,
+            626,
+            784
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The NEL dataset is built from the 2021-07 English Wikipedia dump under the CC BY-SA 3.0 license. We take the hyperlinks in the raw xml dump as entity mentions, and retain at most 128 tokens around the mentions as their context. We take 64 tokens left to the mention and 64 tokens right to the mention by default, and more tokens will be included in one side if the other side does not contain enough",
+        "bbox": [
+            510,
+            790,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10855",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/9a486402c2662152c7a0a4b11b726bcba00bd584ce6c0dc018783c770f912f6c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Hyperparameter</td><td colspan=\"4\">NEL &amp; AIDA w/ NIL</td></tr><tr><td colspan=\"2\">CLINK-bi</td><td colspan=\"2\">CLINK-cross</td></tr><tr><td>Value</td><td>Range</td><td>Value</td><td>Range</td></tr><tr><td>Learning Rate</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td></tr><tr><td>λ</td><td>0.5</td><td></td><td>0.5</td><td></td></tr><tr><td>ε</td><td>0.5</td><td></td><td>0.5</td><td></td></tr><tr><td>Epoch</td><td>4</td><td>{1, 4}</td><td>4</td><td>{1, 4}</td></tr><tr><td>Batch Size</td><td>4</td><td>{1, 4, 8, 16, 32}</td><td>1</td><td>{1, 4, 8, 16, 32}</td></tr><tr><td># Parameters</td><td>670M</td><td>-</td><td>335M</td><td>-</td></tr><tr><td>Training Time</td><td>2 hrs</td><td>-</td><td>3.5 hrs</td><td>-</td></tr><tr><td rowspan=\"3\">Hyperparameter</td><td colspan=\"4\">AIDA w/o NIL, MSNBC &amp; WNED-WIKI</td></tr><tr><td colspan=\"2\">CLINK-bi</td><td colspan=\"2\">CLINK-cross</td></tr><tr><td>Value</td><td>Range</td><td>Value</td><td>Range</td></tr><tr><td>Learning Rate</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td></tr><tr><td>λ</td><td>0.9</td><td>[0.0, 1.0]</td><td>0.9</td><td>[0.0, 1.0]</td></tr><tr><td>ε</td><td>0</td><td>-</td><td>0</td><td>-</td></tr><tr><td>Epoch</td><td>1</td><td>{1, 4}</td><td>1</td><td>{1, 4}</td></tr><tr><td>Batch Size</td><td>4</td><td>{1, 4, 8, 16, 32}</td><td>1</td><td>{1, 4, 8, 16, 32}</td></tr><tr><td># Parameters</td><td>670M</td><td>-</td><td>335M</td><td>-</td></tr><tr><td>Training Time</td><td>4 hrs</td><td>-</td><td>3 hrs</td><td>-</td></tr></table>",
+        "bbox": [
+            203,
+            80,
+            794,
+            414
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 5: The hyperparameters used in the training process.",
+        "bbox": [
+            297,
+            424,
+            695,
+            439
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "tokens. We then discard tokens from both ends to ensure that the context plus the mention do not exceed the 128 token limit. Media files (image, audio) and Lua commands are discarded during preprocess.",
+        "bbox": [
+            112,
+            464,
+            487,
+            544
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.2 Data Selection",
+        "text_level": 1,
+        "bbox": [
+            114,
+            555,
+            278,
+            569
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Entries with the following features are viewed as noise and discarded:",
+        "bbox": [
+            112,
+            576,
+            485,
+            606
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The mention context contains the token  $^{\\star \\star}$  which is usually a list or formula;",
+            "- The mention with only 1 candidate entity, which does not pose much challenge;",
+            "- The mention with more than 20 candidate entities, which is far too challenging;",
+            "- The mention appears as a sub-span of a word;",
+            "- The mention has a probability of over  $50\\%$  of linking to a certain candidate entity, in which case we view the mention as unambiguous."
+        ],
+        "bbox": [
+            136,
+            617,
+            487,
+            819
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.3 Textual Representation Format",
+        "text_level": 1,
+        "bbox": [
+            114,
+            829,
+            410,
+            845
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Textual representation format for bi-encoder:",
+        "bbox": [
+            112,
+            850,
+            448,
+            865
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC = \\left[ \\mathrm {C L S} \\right] C _ {l} \\left[ m _ {\\text {s t a r t}} \\right] m \\left[ m _ {\\text {e n d}} \\right] C _ {r} \\left[ \\mathrm {S E P} \\right]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            139,
+            877,
+            458,
+            895
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE = \\left[ \\begin{array}{c} \\text {C L S} \\end{array} \\right] e _ {t i t l e} \\left[ \\begin{array}{c} m _ {t i t l e} \\end{array} \\right] e _ {d e s c} \\left[ \\begin{array}{c} \\text {S E P} \\end{array} \\right]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            142,
+            898,
+            411,
+            915
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Textual representation format for cross-encoder:",
+        "bbox": [
+            527,
+            464,
+            882,
+            479
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} (C, E) = \\left[ \\begin{array}{c} \\text {C L S} \\end{array} \\right] C _ {l} \\left[ \\begin{array}{c} m _ {\\text {s t a r t}} \\end{array} \\right] m \\left[ \\begin{array}{c} m _ {\\text {e n d}} \\end{array} \\right] C _ {r} \\\\ [ \\mathrm {S E P} ] e _ {t i t l e} [ m _ {t i t l e} ] e _ {d e s c} [ \\mathrm {S E P} ] \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            539,
+            489,
+            848,
+            527
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "where  $C$  represents mention context and  $E$  represents the textual description of candidate mentions.  $[m_{start}],[m_{end}],[m_{title}]$  are special tokens.",
+        "bbox": [
+            507,
+            536,
+            882,
+            585
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Experiment Details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            595,
+            714,
+            612
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We use the BERT-large-uncased model as the encoder base, with parameters initialized from the python transformers library. The models are trained on a single NVIDIA GeForce RTX 3090 GPU. We obtain the AIDA, MSNBC and WNED-WIKI dataset from the BLINK (Wu et al., 2020) repository https://github.com/facebookresearch/BLINK. We trained our model on the AIDA-train split, and evaluated on all three datasets.",
+        "bbox": [
+            507,
+            620,
+            882,
+            778
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The hyperparameter configurations are as follows. Detailed hyperparameters are shown in Table 5.",
+        "bbox": [
+            507,
+            782,
+            882,
+            828
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Typing System",
+        "text_level": 1,
+        "bbox": [
+            509,
+            841,
+            675,
+            858
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C.1 Typing System on NEL",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            742,
+            882
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We use a tree-like typing system with 187 distinct types on NEL. The typing system is build on the",
+        "bbox": [
+            507,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10856",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/972a6a6981a734ca9e7522b7d248d209f0d507aabf718084d6a57010f03fbdea.jpg",
+        "table_caption": [
+            "Table 6: Examples of type lines"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Entity</td><td>Types</td></tr><tr><td>14th Street (Manhattan)</td><td>Road-&gt;RouteOfTransportation-&gt;Infrastructure-&gt;ArchitecturalStructure-&gt;Place</td></tr><tr><td>1958 Copa del Generalíasimo</td><td>SoccerTournament-&gt;Tournament-&gt;SportsEvent-&gt;SocietalEvent-&gt;Event</td></tr><tr><td>ATM (song)</td><td>Song-&gt;MusicalWork-&gt;Work</td></tr><tr><td>Brats (1991 film)</td><td>Film-&gt;Work</td></tr><tr><td>Babe Ruth</td><td>BaseballPlayer-&gt;Athlete-&gt;Person</td></tr></table>",
+        "bbox": [
+            139,
+            107,
+            853,
+            256
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "base of Wikidata types. Table 6 shows some examples of type lines in the system. The most 10 frequent types in NEL are: (Work, Organisation, Place, Event, Person, Activity, FictionalCharacter, Award, Species, MeanOfTransportation).",
+        "bbox": [
+            112,
+            278,
+            489,
+            357
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C.2 Typing System on Traditional Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            370,
+            470,
+            386
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We retain 14 top-level types to make CLINK more transferable on different datasets. These types are: (Other, Person, Place, Work, Organization, Event, Fictional Character, Species, Activity, Device, Topical Concept, Ethnic Group, Food, Disease)",
+        "bbox": [
+            112,
+            393,
+            489,
+            473
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Errors in AIDA",
+        "text_level": 1,
+        "bbox": [
+            112,
+            487,
+            290,
+            502
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 7 demonstrates some mentions in AIDA that are incorrectly linked to NIL. 50 errors are detected among the 300 randomly sampled data in AIDA.",
+        "bbox": [
+            112,
+            514,
+            487,
+            562
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "E Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            576,
+            250,
+            593
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 8 shows some examples predicted by CLINK and BLINK without type information, which reflects how entity types influence the linking result.",
+        "bbox": [
+            112,
+            602,
+            489,
+            651
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In the first example, models with the bi-encoder structure incorrectly take the \"Gates of Heaven\" entry (which is in fact a documentary film) as the linking result, while CLINK-cross notices the context word \"album\" may indicate the entity type type, and correctly links the mention to the album. In the second example, the \"Home Before Dark\" mention actually refers to a 1997 movie like other movies in the context, however the corresponding entry is absent in the English Wikipedia. The CLINK-cross model is able to identify that the mention should be labelled as NIL, where the other models mistakenly link it to the \"Home Before Dark\" entry, which is an album rather than a movie. We observe that the entity types do help measure the similarity",
+        "bbox": [
+            112,
+            653,
+            489,
+            894
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "between mentions and entities, which enhances the performance of CLINK.",
+        "bbox": [
+            507,
+            278,
+            880,
+            310
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>4</sup>https://www.imdb.com/title/tt0116547/",
+        "bbox": [
+            134,
+            903,
+            381,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "10857",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/b471777a34000eb5fdd5d54d23bc44300ff28fdbf4eec1f8f8746e13b7df42de.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Mention Context</td><td>Mention</td><td>Assumed Entity</td></tr><tr><td>Bosnian premier in Turkey for one day visit . ANKARA 1996-08-27</td><td>Turkey</td><td>Turkey (Country)</td></tr><tr><td>EU rejects German call to boycott British lamb . Peter Blackburn BRUS-SELS 1996-08-22</td><td>EU</td><td>European Union</td></tr><tr><td>U.S. F-14 catches fire while landing in Israel . JERUSALEM 1996-08-25A U.S. fighter plane blew a tyre and ...</td><td>F-14</td><td>Grumman F-14 Tomcat</td></tr><tr><td>This is the leading story in the Mozambican press on Monday. Reuters has not verified this story and does not vouch for its accuracy.</td><td>Reuters</td><td>Reuters</td></tr></table>",
+        "bbox": [
+            139,
+            190,
+            858,
+            336
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/50eef25dac563053a11230f6bf2b95c2b0159c9360f8d3dc6c3f4380f5381c26.jpg",
+        "table_caption": [
+            "Table 7: Examples data in AIDA that are incorrectly linked to NIL."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Mention Context:</td><td colspan=\"4\">...singer, Michelle Branch, during her visit in Japan to promote her album Hotel Paper, for a magazine interview and photoshoot. After the release of Gates of Heaven the group a short break and performed in New York City ...</td></tr><tr><td>Model</td><td>BLINK-bi</td><td>CLINK-bi</td><td>BLINK-cross</td><td>CLINK-cross</td></tr><tr><td>Prediction</td><td>Gates of Heaven</td><td>Gates of Heaven</td><td>NIL</td><td>Gates of Heaven (album)</td></tr><tr><td>Mention Context:</td><td colspan=\"4\">...Michael Williams and David Collins founded the company in 1994 focusing on independent features, including Never Met Picasso (1996), Home Before Dark (1997), Six Ways To Sunday (1998), ...</td></tr><tr><td>Model</td><td>BLINK-bi</td><td>CLINK-bi</td><td>BLINK-cross</td><td>CLINK-cross</td></tr><tr><td>Prediction</td><td>Home Before Dark</td><td>Home Before Dark</td><td>Home Before Dark</td><td>NIL</td></tr></table>",
+        "bbox": [
+            117,
+            586,
+            878,
+            766
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Examples of predicted entity by different models. Entity mentions in context are labelled as red and the correct answer is labelled as bold.",
+        "bbox": [
+            112,
+            776,
+            882,
+            804
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "10858",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            129,
+            126,
+            532,
+            143
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section Limitations",
+        "bbox": [
+            151,
+            145,
+            297,
+            156
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            129,
+            168,
+            552,
+            186
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section Limitations",
+        "bbox": [
+            151,
+            187,
+            297,
+            200
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            129,
+            212,
+            695,
+            229
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Abstract and Section 1",
+        "bbox": [
+            149,
+            230,
+            319,
+            243
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            129,
+            256,
+            668,
+            272
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            151,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            115,
+            299,
+            487,
+            316
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section 3",
+        "bbox": [
+            132,
+            321,
+            205,
+            335
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            129,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            151,
+            363,
+            248,
+            379
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            129,
+            388,
+            778,
+            406
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section Ethical Considerations",
+        "bbox": [
+            151,
+            407,
+            383,
+            420
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            129,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section Ethical Considerations",
+        "bbox": [
+            151,
+            498,
+            383,
+            511
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            129,
+            523,
+            880,
+            571
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section Ethical Considerations",
+        "bbox": [
+            151,
+            573,
+            383,
+            586
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            129,
+            598,
+            880,
+            631
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section Ethical Considerations",
+        "bbox": [
+            151,
+            632,
+            383,
+            646
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            129,
+            657,
+            880,
+            739
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section 3",
+        "bbox": [
+            151,
+            740,
+            221,
+            753
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            115,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Section 5",
+        "bbox": [
+            132,
+            787,
+            205,
+            801
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            812,
+            880,
+            845
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Appendix B",
+        "bbox": [
+            151,
+            846,
+            238,
+            860
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "10859",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix B",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 5",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 5"
+        ],
+        "bbox": [
+            129,
+            83,
+            878,
+            281
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Section 3",
+        "bbox": [
+            112,
+            292,
+            877,
+            329
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? The instruction materials contain private information about authors and annotators, which may affect the double-blind review process. We have not yet been permitted to publish it.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Section Ethical Considerations",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Section Ethical Considerations",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? There have not yet established an ethics review board committee in our region.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? These data involve private information about annotators, and we have not yet been permitted to publish it."
+        ],
+        "bbox": [
+            129,
+            340,
+            878,
+            673
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "10860",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2023/Learn to Not Link_ Exploring NIL Prediction in Entity Linking/980725a0-e77c-4bea-bb00-a61ff06440a3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8545ffcf0e0f3578be43e0f46c76037315e6b2fbf6785d7be0b6219d52e2078d
+size 685357

2023/Learn to Not Link_ Exploring NIL Prediction in Entity Linking/full.md

@@ -0,0 +1,469 @@
+# Learn to Not Link: Exploring NIL Prediction in Entity Linking
+
+Fangwei Zhu $^{1,2}$ , Jifan Yu $^{1,2}$ , Hailong Jin $^{1,2}$ , Juanzi Li $^{1,2}$ , Lei Hou $^{1,2*}$  and Zhifang Sui $^{3}$
+
+$^{1}$ Dept. of Computer Sci.&Tech., BNRist, Tsinghua University, Beijing 100084, China  $^{2}$ KIRC, Institute for Artificial Intelligence, Tsinghua University
+
+$^{3}$ MOE Key Lab of Computational Linguistics, Peking University {zfw19@mails.,yujf18@mails,jinhl15@mails.,lixuanzi@,houlei@}tsinghua.edu.cn {szf@}pku.edu.cn
+
+# Abstract
+
+Entity linking models have achieved significant success via utilizing pretrained language models to capture semantic features. However, the NIL prediction problem, which aims to identify mentions without a corresponding entity in the knowledge base, has received insufficient attention. We categorize mentions linking to NIL into Missing Entity and Non-Entity Phrase, and propose an entity linking dataset NEL that focuses on the NIL prediction problem. NEL takes ambiguous entities as seeds, collects relevant mention context in the Wikipedia corpus, and ensures the presence of mentions linking to NIL by human annotation and entity masking. We conduct a series of experiments with the widely used bi-encoder and cross-encoder entity linking models, results show that both types of NIL mentions in training data have a significant influence on the accuracy of NIL prediction. Our code and dataset can be accessed at https://github.com/solitaryzero/NIL_EL.
+
+# 1 Introduction
+
+Entity Linking (EL) aims to map entity mentions in free texts to their corresponding entities in a given Knowledge Base (KB). Entity linking acts as a bridge between unstructured text and structured knowledge, and benefits various downstream tasks like question answering (Luo et al., 2018) and knowledge extraction (Chen et al., 2021).
+
+However, not all entity mentions correspond to a specific entity in the KB that suits the mention context (Ling et al., 2015). Take Table 1 as an example, Peter Blackburn is actually a journalist, and the householder is a common phrase rather than an entity. These two mentions do not refer to any entity in the given KB. The identification of these mentions is referred to as NIL prediction. Therefore, to tackle NIL prediction, the entity linking model
+
+needs to select mentions whose references are absent in KB, and link them to a special placeholder  $\text{NIL}$ . Dredze et al. (2010) states NIL prediction as one of the key issues in entity linking, which may lead to a decrease in the recall of entity linking systems. Meanwhile, the incorrectly linked entities may provide false information to downstream tasks.
+
+There have been some earlier representation learning based researches that take NIL prediction into consideration (Eshel et al., 2017; Lazic et al., 2015; Peters et al., 2019). They identify NIL mentions by setting a vector similarity threshold or viewing NIL as a special entity. Recently, pretrained language model (PLM) based models (Wu et al., 2020; Fitzgerald et al., 2021; Cao et al., 2021) have achieved great success for their great transferability and expandability. However, these models generally assume that there always exists a correct entity for each mention in the knowledge base, which leaves the NIL prediction problem without adequate attention.
+
+Previous entity linking datasets have paid insufficient attention to the NIL prediction problem. For example, some of the previous datasets like AIDA (Hoffart et al., 2011) view it as an auxiliary task, while others like MSNBC (Cucerzan, 2007) and WNED-WIKI (Eshel et al., 2017) does not require NIL prediction at all. There does not yet exist a strong benchmark for the ability on NIL prediction.
+
+In this paper, we propose an entity linking dataset NEL that focuses on the NIL prediction problem. About  $30\%$  of the mentions in NEL do not have their corresponding entity in the candidate entity set, which requires models to identify these mentions rather than linking them to the wrong candidates. In NEL construction, we take ambiguous entities as seeds, and build the dataset by mining mention contexts related to seed entities on the Wikipedia corpus. Then, human annotators are
+
+<table><tr><td colspan="2">Missing Entity</td></tr><tr><td>Mention Context</td><td>EU rejects German call to boycott British lamb. Peter Blackburn BRUS-SELS 1996-08-22</td></tr><tr><td>Peter Blackburn (Bishop)</td><td>Peter Blackburn (d.1616) was a Scottish scholar and prelate. He was the second Protestant Bishop of Aberdeen.</td></tr><tr><td>Peter Blackburn (MP)</td><td>Peter Blackburn (1811 – 20 May 1870) was a British Conservative Party politician.</td></tr><tr><td>Peter Blackburn (Badminton)</td><td>Peter Grant Blackburn (born 25 March 1968) is an Australian badminton player who affiliated with the Ballarat Badminton Association.</td></tr><tr><td colspan="2">Non-Entity Phrase</td></tr><tr><td>Mention Context</td><td>Most Hindus accept that there is a duty to have a family during the house-holder stage of life, as debt to family lineage called Pitra Rin (Father&#x27;s Debt) and so are unlikely to avoid having children altogether ...</td></tr><tr><td>The Householder (Film)</td><td>The Householder (Hindi title: Gharbar) is a 1963 film by Merchant Ivory Productions, with a screenplay by Ruth Prawer Jhabvala ...</td></tr><tr><td>The Householder (Novel)</td><td>The Householder is a 1960 English-language novel by Ruth Prawer Jhab-vala ...</td></tr></table>
+
+Table 1: Example of mentions that should be linked to NIL and their potential candidate entities. Mentions are labeled as red.
+
+asked to identify whether the mentions correspond to a candidate entity or not, and we further perform entity masking to ensure a fair proportion of NIL data of about  $30\%$ .
+
+In NIL prediction, we propose to use the widely used bi-encoder and cross-encoder structures as the model backbone, and further integrate type information by adding an entity typing subtask. We combine semantic and type similarity as the final similarity score, and identify NIL mentions by setting a similarity threshold.
+
+We conduct a series of experiments on both NEL and previous entity linking datasets. Experimental results show that the models suffer from an accuracy drop when taking NIL prediction into consideration, indicating that the accuracy may be inflated without the NIL prediction task, and NEL could better diagnose the performance of different models. We also conducted ablation studies on how type information and NIL examples affect the models. We discover that the entity typing subtask yields better embedding even when type similarity is not used, and both types of NIL examples in training data would boost the ability of NIL prediction.
+
+Our contributions can be concluded as:
+
+- We categorize the NIL prediction problem into two patterns: Missing Entity and Non-Entity Phrase, where the latter one has not received sufficient attention in previous works.
+
+- We propose an entity linking dataset NEL focusing on NIL prediction, which covers two patterns of NIL data and could act as a benchmark for diagnosing the ability of NIL prediction.  
+- We conducted a series of experiments, whose results demonstrate that the accuracy of models may be inflated when not taking NIL prediction into consideration. Meanwhile, both patterns of NIL data in training are essential for triggering the ability of NIL prediction.
+
+# 2 Preliminary
+
+Entity mentions  $M = \{m_i\}$  refer to text spans potentially corresponding to entities in a document  $D = (w_1, w_2, \dots, w_n)$ , where  $w_i$  is either a plain token or a mention. Each mention  $m_i$  may correspond to an entity  $e_i \in E$  in the entity set  $E$  of knowledge base  $\mathcal{KB}$ .
+
+Definition 1. Entity linking aims to find an optimal mapping  $\Gamma : M \Rightarrow E$ , which maps entity mentions  $M = \{m_i\}$  to their corresponding entities  $E = \{e_i\}$ , where  $e_i \in \mathcal{KB}$ .
+
+The NIL prediction problem is to determine whether an entity mention  $m$  is absent from the knowledge base  $\mathcal{KB}$ . When there does not exist a proper entity  $e \in \mathcal{KB}$  for the given mention  $m$ ,
+
+![](images/6e37f93f68c99d676297e6099dddeae0ffc36f5d7c8fa5bda4d4f8b1ea8660ed.jpg)  
+Figure 1: An illustration of how NEL is constructed. We select ambiguous entities as seeds, taking their alias as potential mentions to discover mention contexts from Wikipedia. The entries are annotated by human annotators, and entity masking is performed on some entries to control the portion of NIL data.
+
+the model should link  $m$  to a special NIL entity, indicating that the mention is unlinkable.
+
+Definition 2. Entity linking with NIL prediction aims to find an optimal mapping  $\Gamma : M \Rightarrow E \cup \{NIL\}$ , where NIL is a special placeholder and  $m$  correspond to NIL only when there is no correct answer in the candidate entity set  $E$ .
+
+As demonstrated in Table 1, there exist two situations in real-world entity linking where NIL prediction should be taken into consideration:
+
+- Missing Entity: The mention  $m$  refers to certain entity  $e$  that has not been yet included in  $\mathcal{KB}$ , i.e.  $m \Rightarrow e \notin \mathcal{KB}$ . For example, in the upper half of Table 1, the mention Peter Blackburn refers to a certain journalist, while entries in English Wikipedia include only people with other occupations, leading to the mention linking to NIL.  
+- Non-Entity Phrase: The mention  $m$  refers to a common phrase that is not usually viewed as an entity, i.e.  $m \nRightarrow e$ . For example, the mention the householder in the lower half of Table 1 refers to a concept rather than a film or novel.
+
+# 3 Dataset Construction
+
+There does not yet exist a strong benchmark on NIL prediction. We manually annotated 300 examples
+
+with their mentions linking to NIL from the widely used entity linking dataset AIDA<sup>1</sup>, and discover that about  $10\%$  of these mentions should actually link to an entity. For example, the mention "EU" in "EU rejects German" should be linked to the European Union rather than NIL (See Appendix D for details). Meanwhile, NIL mentions in AIDA fall mostly in the Missing Entity category. The incorrect and imbalanced data for NIL prediction indicates that the importance of NIL prediction is currently underestimated.
+
+In this section, we propose an entity linking dataset named NEL, which focuses on the NIL prediction problem. The construction process is demonstrated in Figure 1.
+
+Unlike normal entity linking data, there does not exist annotated references for mentions linking to NIL, and the portion of NIL data in the text corpus is unknown. Hyperlinks in Wikipedia can be viewed as mentions linking to non-NIL entities, from which we can find the aliases of entities. We assume that when an alias does not appear as a hyperlink in a certain context, it may be identified as a mention linking to NIL. In this way, we collect such contexts as the raw data. The raw data is then annotated by humans to ensure correctness, and we further mask out some answer entities in the candidate set to control the percentage of NIL in
+
+<table><tr><td>Dataset</td><td># Data</td><td>Annotated</td><td>NIL percentage</td><td>%Missing Entity</td><td>%Non-Entity Phrase</td></tr><tr><td>AIDA</td><td>34956</td><td>✓</td><td>20.41%</td><td>73%*</td><td>10%*</td></tr><tr><td>MSNBC</td><td>654</td><td>✓</td><td>0%</td><td>-</td><td>-</td></tr><tr><td>WNED-Wiki</td><td>240000</td><td>×</td><td>0%</td><td>-</td><td>-</td></tr><tr><td>NEL (ours)</td><td>9924</td><td>✓</td><td>33.57%</td><td>17%</td><td>83%</td></tr></table>
+
+Table 2: Statistics of the NEL dataset compared with previous entity linking datasets. *The percentage of two NIL patterns in AIDA is calculated from 300 randomly sampled NIL data, and data with errors do not count as any pattern.
+
+answers.
+
+# 3.1 Data Collection
+
+Levin (1977) states that the title of creative works could be a place, a personal name, or a certain abstract concept like the choric embodiment of some collectivity (The Clouds, The Birds) or stock types (The Alchemist, Le Misanthrope), which would naturally lead to the two situations where a mention links to NIL. The absence of the referenced entity from the KB would lead to Missing Entity, while an abstract concept not viewed as an entity would lead to Non-Entity Phrase.
+
+To discover NIL mentions of both types, we start by taking entities that share an alias with other entities as seeds. We assume that a mention referring to multiple entities has a higher probability of linking to a Missing Entity outside the KB, and the complex meaning of the mention will lead to Non-Entity Phrase. Thus, the aliases of ambiguous seed entities would be good starting points for mining NIL mentions.
+
+Entity Selection We further filter ambiguous entities to remove low-quality seeds. First, we remove noise instances like template pages, and entities with less than 5 hyperlinks are also removed. Meanwhile, we discarded entities with a probability greater than  $50\%$  of being the linking result, as these entities can generally be considered to be unambiguous and lack difficulty. Finally, 1000 entities are sampled as the seed entity set  $E_{s}$ .
+
+We use a typing system based on Wikidata to identify the type of selected entities. We view the instance of relation as the type indicator, and utilize the subclass of relation to build a tree-form type system. The detailed type system can be found in Appendix C.
+
+Mention Discovery We build an alias table from the 2021 Wikipedia dump by extracting alias-entity pairs  $(m,e)$  from internal hyperlinks. All alias
+
+$m$  related to entities in the seed entity set  $E_{s}$  are gathered as the mention set  $M$ . For each mention  $m \in M$ , we look for its occurrence throughout the Wikipedia corpus (whether it appears as a hyperlink or not) to obtain the entry tuple  $(C_{l}, m, C_{r}, E_{m})$ , where  $C_{l}$  and  $C_{r}$  represent contexts left and right to the entity mention  $m$ , and  $E_{m}$  represents the candidate entities set of  $m$ . For each mention  $m$ , we sampled 5 entries where  $m$  appears as a hyperlink and 5 entries where  $m$  appears in plain text to balance the number of positive and negative examples, and a total of 10,000 entries are collected.
+
+# 3.2 Human Annotation and Post-processing
+
+We perform annotation on the above entries with 3 annotators. The annotators are provided with the mention context  $(C_l, m, C_r)$  and candidate entities  $E_m$ . Each candidate entity  $e$  consists of its title, textual description, and Wikipedia URL. The annotators are asked to choose the answer entity  $a \in E_m$  corresponding to the mention  $m$ , or  $a = NIL$  if none of the candidates are correct.
+
+An expert will further investigate entries in which annotators fail to reach a consensus. The expert is a senior annotator with essential knowledge of entity linking, and will confirm the final annotation result after carefully reading through the context and candidate entities. We use the annotation result as the final answer  $a$  if there is an agreement between 3 annotators, and the entity chosen by the expert otherwise. The annotated tuple  $(C_l,m,C_r,E_m,a)$  acts as the final entry of our dataset.
+
+To further simulate the situation where new emerging entities do not appear in knowledge bases, we perform entity masking on positive entries. We randomly sample  $10\%$  entries where  $a \neq NIL$ , and mask the correct entity in the candidate set  $E_{m}$ . In this case, as the correct answer is removed from the candidate list, we have  $a = NIL$ , i.e. the mention  $m$  corresponds to the empty entity NIL.
+
+![](images/187d4d5e1d3f165d1a370301599dc08dc9dc0509daa8512a564f8d0ea2efee08.jpg)  
+Figure 2: The overall structure of PLM-based retrieval models. Candidates which are confusing in the semantic space may be more distinguishable in the type space. Mentions linking to NIL frequently differ from their candidates in their types, so we combine semantic similarity with type similarity for NIL prediction.
+
+![](images/aad5c3b877f42419d4829cd65c3ad0ec9b9b506b272f43060a25e453af09d4b7.jpg)
+
+# 3.3 Dataset Statistics
+
+Table 2 demonstrates the properties of the NEL dataset. NEL includes 6,593 positive examples and 3,331 negative examples, covering 1,000 mentions and 3,840 related entities. Each mention has an average of 3.80 candidate entities. The inter-annotator agreement of NEL is  $94.61\%$ , indicating that the expert calibrated about  $5\%$  of the data. The full dataset is split into train/validation/test sets by the ratio of  $80\%/10\%/10\%$ .
+
+NEL contains a fair number of entries, and is human-annotated to ensure correctness. Compared with previous entity linking datasets, NEL has a higher percentage of NIL data, which could better diagnose the ability of different models in NIL prediction. Meanwhile, mentions linking to NIL in AIDA mostly fall in the Missing Entity situation, while NEL focuses more on the Non-Entity Phrase situation, thus complementing the insufficient attention on Non-Entity Phrase in NIL prediction.
+
+# 4 Entity Linking with NIL prediction
+
+A mention links to NIL when all of its candidate entities fail to match its context. A common paradigm of NIL prediction is to compute the similarity scores between the mention context and candidates, and judge its linking result on the base of similarities.
+
+# 4.1 Scoring Similarities
+
+Bi-encoder and cross-encoder are widely adopted scorer structures, as they are well compatible with pretrained language models. Bi-encoder encodes mention contexts and entities into the same dense vector space, while cross-encoder views similarity scoring as a sequence classification task:
+
+$$
+s _ {b i} (c, e) = \sigma (f (c) \cdot g (e)) \tag {1}
+$$
+
+$$
+s _ {c r o s s} (c, e) = \sigma (\mathbf {W h} ([ c, e ]) + \mathbf {b}) \tag {2}
+$$
+
+where  $f,g,h$  are PLM-based encoders,  $\mathbf{W}$  and  $\mathbf{b}$  are trainable variables, and  $\sigma$  refers to the sigmoid function.
+
+The bi-encoder structure allows precomputing entity embeddings in knowledge bases, which enables efficient retrieval in real-world applications. Compared with bi-encoder, cross-encoder better captures the relation between context and entities with the cross attention mechanism, thus demonstrating higher accuracy.
+
+# 4.2 Integrating Entity Types
+
+Previous entity linking models (Gupta et al., 2017; Onoe and Durrett, 2020; Raiman and Raiman, 2018) have proved that entity types do help models better disambiguate between candidate entities.
+
+The type information can be integrated into bi-encoders and cross-encoders by adding a typing layer. In the bi-encoder structure, the mention types
+
+Table 3: Experimental results on NEL and previous datasets. Non-NAC, NAC, and OAC represent non-NIL accuracy, NIL accuracy and overall accuracy. *Results of GENRE on AIDA w/o NIL, MSNBC, and WNED-WIKI are taken from the original paper (Cao et al., 2021).  
+
+<table><tr><td></td><td colspan="3">NEL (our dataset)</td><td colspan="3">AIDA w/ NIL</td><td>AIDA w/o NIL</td><td>MSNBC</td><td>WNED-WIKI</td></tr><tr><td></td><td>Non-NAC</td><td>NAC</td><td>OAC</td><td>Non-NAC</td><td>NAC</td><td>OAC</td><td>OAC</td><td>OAC</td><td>OAC</td></tr><tr><td>BLINK-bi</td><td>72.27</td><td>88.59</td><td>77.74</td><td>64.54</td><td>69.59</td><td>65.01</td><td>82.61</td><td>70.86</td><td>58.56</td></tr><tr><td>CLINK-bi</td><td>79.24</td><td>79.28</td><td>79.25</td><td>75.98</td><td>66.36</td><td>75.09</td><td>83.26</td><td>73.29</td><td>58.99</td></tr><tr><td>GENRE*</td><td>54.00</td><td>62.84</td><td>56.96</td><td>-</td><td>-</td><td>-</td><td>88.60</td><td>88.10</td><td>71.70</td></tr><tr><td>BLINK-cross</td><td>84.09</td><td>88.89</td><td>85.70</td><td>83.08</td><td>45.16</td><td>79.58</td><td>87.49</td><td>82.02</td><td>69.48</td></tr><tr><td>CLINK-cross</td><td>86.97</td><td>89.19</td><td>87.71</td><td>84.42</td><td>58.53</td><td>82.03</td><td>88.16</td><td>89.70</td><td>72.43</td></tr></table>
+
+$t_c$  and entity types  $t_e$  are predicted separately:
+
+$$
+t _ {c} = \sigma \left(W _ {c} f (c) + b _ {c}\right) \tag {3}
+$$
+
+$$
+t _ {e} = \sigma \left(W _ {e} g (e) + b _ {e}\right) \tag {4}
+$$
+
+while they are simultaneously predicted in the cross-encoder structure:
+
+$$
+\left[ t _ {c}, t _ {e} \right] = \sigma \left(W f ([ c, e ]) + b\right) \tag {5}
+$$
+
+where  $\sigma$  represents the sigmoid function and  $W_{c}, b_{c}, W_{e}, b_{e}, W, b$  are trainable parameters.
+
+To tackle the label imbalance between types, we use the focal loss (Lin et al., 2017) on the typing task:
+
+$$
+\mathcal {L} _ {t} = - \sum_ {i = 1} ^ {n _ {t}} \left(y _ {i} \left(1 - t _ {i}\right) ^ {\gamma} \log t _ {i} + \left(1 - y _ {i}\right) t _ {i} ^ {\gamma} \log \left(1 - t _ {i}\right)\right) \tag {6}
+$$
+
+where  $n_t$  is the total number of types in the type system,  $\gamma$  is a hyperparameter,  $y_i$  is the golden label of the  $i$ -th type and  $t_i$  is the predicted label of the  $i$ -th type. In bi-encoder,  $\mathcal{L}_t$  is the average of loss on  $t_c$  and  $t_e$ , while in cross-encoder  $\mathcal{L}_t$  is directly calculated from  $[t_c, t_e]$ .
+
+We train the semantic encoder with binary classification loss  $\mathcal{L}_s$ , and combine  $\mathcal{L}_s$  with  $\mathcal{L}_t$  as the final loss  $\mathcal{L}$ :
+
+$$
+\mathcal {L} = \mathcal {L} _ {s} + \mathcal {L} _ {t} \tag {7}
+$$
+
+# 4.3 Identifying Mentions Linking to NIL
+
+The type similarity is computed with cosine similarity, and the final score is a weighted sum between type similarity and semantic similarity:
+
+$$
+s _ {t} (c, e) = \cos \left(t _ {c}, t _ {e}\right) \tag {8}
+$$
+
+$$
+s (c, e) = \lambda s _ {s} (c, e) + (1 - \lambda) s _ {t} (c, e) \tag {9}
+$$
+
+where  $\lambda$  is a hyperparameter.
+
+For each entry  $(C_l, m, C_r, E_m, a)$ , we concatenate  $(C_l, m, C_r)$  to form the context  $c$ . In the training step, for each candidate entity  $e \in E_m$ , we
+
+view  $(c,e)$  as a positive example if  $e = a$  , and as a negative example if  $a = N I L$  or  $e\neq a$
+
+During evaluation, the similarity score  $s(c, e)$  is computed between context  $c$  and each candidate entity  $e$ . If there exist entities with a score equal to or higher than the nil threshold  $\epsilon = 0.5$ , we choose the entity with the highest similarity score as the answer; If all entities fail to reach the threshold, then the mention  $m$  links to NIL.
+
+$$
+a = \left\{ \begin{array}{l l} \arg \max  _ {e} s (c, e), & \exists e, s (c, e) \geq \epsilon \\ N I L, & \forall e, s (c, e) <   \epsilon \end{array} \right. \tag {10}
+$$
+
+# 5 Experiments
+
+We conduct a series of experiments on two types of datasets, which test the different ability of entity linking models: (1) NEL that tests the ability of NIL prediction; (2) previous EL datasets that tests the ability of entity disambiguation. We choose the following models for comparison: BLINK (Wu et al., 2020) that uses the bi-encoder and cross-encoder alone to score candidate entities, CLINK that integrates type information with BLINK, and GENRE (Cao et al., 2021) that generates the linking result with a sequence-to-sequence model.
+
+# 5.1 Main Experiment Results
+
+We trained and tested the models on NEL, with BERT-large as the encoder base of BLINK and CLINK, and BART-large as the backbone of GENRE, to observe their ability in NIL prediction. We also experimented on previous entity linking datasets to observe the disambiguation ability of different models. The models are trained on the AIDA-YAGO2-train (Hoffart et al., 2011) dataset, and tested on AIDA-YAGO2-testb, MSNBC (Cucerzan, 2007) and WNED-WIKI (Eshel et al., 2017).
+
+Table 4: Experimental results on the influence of the entity typing task on NEL. OAC indicates the overall accuracy of entity linking. The overall accuracy with typing is achieved without using the type similarity score.  
+
+<table><tr><td>Model</td><td>Ctxt Type Acc.</td><td>Cand Type Acc.</td><td>OAC w/ Typing</td><td>OAC w/o Typing</td></tr><tr><td>Bi-encoder</td><td>83.75</td><td>93.46</td><td>78.35</td><td>77.74</td></tr><tr><td>Cross-encoder</td><td>83.95</td><td>98.01</td><td>87.41</td><td>83.67</td></tr></table>
+
+![](images/02e30405e73002b7378d2bc1a22777bb96a4c24cb2ba4a58180d370f3daabd8f.jpg)  
+Figure 3: Ablation study on the influence of NIL training data. The x-axis indicates the percentage of used NIL data or Non-Entity Phrase data in the training set.
+
+![](images/263bc244634f9c1b1445bb160d05b1135b3a88b60cd36aa8e8fea317e4080aac.jpg)
+
+![](images/24005887099eaae4c374ea6b9eeba60e1897c3e3afef1fec1f89daf1fc351797.jpg)
+
+187 distinct types are used in experiments on NEL, and considering that the entity type distribution may be different across datasets, we use a type system with only 14 top-level entity types on previous datasets to make CLINK more transferable (See Appendix C for details). We retain the same textual representation format with BLINK (see Appendix A), while using 128 as the length of context sequences and entity descriptions. All models are implemented with PyTorch and optimized with the Adam (Kingma and Ba, 2014) optimizer with a learning rate of 1e-5.
+
+Table 3 shows the evaluation results, from which some insights could be discovered:
+
+Type Information Matters. CLINK with cross-encoder structure achieves the highest accuracy on almost all datasets, and is still comparable with GENRE on AIDA, from which we may assert that taking type information into consideration is helpful even without NIL prediction. Meanwhile, on both structures, the overall accuracy of CLINK outperforms BLINK on all datasets, proving that the entity typing task assists both bi-encoder and cross-encoder distinguish correct entities.
+
+Encoder Structure. The cross-encoder structure generally performs better than bi-encoder, but we observe that sometimes bi-encoders show de
+
+cent ability in detecting NIL mentions, especially when type information is not utilized. BLINK-bi achieves the highest NIL accuracy score of 69.59 on AIDA with NIL, and has a score of 88.59 on NEL, which is comparable with the best-performing CLINK-cross. This phenomenon indicates that cross-encoders may be more prone to overfitting, while entity types would alleviate this tendency.
+
+NIL Entries. On the AIDA dataset, we observe that the models generally suffer from a drop in accuracy when taking NIL entries into consideration, and the drop is more obvious in bi-encoders. This may indicate that the performance of models is inflated without NIL prediction, and NIL entries may confuse the models in practical application.
+
+# 5.2 Ablation Study
+
+# 5.2.1 Influence of the Entity Typing Task
+
+We conducted experiments to observe how the entity typing task influences the model. We trained the model with both  $\mathcal{L}_t$  and  $\mathcal{L}_s$  as loss, while setting  $\lambda = 1$  during evaluation to ignore the influence of type similarity scores. Results shown in Table 4 reflects two observations:
+
+First, candidate type predictions benefit from mention context. We observe that when changing the structure from bi-encoder to cross-encoder, the
+
+type prediction accuracy on candidates raises by  $5\%$ , where the accuracy on contexts remains unchanged. This is likely because the context helps narrow down the type range, while the context type prediction generally remains unaffected by entity descriptions.
+
+Second, unifying the entity typing task in the training process leads to higher overall accuracy, even when the type similarity score is not taken into consideration in the evaluation, which can be demonstrated by the improved score of OAC w/ Typing compared to OAC w/o Typing on both models. This may indicate that predicting entity types would help the model learn semantic embeddings with higher quality.
+
+# 5.2.2 Influence of NIL Training Data
+
+Compared with previous datasets like AIDA, NEL contains more NIL mentions of the Non-Entity Phrase type. We trained the models with different numbers of Non-Entity Phrase examples to observe the influence of NIL Training Data.
+
+As demonstrated by Figure 3, all models suffer from a great decline in NIL accuracy when no NIL examples are used during the training stage, and bi-encoder is more prone to the accuracy drop. However, by using only  $25\%$  of Non-Entity Phrase examples in training, the NIL accuracy would recover to a decent level. Further adding NIL examples has little impact on cross-encoder models, but bi-encoder models still constantly benefit from additional data.
+
+Besides, ignoring NIL data with the Non-Entity Phrase type will also harm the NIL accuracy and overall accuracy. Both types of NIL training data are necessary to reach to best performance.
+
+We discover that entity linking models may be unaware of the NIL mentions when there is insufficient training data. A small amount of training data is enough for cross-encoder models to reach a reasonable accuracy, while bi-encoder models constantly benefit from additional training data.
+
+# 6 Related Work
+
+PLM-based Models in Entity Linking. Using pretrained language models (PLM) to capture semantic information is widely adopted in recent entity linking models. BLINK (Wu et al., 2020) marks the mention in context with pre-defined special tokens, and takes BERT as the encoder base. Two structures are adopted by BLINK to handle
+
+different situations: bi-encoder for fast dense retrieval, and cross-encoder for further disambiguation. MOLEMAN (Fitzgerald et al., 2021) searches for similar mention contexts instead of entities, which better captures the diverse aspects an entity reflects in various contexts. GENRE (Cao et al., 2021) finetunes the sequence-to-sequence model BART, directly generating the unique entity name according to the mention context.
+
+Research on NIL Prediction. The NIL prediction problem has been long viewed as an auxiliary task of entity linking. Some entity linking datasets (AIDA (Hoffart et al., 2011), TAC-KBP series (McNamee and Dang, 2009)) take the NIL prediction problem into consideration, while some (ACE and MSNBC) (Ratinov et al., 2011) omit mentions linking to NIL. Some research has already been conducted on the NIL prediction problem. Lazic et al. (2015) and Peters et al. (2019) set a score threshold to filter reasonable candidates, and mentions with no candidate score above the threshold are linked to NIL. Sil and Florian (2016); Kolitsas et al. (2018) views the NIL placeholder as a special entity, and selecting it as the best match indicates that the mention refers to no entities in the given KB. However, recent entity linking models, which use pretrained language models (PLM) as encoder bases, generally take the in-KB setting, which assumes that each mention has a valid golden entity in the KB (Wu et al., 2020).
+
+Entity Type Assisted Entity Linking. Entity types can effectively assist entity linking and have been studied in various works. Gupta et al. (2017) jointly encodes mention context, entity description, and Freebase types with bidirectional LSTM to maximize the cosine similarity. DeepType (Raiman and Raiman, 2018) predicts the type probability of each token and gathers relevant tokens to predict the entity types, which would help eliminate candidates with incompatible types. Onoe and Durrett (2020) views entity types as a training objective rather than a feature, predicting fine-grained Wikipedia category tags to select the most relevant entity.
+
+# 7 Conclusion
+
+In this paper, we propose an entity linking dataset NEL that focuses on the NIL prediction problem. We observe that mentions linking to NIL can be categorized into two patterns: Missing Entity and
+
+Non-Entity Phrase, but the latter one has not been paid sufficient attention. We propose an entity linking dataset NEL that focuses on NIL prediction. The dataset is built upon the Wikipedia corpus by choosing ambiguous entities as seeds and collecting relevant mention contexts. NEL is human-annotated to ensure correctness, and entity masking is further performed to control the percentage of NIL.
+
+We conducted a series of experiments to examine the performance of PLM-based models on different datasets. Experimental results indicate that the accuracy without considering NIL prediction would be inflated. Meanwhile, sufficient data of both NIL types during training is essential to trigger the ability of NIL prediction. In the future, we may further try to integrate entity types into the pretraining process and explore type transfer between datasets.
+
+# Acknowledgements
+
+We thank the anonymous reviewers for their insightful comments. This paper is supported by a grant from the Institute for Guo Qiang, Tsinghua University (2019GQB0003).
+
+# Limitations
+
+Our work still exist some limitations. First, we choose an entity typing system on the base of Wiki-data tags, however, the granularity of the typing system remains to be discussed. A system with too many types would introduce noise to long-tail types, while insufficient types would weaken the disambiguation ability of type similarity. Thus, building a type system with adequate granularity remains a challenge.
+
+Second, we combine the entity typing task with PLM-based semantic encoders, which require a fixed type system and further finetuning. Integrating the entity typing task into the pretraining process may enhance the transferability of the model and remove the dependency on a fixed type system.
+
+Potential Risks. Our proposed dataset NEL centers on ambiguous entities, whose type distribution may not remain the same with other datasets. A potential risk is that the model trained on NEL may experience under-exposure of other entity types, which would damage their transferability and lead to undesired outputs on other datasets.
+
+# Ethics Statement
+
+In this section, we will discuss the ethical considerations of our work.
+
+Licenses and terms. The Wikipedia corpus and Wikidata types are obtained via the Wikimedia dump $^2$ , under the CC BY-SA 3.0 license $^3$ . AIDA, MSNBC, and WNED-WIKI are shared under the CC BY-SA 3.0 license. These datasets have been widely used in entity linking research, and we believe that they have been anonymized and desensitized.
+
+Human Annotation. We recruited 3 human annotators without a background of expertise in annotation, and 1 expert annotator with adequate knowledge in entity linking for checking. These annotators are employed by commercial data annotation companies. We have paid these recruited annotators with adequate rewards under the agreed working time and price. The annotators are well informed about how these annotated data will be used and released, which has been recorded in the contract.
+
+Intended use. NEL is an entity linking dataset focusing on the NIL prediction problem. Researchers are intended to use NEL for examining the ability of NIL prediction of newly created entity linking models. AIDA, MSNBC, and WNED-WIKI are intended for entity linking research, which is compatible with our work.
+
+# References
+
+Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021. Autoregressive entity retrieval. In International Conference on Learning Representations.  
+Lihu Chen, Gaël Varoquaux, and Fabian M Suchanek. 2021. A lightweight neural model for biomedical entity linking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 12657-12665.  
+Silviu Cucerzan. 2007. Large-scale named entity disambiguation based on wikipedia data. In Proceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP-CoNLL), pages 708-716.
+
+Mark Dredze, Paul McNamee, Delip Rao, Adam Gerber, and Tim Finin. 2010. Entity disambiguation for knowledge base population. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 277-285.  
+Yotam Eshel, Noam Cohen, Kira Radinsky, Shaul Markovitch, Ikuya Yamada, and Omer Levy. 2017. Named entity disambiguation for noisy text. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 58-68.  
+Nicholas Fitzgerald, Dan Bikel, Jan Botha, Dan Gillick, Tom Kwiatkowski, and Andrew McCallum. 2021. Moleman: Mention-only linking of entities with a mention annotation network. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 278-285.  
+Nitish Gupta, Sameer Singh, and Dan Roth. 2017. Entity linking via joint encoding of types, descriptions, and context. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2681-2690.  
+Johannes Hoffart, Mohamed Amir Yosef, Ilaria Bordino, Hagen Fürstenau, Manfred Pinkal, Marc Spaniol, Bilyana Taneva, Stefan Thater, and Gerhard Weikum. 2011. Robust disambiguation of named entities in text. In Proceedings of the 2011 conference on empirical methods in natural language processing, pages 782-792.  
+Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.  
+Nikolaos Kolitsas, Octavian-Eugen Ganea, and Thomas Hofmann. 2018. End-to-end neural entity linking. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 519-529.  
+Nevena Lazic, Amarnag Subramanya, Michael Ringgaard, and Fernando Pereira. 2015. Plato: A selective context model for entity resolution. Transactions of the Association for Computational Linguistics, 3:503-515.  
+Harry Levin. 1977. The title as a literary genre. *The Modern language review*, 72(4):xxiii-xxxvi.  
+Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988.  
+Xiao Ling, Sameer Singh, and Daniel S Weld. 2015. Design challenges for entity linking. Transactions of the Association for Computational Linguistics, 3:315-328.
+
+Kangqi Luo, Fengli Lin, Xusheng Luo, and Kenny Zhu. 2018. Knowledge base question answering via encoding of complex query graphs. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2185-2194.  
+Paul McNamee and Hoa Trang Dang. 2009. Overview of the tac 2009 knowledge base population track. In Text analysis conference (TAC), volume 17, pages 111-113.  
+Yasumasa Onoe and Greg Durrett. 2020. Fine-grained entity typing for domain independent entity linking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8576-8583.  
+Matthew E Peters, Mark Neumann, Robert Logan, Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A Smith. 2019. Knowledge enhanced contextual word representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 43-54.  
+Jonathan Raiman and Olivier Raiman. 2018. Deeype: multilingual entity linking by neural type system evolution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.  
+Lev Ratinov, Dan Roth, Doug Downey, and Mike Anderson. 2011. Local and global algorithms for disambiguation to wikipedia. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies, pages 1375-1384.  
+Avirup Sil and Radu Florian. 2016. One for all: Towards language independent named entity linking. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2255-2264.  
+Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Scalable zero-shot entity linking with dense entity retrieval. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6397-6407.
+
+# A Details about the NEL Dataset Construction
+
+# A.1 Corpora
+
+The NEL dataset is built from the 2021-07 English Wikipedia dump under the CC BY-SA 3.0 license. We take the hyperlinks in the raw xml dump as entity mentions, and retain at most 128 tokens around the mentions as their context. We take 64 tokens left to the mention and 64 tokens right to the mention by default, and more tokens will be included in one side if the other side does not contain enough
+
+<table><tr><td rowspan="3">Hyperparameter</td><td colspan="4">NEL &amp; AIDA w/ NIL</td></tr><tr><td colspan="2">CLINK-bi</td><td colspan="2">CLINK-cross</td></tr><tr><td>Value</td><td>Range</td><td>Value</td><td>Range</td></tr><tr><td>Learning Rate</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td></tr><tr><td>λ</td><td>0.5</td><td></td><td>0.5</td><td></td></tr><tr><td>ε</td><td>0.5</td><td></td><td>0.5</td><td></td></tr><tr><td>Epoch</td><td>4</td><td>{1, 4}</td><td>4</td><td>{1, 4}</td></tr><tr><td>Batch Size</td><td>4</td><td>{1, 4, 8, 16, 32}</td><td>1</td><td>{1, 4, 8, 16, 32}</td></tr><tr><td># Parameters</td><td>670M</td><td>-</td><td>335M</td><td>-</td></tr><tr><td>Training Time</td><td>2 hrs</td><td>-</td><td>3.5 hrs</td><td>-</td></tr><tr><td rowspan="3">Hyperparameter</td><td colspan="4">AIDA w/o NIL, MSNBC &amp; WNED-WIKI</td></tr><tr><td colspan="2">CLINK-bi</td><td colspan="2">CLINK-cross</td></tr><tr><td>Value</td><td>Range</td><td>Value</td><td>Range</td></tr><tr><td>Learning Rate</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td><td>1e-5</td><td>{1e-3, 1e-5, 3e-5}</td></tr><tr><td>λ</td><td>0.9</td><td>[0.0, 1.0]</td><td>0.9</td><td>[0.0, 1.0]</td></tr><tr><td>ε</td><td>0</td><td>-</td><td>0</td><td>-</td></tr><tr><td>Epoch</td><td>1</td><td>{1, 4}</td><td>1</td><td>{1, 4}</td></tr><tr><td>Batch Size</td><td>4</td><td>{1, 4, 8, 16, 32}</td><td>1</td><td>{1, 4, 8, 16, 32}</td></tr><tr><td># Parameters</td><td>670M</td><td>-</td><td>335M</td><td>-</td></tr><tr><td>Training Time</td><td>4 hrs</td><td>-</td><td>3 hrs</td><td>-</td></tr></table>
+
+Table 5: The hyperparameters used in the training process.
+
+tokens. We then discard tokens from both ends to ensure that the context plus the mention do not exceed the 128 token limit. Media files (image, audio) and Lua commands are discarded during preprocess.
+
+# A.2 Data Selection
+
+Entries with the following features are viewed as noise and discarded:
+
+- The mention context contains the token  $^{\star \star}$  which is usually a list or formula;  
+- The mention with only 1 candidate entity, which does not pose much challenge;  
+- The mention with more than 20 candidate entities, which is far too challenging;  
+- The mention appears as a sub-span of a word;  
+- The mention has a probability of over  $50\%$  of linking to a certain candidate entity, in which case we view the mention as unambiguous.
+
+# A.3 Textual Representation Format
+
+Textual representation format for bi-encoder:
+
+$$
+C = \left[ \mathrm {C L S} \right] C _ {l} \left[ m _ {\text {s t a r t}} \right] m \left[ m _ {\text {e n d}} \right] C _ {r} \left[ \mathrm {S E P} \right]
+$$
+
+$$
+E = \left[ \begin{array}{c} \text {C L S} \end{array} \right] e _ {t i t l e} \left[ \begin{array}{c} m _ {t i t l e} \end{array} \right] e _ {d e s c} \left[ \begin{array}{c} \text {S E P} \end{array} \right]
+$$
+
+Textual representation format for cross-encoder:
+
+$$
+\begin{array}{l} (C, E) = \left[ \begin{array}{c} \text {C L S} \end{array} \right] C _ {l} \left[ \begin{array}{c} m _ {\text {s t a r t}} \end{array} \right] m \left[ \begin{array}{c} m _ {\text {e n d}} \end{array} \right] C _ {r} \\ [ \mathrm {S E P} ] e _ {t i t l e} [ m _ {t i t l e} ] e _ {d e s c} [ \mathrm {S E P} ] \\ \end{array}
+$$
+
+where  $C$  represents mention context and  $E$  represents the textual description of candidate mentions.  $[m_{start}],[m_{end}],[m_{title}]$  are special tokens.
+
+# B Experiment Details
+
+We use the BERT-large-uncased model as the encoder base, with parameters initialized from the python transformers library. The models are trained on a single NVIDIA GeForce RTX 3090 GPU. We obtain the AIDA, MSNBC and WNED-WIKI dataset from the BLINK (Wu et al., 2020) repository https://github.com/facebookresearch/BLINK. We trained our model on the AIDA-train split, and evaluated on all three datasets.
+
+The hyperparameter configurations are as follows. Detailed hyperparameters are shown in Table 5.
+
+# C Typing System
+
+# C.1 Typing System on NEL
+
+We use a tree-like typing system with 187 distinct types on NEL. The typing system is build on the
+
+Table 6: Examples of type lines  
+
+<table><tr><td>Entity</td><td>Types</td></tr><tr><td>14th Street (Manhattan)</td><td>Road-&gt;RouteOfTransportation-&gt;Infrastructure-&gt;ArchitecturalStructure-&gt;Place</td></tr><tr><td>1958 Copa del Generalíasimo</td><td>SoccerTournament-&gt;Tournament-&gt;SportsEvent-&gt;SocietalEvent-&gt;Event</td></tr><tr><td>ATM (song)</td><td>Song-&gt;MusicalWork-&gt;Work</td></tr><tr><td>Brats (1991 film)</td><td>Film-&gt;Work</td></tr><tr><td>Babe Ruth</td><td>BaseballPlayer-&gt;Athlete-&gt;Person</td></tr></table>
+
+base of Wikidata types. Table 6 shows some examples of type lines in the system. The most 10 frequent types in NEL are: (Work, Organisation, Place, Event, Person, Activity, FictionalCharacter, Award, Species, MeanOfTransportation).
+
+# C.2 Typing System on Traditional Datasets
+
+We retain 14 top-level types to make CLINK more transferable on different datasets. These types are: (Other, Person, Place, Work, Organization, Event, Fictional Character, Species, Activity, Device, Topical Concept, Ethnic Group, Food, Disease)
+
+# D Errors in AIDA
+
+Table 7 demonstrates some mentions in AIDA that are incorrectly linked to NIL. 50 errors are detected among the 300 randomly sampled data in AIDA.
+
+# E Case Study
+
+Table 8 shows some examples predicted by CLINK and BLINK without type information, which reflects how entity types influence the linking result.
+
+In the first example, models with the bi-encoder structure incorrectly take the "Gates of Heaven" entry (which is in fact a documentary film) as the linking result, while CLINK-cross notices the context word "album" may indicate the entity type type, and correctly links the mention to the album. In the second example, the "Home Before Dark" mention actually refers to a 1997 movie like other movies in the context, however the corresponding entry is absent in the English Wikipedia. The CLINK-cross model is able to identify that the mention should be labelled as NIL, where the other models mistakenly link it to the "Home Before Dark" entry, which is an album rather than a movie. We observe that the entity types do help measure the similarity
+
+between mentions and entities, which enhances the performance of CLINK.
+
+<table><tr><td>Mention Context</td><td>Mention</td><td>Assumed Entity</td></tr><tr><td>Bosnian premier in Turkey for one day visit . ANKARA 1996-08-27</td><td>Turkey</td><td>Turkey (Country)</td></tr><tr><td>EU rejects German call to boycott British lamb . Peter Blackburn BRUS-SELS 1996-08-22</td><td>EU</td><td>European Union</td></tr><tr><td>U.S. F-14 catches fire while landing in Israel . JERUSALEM 1996-08-25A U.S. fighter plane blew a tyre and ...</td><td>F-14</td><td>Grumman F-14 Tomcat</td></tr><tr><td>This is the leading story in the Mozambican press on Monday. Reuters has not verified this story and does not vouch for its accuracy.</td><td>Reuters</td><td>Reuters</td></tr></table>
+
+Table 7: Examples data in AIDA that are incorrectly linked to NIL.  
+
+<table><tr><td>Mention Context:</td><td colspan="4">...singer, Michelle Branch, during her visit in Japan to promote her album Hotel Paper, for a magazine interview and photoshoot. After the release of Gates of Heaven the group a short break and performed in New York City ...</td></tr><tr><td>Model</td><td>BLINK-bi</td><td>CLINK-bi</td><td>BLINK-cross</td><td>CLINK-cross</td></tr><tr><td>Prediction</td><td>Gates of Heaven</td><td>Gates of Heaven</td><td>NIL</td><td>Gates of Heaven (album)</td></tr><tr><td>Mention Context:</td><td colspan="4">...Michael Williams and David Collins founded the company in 1994 focusing on independent features, including Never Met Picasso (1996), Home Before Dark (1997), Six Ways To Sunday (1998), ...</td></tr><tr><td>Model</td><td>BLINK-bi</td><td>CLINK-bi</td><td>BLINK-cross</td><td>CLINK-cross</td></tr><tr><td>Prediction</td><td>Home Before Dark</td><td>Home Before Dark</td><td>Home Before Dark</td><td>NIL</td></tr></table>
+
+Table 8: Examples of predicted entity by different models. Entity mentions in context are labelled as red and the correct answer is labelled as bold.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Section Limitations
+
+A2. Did you discuss any potential risks of your work?
+
+Section Limitations
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Abstract and Section 1
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Section 3
+
+B1. Did you cite the creators of artifacts you used?
+
+No response.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Section Ethical Considerations
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Section Ethical Considerations
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Section Ethical Considerations
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Section Ethical Considerations
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Section 3
+
+C Did you run computational experiments?
+
+Section 5
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Appendix B
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix B  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 5  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 5
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Section 3
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? The instruction materials contain private information about authors and annotators, which may affect the double-blind review process. We have not yet been permitted to publish it.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Section Ethical Considerations  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Section Ethical Considerations  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? There have not yet established an ethics review board committee in our region.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? These data involve private information about annotators, and we have not yet been permitted to publish it.

2023/Learn to Not Link_ Exploring NIL Prediction in Entity Linking/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0389752ff8cef2f4b0afb04160e040d857ed2b41ed02f2a3d85eb71c82187d76
+size 764057

2023/Learned Adapters Are Better Than Manually Designed Adapters/55f2333f-7821-49a3-9f81-0d94e944171c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d926b7119d23d46e7b2446665548ec63122969b60da59f4e33bf5ae8b4fd6b54
+size 561241

2023/Learned Adapters Are Better Than Manually Designed Adapters/full.md

@@ -0,0 +1,467 @@
+# Learned Adapters Are Better Than Manually Designed Adapters
+
+Yuming Zhang $^{1}$ , Peng Wang $^{2*}$ , Ming Tan $^{3*}$ , Wei Zhu $^{4*†}$
+
+<sup>1</sup> College of Computer Science and Software Engineering, Shenzhen University
+
+2 Tomorrow Advancing Life
+
+$^{3}$  Southern University of Science and Technology
+
+$^{4}$  East China Normal University
+
+# Abstract
+
+Recently, a series of works have looked into further improving the adapter-based tuning by manually designing better adapter architectures. Understandably, these manually designed solutions are sub-optimal. In this work, we propose the Learned Adapter framework to automatically learn the optimal adapter architectures for better task adaptation of pre-trained models (PTMs). First, we construct a unified search space for adapter architecture designs. In terms of the optimization method on the search space, we propose a simple-yet-effective method, GDNAS, for better architecture optimization. Extensive experiments show that our Learned Adapter framework can outperform the previous parameter-efficient tuning (PETuning) baselines while tuning comparable or fewer parameters. Moreover: (a) the learned adapter architectures are explainable and transferable across tasks. (b) We demonstrate that our architecture search space design is valid.
+
+# 1 Introduction
+
+Increasingly large pre-trained models (Han et al., 2021; Devlin et al., 2019; Peters et al., 2018; Liu et al., 2019b; Radford and Narasimhan, 2018; Raffel et al., 2019) built upon the Transformer architecture (Vaswani et al., 2017) have been emerging and achieving the state-of-the-art (SOTA) results on a variety of downstream tasks (Gao et al., 2023; Zhu et al., 2023; Li et al., 2019; Zhu, 2021b; Zuo et al., 2022; Zhang et al., 2022; Guo et al., 2021b; Zhu et al., 2021a; Sun et al., 2020; Zhu et al., 2019). Despite their effectiveness, these large-scale models also bring the curse of prohibitive computation (Zhu et al., 2021c; Zhu, 2021c; Sun et al., 2022) and storage costs during the adaptations to downstream tasks due to the gradient computation of the whole model and the giant size of the fine-tuned checkpoint.
+
+![](images/c381f6c42b781ff366b7435f76d5f8bb03c20bef34feabb7424488a3da5340b3.jpg)  
+Figure 1: Overall comparison between our Learned Adapter framework and baselines. The  $x$ -axis is the number of tunable parameters, and the  $y$ -axis is the average performance on the GLUE benchmark with RoBERTa-large backbone. The details can be found in Section 5.
+
+Recently, parameter efficient tuning (PETuning) has raised much attention in the research field since it can only train a small portion of PTMs and keep the vast majority of parameters frozen, thus alleviating the computation costs during full finetuning. A series of studies (Houlsby et al., 2019; Pfeiffer et al., 2021; Mahabadi et al., 2021; BenZaken et al., 2021; Hu et al., 2021; Guo et al., 2021a; Li and Liang, 2021; Lester et al., 2021) has verified that PETuning can achieve competitive performance compared to conventional finetuning with very few trainable parameters, resulting in a considerable reduction in model adaptation costs. Adapter-based methods (Houlsby et al., 2019; Pfeiffer et al., 2021; Mahabadi et al., 2021; He et al., 2021) inject newly-introduced layers after or around the attention or feed-forward modules of the Transformer block, and yield promising results by fine-tuning a small portion of the PTM's parameters.
+
+Recently, a branch of recent research has ad
+
+![](images/241adbab414f8f4c0fad70fa5b045d3c058158cc5bb7d59957e1a5899357ea52.jpg)  
+Figure 2: The overall framework of our Learned Adapter.
+
+![](images/4a205de605cc8d6a9b18ecb0a0736f4db5c39a6ccfffccf26359f528d01c2815.jpg)
+
+![](images/1fae0fd2febc84147f08f0766bd24b1bf8a4061867ecb3c24659d3d15075fbda.jpg)
+
+vanced the understanding of adapter-based tuning more deeply and improved the adapters' architectures to improve parameter efficiency further. Adaptable adapters (Moosavi et al., 2022) propose that adapters at different layers should have different activation functions. Thus they fit the rational activation functions to downstream tasks during parameter tuning. AdapterDrop (Rücklé et al., 2020) tries to reduce the number of adapters' parameter number by not inserting adapters on the lower layers. He et al. (2021) bridge connections among different PETuning approaches to form a unified framework and further propose to insert adapters in parallel to the modules of the Transformer block. Jie and Deng (2022) and Sung et al. (2022) propose to add encoding operations between the projection layers of an adapter and achieve better PETuning performances. The above empirical evidence implies that altering the adapters' architecture designs can help to improve the PETuning performances of adapters with even fewer tunable parameters. Predictably, such an optimal architecture is difficult to construct manually and may vary across different PTM backbones and tasks. Therefore, we propose to search for the optimal architecture of adapters automatically.
+
+We present the Learned Adapter framework to search for the optimal architecture of adapters automatically. We first construct a unified search space (Figure 2) that considers various design choices of adapters, including the activation functions, encoding operations, and how the adapters are connected to the PTM backbone. In terms of the specific
+
+methodology for optimization on the search space, make a simple-yet-effective modification to the optimization method in DARTS (Liu et al., 2019a), which is better at identifying the proper components for adapters at different intermediate layers.
+
+We conduct extensive experiments to study the effectiveness of our Learned Adapter framework. The experimental results show that with  $0.068\%$  parameters, we can recover  $99.5\%$  finetuning performances on the GLUE (Wang et al., 2018) benchmark. Moreover, the searched architecture outperforms the manually designed PETuning baselines while tuning fewer parameters. Figure 1 depicts the overall comparison between our Learned Adapter and the baselines. Furthermore, the learned architectures of adapters are transferable across tasks, which significantly strengthens the usefulness of the searched structures. Further experiments demonstrate that our newly proposed search space for adapters is valid.
+
+# 2 Related work
+
+Adapter-based tuning. One of the most important research lines of PETuning is adapter-based tuning. Adapter (Houlsby et al., 2019) inserts adapter modules with bottleneck architecture between every consecutive Transformer (Vaswani et al., 2017) sublayers. AdapterFusion (Pfeiffer et al., 2021) only inserts sequential adapters after the feed-forward module. Adapter-based tuning methods have comparable results with model tuning when only tuning a fraction of the backbone model's parameter number. Due to their amazing results on PETuning, a
+
+branch of literature has investigated the architecture of adapters in search of further improvements. He et al. (2021) analyze a wide range of PETuning methods and show that they are essentially equivalent. They also propose the general architecture of PETuning. AdapterDrop (Rückle et al., 2020) investigates the efficiency of removing adapters from lower layers. Adaptive adapters (Moosavi et al., 2022) investigate the activation functions of adapters and propose to learn the activation functions of adapters via optimizing the parameters of rational functions as a part of the model parameters. Compacter (Mahabadi et al., 2021) uses low-rank parameterized hypercomplex multiplication (Le et al., 2021) to compress adapters' tunable parameters. There is also work (Sung et al., 2022; Jie and Deng, 2022) trying to add different encoding operations, like self-attention operations and convolutions between the bottleneck structure of adapters, and achieve better performances.
+
+Our work complements this branch of literature by investigating: (a) whether and how the adapter architectures affect the PETuning performances, and whether different layers of PTMs need different adapter architectures; (b) whether we can obtain better adapter architectures via neural architecture search.
+
+Other PETuning methods Another main research line of PETuning is the prompt-based tuning that inserts some additional soft prompts into the hidden states instead of injecting new neural modules to PTMs. Prompt tuning (Lester et al., 2021) and P-tuning (Liu et al., 2022) insert a soft prompt to word embeddings only, and can achieve competitive results when applied to supersized PTMs. Prefix-tuning (Li and Liang, 2021) and P-tuning v2 (Liu et al., 2021) insert prompts to every hidden layer of PTM. IDPG (Wu et al., 2022) uses the prompt generator with parameterized hypercomplex multiplication (Le et al., 2021) to generate a soft prompt for every instance. There are also some other popular PETuning methods, such as BitFit (Ben-Zaken et al., 2021) which only tunes the bias terms, LoRA (Hu et al., 2021) which optimizes low-rank decomposition matrices of the weights within self-attention layers.
+
+Neural architecture search In the early attempts, neural architecture search (NAS) requires massive computations, like thousands of GPU days (Zoph and Le, 2017; Zoph et al., 2018; Liu et al., 2018). Recently, a particular group of one-shot NAS, led
+
+by the seminal work DARTS (Liu et al., 2019a) has attracted much attention. DARTS formulates the search space into a super-network that can adjust itself in a continuous space so that the network and architectural parameters can be optimized alternately (bi-level optimization) using gradient descent. A series of literature try to improve the performance and efficiency of DARTS, such as Xie et al. (2019), Chen et al. (2021), Chu et al. (2021), Nayman et al. (2019). SNAS (Xie et al., 2019) reformulate DARTS as a credit assignment task while maintaining the differentiability. Gao et al. (2020) penalize the entropy of the architecture parameters to encourage discretization on the hyper-network. P-DARTS (Chen et al., 2021) analyze the issues during the DARTS bi-level optimization, and propose a series of modifications. PC-DARTS (Xu et al., 2021) reduces the memory cost during search by sampling a portion of the channels in supernetworks. FairDARTS (Chu et al., 2021) change the softmax operations in DARTS into sigmoid and introduce a zero-one loss to prune the architectural parameters. XNAS (Nayman et al., 2019) dynamically wipes out inferior architectures and enhances superior ones. NAS is widely applied in both computer vision and natural language processing, especially in knowledge distillation (Zhu, 2021a; Zhang et al., 2021).
+
+Our work complements the literature by examining the optimization of DARTS on our search space and propose a new training procedure that does not require re-training after discretization.
+
+# 3 Search space of Learned Adapter
+
+# 3.1 Pilot experiments and motivations
+
+In this subsection, we conduct a series of experiments on the RTE (Dagan et al., 2005) and MRPC (Dolan and Brockett, 2005) datasets to demonstrate the necessity of investigating the architecture of adapters. The baseline model  $\mathcal{M}$  is RoBERTa-large model with an parallel adapter at the feed-forward module (FFN adapter) (He et al., 2021). The backbone model is frozen and we only tune the adapters on downstream tasks. The bottleneck dimension is 32 and the activation function is ReLU. The other experimental settings follow Appendix B. We now consider a series of simple modifications to the baseline model.
+
+Modifying the activation function We replace the activation functions of the adapters from ReLU to GeLU, SWISH or Tanh, while keeping the other
+
+<table><tr><td>Model</td><td>RTE</td><td>MRPC</td></tr><tr><td>M</td><td>79.1 (0.5)</td><td>89.3 (0.4)</td></tr><tr><td>Mgelu</td><td>79.3 (0.2)</td><td>89.5 (0.3)</td></tr><tr><td>Mswish</td><td>79.6 (0.3)</td><td>89.2 (0.4)</td></tr><tr><td>Mtanh</td><td>79.2 (0.5)</td><td>88.9 (0.6)</td></tr><tr><td>Msa</td><td>79.5 (0.3)</td><td>89.6 (0.2)</td></tr><tr><td>Mconv</td><td>79.5 (0.6)</td><td>89.4 (0.5)</td></tr><tr><td>M attenu</td><td>79.0 (0.4)</td><td>89.1 (0.3)</td></tr><tr><td>Mblock</td><td>79.6 (0.5)</td><td>89.4 (0.3)</td></tr></table>
+
+Table 1: Results of the pilot experiments.
+
+settings unchanged. The three modified models are denoted as  $\mathcal{M}_{gelu}$ ,  $\mathcal{M}_{swish}$  and  $\mathcal{B}_{tanh}$ , respectively.
+
+Adding encoding operations We add a self-attention operation (Vaswani et al., 2017) or a convolutional operation of kernel size 3 after the downprojection and before the activation function. The two variants of model  $\mathcal{M}$  are denoted as  $\mathcal{M}_{sa}$  and  $\mathcal{M}_{conv}$ , respectively. Since extra operations introduce more parameters, we reduce the bottleneck dimension of  $\mathcal{M}_{sa}$  and  $\mathcal{M}_{conv}$  to 24 to ensure fair comparison.
+
+Alternative adapter placements Instead of inserting the adapter around the FFN module of the transformer block, we now consider: (1)  $\mathcal{M}_{\text {attn }}$  inserts the adapters at the attention modules (attn adapter); (2)  $\mathcal{M}_{\text {block }}$  inserts the adapters around the entire transformer block (block adapter). Note that the setting of block adapters is theoretically supported by the general framework of PETuning in He et al. (2021) but not considered by the previous work. In this work, we will demonstrate the usefulness of block adapters via experiments.
+
+Table 1 reports the experimental results of the above models. The evaluation metrics for the RTE and MRPC tasks are introduced by Appendix A.2. We can see that the four simple modifications to the baseline model,  $\mathcal{M}_{gelu}$ ,  $\mathcal{M}_{sa}$ ,  $\mathcal{M}_{conv}$  and  $\mathcal{M}_{block}$ , can slightly outperform  $\mathcal{M}$ , demonstrating that the adapter architectures are essential for adapter tuning, and it is promising to design better adapter architectures for better adapter tuning performances.
+
+The pilot experiments raise a vital research question: What are the optimal architectures for adapters? Obviously, such an optimal architecture will be different across tasks and PTM models and even across different intermediate layers of a PTM, making it impossible for manual designs. We are motivated to investigate the problem of optimizing the architectures of adapters via neural architecture
+
+search.
+
+# 3.2 General architecture of adapters
+
+As depicted in Figure 2, we now construct the search space of the Learned Adapter. The adapter is a bottleneck architecture with bottleneck dimension  $r$ , consisting of down-projection layer  $\mathrm{MLP}_d$ , an activation function  $g_1$ , an encoder layer  $\mathrm{Enc}$ , another activation function  $g_2$  and finally a up-projection layer  $\mathrm{MLP}_u$ . Formally, the hidden states  $h_x$  goes through the adapter and becomes
+
+$$
+h _ {x} ^ {(A)} = \operatorname {M L P} _ {u} \left(g _ {2} \left(\operatorname {E n c} \left(g _ {1} \left(\operatorname {M L P} _ {d} \left(h _ {x}\right)\right)\right)\right)\right). \tag {1}
+$$
+
+Following He et al. (2021), the hidden representation  $h_x$  will also go through the backbone's certain encoding module BEnc, and the adapted hidden states will become  $h_x' = \mathrm{BEnc}(h_x) + h_x^{(A)}$ . Following (Wu et al., 2022; Mahabadi et al., 2021; Le et al., 2021), we employ the parameterized hypercomplex multiplication (PHM) layer (Le et al., 2021) with parameter  $n$  to reduce the parameters of  $\mathrm{MLP}_d$  and  $\mathrm{MLP}_u$ . The PHM layer has a parameter complexity of  $\mathcal{O}(rd / n)$ , reducing the parameters of the projection layers by at most  $\frac{1}{n}$ .
+
+# 3.3 Search space
+
+We now formally introduce the search space of our Learned Adapter framework. The whole search space contains three types of search cells as shown in Figure 2:
+
+Activation Search Cell The Activation Search Cell is designated to choose the proper activation functions  $g_{1}$  and  $g_{1}$  from several candidates. Similar to So et al. (2019), the collection of candidate activation functions considered is: (a) ReLU (Agarap, 2018); (b) GeLU (Hendrycks and Gimpel, 2016); (c) SWISH (Ramachandran et al., 2017); (d) Tanh (Krizhevsky et al., 2012); (e) NulAct, which means no activation function and not to make changes to the input.
+
+Encoder Search Cell As is shown in Figure 2, different from (Wang et al., 2020; Zhu et al., 2021b), we construct our encoder cell as a simple DAG with a single edge. Our collection of encoder operations consists of the following four groups: (a) 1-d convolutional layers, with stride 1, same padding, output filters equal to the input's dimension, and kernel size equal to 1, 3, 5, or 7 (denoted as conv_k,  $k = 1, 3, 5, 7$ ). (b) Multi-head self-attention (MHA) layers (Vaswani et al., 2017), with head size equal to 2 or 8 (denoted as mha_k,
+
+$k = 2,8)$ . (c) Skip connection (He et al., 2015), denoted as skip-connect. (d) The null encoding operation that multiplies zero tensors to the input (null).<sup>1</sup>
+
+Adapter Placement Search Cell This search cell is designated to decide the placement of the adapter in an intermediate transformer block. We consider three candidate placements shown in Figure 2:² (a) FFN adapter, that is, to insert the adapter in parallel to the feed-forward module; (b) Attn adapter, parallel to the self-attention module; (c) Block adapter inserts the adapter in parallel to the whole transformer block. This placement option is supported by the theoretical analysis of He et al. (2021) but has not been considered by the literature. In the experiments, we will show that including the above three choices for adapter placements are necessary.
+
+Note that the above three search cells are single-edge DAGs. Following Pham et al. (2018); Wang et al. (2020); Zhu et al. (2021b); Zhu (2021a); Zhu et al. (2021d), we consider the macro search space, that is, different adapter architectures are learned for different intermediate layers. Intuitively, the macro search space allows for idiosyncratic architectures for different intermediate layers, leading to easier model adaptation.
+
+Despite the simple structures of search cell DAGs compared to the general NAS literature, our macro search space can result in  $6.38\mathrm{e} + 90$  combinations of different adapter architectures across different intermediate layers of the PTM backbones. Note that our search space contains the adapter architectures from Sung et al. (2022); Jie and Deng (2022) as special cases.
+
+# 4 Search Method
+
+# 4.1 Preliminaries on DARTS
+
+Assume there is a pre-defined space of operations denoted by  $\mathcal{O}$ , where each element,  $o(\cdot)$ , denotes a neural network operation, like convolutional operation, self-attention, activation, etc. DARTS (Liu et al., 2019a) initialize a hypernetwork in which each block is a search cell, that is, a fully connected directed acyclic graph (DAG) with N nodes. Let  $(i,j)$  denote a pair of
+
+nodes in the DAG. The core idea of DARTS is to use a weighted sum to include all  $|\mathcal{O}|$  operations in  $\mathcal{O}$ ,  $f_{i,j}(z_i) = \sum_{o\in \mathcal{O}}a_{i,j}^o\cdot o(z_i)$ , where  $a_{i,j}^{o} = \frac{\exp\alpha_{i,j}^{o}}{\sum_{o^{\prime}\in\mathcal{O}}\exp\alpha_{i,j}^{o^{\prime}}}$ ,  $z_{i}$  denotes the output of the  $i$ -th node, and  $\alpha_{i,j}^{o}$  is the architectural parameters that represents the weight (or the importance score) of  $o(\cdot)$  in edge  $(i,j)$ . The output of a node is the sum of all input flow, i.e.,  $z_{j} = \sum_{i < j}f_{i,j}(z_{i})$ . The output of the entire cell is formed by summing the last two nodes.
+
+This design makes the entire framework differentiable to both layer weights and architectural parameters  $\alpha_{i,j}^{o}$  so that it is possible to perform architecture search in an end-to-end fashion. After the search process is completed, the discretization procedure extracts the final sub-network by selecting the best operation on each edge and dropping the lower-score operations. And the final sub-network will train on the original train set with randomly initialized parameters.
+
+# 4.2 Discussion on the search method
+
+The standard optimization method for the above framework is the bi-level optimization proposed in DARTS (Liu et al., 2019a). However, there are recent works arguing that the single-level optimization method could also work for the DARTS framework. As pointed out by Bi et al. (2019) and Bi et al. (2020), bi-level optimization suffers considerable inaccuracy of gradient estimation and the potential instability can increase with the complexity of the search space. And Bi et al. (2020) conduct experiments to demonstrate that one-level optimization performs comparably with bi-level optimization but with better efficiency. Their experiments are conducted mainly on computer vision benchmarks like CIFAR-10 (Krizhevsky et al., 2012). In this work, we would like to investigate which optimization method is better under our framework.
+
+Note that the original DARTS requires one to re-train the learned networks from scratch after the search procedure, which definitely introduce additional computation costs. In this work, we propose to gradually discretize the hyper-network and obtain a sub-network without re-training. We will refer to this method as gradually discretizing neural architecture search (GDNAS). We first train the complete hyper-network  $M$  for  $K_{1}$  epochs. Then we select a edge  $e$  in the search space to discretize (for example, the edge in the encoder cell). Dis
+
+cretization simply means selecting the operation  $o_{*}^{e}$  with highest architectural parameter, and drop the other operations. Now we have obtain a new reduced hyper-network  $M$ . The discretized edge may cause the performance of the hyper-network to drop significantly, so we further finetune the hyper-network  $M$  for  $K_{2}$  epochs.
+
+In addition to the advantage of not to retrain the learned network, GDNAS retains the knowledge in the hyper-network, and obtain the performance gains against the re-trained sub-network. This is analogous to the model pruning literature, where a network pruned from a larger one is usually better than the network trained from scratch (Liang et al., 2021).
+
+# Algorithm 1: GDNAS
+
+Input: A hyper-network  $M$ , all edges  $\mathcal{E}$  on hyper-network  $M$ ;  
+Output: Set of selected operations  $\{o_e^*\}_{e\in \mathcal{E}}$   
+Data: Training set  $D_{train}$ , a batch of validation data  $B_{val}$   
+1 Train hyper-network  $M$  on the training set  $D_{train}$  for  $K_{1}$  epochs;  
+2 for edge  $e$  in  $\mathcal{E}$  do  
+3 Select the best operation  $o_e^* \gets \arg \max_o o_e^*$ ;  
+4 Discretize edge  $e$  of hyper-network  $M$  by only keeping  $o_e^*$ ;  
+5 Further train the hyper-network  $M$  on  $D_{train}$  for  $K_{2}$  epochs;
+
+# 5 Experiments
+
+# 5.1 Evaluation datasets
+
+We evaluate the performance of the methods on the GLUE (Wang et al., 2018) benchmarks. These benchmarks cover multiple tasks of paraphrase detection (MRPC, QQP), sentiment classification (SST-2), natural language inference (MNLI, RTE, QNLI), linguistic acceptability (CoLA).
+
+Since the original test sets of the GLUE benchmark are not publicly available, we follow Zhang et al. (2020) and Mahabadi et al. (2021) to construct the train/dev/test splits as follows to ensure a fair comparison: (a) for datasets with fewer than 10k samples (RTE, MRPC, STS-B, CoLA), we divide
+
+the original validation set in half, using one half for validation and the other for testing. (b) for larger datasets, we split 1k samples from the training set as the development set, and use the original development set as the test set. The detailed statistics and evaluation metrics of the GLUE benchmark is presented in Table 7 of Appendix A.
+
+# 5.2 Experiment Settings
+
+We run all the experiments on NVIDIA V100 32GB GPUs. We mainly evaluate our method on the GLUE benchmarks with the RoBERTa-large (Liu et al., 2019c) backbone model. We also evaluate our framework with the DeBERTa-large (He et al., 2020) and GPT2-large (Radford et al., 2019) backbones. We use the HugginFace Transformers (Wolf et al., 2020) for implementing all the methods. Unless otherwise specified, GDNAS will adopt the bi-level optimization method of DARTS. For GDNAS' discretization procedure, we set  $K_{1} = 5$  and  $K_{2} = 0.5$  on large datasets (SST-2, QNLI, QQP and MNLI), and  $K_{1} = 20$  and  $K_{2} = 2$  on low-resource datasets. And the batch size is set to 128 for dataset with more than 10k training samples, and 32 otherwise. For Learned Adapter, we set the bottleneck dimension  $r$  to 32 and select at most one adapter at each transformer layer. For the PHM layers, we use the PyTorch implementation of Le et al. (2021) and set  $n$  to 8. We run each task under 5 different random seeds and report the average performance and standard deviation. More details of the experimental settings are put in Appendix B.
+
+# 5.3 Baselines
+
+We compare our Learned Adapter framework with the current SOTA baseline methods.
+
+Fine-tune The traditional fine-tuning methods that train all parameters in the PTM backbone.
+
+Adapter-based tuning For adapter-based tuning methods, we compare with: (1) Adapter (Houlsby et al., 2019); (2) Compacter (Mahabadi et al., 2021); (3) Parallel Adapter proposed by (He et al., 2021) added on the FFN module; (4) LST (Sung et al., 2022). We re-implement Parallel Adapter with PHM projection layers  $(n = 8)$ .
+
+Prompt-based tuning For prompt-based tuning methods, we compare with (1) Prompt Tuning (Lester et al., 2021), (2) P-tuning v2 (Liu et al., 2021). The number of prompt tokens in these methods is set to 20.
+
+<table><tr><td>Method</td><td>Tunable 
+Params</td><td>CoLA 
+(mcc)</td><td>SST-2 
+(acc)</td><td>MRPC 
+(acc-f1)</td><td>QQP 
+(acc-f1)</td><td>STS-B 
+(corr)</td><td>MNLI 
+(acc)</td><td>QNLI 
+(acc)</td><td>RTE 
+(acc)</td><td>Avg</td></tr><tr><td colspan="11">Baselines</td></tr><tr><td>RoBERTa-large</td><td>355M</td><td>65.3</td><td>95.4</td><td>91.5</td><td>90.3</td><td>91.8</td><td>89.8</td><td>94.5</td><td>80.1</td><td>87.3</td></tr><tr><td>Prompt Tuning</td><td>21K</td><td>54.6 (2.5)</td><td>90.9 (0.5)</td><td>74.8 (1.3)</td><td>87.4 (0.5)</td><td>90.1 (0.3)</td><td>83.4 (0.2)</td><td>92.4 (0.3)</td><td>68.7 (1.9)</td><td>80.3</td></tr><tr><td>P-tuning v2</td><td>985K</td><td>57.3 (2.1)</td><td>92.3 (0.3)</td><td>86.9 (1.2)</td><td>88.1 (0.4)</td><td>90.4 (0.1)</td><td>87.3 (0.2)</td><td>92.9 (0.4)</td><td>77.2 (2.1)</td><td>84.1</td></tr><tr><td>BitFit</td><td>273K</td><td>59.2 (0.9)</td><td>94.1 (0.3)</td><td>88.6 (0.8)</td><td>88.5 (0.6)</td><td>91.1 (0.3)</td><td>87.3 (0.1)</td><td>93.2 (0.2)</td><td>78.6 (1.4)</td><td>85.1</td></tr><tr><td>LoRA</td><td>778K</td><td>59.7 (1.4)</td><td>93.6 (0.1)</td><td>88.9 (0.7)</td><td>88.3 (0.4)</td><td>90.9 (0.2)</td><td>87.9 (0.2)</td><td>93.2 (0.1)</td><td>78.8 (1.3)</td><td>85.2</td></tr><tr><td>UNIPELT</td><td>1.4M</td><td>61.5 (1.7)</td><td>93.6 (0.2)</td><td>89.2 (0.8)</td><td>88.7 (0.4)</td><td>91.0 (0.5)</td><td>87.5 (0.2)</td><td>92.9 (0.1)</td><td>79.1 (0.9)</td><td>85.4</td></tr><tr><td>Adapter</td><td>1.6M</td><td>63.1 (1.4)</td><td>93.8 (0.1)</td><td>87.5 (0.8)</td><td>88.7 (0.4)</td><td>90.9 (0.4)</td><td>88.5 (0.2)</td><td>93.2 (0.1)</td><td>78.9 (1.3)</td><td>85.6</td></tr><tr><td>LST</td><td>1.7M</td><td>63.6 (1.5)</td><td>93.9 (0.2)</td><td>89.2 (0.7)</td><td>88.7 (0.2)</td><td>91.2 (0.4)</td><td>88.3 (0.1)</td><td>93.0 (0.2)</td><td>79.2 (0.9)</td><td>85.9</td></tr><tr><td>Parallel Adapters</td><td>279K</td><td>63.8 (1.5)</td><td>94.2 (0.3)</td><td>89.3 (0.5)</td><td>88.9 (0.6)</td><td>91.1 (0.4)</td><td>88.3 (0.1)</td><td>93.2 (0.1)</td><td>79.1 (0.5)</td><td>86.0</td></tr><tr><td>Compacter</td><td>279K</td><td>63.7 (0.9)</td><td>94.2 (0.4)</td><td>89.1 (0.7)</td><td>88.6 (0.3)</td><td>90.8 (0.2)</td><td>88.2 (0.2)</td><td>92.8 (0.2)</td><td>78.7 (0.9)</td><td>85.8</td></tr><tr><td colspan="11">Our proposed methods</td></tr><tr><td>Learned Adapter</td><td>294K</td><td>64.3 (0.9)</td><td>94.9 (0.3)</td><td>89.8 (0.9)</td><td>89.2 (0.3)</td><td>91.3 (0.2)</td><td>88.6 (0.2)</td><td>93.5 (0.2)</td><td>80.4 (1.3)</td><td>86.5</td></tr></table>
+
+Table 2: The Overall comparison on the GLUE benchmark with RoBERTa-large backbone. We report mean and standard deviation of performance over 5 random seeds. Bold and Underline indicate the best and the second best results. The metric for each task is explained in Appendix A.2.
+
+Other PETuning methods We also compare: (1) BitFit (Ben-Zaken et al., 2021); (3) LoRA (Hu et al., 2021); (4) UNIPET (Mao et al., 2021) combines different types of PETuning methods in an automatic manner.
+
+We implement Aadpter, BitFit, and LoRA using the OpenDelta library. Other baselines are implemented using their open-sourced codes with their default settings. For a fair comparison, we do not use supplementary training like Wu et al. (2022) to enhance performance.
+
+# 5.4 Results on the GLUE benchmark
+
+Table 2 shows the results on GLUE with RoBERTa-large. Our Learned Adapter framework, outperforms the previous PETuning methods and notably preserves  $99.4\%$  performance of the full-model fine-tuning method while only tuning 240K to 300K parameters.
+
+We can observe from Table 2 that: (a) Note that our Learned Adapter framework obtains further improvements by automatically designing adapter architectures for different intermediate layers of the PTM. (b) Note that although we add encoding operations in adapters, the total tunable parameters of the Learned Adapter in the macro setting are fewer than Compacter since our framework can automatically drop adapters on certain layers when necessary.
+
+# 5.5 Further analysis
+
+Explanations of the searched architectures To understand the searched adapter architectures under our Learned Adapter framework, we present the
+
+learned adapter architectures on the RTE and SST-2 tasks on Table 9 and 10 in Appendix D, respectively. From the learned adapter architectures, we can observe that: (a) The adapter architecture varies for different layers, showing that different layers require different adapter architectures. (b) On each task, Learned Adapter chooses the null encoding operation on 3-5 intermediate layers, meaning to drop the adapters on these layers. (c) Regarding the adapter placement choices, we find that on each task, all three placement candidates, FFN adapter, Attn adapter, and Block adapter, are selected. This observation demonstrates that introducing block adapters into our search space is necessary. (d) Most adapters select the convolutional operations, and multi-head self-attention operations tend to occur in adapters of deeper layers. (e) around half of the learned adapters choose the NullAct for the second activation function  $g_{2}$ . Furthermore, we observe that there are adapters on the deeper transformer layers that requires no activation function but an encoder operation, demonstrating novel design patterns for adapters.
+
+Exploring the limit of parameter efficiency To explore the limit of parameter efficiency, we train the Learned Adapter, and Compacter (Mahabadi et al., 2021) with different rank parameters  $n \in \{1,2,4,8,16,32\}$ . Note that in the main experiments, we set  $n$  equal to 4. With larger  $n$ , the parameters of adapters will increase proportionally. In Figure 3, we demonstrate the results of the RTE and SST-2 tasks. We can see that the advantages of our Learned Adapter framework become more prominent with lower tunable parameter budgets. The results demonstrate that our framework can
+
+<table><tr><td rowspan="2">Source</td><td colspan="2">Target</td></tr><tr><td>RTE</td><td>SST-2</td></tr><tr><td>RTE</td><td>80.4 (1.3)</td><td>94.6 (0.1)</td></tr><tr><td>SST-2</td><td>80.3 (1.4)</td><td>94.9 (0.3)</td></tr></table>
+
+Table 3: Architecture transfer from source datasets to target datasets. The target datasets are in the column names, and the source datasets are in the row names.  
+
+<table><tr><td>Search space</td><td>RTE (acc)</td><td>SST-2 (acc)</td></tr><tr><td>S</td><td>80.4 (1.3)</td><td>94.9 (0.3)</td></tr><tr><td>S1</td><td>80.1 (1.5)</td><td>94.6 (0.3)</td></tr><tr><td>S2</td><td>79.8 (1.0)</td><td>94.4 (0.4)</td></tr><tr><td>Parallel Adapter</td><td>79.1 (0.5)</td><td>94.2 (0.3)</td></tr></table>
+
+effectively deliver the most proper architectures under the given parameter budgets and boost the performance of adapter-based tuning.
+
+Architectures' Transferability We now evaluate the transferability of the searched structures by the Learned Adapter. The RTE and SST-2 datasets are used as source and target datasets. We search the adapter architectures on the source dataset and train the searched architectures from scratch on the target task, and report the average and standard deviation of scores over 5 random seeds on table 3. We can see from Table 3 that the searched architectures are highly transferable. The transferred architectures can already achieve comparable or better performances than most baseline models (Table 2). The transferability guarantees the reusability of the searched adapter architectures.
+
+Ablation studies on the search space We now conduct an ablation study of our search space by reducing our search space  $S$  to a singleton step-by-step: (a) reduce the activation cells by only keeping the ReLU activation for  $g_{1}$  and the NullAct for  $g_{2}$ $(S_{1})$ ; (b) further reduce the encoder cell to only include skip-connect  $(S_{2})$ ; (c) further reduce the adapter placement cell to only include FFN adapter, and now the search space only contains Parallel Adapter (He et al., 2021). Table 4 reports the search results on different search spaces, showing that that dropping any components of the whole search space results in performance losses. The results demonstrate that each search cell in our search space design is necessary and beneficial.
+
+Working with other PTM backbones To verify the general applicability of our Learned Adapter
+
+Table 4: Experimental results for the ablation study of our Learned Adapter search space.  
+
+<table><tr><td>Method</td><td>Tunable 
+Params</td><td>RTE 
+(acc)</td><td>SST-2 
+(acc)</td></tr><tr><td colspan="4">DeBERTa-large backbone</td></tr><tr><td>Fine-tuning</td><td>406M</td><td>82.1 (1.1)</td><td>95.7 (0.2)</td></tr><tr><td>Adapters</td><td>1.6M</td><td>80.7 (1.3)</td><td>93.8 (0.1)</td></tr><tr><td>Parallel Adapters</td><td>279K</td><td>81.5 (1.2)</td><td>93.9 (0.3)</td></tr><tr><td>Learned Adapter</td><td>245K</td><td>82.0 (0.9)</td><td>94.5 (0.3)</td></tr><tr><td colspan="4">GPT2-large backbone</td></tr><tr><td>Fine-tuning</td><td>774M</td><td>79.5 (0.8)</td><td>95.5 (0.1)</td></tr><tr><td>Adapters</td><td>2.7M</td><td>78.1 (0.9)</td><td>93.9 (0.1)</td></tr><tr><td>Parallel Adapters</td><td>462K</td><td>78.5 (1.2)</td><td>93.8 (0.1)</td></tr><tr><td>Learned Adapter</td><td>482K</td><td>79.1 (0.6)</td><td>94.2 (0.2)</td></tr></table>
+
+Table 5: Results on 2 GLUE tasks using DeBERTa-large and GPT2-large models as the backbone. Bold indicates the best PETuning results.
+
+framework, we also conduct experiments on two other widely used PTM backbones, DeBERTa-large (He et al., 2020), and GPT2-large (Radford et al., 2019). The results are shown in Table 5. Our Learned Adapter successfully outperforms the adapter-based tuning baselines on both pre-trained backbones. This result enhances the reliability of our framework.
+
+We now validate our Learned Adapter framework on other pre-trained backbone: DeBERTa-large (He et al., 2020) and GPT2-large (Radford et al., 2019). The results are presented in Table 5.
+
+# 5.6 Discussions on the search method
+
+Search efficiency of GDNAS We use the RTE task to demonstrate the search efficiency. Running the RTE task with DARTS takes 1.5h (70.5min for bi-level optimization for 25 epochs and 21.6min for re-training with 25 epochs). Since GDNAS does not require re-training, it requires 1.2h (73.3min for training the hyper-network for  $k_{1} + 3*K_{2} = 26$  epochs). Our method consumes around three times the training time of Parallel Adapter (He et al., 2021), which is affordable compared to manually designing different architectures and running numerous evaluations.
+
+Ablation study of search methods We now run Learned Adapter with GDNAS with single-level optimization, the original DARTS (Liu et al., 2019a) and ENAS (Cai et al., 2018). The results are shown in Table 6. The results demonstrate that our GDNAS are effective in discovering better adapter architectures. In addition, the results demonstrate that bi-level optimization obtains slightly better results.
+
+Performance on a NAS benchmark To further
+
+![](images/4c888ef99f557d4305707afded8b40fe105d7428728493201240466644a83bc3.jpg)  
+(a) RTE
+
+![](images/35eb84d134582a80016aeb32d56e67c6d2686d7f0f9a4b02c35440ee420dc8b7.jpg)  
+(b) SST-2  
+Figure 3: Performances under different PHM rank  $n$ . The  $x$ -axis represents the number of tunable parameters, and the  $y$ -axis represents the performance. The performance of full-model fine-tuning is in dotted horizontal line.
+
+<table><tr><td>Search method</td><td>RTE (acc)</td><td>SST-2 (acc)</td></tr><tr><td>GDNAS</td><td>80.4 (1.3)</td><td>94.9 (0.3)</td></tr><tr><td>GDNAS (single-level)</td><td>80.2 (0.9)</td><td>94.7 (0.2)</td></tr><tr><td>DARTS</td><td>79.8 (1.2)</td><td>94.5 (0.5)</td></tr><tr><td>ENAS</td><td>80.3 (1.0)</td><td>94.3 (0.4)</td></tr></table>
+
+Table 6: Experimental results for the ablation study of the search methods.
+
+validate our GDNAS method, we conduct experiments on the widely studied neural architecture search benchmark dataset, CIFAR-10 (Krizhevsky, 2009). The results are reported in Table 8 of Appendix C. Our GDNAS method achieves  $2.52\%$  test error, which manageable search cost of 0.6 GPU days.
+
+# 6 Conclusion
+
+In this work, we propose the Learned Adapter framework, which automatically optimizes the adapter architectures. First, we design a unified search space for adapters, taking into account the recent works of manual adapter designs. Second, in light of the issues in the DARTS method, we propose a novel GDNAS method that can deliver better adapter architectures and requires no re-training of the learned adapter architectures. We run extensive experiments and analyses on the GLUE benchmark, demonstrating that our Learned Adapter framework can achieve better tuning performances than the baselines while maintaining parameter efficiency.
+
+# Limitations
+
+We showed that our proposed method can greatly improve the performance of parameter efficient tuning on diverse NLU tasks and three different pre
+
+trained models (i.e., RoBERTa-large, DeBERTa-large and GPT2-large). However, we acknowledge the following limitations: (a) the more super-sized pretrained models with tens of billions of or more parameters were not studied due to limited computation resources. (b) Other tasks in natural language processing, like the text generation tasks, were also not considered. But our framework can be easily transferred to other backbone architectures and different types of tasks. It would be of interest to investigate if the superiority of our method holds for other backbone models and types of tasks. And we will explore it in future work.
+
+# Ethics Statement
+
+The finding and proposed method aims to improve the adapter based tuning in terms of tuning parameters and performances. The used datasets are widely used in previous work and, to our knowledge, do not have any attached privacy or ethical issues.
+
+# References
+
+Abien Fred Agarap. 2018. Deep learning using rectified linear units (relu). ArXiv, abs/1803.08375.  
+Elad Ben-Zaken, Shauli Ravfogel, and Yoav Goldberg. 2021. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. ArXiv, abs/2106.10199.  
+Kaifeng Bi, Changping Hu, Lingxi Xie, Xin Chen, Longhui Wei, and Qi Tian. 2019. Stabilizing DARTS with Amended Gradient Estimation on Architectural Parameters. arXiv e-prints, page arXiv:1910.11831.  
+Kaifeng Bi, Lingxi Xie, Xin Chen, Longhui Wei, and Qi Tian. 2020. Gold-nas: Gradual, one-level, differentiable. ArXiv, abs/2007.03331.
+
+Han Cai, Tianyao Chen, Weinan Zhang, Yong Yu, and Jun Wang. 2018. Efficient architecture search by network transformation. In AAAI.  
+Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. 2021. Progressive darts: Bridging the optimization gap for nas in the wild. ArXiv, abs/1912.10952.  
+Xiangxiang Chu, Bo Zhang, Ruijun Xu, and Jixiang Li. 2021. Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12219-12228.  
+Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. pages 177-190.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+William B. Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the Third International Workshop on Paraphrasing (IWP2005).  
+Xiangxiang Gao, Wei Zhu, Jiasheng Gao, and Congrui Yin. 2023. F-pabee: Flexible-patience-based early exiting for single-label and multi-label text classification tasks. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE.  
+Yuan Gao, Haoping Bai, Zequn Jie, Jiayi Ma, Kui Jia, and Wei Liu. 2020. Mtl-nas: Task-agnostic neural architecture search towards general-purpose multitask learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11540-11549.  
+Demi Guo, Alexander Rush, and Yoon Kim. 2021a. Parameter-efficient transfer learning with diff pruning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4884-4896, Online. Association for Computational Linguistics.  
+Zhao Guo, Yuan Ni, Keqiang Wang, Wei Zhu, and Guotong Xie. 2021b. Global attention decoder for Chinese spelling error correction. In *Findings of the Association for Computational Linguistics: ACLIJCNLP* 2021, pages 1419–1428, Online. Association for Computational Linguistics.  
+Xu Han, Zhengyan Zhang, Ning Ding, Yuxian Gu, Xiao Liu, Yuqi Huo, Jiezhong Qiu, Liang Zhang, Wentao Han, Minlie Huang, Qin Jin, Yanyan Lan, Yang Liu,
+
+Zhiyuan Liu, Zhiwu Lu, Xipeng Qiu, Ruihua Song, Jie Tang, Ji rong Wen, Jinhui Yuan, Wayne Xin Zhao, and Jun Zhu. 2021. Pre-trained models: Past, present and future. ArXiv, abs/2106.07139.  
+Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. 2021. Towards a unified view of parameter-efficient transfer learning. ArXiv, abs/2110.04366.  
+Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778.  
+Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. ArXiv, abs/2006.03654.  
+Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (gelus). arXiv: Learning.  
+Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.  
+Shibo Jie and Zhifang Deng. 2022. Convolutional by-passes are better vision transformer adapters. ArXiv, abs/2207.07039.  
+A. Krizhevsky. 2009. Learning multiple layers of features from tiny images.  
+Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. 2012. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60:84 - 90.  
+Tuan Le, Marco Bertolini, Frank No'e, and Djork-Arné Clevert. 2021. Parameterized hypercomplex graph neural networks for graph classification. In International Conference on Artificial Neural Networks.  
+Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691.  
+Hector J. Levesque, Ernest Davis, and L. Morgenstern. 2011. The winograd schema challenge. In International Conference on Principles of Knowledge Representation and Reasoning.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190.
+
+Xiepeng Li, Zhexi Zhang, Wei Zhu, Zheng Li, Yuan Ni, Peng Gao, Junchi Yan, and Guotong Xie. 2019. Pingan smart health and SJTU at COIN - shared task: utilizing pre-trained language models and commonsense knowledge in machine reading tasks. In Proceedings of the First Workshop on Commonsense Inference in Natural Language Processing, pages 93-98, Hong Kong, China. Association for Computational Linguistics.  
+Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. 2021. Pruning and Quantization for Deep Neural Network Acceleration: A Survey. arXiv e-prints, page arXiv:2101.09671.  
+Chenxi Liu, Barret Zoph, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Loddon Yuille, Jonathan Huang, and Kevin P. Murphy. 2018. Progressive neural architecture search. In ECCV.  
+Hanxiao Liu, Karen Simonyan, and Yiming Yang. 2019a. Darts: Differentiable architecture search. ArXiv, abs/1806.09055.  
+Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2021. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. ArXiv, abs/2110.07602.  
+Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fin-tuning across scales and tasks. In Annual Meeting of the Association for Computational Linguistics.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019b. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019c. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692.  
+Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. 2021. Compacter: Efficient low-rank hypercomplex adapter layers. In NeurIPS.  
+Yuning Mao, Lambert Mathias, Rui Hou, Amjad Almahairi, Hao Ma, Jiawei Han, Wen tau Yih, and Madian Khabsa. 2021. Unipelt: A unified framework for parameter-efficient language model tuning. ArXiv, abs/2110.07577.  
+Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting, and Iryna Gurevych. 2022. Adaptable adapters. In North American Chapter of the Association for Computational Linguistics.  
+Niv Nayman, Asaf Noy, T. Ridnik, Itamar Friedman, Rong Jin, and Lihi Zelnik-Manor. 2019. Xnas: Neural architecture search with expert advice. *ArXiv*, abs/1906.08031.
+
+Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In North American Chapter of the Association for Computational Linguistics.  
+Jonas Pfeiffer, Aishwarya Kamath, Andreas Rückle, Kyunghyun Cho, and Iryna Gurevych. 2021. AdapterFusion: Non-destructive task composition for transfer learning. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 487-503, Online. Association for Computational Linguistics.  
+Hieu Pham, Melody Y. Guan, Barret Zoph, Quoc V. Le, and Jeff Dean. 2018. Efficient neural architecture search via parameter sharing. In ICML.  
+Alec Radford and Karthik Narasimhan. 2018. Improving language understanding by generative pretraining.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. *ArXiv*, abs/1910.10683.  
+Prajit Ramachandran, Barret Zoph, and Quoc V. Le. 2017. Swish: a self-gated activation function. arXiv: Neural and Evolutionary Computing.  
+Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V. Le. 2018. Regularized evolution for image classifier architecture search. In AAAI Conference on Artificial Intelligence.  
+Andreas Rückle, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2020. Adapterdrop: On the efficiency of adapters in transformers. In Conference on Empirical Methods in Natural Language Processing.  
+David R. So, Chen Liang, and Quoc V. Le. 2019. The evolved transformer. ArXiv, abs/1901.11117.  
+Haixia Sun, Jin Xiao, Wei Zhu, Yilong He, Sheng Zhang, Xiaowei Xu, Li Hou, Jiao Li, Yuan Ni, and Guotong Xie. 2020. Medical knowledge graph to enhance fraud, waste, and abuse detection on claim data: Model development and performance evaluation. JMIR Med Inform, 8(7):e17653.  
+Tianxiang Sun, Xiangyang Liu, Wei Zhu, Zhichao Geng, Lingling Wu, Yilong He, Yuan Ni, Guotong Xie, Xuanjing Huang, and Xipeng Qiu. 2022. A simple hash-based early exiting approach for language understanding and generation. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 2409-2421, Dublin, Ireland. Association for Computational Linguistics.
+
+Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. 2022. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. ArXiv, abs/2206.06522.  
+Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. ArXiv, abs/1706.03762.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In BlackboxNLP@EMNLP.  
+Yujing Wang, Yaming Yang, Yiren Chen, Jing Bai, Ce Zhang, Guinan Su, Xiaoyu Kou, Yunhai Tong, Mao Yang, and Lidong Zhou. 2020. Textnas: A neural architecture search space tailored for text representation. In AAAI.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Zhuofeng Wu, Sinong Wang, Jiatao Gu, Rui Hou, Yuxiao Dong, V. G. Vinod Vydiswaran, and Hao Ma. 2022. Idpg: An instance-dependent prompt generation method. In North American Chapter of the Association for Computational Linguistics.  
+Sirui Xie, Hehui Zheng, Chunxiao Liu, and Liang Lin. 2018. Snas: Stochastic neural architecture search. ArXiv, abs/1812.09926.  
+Sirui Xie, Hehui Zheng, Chunxiao Liu, and Liang Lin. 2019. Snas: Stochastic neural architecture search. ArXiv, abs/1812.09926.  
+Yuhui Xu, Lingxi Xie, Wenrui Dai, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Hongkai Xiong, and Qi Tian. 2021. Partially-connected neural architecture search for reduced computational redundancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:2953-2970.  
+Arber Zela, Thomas Elsken, Tonmoy Saikia, Yassine Marrakchi, Thomas Brox, and Frank Hutter. 2019. Understanding and robustifying differentiable architecture search. ArXiv, abs/1909.09656.  
+Tianyi Zhang, Felix Wu, Arzoo Katiyar, Kilian Q. Weinberger, and Yoav Artzi. 2020. Revisiting few-sample bert fine-tuning. *ArXiv*, abs/2006.05987.
+
+Zhen Zhang, Wei Zhu, Jinfan Zhang, Peng Wang, Rize Jin, and Tae-Sun Chung. 2022. PCEE-BERT: Accelerating BERT inference via patient and confident early exiting. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 327-338, Seattle, United States. Association for Computational Linguistics.  
+Zhexi Zhang, Wei Zhu, Junchi Yan, Peng Gao, and Guotong Xie. 2021. Automatic student network search for knowledge distillation. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2446-2453. IEEE.  
+Hongpeng Zhou, Minghao Yang, Jun Wang, and Wei Pan. 2019. Bayesnas: A bayesian approach for neural architecture search. *ArXiv*, abs/1905.04919.  
+Wei Zhu. 2021a. Autonlu: Architecture search for sentence and cross-sentence attention modeling with redesigned search space. In Natural Language Processing and Chinese Computing: 10th CCF International Conference, NLPCC 2021, Qingdao, China, October 13-17, 2021, Proceedings, Part I 10, pages 155-168. Springer.  
+Wei Zhu. 2021b. AutoRC: Improving BERT based relation classification models via architecture search. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: Student Research Workshop, pages 33-43, Online. Association for Computational Linguistics.  
+Wei Zhu. 2021c. LeeBERT: Learned early exit for BERT with cross-level optimization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2968-2980, Online. Association for Computational Linguistics.  
+Wei Zhu, Yilong He, Ling Chai, Yunxiao Fan, Yuan Ni, Guotong Xie, and Xiaoling Wang. 2021a. paht_nlp @ MEDIQA 2021: Multi-grained query focused multi-answer summarization. In Proceedings of the 20th Workshop on Biomedical Language Processing, pages 96-102, Online. Association for Computational Linguistics.  
+Wei Zhu, Yuan Ni, Xiaoling Wang, and Guotong Xie. 2021b. Discovering better model architectures for medical query understanding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Papers, pages 230-237, Online. Association for Computational Linguistics.  
+Wei Zhu, Peng Wang, Xiaoling Wang, Yuan Ni, and Guotong Xie. 2023. Acf: Aligned contrastive fin-tuning for language and vision tasks. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5.
+
+<table><tr><td>Category</td><td>Datasets</td><td>ltrainl</td><td>ldevl</td><td>ltestl</td><td>|y|</td><td>Type</td><td>Labels</td></tr><tr><td rowspan="2">Single-sentence</td><td>SST-2</td><td>66349</td><td>1000</td><td>872</td><td>2</td><td>sentiment</td><td>positive, negative</td></tr><tr><td>CoLA</td><td>8551</td><td>521</td><td>522</td><td>2</td><td>linguistic acceptability</td><td>acceptable, not acceptable</td></tr><tr><td rowspan="6">Sentence-pair</td><td>MNLI</td><td>391702</td><td>1000</td><td>19647</td><td>3</td><td>NLI</td><td>entailment, neutral, contradiction</td></tr><tr><td>MRPC</td><td>2668</td><td>1000</td><td>408</td><td>2</td><td>paraphrase</td><td>equivalent, not equivalent</td></tr><tr><td>QNLI</td><td>103743</td><td>1000</td><td>5463</td><td>2</td><td>NLI</td><td>entailment, not entailment</td></tr><tr><td>QQP</td><td>362846</td><td>1000</td><td>40430</td><td>2</td><td>paraphrase</td><td>equivalent, not equivalent</td></tr><tr><td>RTE</td><td>1490</td><td>1000</td><td>277</td><td>2</td><td>NLI</td><td>entailment, not entailment</td></tr><tr><td>STS-B</td><td>5750</td><td>750</td><td>751</td><td>-</td><td>semantic similarity</td><td>-</td></tr></table>
+
+Table 7: The dataset statistics of the GLUE benchmark tasks evaluated in this work. For MNLI task, the number of samples in the test set is summed by matched and mismatched samples.  $|\mathcal{V}|$  is the number of classes for a classification task.
+
+Wei Zhu, Xiaoling Wang, Yuan Ni, and Guo Tong Xie. 2021c. Gaml-bert: Improving bert early exiting by gradient aligned mutual learning. In EMNLP.
+
+Wei Zhu, Xiaoling Wang, Yuan Ni, and Guotong Xie. 2021d. Autotrans: Automating transformer design via reinforced architecture search. In *Natural Language Processing and Chinese Computing*, pages 169–182, Cham. Springer International Publishing.
+
+Wei Zhu, Xiaofeng Zhou, Keqiang Wang, Xun Luo, Xiepeng Li, Yuan Ni, and Guotong Xie. 2019. PANLP at MEDIQA 2019: Pre-trained language models, transfer learning and knowledge distillation. In Proceedings of the 18th BioNLP Workshop and Shared Task, pages 380-388, Florence, Italy. Association for Computational Linguistics.
+
+Barret Zoph and Quoc V. Le. 2017. Neural architecture search with reinforcement learning. *ArXiv*, abs/1611.01578.
+
+Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. 2017. Learning transferable architectures for scalable image recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8697-8710.
+
+Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. 2018. Learning transferable architectures for scalable image recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8697-8710.
+
+Yuhui Zuo, Wei Zhu, and Guoyong GUET Cai. 2022. Continually detection, rapidly react: Unseen rumors detection based on continual prompt-tuning. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3029-3041, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.
+
+# A Datasets and evaluation metrics
+
+# A.1 Dataset splits
+
+The detailed statistics of the GLUE benchmark is presented in Table 7.
+
+# A.2 Evaluation metrics
+
+For MNLI, we report the average of the accuracy scores on the matched and mis-matched test set. For MRPC and QQP, we report acc-f1, which is the average of accuracy and F1 scores. For STS-B, we report corr, which denotes the average of the Pearson and Spearman correlation coefficients. For CoLA, we report mcc, which is the Matthews correlation. For all other tasks, we report accuracy.
+
+# B Appendix for experimental settings
+
+Details of experiments We run our methods and baseline method on the GLUE (Wang et al., 2018) following the previous works. All datasets are downloaded via the HuggingFace Datasets library. Since the test split of these tasks are invisible to the researchers, we randomly split off 1k samples from the training set as validation set for large datasets(QQP, QNLI, SST2, MNLI), and use the remaining as the training set, and use the original validation set as the test set. For other datasets, we randomly split the original validation set in half as the validation set and the test set, and use the original train set as our train set. The same dataset is split differently with different random seeds.
+
+For each experiment setting, we repeat the experiment with 5 seeds, and report the average score and standard deviation. In all experiments, the maximum sequence length is 128 for the tasks. We mainly use the RoBERTa-large model (355M parameters) as the backbone model, but we also adopt the GPT2-large and DeBERTa-large the backbone models for ablation studies. We freeze the pretrained parameters in all experiments except full-model finetuning. We use AdamW as the optimizer with a linear learning rate decay schedule and  $6\%$  of the training steps for warm-up.
+
+We train the hyper-network on the train set  $D_{train}$  following (Liu et al., 2019a). For training epochs, we set  $K_{1} = 5$  and  $K_{2} = 1$  on large datasets (SST-2, QNLI, QQP and MNLI), and  $K_{1} = 20$  and  $K_{2} = 5$  on low-resource datasets. We will run the search procedure once for each task.
+
+After the hyper-network is fully discretized, instead of retraining from scratch, we further train the remained network for  $K_{2}$  epochs, and we evaluate the model on the dev set and save the model checkpoint every  $I_{eval}$  steps. The best checkpoint on the dev set is used to run predictions on the test set. We report the average scores on the test set and standard deviations across 5 random seeds.
+
+Other hyper-parameters We do pilot experiments on SST-2 using learning rates in  $\{2\mathrm{e} - 5,5\mathrm{e} - 5,$ $1\mathrm{e} - 4,2\mathrm{e} - 4\}$  , and find that 1e-4 performs the best. For fine-tuning, we try learning rates in  $\{1\mathrm{e} - 5,2\mathrm{e} -$ $5,5\mathrm{e} - 5\}$  and find that 2e-5 performs the best. The number of training epochs for the baselines is set as  $K = 5$  on large datasets (SST-2, QNLI, QQP and MNLI), and  $K = 20$  on smaller datasets. We apply these hyper-parameters to all baselines, and no further hyperparameter-tuning are conducted. Therefore, the comparison is fair for all methods.
+
+# C Experimental results on the CIFAR-10 task
+
+To further validate that our GDNAS method can obtain better search performances than the DNAS baselines, we now conduct experiments in the general NAS setting. Following DARTS (Liu et al., 2019a), we conduct neural architecture search on the CIFAR-10 dataset (Krizhevsky, 2009) based on the search space of DARTS. We keep all the search settings identical to DARTS. We first train the hyper-network with frozen architectural weights for 50 epochs. After selecting the operation on an edge, we tune the hyper-net for 8 epochs to let the modified hyper-network to adjust. Following DARTS (Liu et al., 2019a), we run the search and architecture selection phase with four random seeds and report both the best and average test errors of the obtained architectures. The results are reported in Table 8, which compares GDNAS with the DNAS baseline methods. Our GDNAS method achieves  $2.52\%$  test error, which manageable search cost of 0.6 GPU days. The results of GDNAS is comparable to other method (like P-DARTS (Chen et al., 2021)) with more complex
+
+procedures.
+
+# D Learned architectures on the GLUE tasks
+
+In this section, we present the learned adapter architectures on the RTE and SST-2 tasks. The learned adapter architecture are presented in Table 9, 10, respectively.
+
+<table><tr><td>Architecture</td><td>Test Error (%)</td><td>Params (M)</td><td>Search Cost (GPU days)</td><td>Search Framework</td></tr><tr><td>NASNet-A (Zoph et al., 2017)</td><td>2.65</td><td>3.3</td><td>2000</td><td>RL</td></tr><tr><td>AmoebaNet (Real et al., 2018)</td><td>3.34</td><td>3.2</td><td>3150</td><td>evolution</td></tr><tr><td>ENAS (Pham et al., 2018)</td><td>2.89</td><td>4.6</td><td>0.5</td><td>RL</td></tr><tr><td>DARTS (Liu et al., 2019a)</td><td>3.02</td><td>3.3</td><td>0.4</td><td>DNAS</td></tr><tr><td>SNAS (Xie et al., 2018)</td><td>2.85</td><td>2.8</td><td>1.5</td><td>DNAS</td></tr><tr><td>BayesNAS (Zhou et al., 2019)</td><td>2.81</td><td>3.4</td><td>0.2</td><td>DNAS</td></tr><tr><td>P-DARTS (Chen et al., 2021)</td><td>2.50</td><td>3.4</td><td>0.3</td><td>DNAS</td></tr><tr><td>R-DARTS (Zela et al., 2019)</td><td>2.95</td><td>-</td><td>1.6</td><td>DNAS</td></tr><tr><td>GDNAS (ours)</td><td>2.52</td><td>3.3</td><td>0.6</td><td>DNAS</td></tr></table>
+
+Table 8: The search results on the CIFAR-10 task.  
+
+<table><tr><td>Layer index</td><td>Adapter placement</td><td>Activation g1</td><td>Activation g2</td><td>Encoder operation 1</td></tr><tr><td>1</td><td>FFN</td><td>elu</td><td>null_ACT</td><td>mha_8</td></tr><tr><td>2</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>3</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>4</td><td>FFN</td><td>elu</td><td>null_ACT</td><td>mha_2</td></tr><tr><td>5</td><td>Block</td><td>tanh</td><td>null_ACT</td><td>mha_2</td></tr><tr><td>6</td><td>FFN</td><td>gelu_new</td><td>leaky_relu</td><td>conv_3</td></tr><tr><td>7</td><td>FFN</td><td>null_ACT</td><td>tanh</td><td>mha_2</td></tr><tr><td>8</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>9</td><td>Attn</td><td>elu</td><td>relu</td><td>conv_3</td></tr><tr><td>10</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>11</td><td>Attn</td><td>gelu_new</td><td>relu</td><td>conv_1</td></tr><tr><td>12</td><td>Attn</td><td>gelu_new</td><td>relu</td><td>conv_3</td></tr><tr><td>13</td><td>Block</td><td>relu</td><td>leaky_relu</td><td>conv_1</td></tr><tr><td>14</td><td>Attn</td><td>swish</td><td>relu</td><td>conv_3</td></tr><tr><td>15</td><td>FFN</td><td>leaky_relu</td><td>relu</td><td>conv_3</td></tr><tr><td>16</td><td>Block</td><td>leaky_relu</td><td>relu</td><td>conv_5</td></tr><tr><td>17</td><td>FFN</td><td>leaky_relu</td><td>relu</td><td>conv_1</td></tr><tr><td>18</td><td>Attn</td><td>leaky_relu</td><td>null_ACT</td><td>skip_connect</td></tr><tr><td>19</td><td>FFN</td><td>relu</td><td>relu</td><td>conv_3</td></tr><tr><td>20</td><td>FFN</td><td>gelu_new</td><td>null_ACT</td><td>conv_3</td></tr><tr><td>21</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>22</td><td>Block</td><td>tanh</td><td>null_ACT</td><td>mha_8</td></tr><tr><td>23</td><td>Attn</td><td>tanh</td><td>null_ACT</td><td>conv_3</td></tr><tr><td>24</td><td>FFN</td><td>tanh</td><td>null_ACT</td><td>mha_2</td></tr></table>
+
+Table 9: The learned adapter architectures on the RTE task when the PTM backbone is RoBERTa-large. If an adapter's architecture contains only "-", it means our Learned Adapter framework choose the null encoder operation, and equivalently, dropping this layer's adapter.
+
+<table><tr><td>Layer index</td><td>Adapter placement</td><td>Activation g1</td><td>Activation g2</td><td>Encoder operation 1</td></tr><tr><td>1</td><td>Attn</td><td>null_ACT</td><td>null_ACT</td><td>conv_5</td></tr><tr><td>2</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>3</td><td>FFN</td><td>elu</td><td>null_ACT</td><td>mha_2</td></tr><tr><td>4</td><td>Attn</td><td>tanh</td><td>null_ACT</td><td>mha_2</td></tr><tr><td>5</td><td>Attn</td><td>elu</td><td>null_ACT</td><td>skip_connect</td></tr><tr><td>6</td><td>FFN</td><td>gelu_new</td><td>relu</td><td>conv_5</td></tr><tr><td>7</td><td>Block</td><td>leaky_relu</td><td>leaky_relu</td><td>conv_3</td></tr><tr><td>8</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>9</td><td>Block</td><td>leaky_relu</td><td>leaky_relu</td><td>conv_5</td></tr><tr><td>10</td><td>FFN</td><td>relu</td><td>relu</td><td>conv_3</td></tr><tr><td>11</td><td>Block</td><td>tanh</td><td>null_ACT</td><td>mha_8</td></tr><tr><td>12</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>13</td><td>Block</td><td>tanh</td><td>null_ACT</td><td>skip_connect</td></tr><tr><td>14</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>15</td><td>FFN</td><td>swish</td><td>swish</td><td>conv_1</td></tr><tr><td>16</td><td>Block</td><td>relu</td><td>relu</td><td>conv_1</td></tr><tr><td>17</td><td>FFN</td><td>relu</td><td>swish</td><td>conv_3</td></tr><tr><td>18</td><td>FFN</td><td>leaky_relu</td><td>relu</td><td>conv_3</td></tr><tr><td>19</td><td>Attn</td><td>gelu_new</td><td>relu</td><td>mha_8</td></tr><tr><td>20</td><td>FFN</td><td>swish</td><td>relu</td><td>mha_8</td></tr><tr><td>21</td><td>Attn</td><td>swish</td><td>null_ACT</td><td>conv_3</td></tr><tr><td>22</td><td>Attn</td><td>elu</td><td>null_ACT</td><td>conv_3</td></tr><tr><td>23</td><td>Attn</td><td>null_ACT</td><td>null_ACT</td><td>mha_8</td></tr><tr><td>24</td><td>Attn</td><td>elu</td><td>null_ACT</td><td>mha_2</td></tr></table>
+
+Table 10: The learned adapter architectures on the SST-2 task when the PTM backbone is RoBERTa-large. If an adapter's architecture contains only "-", it means our Learned Adapter framework choose the null encoder operation, and equivalently, dropping this layer's adapter.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Left blank.
+
+A2. Did you discuss any potential risks of your work?
+
+Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Left blank.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+No response.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+No response.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+No response.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+No response.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+No response.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+No response.
+
+C Did you run computational experiments?
+
+Section 6
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Section 6.2
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 6.2, Appendix C  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 6.4, Section 6.5, Section 6.6  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 6.2
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response.

2023/Learned Adapters Are Better Than Manually Designed Adapters/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9414813b5253e7e4027dd2771270b125f84c201de5e272b9841db3cf5701d95
+size 635267