Add Batch 25135a76-58d1-4c0e-ab0c-13934f179624 data

.gitattributes

@@ -4463,3 +4463,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/Factuality[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]A[[:space:]]Survey/5966e9c3-d456-4370-b820-b27e7c7acf15_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/FairFlow_[[:space:]]Mitigating[[:space:]]Dataset[[:space:]]Biases[[:space:]]through[[:space:]]Undecided[[:space:]]Learning[[:space:]]for[[:space:]]Natural[[:space:]]Language[[:space:]]Understanding/eb214dba-d827-43d4-825c-ee36dfc78e4d_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Fairer[[:space:]]Preferences[[:space:]]Elicit[[:space:]]Improved[[:space:]]Human-Aligned[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Judgments/0f4380f4-940d-47f6-93d2-de8dc29bea1b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fill[[:space:]]In[[:space:]]The[[:space:]]Gaps_[[:space:]]Model[[:space:]]Calibration[[:space:]]and[[:space:]]Generalization[[:space:]]with[[:space:]]Synthetic[[:space:]]Data/94f56978-61ff-41c3-b3a1-8b022f290de8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Filtered[[:space:]]Direct[[:space:]]Preference[[:space:]]Optimization/e2599cf1-1b15-4843-9792-84fb610259c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FinDVer_[[:space:]]Explainable[[:space:]]Claim[[:space:]]Verification[[:space:]]over[[:space:]]Long[[:space:]]and[[:space:]]Hybrid-content[[:space:]]Financial[[:space:]]Documents/3adc16a8-3414-4d00-95ba-3c3dbbf67df2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Finding[[:space:]]Blind[[:space:]]Spots[[:space:]]in[[:space:]]Evaluator[[:space:]]LLMs[[:space:]]with[[:space:]]Interpretable[[:space:]]Checklists/e1a38a24-c9fd-4bd0-b41f-2477351b5a81_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-Grained[[:space:]]Detection[[:space:]]of[[:space:]]Solidarity[[:space:]]for[[:space:]]Women[[:space:]]and[[:space:]]Migrants[[:space:]]in[[:space:]]155[[:space:]]Years[[:space:]]of[[:space:]]German[[:space:]]Parliamentary[[:space:]]Debates/ec6157f5-fff1-423b-b564-2664514ae3c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-Grained[[:space:]]Prediction[[:space:]]of[[:space:]]Reading[[:space:]]Comprehension[[:space:]]from[[:space:]]Eye[[:space:]]Movements/6bd75cfc-d816-4947-81b0-c69bbdffc4eb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-Tuning[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Translate_[[:space:]]Will[[:space:]]a[[:space:]]Touch[[:space:]]of[[:space:]]Noisy[[:space:]]Data[[:space:]]in[[:space:]]Misaligned[[:space:]]Languages[[:space:]]Suffice_/8e3a98d3-e6ed-4c7a-afff-0b0d7bc4b4fb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-Tuning[[:space:]]and[[:space:]]Prompt[[:space:]]Optimization_[[:space:]]Two[[:space:]]Great[[:space:]]Steps[[:space:]]that[[:space:]]Work[[:space:]]Better[[:space:]]Together/b6b2cc5a-f4ec-482a-8da6-09a5acae609f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-Tuning[[:space:]]or[[:space:]]Retrieval_[[:space:]]Comparing[[:space:]]Knowledge[[:space:]]Injection[[:space:]]in[[:space:]]LLMs/76375409-778a-4580-941a-496a79ee2e4d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-grained[[:space:]]Pluggable[[:space:]]Gradient[[:space:]]Ascent[[:space:]]for[[:space:]]Knowledge[[:space:]]Unlearning[[:space:]]in[[:space:]]Language[[:space:]]Models/32c59d2b-6d4d-44b6-9280-1a55f1a54257_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FineCops-Ref_[[:space:]]A[[:space:]]new[[:space:]]Dataset[[:space:]]and[[:space:]]Task[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Compositional[[:space:]]Referring[[:space:]]Expression[[:space:]]Comprehension/614bac07-8ac6-4550-8891-6e102a43e942_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Finer_[[:space:]]Investigating[[:space:]]and[[:space:]]Enhancing[[:space:]]Fine-Grained[[:space:]]Visual[[:space:]]Concept[[:space:]]Recognition[[:space:]]in[[:space:]]Large[[:space:]]Vision[[:space:]]Language[[:space:]]Models/f021ece7-5771-4856-af01-95f6483b3038_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/First[[:space:]]Heuristic[[:space:]]Then[[:space:]]Rational_[[:space:]]Dynamic[[:space:]]Use[[:space:]]of[[:space:]]Heuristics[[:space:]]in[[:space:]]Language[[:space:]]Model[[:space:]]Reasoning/43229adf-3f75-4880-b23e-4cdff23b7521_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fisher[[:space:]]Information-based[[:space:]]Efficient[[:space:]]Curriculum[[:space:]]Federated[[:space:]]Learning[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/7f50aeaa-f5d8-4f5b-a3ce-fc41cc92c39d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fishing[[:space:]]for[[:space:]]Magikarp_[[:space:]]Automatically[[:space:]]Detecting[[:space:]]Under-trained[[:space:]]Tokens[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/353e6d6b-cd52-4741-aeae-d8b632f3e322_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Flee[[:space:]]the[[:space:]]Flaw_[[:space:]]Annotating[[:space:]]the[[:space:]]Underlying[[:space:]]Logic[[:space:]]of[[:space:]]Fallacious[[:space:]]Arguments[[:space:]]Through[[:space:]]Templates[[:space:]]and[[:space:]]Slot-filling/86053e3b-8472-48b0-abf2-a03b8bc0d293_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FlipGuard_[[:space:]]Defending[[:space:]]Preference[[:space:]]Alignment[[:space:]]against[[:space:]]Update[[:space:]]Regression[[:space:]]with[[:space:]]Constrained[[:space:]]Optimization/31a77215-9255-4c25-9703-ebd3393bf38a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Focused[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Stable[[:space:]]Many-Shot[[:space:]]Learners/72712a40-d5de-47cf-a21d-295ee387be42_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FoodieQA_[[:space:]]A[[:space:]]Multimodal[[:space:]]Dataset[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Understanding[[:space:]]of[[:space:]]Chinese[[:space:]]Food[[:space:]]Culture/daf7be80-c1a3-4ece-8be4-8fcfcbe53417_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fool[[:space:]]Me[[:space:]]Once_[[:space:]]Contrasting[[:space:]]Textual[[:space:]]and[[:space:]]Visual[[:space:]]Explanations[[:space:]]in[[:space:]]a[[:space:]]Clinical[[:space:]]Decision-Support[[:space:]]Setting/7de78527-0f0a-4bf9-8bcf-5d732197f0fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Forgetting[[:space:]]Curve_[[:space:]]A[[:space:]]Reliable[[:space:]]Method[[:space:]]for[[:space:]]Evaluating[[:space:]]Memorization[[:space:]]Capability[[:space:]]for[[:space:]]Long-Context[[:space:]]Models/25d26c46-87f2-43c6-914a-6e8826c6dfc5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Formality[[:space:]]is[[:space:]]Favored_[[:space:]]Unraveling[[:space:]]the[[:space:]]Learning[[:space:]]Preferences[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]on[[:space:]]Data[[:space:]]with[[:space:]]Conflicting[[:space:]]Knowledge/f8f11c7a-4889-4553-864b-3eece86f93f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Foundational[[:space:]]Autoraters_[[:space:]]Taming[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Better[[:space:]]Automatic[[:space:]]Evaluation/e31f3770-eda0-4386-b633-318a4321988b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Free[[:space:]]your[[:space:]]mouse![[:space:]]Command[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Generate[[:space:]]Code[[:space:]]to[[:space:]]Format[[:space:]]Word[[:space:]]Documents/b395b777-ba32-4010-829a-244e3ec52179_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/From[[:space:]]Bottom[[:space:]]to[[:space:]]Top_[[:space:]]Extending[[:space:]]the[[:space:]]Potential[[:space:]]of[[:space:]]Parameter[[:space:]]Efficient[[:space:]]Fine-Tuning/e997327a-62fe-44f7-b2c0-29be58fb2ee1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/From[[:space:]]Descriptive[[:space:]]Richness[[:space:]]to[[:space:]]Bias_[[:space:]]Unveiling[[:space:]]the[[:space:]]Dark[[:space:]]Side[[:space:]]of[[:space:]]Generative[[:space:]]Image[[:space:]]Caption[[:space:]]Enrichment/1e4435f4-fc4b-4a9c-ad82-7a02eed5ac5e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/From[[:space:]]Insights[[:space:]]to[[:space:]]Actions_[[:space:]]The[[:space:]]Impact[[:space:]]of[[:space:]]Interpretability[[:space:]]and[[:space:]]Analysis[[:space:]]Research[[:space:]]on[[:space:]]NLP/925e5c64-8ce6-45f2-883e-d126d1ec9f2f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/From[[:space:]]LLMs[[:space:]]to[[:space:]]MLLMs_[[:space:]]Exploring[[:space:]]the[[:space:]]Landscape[[:space:]]of[[:space:]]Multimodal[[:space:]]Jailbreaking/8aeab8cb-0106-4998-bb30-b5aece2b8cad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/From[[:space:]]Local[[:space:]]Concepts[[:space:]]to[[:space:]]Universals_[[:space:]]Evaluating[[:space:]]the[[:space:]]Multicultural[[:space:]]Understanding[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/0c5d10c1-7000-432a-b8bb-cc9ff528d191_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/From[[:space:]]RAG[[:space:]]to[[:space:]]Riches_[[:space:]]Retrieval[[:space:]]Interlaced[[:space:]]with[[:space:]]Sequence[[:space:]]Generation/82c5a604-c678-4e97-b8f7-759f78703b5c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/From[[:space:]]the[[:space:]]Least[[:space:]]to[[:space:]]the[[:space:]]Most_[[:space:]]Building[[:space:]]a[[:space:]]Plug-and-Play[[:space:]]Visual[[:space:]]Reasoner[[:space:]]via[[:space:]]Data[[:space:]]Synthesis/bb6b7b4a-78a9-49d0-8843-15d0f34a1b47_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fuse[[:space:]]to[[:space:]]Forget_[[:space:]]Bias[[:space:]]Reduction[[:space:]]and[[:space:]]Selective[[:space:]]Memorization[[:space:]]through[[:space:]]Model[[:space:]]Fusion/be6d54e6-7daf-4a2e-bf62-d2b2061cfb7f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FuseGen_[[:space:]]PLM[[:space:]]Fusion[[:space:]]for[[:space:]]Data-generation[[:space:]]based[[:space:]]Zero-shot[[:space:]]Learning/d2d99be7-1dd8-4351-90f3-9ceee7efe0dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GAMA_[[:space:]]A[[:space:]]Large[[:space:]]Audio-Language[[:space:]]Model[[:space:]]with[[:space:]]Advanced[[:space:]]Audio[[:space:]]Understanding[[:space:]]and[[:space:]]Complex[[:space:]]Reasoning[[:space:]]Abilities/320ce742-4fcd-4eda-8bd8-fdc71fc44d28_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GDPO_[[:space:]]Learning[[:space:]]to[[:space:]]Directly[[:space:]]Align[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Diversity[[:space:]]Using[[:space:]]GFlowNets/ce5557c5-573c-45a7-8ad3-adf3f88bded9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GDTB_[[:space:]]Genre[[:space:]]Diverse[[:space:]]Data[[:space:]]for[[:space:]]English[[:space:]]Shallow[[:space:]]Discourse[[:space:]]Parsing[[:space:]]across[[:space:]]Modalities,[[:space:]]Text[[:space:]]Types,[[:space:]]and[[:space:]]Domains/bfbd5a7b-9727-4f4e-a47a-bda9922888ce_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GENRA_[[:space:]]Enhancing[[:space:]]Zero-shot[[:space:]]Retrieval[[:space:]]with[[:space:]]Rank[[:space:]]Aggregation/ad0e4b38-790e-4d29-89a2-674b5c785cc1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GLaPE_[[:space:]]Gold[[:space:]]Label-agnostic[[:space:]]Prompt[[:space:]]Evaluation[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/f19c59ad-b142-4ce3-a3d3-3c92e8e34dbb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GOME_[[:space:]]Grounding-based[[:space:]]Metaphor[[:space:]]Binding[[:space:]]With[[:space:]]Conceptual[[:space:]]Elaboration[[:space:]]For[[:space:]]Figurative[[:space:]]Language[[:space:]]Illustration/94abb958-9799-4f08-9e2c-bc75978e27ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GPT[[:space:]]vs[[:space:]]RETRO_[[:space:]]Exploring[[:space:]]the[[:space:]]Intersection[[:space:]]of[[:space:]]Retrieval[[:space:]]and[[:space:]]Parameter-Efficient[[:space:]]Fine-Tuning/8d56c3c8-1c1a-4adc-9826-ce75b6a116c2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GPT-4[[:space:]]Jailbreaks[[:space:]]Itself[[:space:]]with[[:space:]]Near-Perfect[[:space:]]Success[[:space:]]Using[[:space:]]Self-Explanation/2bd5553d-4e22-4604-96ab-35df9a73eddd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GRASS_[[:space:]]Compute[[:space:]]Efficient[[:space:]]Low-Memory[[:space:]]LLM[[:space:]]Training[[:space:]]with[[:space:]]Structured[[:space:]]Sparse[[:space:]]Gradients/b0f09f03-3494-47b5-ba93-f970534f8f5c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GRIZAL_[[:space:]]Generative[[:space:]]Prior-guided[[:space:]]Zero-Shot[[:space:]]Temporal[[:space:]]Action[[:space:]]Localization/f19d56d3-e31c-41f7-a76a-6e9c0576c865_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Game[[:space:]]on[[:space:]]Tree_[[:space:]]Visual[[:space:]]Hallucination[[:space:]]Mitigation[[:space:]]via[[:space:]]Coarse-to-Fine[[:space:]]View[[:space:]]Tree[[:space:]]and[[:space:]]Game[[:space:]]Theory/c44fa5a4-7478-42fa-b463-4f648fc2c3d4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Generalizing[[:space:]]Clinical[[:space:]]De-identification[[:space:]]Models[[:space:]]by[[:space:]]Privacy-safe[[:space:]]Data[[:space:]]Augmentation[[:space:]]using[[:space:]]GPT-4/959d8b4f-45a7-436d-81bb-a1b4bdda09e3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Generate-on-Graph_[[:space:]]Treat[[:space:]]LLM[[:space:]]as[[:space:]]both[[:space:]]Agent[[:space:]]and[[:space:]]KG[[:space:]]for[[:space:]]Incomplete[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Question[[:space:]]Answering/b1ac20a0-097c-4105-938a-ac198db738ad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Generating[[:space:]]Demonstrations[[:space:]]for[[:space:]]In-Context[[:space:]]Compositional[[:space:]]Generalization[[:space:]]in[[:space:]]Grounded[[:space:]]Language[[:space:]]Learning/dfb780fe-085d-4326-87ee-59a801729592_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Generation[[:space:]]with[[:space:]]Dynamic[[:space:]]Vocabulary/1ecbbca4-6ed5-4114-9921-66c80ed68bd1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Generative[[:space:]]Models[[:space:]]for[[:space:]]Automatic[[:space:]]Medical[[:space:]]Decision[[:space:]]Rule[[:space:]]Extraction[[:space:]]from[[:space:]]Text/d4595d92-e8e9-43bd-9107-dfa1d077f79d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Generative[[:space:]]Subgraph[[:space:]]Retrieval[[:space:]]for[[:space:]]Knowledge[[:space:]]Graph–Grounded[[:space:]]Dialog[[:space:]]Generation/58a5dce7-c58e-4b5e-a47a-076016328e9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GeoGPT4V_[[:space:]]Towards[[:space:]]Geometric[[:space:]]Multi-modal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Geometric[[:space:]]Image[[:space:]]Generation/68f7c382-6c46-4efd-afe3-4447c74011c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Getting[[:space:]]More[[:space:]]from[[:space:]]Less_[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Good[[:space:]]Spontaneous[[:space:]]Multilingual[[:space:]]Learners/9a8de092-9282-4d22-8f04-ca390d7e5882_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Getting[[:space:]]The[[:space:]]Most[[:space:]]Out[[:space:]]of[[:space:]]Your[[:space:]]Training[[:space:]]Data_[[:space:]]Exploring[[:space:]]Unsupervised[[:space:]]Tasks[[:space:]]for[[:space:]]Morphological[[:space:]]Inflection/83c11fda-53fe-46e5-97e7-66838a346b9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Global[[:space:]]Reward[[:space:]]to[[:space:]]Local[[:space:]]Rewards_[[:space:]]Multimodal-Guided[[:space:]]Decomposition[[:space:]]for[[:space:]]Improving[[:space:]]Dialogue[[:space:]]Agents/72b8b266-aa00-4d8d-9547-276c454986ac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GlobeSumm_[[:space:]]A[[:space:]]Challenging[[:space:]]Benchmark[[:space:]]Towards[[:space:]]Unifying[[:space:]]Multi-lingual,[[:space:]]Cross-lingual[[:space:]]and[[:space:]]Multi-document[[:space:]]News[[:space:]]Summarization/d78b250e-32ea-4a58-ac0f-9ac8b33d0111_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GlossLM_[[:space:]]A[[:space:]]Massively[[:space:]]Multilingual[[:space:]]Corpus[[:space:]]and[[:space:]]Pretrained[[:space:]]Model[[:space:]]for[[:space:]]Interlinear[[:space:]]Glossed[[:space:]]Text/abc2efa5-34fe-4dfb-b916-b69bfec1ba8c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Glue[[:space:]]pizza[[:space:]]and[[:space:]]eat[[:space:]]rocks[[:space:]]-[[:space:]]Exploiting[[:space:]]Vulnerabilities[[:space:]]in[[:space:]]Retrieval-Augmented[[:space:]]Generative[[:space:]]Models/000b0382-bdb5-460c-8e24-46dbba2ee9f8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Gold[[:space:]]Panning[[:space:]]in[[:space:]]Vocabulary_[[:space:]]An[[:space:]]Adaptive[[:space:]]Method[[:space:]]for[[:space:]]Vocabulary[[:space:]]Expansion[[:space:]]of[[:space:]]Domain-Specific[[:space:]]LLMs/a2cc71e6-8f10-4dbf-8e26-be2359ad4070_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GoldCoin_[[:space:]]Grounding[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Privacy[[:space:]]Laws[[:space:]]via[[:space:]]Contextual[[:space:]]Integrity[[:space:]]Theory/922de50f-4b77-4382-af5b-4217638b10e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GottBERT_[[:space:]]a[[:space:]]pure[[:space:]]German[[:space:]]Language[[:space:]]Model/0681f00e-8c5d-45c0-a978-ac636ff1bb69_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Granular[[:space:]]Privacy[[:space:]]Control[[:space:]]for[[:space:]]Geolocation[[:space:]]with[[:space:]]Vision[[:space:]]Language[[:space:]]Models/ed97918e-e197-4705-a390-905e4aa9b9eb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Grasping[[:space:]]the[[:space:]]Essentials_[[:space:]]Tailoring[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Relation[[:space:]]Extraction/fbddf759-47eb-4dc6-9109-a37c26aac679_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Grounding[[:space:]]Language[[:space:]]in[[:space:]]Multi-Perspective[[:space:]]Referential[[:space:]]Communication/b2765753-cc2b-4149-a638-47dfdb861d6b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/GuardBench_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Benchmark[[:space:]]for[[:space:]]Guardrail[[:space:]]Models/b0d2775e-8103-4d6d-96cc-730625769a2f_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/Fill In The Gaps_ Model Calibration and Generalization with Synthetic Data/94f56978-61ff-41c3-b3a1-8b022f290de8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08d5c5cc0309edef2cd93ac1d9409a5d5269e7c042b3f34191930b41298105af
+size 1237928

2024/Fill In The Gaps_ Model Calibration and Generalization with Synthetic Data/full.md

@@ -0,0 +1,396 @@
+# Fill In The Gaps: Model Calibration and Generalization with Synthetic Data
+
+Yang Ba $^{1}$ , Michelle V. Mancenido $^{2}$ , and Rong Pan $^{1}$
+
+$^{1}$ School of Computing and Augmented Intelligence, Arizona State University
+
+$^{2}$ School of Mathematical and Natural Sciences, Arizona State University
+
+$^{1}$ yangba@asu.edu, Rong.Pan@asu.edu,  $^{2}$ mmanceni@asu.edu
+
+# Abstract
+
+As machine learning models continue to swiftly advance, calibrating their performance has become a major concern prior to practical and widespread implementation. Most existing calibration methods often negatively impact model accuracy due to the lack of diversity of validation data, resulting in reduced generalizability. To address this, we propose a calibration method that incorporates synthetic data without compromising accuracy. We derive the expected calibration error (ECE) bound using the Probably Approximately Correct (PAC) learning framework. Large language models (LLMs), known for their ability to mimic real data and generate text with mixed class labels, are utilized as a synthetic data generation strategy to lower the ECE bound and improve model accuracy on real test data. Additionally, we propose data generation mechanisms for efficient calibration. Testing our method on four different natural language processing tasks, we observed an average up to  $34\%$  increase in accuracy and  $33\%$  decrease in ECE.
+
+# 1 Introduction
+
+Natural Language Processing (NLP) models have fundamentally advanced the syntactic and semantic analysis, information retrieval, and automated generation of textual data. State-of-the-art (SOTA) models (e.g., transformers (Vaswani et al., 2017), BERT (Devlin et al., 2019), and RoBERTa (Liu et al., 2019)) have excelled in practical, user-centric applications such as automated customer support chatbots, personalized content curation, and real-time multilingual text translation. Other NLP models, which are typically trained for a specialized use context, have also been developed and fine-tuned for numerous downstream tasks, including sentiment analysis, named entity recognition (NER), and text classification, as parts of a decision-support system (DSS). Powered by deep learning algorithms, these classification models
+
+have achieved remarkable levels of performance in terms of their accuracy, F1 scores, and AUCs (Li et al., 2020; Cohan et al., 2019).
+
+As machine learning philosophies continue to evolve, growing attention is placed on metrics beyond simple classification accuracy. In recent years, socially responsible artificial intelligence (AI) has been strongly advocated by algorithmic regulatory frameworks (e.g., the US Algorithmic Accountability Act (Donovan et al., 2018)), especially in safety-critical domains, such as healthcare (Pfohl et al., 2022) and law enforcement (Salvador et al., 2021). Some key pillars of socially responsible AI include accountability, transparency, and robustness (Cooper et al., 2022). Ensuring a calibrated ML model accountable for its decision means that it must provide clear justifications for any decision being made, while transparency requires that these justifications are understandable and interpretable (Kadavath et al., 2022); additionally, robustness requires that the ML model performs consistently well under various conditions. In classification tasks, these requirements can be addressed by properly managing model output uncertainty, i.e., quantifying, calibrating, and communicating the proper confidence level associated with each prediction to the end user. Among the three aspects of uncertainty management, calibration directly improves model performance by ensuring that model predictions are congruent with empirically observed outcomes.
+
+AI risk management is an emerging field that emphasizes understanding the limitations of model predictions. Model calibration techniques are used to address the fact that high accuracy does not always mean high confidence in a model's predictions. For example, consider a classifier trained to recognize handwritten digits. This model might achieve high accuracy on a test set, but it also provides the predicted probability for each class, which reflects its level of uncertainty. If it classi
+
+![](images/203312c013b289d1534d008c2b8e2fb2bb3f7d85c0873184870a8487404500f2.jpg)  
+Figure 1: The framework of our proposed method involves initially training a downstream model using real data to identify poorly calibrated bins. Data instances,  $x_{i}$ , and their prediction probabilities,  $\hat{P}(y_{i}|x_{i})$ , from these bins are then fed into large language models (LLMs) to generate synthetic data,  $x_{i}^{syn}$ , along with their corresponding probabilities,  $\hat{P}(y_{i}|x_{i}^{syn})$ . This synthetic data, combined with the real data, is used to retrain the downstream model, thereby improving calibration outputs without compromising model performance.
+
+fies a digit as a '3' with  $70\%$  probability and as an '8' with  $30\%$  probability, it indicates that while the model predicts '3', it lacks high confidence. Understanding this prediction uncertainty has several key benefits: (1) refining decision-making thresholds to improve overall model performance; (2) adjusting models to perform well under different conditions and data distributions; (3) reducing the black-box nature of machine learning models, fostering greater transparency and trust; and (4) enabling more consistent and reliable decision-making, particularly in risk-sensitive applications where errors can have significant consequences.
+
+Modern deep learning neural networks (NN), however, have been shown to be often miscalibrated i.e., while the NN model performs well in classification, the uncertainty around predictions is also high (Wang et al., 2021; Minderer et al., 2021). NLP models trained on classification tasks (such as sentiment analysis) are built on deep learning algorithms, with many hidden layers and regularization steps, and consequently, numerous hyperparameters to be tuned. Recent work has shown the association between increased depth and/or width of NN layers, and miscalibrated outcomes (Guo et al., 2017). Model calibration becomes worse in data-scarce scenarios, where the fraction of events predicted does not align with actual outcomes because the amount of data available at hand may not be sufficient enough to be representative across different classes. Data augmentation approaches (e.g., the mixup approach (Zhang et al., 2018; Thulasidasan et al., 2019)) and the associated model calibration problems have been discussed in literature (Wen et al., 2021). But, theories for understanding the association between model performance and calibration are still lacking.
+
+This work is motivated by a recently published paper (Sahu et al., 2023), in which LLMs are utilized to generate synthetic data close to the decision boundary to sharpen the discrimination power of the classifier and increase model accuracy. The LLM-generated synthetic data leverage the capability of LLMs in providing both realistic and diverse datasets, which potentially increases the ML model's generalizability on out-of-distribution data. The application of synthetic data has been explored as an augmented training set (Van Breugel et al., 2023), validation set (Shoshan et al., 2023), or test set (van Breugel et al., 2023) to improve ML model performance. However, applying synthetic data to address model calibration has barely been explored. We aim to use LLM-generated synthetic text data to fine-tune downstream binary classification tasks to reduce expected calibration error (ECE) without sacrificing the ML model's accuracy.
+
+Our approach is derived from the Probably Approximately Correct (PAC) learning framework (Valiant, 1984). We prove that reducing both calibration and misclassification errors can be achieved simultaneously, and we establish the necessity of generating synthetic data for enhancing model calibration. This approach is validated on real-world text datasets. The synthetic data generation process is accomplished using open-source Large Language Models - Llama 2 (Touvron et al., 2023). Figure 1 illustrates our proposed framework to improve model calibration and generalization via synthetic data for natural language classification tasks.
+
+The contributions of this work can be summarized as follows:
+
+1. We derive the Expected Calibration Error bound to explore the possibility of achieving
+
+both high accuracy and low ECE.
+
+2. We propose a strategy for fixing calibration errors and filling the gaps in the reliability diagram.  
+3. We demonstrate the effectiveness of purposefully augmenting LLM-generated synthetic data into the training set to achieve ML model prediction uncertainty calibration.
+
+# 2 Calibration Concept
+
+This section introduces some basic concepts related to model calibration, which would lay a foundation to derive our methodology.
+
+Expected Calibration Error (ECE). ECE is a widely-used metric to evaluate how well a model's predicted probabilities (confidence) align with its actual outcomes (accuracy) (Guo et al., 2017). It is calculated by segmenting the full range of predicted probabilities into  $M$  equal bins and sorting predictions into these bins based on their confidence. Within each bin, the model's accuracy (the fraction of correct predictions) and average predicted confidence are computed. Let  $B_{m}$  be the set of examples in the  $m^{th}$  bin, whose accuracy and confidence are:
+
+$$
+\begin{array}{l} \operatorname {A c c} \left(B _ {m}\right) = \frac {1}{\left| B _ {m} \right|} \sum_ {x _ {i} \in B _ {m}} 1 \left(\hat {y} _ {i} = y _ {i}\right), \\ \operatorname {C o n f} \left(B _ {m}\right) = \frac {1}{\left| B _ {m} \right|} \sum_ {x _ {i} \in B _ {m}} \hat {p} _ {i} \tag {1} \\ \end{array}
+$$
+
+where  $1(\hat{y}_i = y_i)$  is an indicator function that is equal to 1 if  $\hat{y}_i = y_i$  and 0 otherwise;  $\hat{p}_i$  is the predicted probability associated with the instance  $x_i$ . The concept of accuracy here is based on each class of labels, which is a subset of the well-accepted model evaluation metric: accuracy. The Expected Calibration Error (ECE) given  $n$  examples is defined by taking a weighted average of the absolute differences between the bin's confidence and its accuracy:
+
+$$
+\mathrm {E C E} = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {A c c} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m}\right) \right| \tag {2}
+$$
+
+There exist some variants of ECE, like MCE (Guo et al., 2017), ACE(Nixon et al., 2019), and other metrics to quantify calibration like brier score (Rufibach, 2010), however, in this paper, the calibration error refers to ECE.
+
+Reliability Diagram. Reliability diagram (see Figure 2a) is a tool to visualize the model calibration.  $Conf(B_m)$  and  $Acc(B_m)$  represent the x-axis and y-axis of the diagram respectively for bin  $B_m$ . The diagonal line denotes perfectly calibrated and any deviations from this diagonal line indicate a model's miscalibration. Therefore, the miscalibration can be divided as above the line (underconfidence:  $Acc(B_m) > Conf(B_m)$ ) and under the line (overconfidence:  $Acc(B_m) < Conf(B_m)$ ) areas. We use a reliability diagram to find out the target bins where synthetic data are needed to fill in.
+
+# 3 Methodology
+
+In this section, we utilize the Probably Approximately Correct (PAC) learning framework to derive the expected calibration error(ECE) bound and discuss the benefits of using synthetic data to improve models' calibration and generalization. Moreover, we use a toy sample to demonstrate our methodology.
+
+# 3.1 From PAC Learning to Expected Calibration Error Bound
+
+Probably Approximately Correct (PAC) learning (Valiant, 1984) offers a theoretical framework that establishes the bounds on learning model parameters with specified levels of error and confidence, relating model accuracy to confidence level and sample size. According to Hoeffding's inequality,
+
+$$
+P (| E (h) - E (h ^ {*}) | > \epsilon) \leq 2 \exp (- 2 \epsilon^ {2} n)
+$$
+
+where  $E(h)$  denotes the true error of the hypothesis  $h$  on unseen data and  $E(h^{*})$  denotes the empirical error of the hypothesis  $h$  on the training data. Let  $\delta^1$  be the confidence level and make  $\delta = 2\exp (-2\epsilon_a^2 n)$ . This inequality presents the maximum allowable difference between the true and empirical errors based on a given sample size  $n$  and desired uncertainty level  $\delta$ . Thus, we get the minimal sample size to make the hypothesis true considering error difference  $\epsilon$  and confidence  $\delta$  is given by  $n = \log (2 / \delta) / (2\epsilon^2)$ .
+
+Now we derive the ECE bound from the PAC learning framework. First, we extend the definition of accuracy and confidence in Equation (1) from
+
+bin-wise to data-wise. Then we have
+
+$$
+\operatorname {A c c} (X) = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \operatorname {A c c} \left(B _ {m}\right),
+$$
+
+$$
+\operatorname {C o n f} (X) = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \operatorname {C o n f} \left(B _ {m}\right)
+$$
+
+The dataset is denoted by  $\{X,y\}_{i}^{n}$ , where  $X$  denotes the feature space,  $X\in \{X_{1},\ldots ,X_{n}\}$  and  $y$  is the label  $y\in \{y_1,\dots ,y_n\}$ . Based on Hoeffding's inequality, we have
+
+$$
+P \left(\left| A c c (X) - A c c \left(X ^ {*}\right) \right| > \epsilon_ {a}\right) \leq 2 \exp \left(- 2 \epsilon_ {a} ^ {2} n\right) \tag {3}
+$$
+
+where  $Acc(X)$  denotes the expected accuracy in the model and  $Acc(X^{*})$  is the observed accuracy of training data.  $\epsilon_{a}$  is the error for accuracy and we let  $\delta_{a} = 2\exp (-2\epsilon_{a}^{2}n)$ .
+
+# Proposition - Expected Calibration Error
+
+Bound. Given  $n$  training samples, if the probability of the difference between the expected model parameter and its estimated value being less than  $\epsilon_{a}$  is at least  $(1 - \delta_{a})\%$ , then the probability of the difference between the expected calibration error and the estimated calibration error in the training samples being less than  $\epsilon_{ECE}$  is at least  $(1 - \delta_{ECE})\%$ . Here,  $\delta_{ECE} = 2\delta_{a}$ , and  $\epsilon_{ECE} = \epsilon_{a} + |Conf(X) - Conf(X^{*})| = \epsilon_{a} + \sum_{m=1}^{M} \frac{|B_{m}|}{n} |Conf(B_{m}) - Conf(B_{m}^{*})|$ .
+
+A shortened proof (detailed proof is provided in Appendix B) is given below:
+
+By deriving from the left side of equation (3), we get:
+
+$$
+\begin{array}{l} P \left(\left| \operatorname {A c c} (X) - \operatorname {A c c} \left(X ^ {*}\right) \right| > \epsilon_ {a}\right) \\ \geq P \left(\operatorname {E C E} (X) - \operatorname {E C E} \left(X ^ {*}\right) > \epsilon_ {a} + \left| \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) \right|\right) \tag {4} \\ \end{array}
+$$
+
+Combined with the right side of equation (3):
+
+$$
+\begin{array}{l} P \left(\operatorname {E C E} (X) - \operatorname {E C E} \left(X ^ {*}\right) > \epsilon_ {a} + \left| \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) \right|\right) \\ \leq 2 \exp (- 2 \epsilon_ {a} ^ {2} n) \\ \Rightarrow P \left(\left| \operatorname {E C E} (X) - \operatorname {E C E} \left(X ^ {*}\right) \right| > \epsilon_ {E C E}\right) \\ \leq 4 \exp (- 2 \left(\epsilon_ {E C E} - \left| \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) \right|\right) ^ {2} n) \tag {5} \\ \end{array}
+$$
+
+where  $\epsilon_{ECE} = \epsilon_a + |\operatorname{Conf}(X) - \operatorname{Conf}(X^*)| = \epsilon_a + \sum_{m=1}^{M} \frac{|B_m|}{n} |\operatorname{Conf}(B_m) - \operatorname{Conf}(B_m^*)|$ . And  $n_{ECE} = \log(4 / \delta_{ECE}) / (2(\epsilon_{ECE} - |\operatorname{Conf}(X) - \operatorname{Conf}(X^*)|)^2)$ .
+
+Hoeffding's inequality holds on Bernoulli random variables and accuracy is computed by counting correct predictions. By introducing the
+
+concept of uncertainty  $\delta$ , we obtain the relationship among error, uncertainty, and sample size for PAC learning. Then we can derive the ECE bound from the same inequality and ECE is a random variable with a value in [0, 1]. Finally, given a sample size, we have the relationship among  $\epsilon_{a}, \epsilon_{ECE}, \delta_{a}, \delta_{ECE}$ .
+
+Remark 1: Since the difference between true prediction probabilities and the estimated prediction probabilities exists, given the same data points to train a model, the error for ECE is larger than the error for accuracy compared with the true metrics and the uncertainty level for ECE is two times that for accuracy.
+
+# Remark 2: Based on  $n_{ECE} = \log (4 / \delta_{ECE})$
+
+$/ (2(\epsilon_{ECE} - |Conf(X) - Conf(X^{*})|)^{2}),$  increasing the amount of training data will result in smaller error  $\epsilon_{ECE}$  and lower uncertainty level  $\delta_{ECE}$ ; similar to the effects on  $\epsilon_{a}$  and  $\delta_{a}$ .
+
+Remark 3: Since  $\epsilon_{ECE} = \epsilon_a + |Conf(X) - Conf(X^*)|$ , reducing  $|Conf(X) - Conf(X^*)|$  aligns the ECE error with the accuracy error. This alignment helps models achieve both good calibration and better generalization.
+
+Remark 1 explains why some neural networks own a good performance but are more likely to be ill-calibrated. Remark 2 indicates that increasing the amount of training data can both improve model generalization and lower expected calibration error. When the training data is insufficient, synthetic data is the natural option to augment the training data size. Remark 3 provides insights into what kind of synthetic data is needed to fix the calibration issue. The newly added synthetic data should reduce the difference between predicted probabilities and the true probabilities. We cannot know the true difference of  $|\mathrm{Conf}(X) - \mathrm{Conf}(X^{*})|$ , but we can decrease it by reducing  $|\mathrm{Conf}(B_m) - \mathrm{Conf}(B_m^*)|$  in bins that display the gaps given that  $\mathrm{Conf}(X) = \sum_{m=1}^{M} \frac{|B_m|}{n} \mathrm{Conf}(B_m)$ , as we know where the perfect calibration line is for each bin. Therefore, we can manipulate the prediction probability of synthetic data to minimize the difference. In other words, synthetic data is applied to fill the gaps against the perfect calibration. That is, we try to decrease ECE by using synthetic data to lower the ECE bound.
+
+![](images/135d77ccb12d21a893fca69a5a700481ae655f6efb471c1f1c1f8954868ed0e5.jpg)  
+(a) Reliability diagram with the target bin
+
+![](images/e0ae559948401a1aad97275e998518543f02805e85010c01e823555fef7f103e.jpg)  
+(b) Move generated synthetic data away from DB  
+Figure 2: Generating synthetic data to address miscalibration gaps. In (a), the target bin for calibration is identified as Low Probability & Overconfidence. Synthetic data is generated away from the decision boundary in (b).
+
+# 3.2 Synthetic Data Generation Strategy
+
+Synthetic data generation consists of two stages: First, we specify the gaps against the perfect calibration line in the reliability diagram. Bins over the line are underconfident while those under the line are overconfident. The data points in those bins are the target data samples for synthetic data generation. With the predicted probability 0.5 as the cutoff, we categorize the reliability diagram (Figure 2a) into four scenarios: Low Probability & Over Confidence, Low Probability & Under Confidence, High Probability & Over Confidence, and High Probability & Under Confidence, as shown in Table 1.
+
+Next, LLMs, which serve as text generators, are used to create synthetic text data. Since LLMs are trained on diverse and extensive data spanning a wide range of sources, we can distill the knowledge from LLMs to generate synthetic data that is considered out-of-distribution of training data. We ask LLMs to imitate the classifier we trained by generating similar instances using data samples we collected from the target bin. Specifically, we pass the data instance  $x_{i}$  and  $\hat{P}(y_{i}|x_{i})$  from a trained classifier to LLMs and ask it to generate a similar instance  $x_{i}^{\text{syn}}$  with  $\hat{P}(y_{i}|x_{i}^{\text{syn}})$ , where  $|\hat{P}(y_{i}|x_{i}) - \hat{P}(y_{i}|x_{i}^{\text{syn}})| = |\operatorname{Conf}(B_{m}) - \operatorname{Conf}(B_{m}^{*})|$ . For example, suppose there are  $n_{\text{bins}}$  bins, if the target text is from  $m^{\text{th}}$  bin and  $|\operatorname{Conf}(B_{m}) - \operatorname{Conf}(B_{m}^{*})|$  is  $\alpha$ , then we will ask LLMs to generate the synthetic texts that have the  $\frac{m}{n_{\text{bins}}} \pm \alpha$  probability belonging to one class and the  $1 - \left(\frac{m}{n_{\text{bins}}} \pm \alpha\right)$  probability for the other class.
+
+To illustrate our method, in Figure 2b a hidden predicted probability line is shown orthogonal to the estimated decision boundary. Data points close
+
+<table><tr><td></td><td>Over Confidence</td><td>Under Confidence</td></tr><tr><td>Low Probability (P(yi|x̂i) ≤ 0.5)</td><td>Decrease predicted prob (Move away from DB)</td><td>Increase predicted prob (Move towards DB)</td></tr><tr><td>High Probability (P(yi|x̂i) &gt; 0.5)</td><td>Increase predicted prob (Move towards DB)</td><td>Increase predicted prob (Move away from DB)</td></tr></table>
+
+Table 1: Synthetic Data Generation Strategy (DB: Decision Boundary). Refer to Appendix A for the prompts used for data generation across different scenarios.
+
+to this decision boundary would be predicted with around 0.5 probability (the softmax output), while data points at the two ends of this line could be predicted with close to 0.1 or 0.9 probability, respectively. Suppose that our targeted bin has a confidence of 0.3 and it is over-confident as shown in Figure 2a (highlighted in a purple rectangular). The gap between the empirical and theoretical uncertainty values is shown in red. There are two possible solutions to fill the gap: 1) increasing the number of incorrect predictions, thus raising the blue bar that represents the empirical inaccurate prediction percentage; or 2) moving this bin to the left, into the bin with a smaller uncertainty value. We use the second solution to align the miscalibration bins because the first solution could harm the accuracy of the classifier.
+
+In Figure 2b, the target data points are circled in black and the synthetic data are in the blue circle, which are generated based on the generation strategy in Table 1. Since the synthetic data shares a similar feature space and the same labels as the target data samples, the retrained classifier would predict them as the same class but with smaller probabilities. This makes it more likely for the synthetic data to be assigned to the same bin as the original target data samples. In this way, we push
+
+![](images/cdb9457988f9d395abc44bc9c6914e4d79a7c41373d9a12f8044bc4bcfb40abf.jpg)  
+(a)
+
+![](images/01bd6927f0a152c232b54a08647799cf14c2f6b15d0e878688778bddd1068bd4.jpg)  
+(b)  
+Figure 3: The iterative process of enhancing the accuracy and calibration of a 1D logistic regression model is demonstrated. Initially, the model is fitted using observed data (a), followed by the creation of its reliability diagram to identify poorly calibrated bins (b). Next, synthetic data points are strategically added to two targeted bins and the model is refitted. This iterative approach results in the model closely approximating the true logistic curve (c), thereby improving the calibration in the reliability diagram (d).
+
+![](images/421529113a3535189e4f889987d554ab7c9dca6de027db3e9340ab67c7655de3.jpg)  
+(c)
+
+![](images/b5073e50fbe879a81ca4201bfb39e34d44ac33937112911edb848dc66beea642.jpg)  
+(d)
+
+the predicted probability of this bin away from the decision boundary and reduce the difference  $|\operatorname{Conf}(B_m) - \operatorname{Conf}(B_m^*)|$ .
+
+Utilizing LLM, we employ a two-stage approach to ensure both the fidelity and diversity of the synthetic data generated. We obtain the instance  $x_{i}^{syn}$  with probability  $\hat{P}(y_{i}^{syn}|x_{i}^{syn})$  in the first stage and relabel it via LLMs to ensure it belongs to the same class of  $x_{i}$  in the second stage. Since step 1 mixes up information from two labels to some degree, it enhances data diversity. The relabeling of the second stage confirms that the generated texts belong to the correct label, which guarantees its fidelity.
+
+# 3.3 Toy Example
+
+We use a 1D logistic regression classifier as an example to demonstrate that adding appropriate synthetic data in the target bins can produce a better-calibrated and more accurate model. Parameters of the true model are defined:  $\beta_0 = -1$  and  $\beta_{1} = 2$ . We randomly simulate 300 data points from the range between -10 and 10 and classify them based on the true model as the label. A logistic regression model is fitted on these data points. The fitted parameters are  $\beta_0 = -0.06$  and  $\beta_{1} = 1.13$ . The model achieves an accuracy of 0.95 and an ECE of 0.0405 (Figure 3a).
+
+Figure 3b shows us that the fitted logistic regression is overconfident about its predictions in the  $2^{nd}$  bin and  $4^{th}$  bin. Now we target these two bins to generate some synthetic data points to fill the gap. The function we used to generate synthetic data points is a left-sided truncated normal distribution, whose parameters are:  $\mu = \mu_{Bin_i}$ ,  $SD = SD_{Bin_i}$ ,  $\mathfrak{n} = |Bin_i|$ ,  $i = \{2,4\}$ . We add new data points step by step to see how the logistic curve changes: 1) add synthetic data of the  $2^{nd}$  bin, 2) then add synthetic data of the  $4^{th}$  bin based on pre
+
+viously added data points. See the parameters and performance for newly fitted models below:
+
+- original data (original fitted model):  $\beta_0 = -0.06$  and  $\beta_{1} = 1.13$ , ACC: 0.95, ECE: 0.0405;  
+- synthetic data in  $2^{nd}$  bin (newly fitted model):  $\beta_0 = -0.339$  and  $\beta_{1} = 1.2627$ , ACC: 0.95327, ECE: 0.0424;  
+- synthetic data in  $2^{nd}$  bin +  $4^{th}$  bin (newly fitted model 2):  $\beta_0 = -0.2558$  and  $\beta_{1} = 1.2953$ , ACC: 0.9469, ECE: 0.0366.
+
+Figure 3c and Figure 3d illustrate that incorporating synthetic data generated from overconfident bins shifts the fitted logistic curve towards the standard logistic curve, resulting in a more accurately calibrated model.  
+
+<table><tr><td>Dataset</td><td>Classes</td><td>Balanced</td><td>Class Prop.</td><td>#Train/#Test</td></tr><tr><td>TC</td><td>2</td><td>No</td><td>65:35</td><td>3104/345</td></tr><tr><td>SUBJ</td><td>2</td><td>Yes</td><td>50:50</td><td>8000/2000</td></tr><tr><td>B77</td><td>2</td><td>Yes</td><td>44:56</td><td>177/80</td></tr><tr><td>SE</td><td>2</td><td>No</td><td>35:65</td><td>3012/750</td></tr><tr><td>Arxiv</td><td>2</td><td>Yes</td><td>50:50</td><td>4800/1200</td></tr><tr><td>Medical</td><td>2</td><td>Yes</td><td>50:50</td><td>2662/1060</td></tr></table>
+
+Table 2: Dataset Statistics
+
+# 4 Experiment
+
+# 4.1 Dataset
+
+We employ four datasets on text classification tasks across multiple domains with varying sample sizes and proportions of class. To better evaluate our approach, we select two balanced and two imbalanced datasets, respectively, and the sample size varies from hundreds to thousands. The Complaints dataset (TC) (Preoticiuc-Pietro et al.,
+
+<table><tr><td>Dataset</td><td colspan="2">Input</td><td>Prompt</td><td>Generated Text (xi syn, P(yi | xi syn))</td><td>Label (Y)</td></tr><tr><td rowspan="3">SE</td><td>xi</td><td>The zoom function on this camera is so loud that sometimes you will be unable to use it if you find yourself in a situation where you must be quiet.</td><td rowspan="3">An example xi which belongs 75% to negative and 25% to positive (based on a classifier&#x27;s categorization). Now I ask you to act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs 55% to negative and 45% to positive.</td><td rowspan="3">I love how this router can handle a large network, but the price is a bit steep for my taste. (55% negative, 45% positive)</td><td rowspan="3">negative</td></tr><tr><td>P(yi | xi)</td><td>0.75</td></tr><tr><td>miscalibration type</td><td>overconfidence</td></tr></table>
+
+Table 3: An example of generating synthetic data via LLM. As an example, we use the SE dataset; for information on other datasets, see Table 8 in Appendix D. Input contains the original text  $(x_{i})$  and the average predictive probability of the bin it comes from  $(\hat{P}(y_{i}|x_{i}))$ . Generated Text is the one after the relabeling process. Note: during re-fine tuning of the downstream model, we exclude  $\hat{P}(y_{i}|x_{i}^{syn}) - (55\%$  negative,  $45\%$  positive) and retain only  $x_{i}^{syn}$  in the dataset.
+
+2019) contains 3K tweets regarding product reviews, which are categorized as complaints or not-a-complaints. The SUBJECTivity dataset (SUBJ) (Pang and Lee, 2004) is a benchmark dataset that contains 10K objective/subjective movie reviews. Banking77 (B77) (Casanueva et al., 2020) is a dataset comprising fine-grained intents within the banking domain featuring multiple classes. For our study, we select instances from two of these classes, which makes the size of the dataset relatively small. SentEval (SE) (Hu and Liu, 2004) contains 3K data used for sentiment analysis tasks.
+
+Additionally, to discover whether the pre-trained knowledge of LLMs is a crucial element in determining the performance of generated synthetic data on downstream tasks. We choose two newly-released datasets that are unlikely to be a part of the training data of LLMs we used in the paper (Llama-2): Arxiv-10 (Farhangi et al., 2022) and Medical (Fansi Tchango et al., 2022). We pick 'cs' and 'stat' classes, and randomly sample  $30\%$  from the entire data in Arxiv-10. Its task is to classify the subject based on the title of a paper. Medical is the dataset in a medical diagnosis domain categorizing a specific disease based on a patient's symptoms, where "Influenza" and "Anaphylaxis" in our experiment. See detailed statistics of these datasets in Table 2.
+
+# 4.2 Training
+
+All datasets are split into training, validation, and test sets. For the TC dataset, we split the entire data into training, validation, and test sets with a ratio of 80:10:10. For other datasets, the validation sets are created by randomly sampling  $20\%$  from
+
+the training set. All experiments are evaluated on their original test sets. We fine-tune  $\mathrm{BERT}_{\mathrm{base}}$  (Devlin et al., 2019) models for text classification by adding a dropout and softmax layer following the pre-trained structure. We train each model 5 epochs and apply a 1e-6 learning rate and  $50\%$  dropout.
+
+Step 1. After completing the training process, we calculate the reliability diagram and the difference  $(D)$  between the proportion of positive labels and the mean predicted values for each bin based on the validation set. If the absolute value of  $D_{m}$  in  $m^{th}$  bin is larger than the threshold 0.03, the data in the  $m^{th}$  bin will be selected to generate synthetic data.
+
+Step 2. We use LLMs to generate synthetic data based on the texts in the target bin from Step 1. To explore the effect of the number of bins (M), we select three scenarios by setting  $M = 10, 15, 20$ .
+
+# 4.3 Synthetic Data Generation
+
+Synthetic text generation is performed using version Llama-2-7b-chat-hf of Llama 2 at a temperature  $T = 0.1$ . We apply the two-stage and three-shot learning generation method proposed in the paper (Sahu et al., 2023) to guarantee diversity and authenticity. First, we define each label and provide three examples for each one (Appendix C). Then, we present the example text from the previous selection stage along with the predicted probability of this example that was extracted from the trained BERT<sub>base</sub> classifier. We then instruct llama 2 to act as the base classifier to generate three similar texts that could be classified with specific probability requirements. Next, we instruct llama 2
+
+<table><tr><td rowspan="2">Metric</td><td colspan="2">TC</td><td colspan="2">SUBJ</td><td colspan="2">B77</td><td colspan="2">SE</td><td colspan="2">Arxiv</td><td colspan="2">Medical</td></tr><tr><td>ACC</td><td>ECE</td><td>ACC</td><td>ECE</td><td>ACC</td><td>ECE</td><td>ACC</td><td>ECE</td><td>ACC</td><td>ECE</td><td>ACC</td><td>ECE</td></tr><tr><td>Baseline</td><td>0.867(0.00)</td><td>0.058(0.02)</td><td>0.955(0.01)</td><td>0.034(0.01)</td><td>0.708(0.12)</td><td>0.234(0.04)</td><td>0.884(0.01)</td><td>0.06(0.00)</td><td>0.805(0.00)</td><td>0.105(0.01)</td><td>0.864(0.00)</td><td>0.051(0.01)</td></tr><tr><td>Isotonic</td><td>0.871(0.00)</td><td>0.082(0.01)</td><td>0.959(0.00)</td><td>0.027(0.01)</td><td>0.850(0.02)</td><td>0.063(0.01)</td><td>0.890(0.01)</td><td>0.058(0.01)</td><td>0.812(0.01)</td><td>0.114(0.01)</td><td>0.869(0.01)</td><td>0.069(0.01)</td></tr><tr><td>Platt scaling</td><td>0.863(0.01)</td><td>0.086(0.01)</td><td>0.955(0.01)</td><td>0.029(0.00)</td><td>0.846(0.03)</td><td>0.207(0.03)</td><td>0.888(0.01)</td><td>0.068(0.00)</td><td>0.807(0.01)</td><td>0.122(0.00)</td><td>0.869(0.01)</td><td>0.065(0.01)</td></tr><tr><td>MC dropout</td><td>0.868(0.02)</td><td>0.054(0.01)</td><td>0.952(0.01)</td><td>0.032(0.01)</td><td>0.821(0.23)</td><td>0.274(0.14)</td><td>0.876(0.01)</td><td>0.050(0.02)</td><td>0.799(0.01)</td><td>0.058(0.04)</td><td>0.871(0.01)</td><td>0.070(0.01)</td></tr><tr><td>Temp scaling</td><td>0.867(0.01)</td><td>0.049(0.01)</td><td>0.955(0.01)</td><td>0.026(0.01)</td><td>0.708(0.12)</td><td>0.253(0.17)</td><td>0.884(0.01)</td><td>0.038(0.00)</td><td>0.805(0.00)</td><td>0.070(0.01)</td><td>0.864(0.00)</td><td>0.056(0.01)</td></tr><tr><td>10 bins</td><td colspan="12"></td></tr><tr><td>Synthesis</td><td>0.867(0.01)</td><td>0.053(0.01)</td><td>0.960(0.01)</td><td>0.027(0.01)</td><td>0.625(0.07)</td><td>0.255(0.10)</td><td>0.871(0.00)</td><td>0.055(0.02)</td><td>0.815(0.01)</td><td>0.077(0.03)</td><td>0.873(0.01)</td><td>0.048(0.01)</td></tr><tr><td>Synthesis+</td><td>0.886(0.01)</td><td>0.046(0.01)</td><td>0.961(0.00)</td><td>0.03(0.00)</td><td>0.792(0.20)</td><td>0.231(0.03)</td><td>0.889(0.01)</td><td>0.064(0.00)</td><td>0.808(0.01)</td><td>0.099(0.01)</td><td>0.871(0.00)</td><td>0.047(0.01)</td></tr><tr><td>15 bins</td><td colspan="12"></td></tr><tr><td>Synthesis</td><td>0.879(0.01)</td><td>0.049(0.01)</td><td>0.961(0.00)</td><td>0.026(0.00)</td><td>0.800(0.11)</td><td>0.224(0.08)</td><td>0.904(0.00)</td><td>0.04(0.00)</td><td>0.802(0.00)</td><td>0.096(0.01)</td><td>0.875(0.00)</td><td>0.052(0.00)</td></tr><tr><td>Synthesis+</td><td>0.881(0.01)</td><td>0.050(0.01)</td><td>0.9605(0.00)</td><td>0.024(0.00)</td><td>0.863(0.09)</td><td>0.203(0.10)</td><td>0.901(0.01)</td><td>0.055(0.01)</td><td>0.824(0.01)</td><td>0.087(0.01)</td><td>0.879(0.01)</td><td>0.055(0.01)</td></tr><tr><td>20 bins</td><td colspan="12"></td></tr><tr><td>Synthesis</td><td>0.883(0.00)</td><td>0.046(0.01)</td><td>0.959(0.00)</td><td>0.027(0.00)</td><td>0.808(0.12)</td><td>0.180(0.07)</td><td>0.900(0.00)</td><td>0.048(0.00)</td><td>0.818(0.01)</td><td>0.089(0.01)</td><td>0.871(0.01)</td><td>0.054(0.00)</td></tr><tr><td>Synthesis+</td><td>0.890(0.00)</td><td>0.046(0.01)</td><td>0.959(0.00)</td><td>0.026(0.00)</td><td>0.950(0.04)</td><td>0.224(0.03)</td><td>0.896(0.01)</td><td>0.049(0.01)</td><td>0.820(0.00)</td><td>0.075(0.00)</td><td>0.867(0.01)</td><td>0.046(0.01)</td></tr></table>
+
+Table 4: Model Performance and Calibration on Real Test Data. Highlighted values considered both ACC and ECE and weigh more on ECE.
+
+to relabel the generated texts to ensure them belong to the "correct" class. Table 3 illustrates the inputs, prompts, and outputs for generating synthetic data using the SE dataset. Additional prompts for different scenarios can be found in Appendix D.
+
+# 4.4 Evaluation
+
+Results are assessed on real test set. Baseline refers to the results trained on the model in Step 1. Suppose we have a total of  $N$  original data points in the training and validation set, and there are  $S_{1}$  data points in target bins from the validation set. Let LLMs generate  $S_{2}$  synthetic data points, and  $S_{2} = S_{1}$ . Synthesis refers to the results that we retrain the model by replacing  $S_{1}$  original data points with  $S_{2}$  synthetic data points. Synthesis+ indicates that we add  $S_{2}$  synthetic samples into the original  $N$  data points.
+
+In addition to the baseline, we also compare the performance of our methods against some widely used model calibration techniques. Isotonic regression (Zadrozny and Elkan, 2001) employs a non-parametric method that adjusts predicted probabilities to align with observed outcomes, and Platt scaling (Platt et al., 1999) fits a logistic regression model to calibrate classifier scores based on predicted probabilities. Monte Carlo dropout (Gal and Ghahramani, 2016) randomly masks nodes to
+
+estimate the probability distribution. In our paper, the model makes 10 predictions for each instance and each time with a different dropout mask. Temperature scaling (Guo et al., 2017) works by dividing the pre-softmax output by a temperature  $T$  and the optimal value of  $T$  is estimated by the validation dataset. All experiments use the same BERT<sub>base</sub> model parameters.
+
+# 4.5 Results
+
+We run each experiment for three random seeds and report the average value (with standard deviation in brackets) of accuracy and ECE in Table 4. By adding synthetic data with a size of  $7\% -18\%$  2 of the training set, we would have a  $21 - 33\%$  ECE decrease. Taking both accuracy and ECE into account, our synthetic data replacement (synthesis) and synthetic data add-on (synthesis+) methods outperform other calibration approaches in five out of six datasets. Temperature scaling can sometimes achieve lower ECE, but a key disadvantage is that it doesn't affect accuracy. On the other hand, while dropout can improve model calibration, it carries the risk of reducing accuracy, as seen in the results on the Arxiv-10 dataset. We also observe a
+
+clear association of a larger bin number with lower  $\mathrm{ECE}^3$ . In addition, the results of our approach have smaller variances compared with those of the baseline.
+
+Whether imbalanced datasets (TC and SE) or balanced datasets (SUBJ, B77, Arxiv, and Medical), improvements in uncertainty calibration are fairly comparable on average. It is also shown that even though the baseline model for the SUBJ dataset already has outstanding accuracy, our approach can still make the model better calibrated without degrading the model's classification performance. Results from B77 have a larger variance due to its smaller data size.
+
+# 4.6 Ablation Study
+
+To discuss if the LLM's self-calibration capability strongly impacts our approach, we instruct Llama2-7b-chat-hf with few-shot learning and set  $top_{k} = 1$  in which we obtain the conditional probability of one class  $P(\text{label} \mid \text{text})$ . Then we compute the accuracy and ECE from LLMs.
+
+<table><tr><td></td><td>LLMACC</td><td>LLMECE</td><td>SynACC(%)</td><td>SynECE(%)</td></tr><tr><td>LLMACC</td><td>1</td><td>-0.737</td><td>0.592</td><td>-0.566</td></tr><tr><td>LLMECE</td><td>-0.737</td><td>1</td><td>-0.026</td><td>0.423</td></tr></table>
+
+Table 5: Pearson Correlation Coefficient on six datasets.  $\mathrm{Syn}_{ACC}(\%)$  and  $\mathrm{Syn}_{ECE}(\%)$  denotes the percentage of the downstream model's accuracy increases and how much percentage of the expected calibration error is decreased respectively.
+
+In Table 5, we didn't observe a strong negative correlation coefficient between  $\mathrm{LLM}_{ACC}$  and  $\mathrm{Syn}_{ECE}(\%)$ , indicating there is no empirical evidence that shows the calibration ability of LLMs determines the application of our proposed methods. Additionally, we found a moderate positive association between the llama's accuracy and the accuracy improvement in downstream tasks. This suggests that the prediction accuracy of LLMs, rather than calibration capability, plays a more important role in downstream models' performance. Therefore, using advanced LLMs (such as Llama 3.2) or fine-tuning LLMs to incorporate domain knowledge could yield better performance when applying our approach.
+
+# 5 Related Work
+
+Model calibration has emerged as an open challenge in machine learning as concerns arise regarding the responsible and ethical use of ML-enabled systems. Several methods have been proposed, including Platt Scaling (Platt et al., 1999), Isotonic regression (Zadrozny and Elkan, 2001), among others. Both of them somewhat change the predicted probability, which could lower the predicted accuracy. In the computer vision field, a mixup method (Zhang et al., 2018) has been proposed to overcome the shortcomings of data scarcity. It combines two instances from the original dataset with different proportions. A follow-up paper (Wen et al., 2021) investigates the computer vision task calibration by using the mixup approach and concludes that this approach could impair the model calibration. However, the calibration issue in the NLP field has rarely been discussed.
+
+Several benefits from using synthetic data have been explored in (Sahu et al., 2023). It is found that ML prediction accuracy can be improved significantly by adding synthetic data generated near the decision boundary. On the other hand, a recent paper (Li et al., 2023) investigates on potential and limitations of synthetic data generated from LLM for text classification tasks and concludes that while synthetic data can be beneficial in certain scenarios, it does not consistently enhance model performance. Our research is different from theirs in that we provide a strategy that enables both good generalization and uncertainty calibration.
+
+# 6 Conclusion
+
+In the era of large models, we believe smaller models still hold tremendous values in, e.g., edge computing and specialized downstream machine learning tasks. We derive the expected calibration error bound for ML models and explore the possibility of leveraging synthetic data to mitigate calibration error. Through empirical validation with text classification tasks, we demonstrate the usefulness of our method; that is, by harnessing the power of LLMs, purposefully generated synthetic data can be utilized to train smaller downstream NLP tasks, achieving both strong classification performance and calibration error reduction.
+
+# Limitations and Future Work
+
+While increasing the sample size generally helps reduce the Expected Calibration Error (ECE), sim
+
+ply adding more synthetic data may not always lead to optimal model performance, as excessive synthetic data can cause overfitting. Therefore, the focus should be on generating high-quality data and strategically identifying instances that require better calibration. It's also important to note that the primary objective of this paper is not to compare or evaluate the capabilities of Large Language Models (LLMs). Rather, we assume that an updated, optimally performing LLM could generate higher-quality synthetic data, which could, in turn, enhance the accuracy of downstream tasks and improve model calibration using our proposed methodology.
+
+In our experiments, we applied a 0.03 threshold to filter out ill-calibrated bins, leaving room for future work to investigate how varying cutoff values might influence calibration enhancement. While our method focuses on text classification applications, there is potential to extend this approach to other downstream NLP tasks. Additionally, future research could explore the use of generative models beyond Large Language Models (LLMs), broadening the scope of applicability. Finally, extending our method to multi-class classification models is proposed as an area for future work.
+
+# References
+
+Inigo Casanueva, Tadas Temčinas, Daniela Gerz, Matthew Henderson, and Ivan Vulić. 2020. Efficient intent detection with dual sentence encoders. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 38-45, Online. Association for Computational Linguistics.  
+Arman Cohan, Iz Beltagy, Daniel King, Bhavana Dalvi, and Dan Weld. 2019. Pretrained language models for sequential sentence classification. 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 3693-3699, Hong Kong, China. Association for Computational Linguistics.  
+A Feder Cooper, Emanuel Moss, Benjamin Laufer, and Helen Nissenbaum. 2022. Accountability in an algorithmic society: relationality, responsibility, and robustness in machine learning. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 864-876.  
+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.  
+Joan Donovan, Robyn Caplan, Jeanna Matthews, and Lauren Hanson. 2018. Algorithmic accountability: A primer.  
+Arsene Fansi Tchango, Rishab Goel, Zhi Wen, Julien Martel, and Joumana Ghosn. 2022. Ddxplus: A new dataset for automatic medical diagnosis. Advances in neural information processing systems, 35:31306-31318.  
+Ashkan Farhangi, Ning Sui, Nan Hua, Haiyan Bai, Arthur Huang, and Zhishan Guo. 2022. Protoformer: Embedding prototypes for transformers. In Advances in Knowledge Discovery and Data Mining: 26th Pacific-Asia Conference, PAKDD 2022, Chengdu, China, May 16–19, 2022, Proceedings, Part I, pages 447–458.  
+Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR.  
+Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International Conference on Machine Learning, pages 1321-1330. PMLR.  
+Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 168-177.  
+Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221.  
+Yuliang Li, Jinfeng Li, Yoshihiko Suhara, AnHai Doan, and Wang-Chiew Tan. 2020. Deep entity matching with pre-trained language models. arXiv preprint arXiv:2004.00584.  
+Zhuoyan Li, Hangxiao Zhu, Zhuoran Lu, and Ming Yin. 2023. Synthetic data generation with large language models for text classification: Potential and limitations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10443-10461, Singapore. 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. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
+
+Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. 2021. Revisiting the calibration of modern neural networks. Advances in Neural Information Processing Systems, 34:15682-15694.  
+Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. 2019. Measuring calibration in deep learning. In CVPR workshops, volume 2.  
+Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL-04), pages 271-278, Barcelona, Spain.  
+Stephen Pfohl, Yizhe Xu, Agata Forcyiarz, Nikolaos Ignatiadis, Julian Genkins, and Nigam Shah. 2022. Net benefit, calibration, threshold selection, and training objectives for algorithmic fairness in healthcare. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1039-1052.  
+John Platt et al. 1999. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large-Margin Classifiers, 10(3):61-74.  
+Daniel Preotciuc-Pietro, Mihaela Gaman, and Nikolaos Aletras. 2019. Automatically identifying complaints in social media. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5008-5019, Florence, Italy. Association for Computational Linguistics.  
+Kaspar Rufibach. 2010. Use of brier score to assess binary predictions. Journal of Clinical Epidemiology, 63(8):938-939.  
+Gaurav Sahu, Olga Vechtomova, Dzmitry Bahdanau, and Issam Laradji. 2023. PromptMix: A class boundary augmentation method for large language model distillation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5316-5327, Singapore. Association for Computational Linguistics.  
+Tiago Salvador, Stephanie Cairns, Vikram Voleti, Noah Marshall, and Adam Oberman. 2021. Faircal: Fairness calibration for face verification. arXiv preprint arXiv:2106.03761.  
+Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, Matan Fintz, and Gerard Medioni. 2023. Synthetic data for model selection. In International Conference on Machine Learning, pages 31633-31656. PMLR.  
+Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. 2019. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 32.
+
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Leslie G Valiant. 1984. A theory of the learnable. Communications of the ACM, 27(11):1134-1142.  
+Boris Van Breugel, Zhaozhi Qian, and Mihaela Van Der Schaar. 2023. Synthetic data, real errors: how (not) to publish and use synthetic data. In International Conference on Machine Learning, pages 34793-34808. PMLR.  
+Boris van Breugel, Nabeel Seedat, Fergus Imrie, and Mihaela van der Schaar. 2023. Can you rely on your model evaluation? Improving model evaluation with synthetic test data. In *Thirty-seventh Conference on Neural Information Processing Systems*.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems, 30.  
+Deng-Bao Wang, Lei Feng, and Min-Ling Zhang. 2021. Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. Advances in Neural Information Processing Systems, 34:11809-11820.  
+Yeming Wen, Ghassen Jerfel, Rafael Muller, Michael W Dusenberry, Jasper Snoek, Balaji Lakshminarayanan, and Dustin Tran. 2021. Combining ensembles and data augmentation can harm your calibration. In International Conference on Learning Representations.  
+Bianca Zadrozny and Charles Elkan. 2001. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML '01: Proceedings of the Eighteenth International Conference on Machine Learning, pages 609-616. ACM.  
+Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. 2018. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations.
+
+# A Appendix
+
+Our code is implemented based on Pytorch 2.2.1 and the pre-trained Bertbase model is downloaded from the huggingface library. Both llama2 and Bertbase run on Nvidia P100 GPUs.
+
+<table><tr><td></td><td>Parameters</td></tr><tr><td>optimizer</td><td>Adam</td></tr><tr><td>max length</td><td>512</td></tr><tr><td>embedding dim</td><td>768</td></tr><tr><td>batch size</td><td>32</td></tr><tr><td>learning rate</td><td>1e-6</td></tr><tr><td>dropout</td><td>0.5</td></tr><tr><td>epoch</td><td>5</td></tr></table>
+
+Table 6: Model Parameters (Bertbase)
+
+We provide the code used to generate the corresponding prompts based on the scenarios to which the bins in Table 1 belong.
+
+```python
+def gen_prompt(conf, diff, label0, label1, indicator):
+    '''diff: the gap against the perfect calibration line.
+    indicator: based on table 1, which the bin belongs to.
+    revised_conf: set a bound if the gap is too large
+    that makes the generated instances assigned correct labels.''
+    if 'low' in indicator and 'under' in indicator:
+        revised_conf = 45 if conf + diff >= 50 else conf + diff
+        generation_prompt = f"which belongs {100-conf} % to {label0} and {conf} % to {label1}
+        (based on a classifier's categorization).
+    Now I ask you to act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs {100-revised_conf} % to {label0} and {revised_conf} % to {label1}.
+    if 'low' in indicator and 'over' in indicator:
+        revised_conf = 5 if conf - diff <= 0 else conf - diff
+        generation_prompt = f"which belongs {100-conf} % to {label0} and {conf} % to {label1}
+        (based on a classifier's categorization).
+    Now I ask you to act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs {100-revised_conf} % to {label0} and {revised_conf} % to {label1}
+```
+
+(no explanation):"
+
+```txt
+if 'high' in indicator and 'under' in indicator: revised_conf  $= 95$  if conf  $^+$  diff  $\geq 100$  else conf  $^+$  diff generation_prompt  $=$  f"which belongs {100-conf}\% to{label0}and{conf}\% to{ label1} (based on a classifier's categorization). Now I ask you to act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs {100-revised_conf}\% to {label0} and{revised_conf}\% to{label1} (no explanation):
+```
+
+```txt
+if 'high' in indicator and 'over' in indicator: revised_conf  $= 55$  if conf - diff <= 50 else conf - diff generation_prompt = f"which belongs {100-conf}\% to {label0} and {conf}\% to { label1} (based on a classifier's categorization). Now I ask you to act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs {100-revised_conf}\% to {label0} and {revised_conf}\% to {label1} (no explanation):
+```
+
+return generation_prompt
+
+# B Appendix
+
+# The Expected Calibration Bound Proof:
+
+From equation (1) in section 2, we extend the definition of accuracy and confidence from bin-wise to data-wise:
+
+$$
+\operatorname {A c c} (X) = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \operatorname {A c c} \left(B _ {m}\right), \quad \operatorname {C o n f} (X) = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \operatorname {C o n f} \left(B _ {m}\right)
+$$
+
+correspondingly,
+
+$$
+\operatorname {A c c} \left(X ^ {*}\right) = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} ^ {*} \right|}{n} \operatorname {A c c} \left(B _ {m} ^ {*}\right), \quad \operatorname {C o n f} \left(X ^ {*}\right) = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} ^ {*} \right|}{n} \operatorname {C o n f} \left(B _ {m} ^ {*}\right)
+$$
+
+According to Hoeffding's inequality, we have:
+
+$$
+P \left(\left| A c c (X) - A c c \left(X ^ {*}\right) \right| > \epsilon_ {a}\right) \leq 2 \exp \left(- 2 \epsilon_ {a} ^ {2} n\right).
+$$
+
+where,  $Acc(X)$  means the expected accuracy in the model;  $Acc(X^{*})$  is the observed accuracy of training data.  $\epsilon_{a}$  is the error for accuracy and we let  $\delta_{a} = 2\exp (-2\epsilon_{a}^{2}n)$ . The derivation from the left side of the inequality:
+
+$$
+\begin{array}{l} P \left(\left| \operatorname {A c c} (X) - \operatorname {A c c} \left(X ^ {*}\right) \right| > \epsilon_ {a}\right) \\ = P \left(\left| \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \operatorname {A c c} \left(B _ {m}\right) - \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \operatorname {A c c} \left(B _ {m} ^ {*}\right) \right| > \epsilon_ {a}\right) \\ = P \left(\sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {A c c} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m}\right) + \operatorname {C o n f} \left(B _ {m}\right) + \operatorname {C o n f} \left(B _ {m} ^ {*}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) - \operatorname {A c c} \left(B _ {m} ^ {*}\right) \right| > \epsilon_ {a}\right) \\ = P \left(\sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {A c c} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m}\right) - \left[ \operatorname {A c c} \left(B _ {m} ^ {*}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) \right] + \operatorname {C o n f} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) \right| > \epsilon_ {a}\right) \\ \geq P \left(\sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {A c c} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m}\right) - \left[ \operatorname {A c c} \left(B _ {m} ^ {*}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) \right] \right| - \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {C o n f} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) \right| > \epsilon_ {a}\right) \\ = P \left(\sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {A c c} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m}\right) - \left[ \operatorname {A c c} \left(B _ {m} ^ {*}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) \right] \right| > \epsilon_ {a} + \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {C o n f} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) \right|\right) \\ \geq P \left(\sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {A c c} \left(B _ {m}\right) - \operatorname {C o n f} \left(B _ {m}\right) \right| - \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{n} \left| \operatorname {A c c} \left(B _ {m} ^ {*}\right) - \operatorname {C o n f} \left(B _ {m} ^ {*}\right) \right| > \epsilon_ {a} + \left| \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) \right|\right) \\ = P \left(\operatorname {E C E} (X) - \operatorname {E C E} \left(X ^ {*}\right) > \epsilon_ {a} + \left| \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) \right|\right) \\ \end{array}
+$$
+
+Combined with the right side of the inequality:
+
+$$
+\begin{array}{l} P \left(\operatorname {E C E} (X) - \operatorname {E C E} \left(X ^ {*}\right) > \epsilon_ {a} + \left| \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) \right|\right) \leq 2 \exp \left(- 2 \epsilon_ {a} ^ {2} n\right) \\ = P \left(| \mathrm {E C E} (X) - \mathrm {E C E} \left(X ^ {*}\right) | > \epsilon_ {a} + | \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) |\right) \leq 4 \exp \left(- 2 \epsilon_ {a} ^ {2} n\right) \\ = P \left(\left| \mathrm {E C E} (X) - \mathrm {E C E} \left(X ^ {*}\right) \right| > \epsilon_ {E C E}\right) \leq 4 \exp \left(- 2 \epsilon_ {a} ^ {2} n\right) \\ = P \left(\left| \operatorname {E C E} (X) - \operatorname {E C E} \left(X ^ {*}\right) \right| > \epsilon_ {E C E}\right) \leq 4 \exp \left(- 2 \left(\epsilon_ {E C E} - \left| \operatorname {C o n f} (X) - \operatorname {C o n f} \left(X ^ {*}\right) \right|\right) ^ {2} n\right) \\ \end{array}
+$$
+
+where  $\epsilon_{ECE} = \epsilon_a + |\operatorname{Conf}(X) - \operatorname{Conf}(X^*)| = \epsilon_a + \sum_{m=1}^{M} \frac{|B_m|}{n} |\operatorname{Conf}(B_m) - \operatorname{Conf}(B_m^*)|$ .
+
+# C Appendix
+
+<table><tr><td colspan="2">System Prompt</td></tr><tr><td>TC</td><td>Consider the task of classifying between the following classes (along with some examples):1. complaint, which is about customer inquiries on a state of affairs, product, organization or event to express a negative mismatch between reality and expectations. Some examples of utterances include:- Dear @nvidia, I don&#x27;t think I should have to roll back to driver v270.61 to make my games work, and my desktop not glitch out.- @FC_Help him order is 913181 did you revise the money? if you did.. how about the shipping ?- @FC_Help Will you be getting the wendy cotton v neck dress in pavlova back in stock on the site?2. not_complaint, which is the opposite of complaint mentioned above, about customer regular or normal inquiries on a state of affairs, product, organization or event without any expression related to a negative mismatch between reality and expectations. Some examples of utterances include:- @FC_Help How can I get a hold of you so we can discuss the problem I am having with my coat ?- @FC_Help I need to check my order.- @FC_Help looking for &quot;bright carol&quot; or &quot;stained glass&quot; dress. do you have these in stock anymore?</td></tr><tr><td>SUBJ</td><td>Consider the task of classifying between the following classes (along with some examples):1. objective, which is assigned to text that presents factual information, descriptions, or statements without personal opinions, emotions, or bias. It focuses on delivering facts or information that is independent of the writer&#x27;s personal feelings or beliefs. Some examples of utterances include:- &quot;nicklas passes out , and the next day when he returns to school he notices that nobody seems to notice him.&quot; - &quot;when reuben buys a black-market cure for his unusual chest complaint, jenny is forced to make a terrible sacrifice.&quot; - &quot;raj has always had a unrequited childhood crush on a friend named Tina, but Tina&#x27;s best friend pooja has always had a crush on raj.&quot;2. subjective, which is applied to text that expresses personal opinions, feelings, beliefs, or thoughts. It often includes evaluative language, personal experiences, or interpretations, reflecting the writer&#x27;s personal stance or emotional reaction. Some examples of utterances include:- &quot;for its seriousness, high literary aspirations and stunning acting, the film can only be applauded.&quot; - &quot;an inegant combination of two unrelated shorts that falls far short of the director&#x27;s previous work in terms of both thematic content and narrative strength.&quot; - &quot;what&#x27;s needed so badly but what is virtually absent here is either a saving dark humor or the feel of poetic tragedy.&quot;</td></tr><tr><td>B77</td><td>Consider the task of classifying between the following classes (along with some examples):1. age_limit, which is about customer inquiries on age-related restrictions for opening a bank account. Some examples of utterances include:- Can I get an account for my son?- Can my teenager have an account?- How young can I be to open my own account?2. atm_support, which is about users asking how to use an ATM, where to find one, or any other clarifications about a transaction at an ATM. Some examples of utterances include:- Is the closest ATM to me within 2 miles?- Are there only certain ATM machines where I can use this card?- Do you know the closest ATM?</td></tr><tr><td>SE</td><td>Consider the task of classifying between the following classes (along with some examples):1. negative, which is assigned to content that expresses negative feelings, emotions, or attitudes. Examples include statements of dissatisfaction, sadness, anger, or criticism. Some examples of utterances include:- &quot;even with newborn diapers it filled way too fast.&quot; - &quot;bluetooth does not work on this phone.&quot; - &quot;also, some other mp3 players such as the nitrus allow you to play wma (windows media audio) files, whereas the ipod does not.&quot;2. positive, which is applied to content that expresses positive feelings, emotions, or attitudes. Examples include statements of happiness, satisfaction, praise, or optimism. Some examples of utterances include:- &quot;4 megapixels is enough for anybody and the photo quality is awesome.&quot; - &quot;an extra gig of room, fm radio, voice recorder, pim applications with sync to outlook.&quot; - &quot;all the buttons &amp; necessary lil&#x27; gadgets are on the sides of the player which is nice for when you are holding it in the palm of your hand.&quot;</td></tr></table>
+
+Table 7: System Prompt for Data Generation
+
+# D Appendix
+
+<table><tr><td>Dataset</td><td colspan="2">Input</td><td>Prompt</td><td>Generated Text (xi^syn, \(\hat{P}(y_i|x_i^{syn})\))</td><td>Label (Y)</td></tr><tr><td rowspan="3">TC</td><td>\(x_i\)</td><td>@UbisoftSupport When will u guys fix the jager glitch?</td><td rowspan="3">An example \(x_i\) which belongs 75% to complaint and 25% to not_complaint (based on a classifier&#x27;s categorization). Now I ask you act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs 65% to complaint and 35% to not_complaint.</td><td rowspan="3">Hey @UbisoftSupport, I&#x27;m still experiencing the jager glitch and it&#x27;s really affecting my gameplay. Can you please fix it asap? (65% complaint, 35% not_complaint)</td><td rowspan="3">complaint</td></tr><tr><td>\(\hat{P}(y_i|x_i)\)</td><td>0.75</td></tr><tr><td>miscalibration type</td><td>overconfidence</td></tr><tr><td rowspan="3">SUBJ</td><td>\(x_i\)</td><td>the gripping story of the dutch bettien martens, based on true facts.</td><td rowspan="3">An example \(x_i\) which belongs 95% to objective and 5% to subjective (based on a classifier&#x27;s categorization). Now I ask you act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs 70% to objective and 30% to subjective.</td><td rowspan="3">The remarkable tale of Bettien Martens serves as a testament to the enduring nature of hope and determination, even in the face of unimaginable hardship. (70% objective, 30% subjective)</td><td rowspan="3">objective</td></tr><tr><td>\(\hat{P}(y_i|x_i)\)</td><td>0.10</td></tr><tr><td>miscalibration type</td><td>underconfidence</td></tr><tr><td rowspan="3">B77</td><td>\(x_i\)</td><td>Can my teenager have an account?</td><td rowspan="3">An example \(x_i\) which belongs 65% to age_limit and 35% to atm_support (based on a classifier&#x27;s categorization). Now I ask you act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs 85% age_limit and 15% to atm_support.</td><td rowspan="3">Can my toddler open their own bank account? (85% age_limit and 15% to atm_support)</td><td rowspan="3">age_limit</td></tr><tr><td>\(\hat{P}(y_i|x_i)\)</td><td>0.65</td></tr><tr><td>miscalibration type</td><td>underconfidence</td></tr><tr><td rowspan="3">SE</td><td>\(x_i\)</td><td>The zoom function on this camera is so loud that sometimes you will be unable to use it if you find yourself in a situation where you must be quiet.</td><td rowspan="3">An example \(x_i\) which belongs 75% to negative and 25% to positive (based on a classifier&#x27;s categorization). Now I ask you to act as that classifier and based on this example, generate a diverse set of 3 short utterances where each utterance belongs 55% to negative and 45% to positive.</td><td rowspan="3">I love how this router can handle a large network, but the price is a bit steep for my taste. (55% negative, 45% positive)</td><td rowspan="3">negative</td></tr><tr><td>\(\hat{P}(y_i|x_i)\)</td><td>0.25</td></tr><tr><td>miscalibration type</td><td>overconfidence</td></tr></table>
+
+Table 8: An example of generating synthetic data via LLM. Input contains the original text  $(x_{i})$  and the average predictive probability of the bin it comes from  $(\hat{P}(y_{i}|x_{i}))$ . Generated Text is the one after the relabeling process. Note: during re-fine tuning of the downstream model, we exclude  $\hat{P}(y_{i}|x_{i}^{syn})$  and retain only  $x_{i}^{syn}$  in the dataset.

2024/Fill In The Gaps_ Model Calibration and Generalization with Synthetic Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d22978bcebcc568be20d1671f154019ce5a901abe976cc936f8353ab5796d9e6
+size 1172984

2024/Filtered Direct Preference Optimization/e2599cf1-1b15-4843-9792-84fb610259c0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b083160721afd68ab4e1d8b91bd7f787e65ccd8133018de21efa43d7b531fc2
+size 966954

2024/Filtered Direct Preference Optimization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8add88b3dee29ffb7d87a07fb0e871ce9980ad23c448b8a56ce5fca5402c651
+size 5311783

2024/FinDVer_ Explainable Claim Verification over Long and Hybrid-content Financial Documents/3adc16a8-3414-4d00-95ba-3c3dbbf67df2_content_list.json

@@ -0,0 +1,1820 @@
+[
+    {
+        "type": "text",
+        "text": "FINDVER: Explainable Claim Verification over Long and Hybrid-Content Financial Documents",
+        "text_level": 1,
+        "bbox": [
+            200,
+            89,
+            800,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yilun Zhao Yitao Long Yuru Jiang Chengye Wang Weiyuan Chen Hongjun Liu Yiming Zhang Xiangru Tang Chen Zhao Arman Cohan",
+        "bbox": [
+            159,
+            154,
+            838,
+            189
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yale NLP - FINDVER Team",
+        "bbox": [
+            383,
+            197,
+            616,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce FINDVER, a comprehensive benchmark specifically designed to evaluate the explainable claim verification capabilities of LLMs in the context of understanding and analyzing long, hybrid-content financial documents. FINDVER is divided into three subsets: information extraction, numerical reasoning, and knowledge-intensive reasoning—each addressing common scenarios encountered in real-world financial contexts. We assess a broad spectrum of LLMs under long-context and RAG settings. Our results show that even the current best-performing system, Claude-3.5-Sonnet, significantly lags behind human experts. Our detailed findings and insights highlight the strengths and limitations of existing LLMs in this new task. We believe FINDVER can serve as a valuable benchmark for evaluating LLM capabilities in claim verification over complex, expert-domain documents.",
+        "bbox": [
+            141,
+            286,
+            460,
+            570
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$\\mathbb{O}$  github.com/yilunzhao/FinDVer",
+        "bbox": [
+            152,
+            575,
+            436,
+            590
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            606,
+            258,
+            621
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In today's information explosion era, the responsibility of verifying the truthfulness of the item is often passed on to the audience. unverified claims about a company's financial performance frequently circulate in online media, potentially misleading investors. Therefore, it is crucial to verify these claims using the companies' original financial documents (i.e., earnings reports and regulatory filings). Recent advancements in Large Language Models (LLMs) have attracted significant attention due to their capabilities in solving a broad range of tasks (Touvron et al., 2023b; Jiang et al., 2023b; OpenAI, 2023a). However, it remains particularly difficult for applying them to document-grounded claim verification in real-world financial domains due to the following two reasons:",
+        "bbox": [
+            112,
+            631,
+            489,
+            888
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "First, financial documents are typically long, intricate and dense, and they include both quantita",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c7ea5953a21de17fd7ca8addbe3db246487ff65e8d9360c84f983e220cf93462.jpg",
+        "image_caption": [
+            "Figure 1: An example from the numerical reasoning subset of the FINDVER benchmark. To verify the claim, the LLM is required to first locate claim-relevant data points within long and hybrid-content financial documents, and then apply numerical reasoning over the extracted data points for claim verification."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            258,
+            878,
+            437
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tive tables and qualitative text (Chen et al., 2021; Zhu et al., 2021; Zhao et al., 2022, 2023d; Koncel-Kedziorski et al., 2024). Extracting and analyzing claim-relevant data from these documents requires complicated document comprehension abilities and professional knowledge in financial domains. Moreover, the type of reasoning involved encompasses various unique aspects that are less studied, necessitating a dedicated approach to evaluation and application.",
+        "bbox": [
+            507,
+            563,
+            884,
+            722
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Second, in the financial domain, where decisions often involve significant stakes, it is often critical to provide clear and comprehensible rationales for any claim verification decisions (Atanasova et al., 2020, 2023). However, existing context-grounded claim verification benchmarks (Chen et al., 2020; Kamoi et al., 2023; Lu et al., 2023; Glockner et al., 2024) primarily focus on the task of entailment classification and do not evaluate the reasoning process. This hinders the practical application and evaluation of LLMs in real-world scenarios.",
+        "bbox": [
+            507,
+            726,
+            882,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In response to the aforementioned pressing need,",
+        "bbox": [
+            526,
+            904,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "14739",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14739-14752 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            152,
+            945,
+            843,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/dd8b16d0459fd173a819982d8f76c1c4bf2b5e60337cbe53fb3becb953067531.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Input Context</td><td>Annotation / Data Creation</td><td># Label</td><td>w. Expla-\nnation?</td><td>Reasoning-\nIntensive?</td></tr><tr><td>PubMedTab (Akhtar et al., 2022)</td><td>Wikipedia table</td><td>Crowd-sourced</td><td>4</td><td>X</td><td>X</td></tr><tr><td>TABFACT (Chen et al., 2020)</td><td>Wikipedia table</td><td>Crowd-sourced</td><td>2</td><td>X</td><td>✓</td></tr><tr><td>INFOTABS (Gupta et al., 2020)</td><td>Wikipedia table</td><td>Crowd-sourced</td><td>3</td><td>X</td><td>✓</td></tr><tr><td>SCITAB (Lu et al., 2023)</td><td>Scientific table</td><td>Expert &amp; InstructGPT</td><td>3</td><td>X</td><td>✓</td></tr><tr><td>HOVER (Jiang et al., 2020)</td><td>Wikipedia articles</td><td>Crowd-sourced</td><td>2</td><td>X</td><td>X</td></tr><tr><td>DOCNLI (Yin et al., 2021)</td><td>News article</td><td>From summarization datasets</td><td>2</td><td>X</td><td>X</td></tr><tr><td>ContractNLI (Koreeda et al., 2021)</td><td>Contract</td><td>Expert &amp; Crowd-sourced</td><td>2</td><td>X</td><td>X</td></tr><tr><td>LLM-AGGREFACT (Tang et al., 2024)</td><td>Doc from various domains</td><td>From existing benchmarks</td><td>2</td><td>X</td><td>X</td></tr><tr><td>WICE (Kamoi et al., 2023)</td><td>Wikipedia article</td><td>Crowd-sourced</td><td>3</td><td>X</td><td>X</td></tr><tr><td>AMBIFC (Glockner et al., 2024)</td><td>Wikipedia article</td><td>Crowd-sourced</td><td>3</td><td>X</td><td>X</td></tr><tr><td>CLAIMDECOMP (Chen et al., 2022a)</td><td>Political article</td><td>Expert</td><td>6</td><td>X</td><td>X</td></tr><tr><td>SCIFACT (Wadden et al., 2020)</td><td>Scientific paper abstracts</td><td>Expert</td><td>2</td><td>X</td><td>✓</td></tr><tr><td>LIAR++ (Russo et al., 2023)</td><td>Political article</td><td>From fact-check website</td><td>2</td><td>✓</td><td>X</td></tr><tr><td>FullFact (Russo et al., 2023)</td><td>Web page</td><td>From fact-check website</td><td>2</td><td>✓</td><td>X</td></tr><tr><td>PUBHEALTH (Akhtar et al., 2022)</td><td>Health Web page</td><td>From fact check website</td><td>4</td><td>✓</td><td>X</td></tr><tr><td>FINDVER (ours)</td><td>Long financial doc with tables</td><td>Expert</td><td>2</td><td>✓</td><td>✓</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            884,
+            337
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Comparison between FINDVER and existing context-grounded claim verification datasets. FINDVER is distinguished by four unique characteristics: (1) Expert Annotation: It is annotated by financial experts to ensure high data quality; (2) Complex Document Comprehension: It requires interpreting a mix of textual and tabular data within a long-context financial document; (3) Examination on Reasoning-Process Explanation: It enhances claim verifications with detailed explanations about the reasoning process, increasing the benchmark's practical value; and (4) Diverse Reasoning for Real-world Scenarios: It incorporates various reasoning challenges, such as extracting complicated information, performing numerical calculations, applying external professional knowledge, and conducting comparative analyses. Accordingly, we divide the benchmark into three focused subsets, each tailored to mirror distinct real-world financial analysis scenarios.",
+        "bbox": [
+            110,
+            346,
+            884,
+            475
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "we present FINDVER, a comprehensive and domain expert-annotated explainable claim verification benchmark that first explores in the context of financial documents. The LLMs are tasked with generating explanations of their reasoning to verify claims labeled as \"entailed\" or \"refuted\", based on the information in the provided document, which contains both textual and tabular data. To identify the common reasoning-intensive scenarios in claim verification based on financial documents, we engage with domain experts and conducted a preliminary study. This helped us determine three key types of scenarios that frequently arise in real-world settings: information extraction, numerical reasoning, and knowledge-intensive reasoning. For each scenario, we construct an evaluation set. Each example in our dataset is annotated with detailed supporting evidence and step-by-step reasoning-process explanations.",
+        "bbox": [
+            110,
+            500,
+            489,
+            804
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We evaluate a wide spectrum of open- and closed-source LLMs, specifically, 19 models from 10 organizations. The documents in our benchmark are exceedingly long; therefore, we employ two widely adopted real-world application settings—retrieval augmented generation (RAG) and long-context—in this study. The experiment",
+        "bbox": [
+            112,
+            809,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tal results indicate that even the existing best-performing LLM (i.e., Claude-3.5-Sonnet) still significantly lags behind human experts  $(77.2\\%$  versus  $93.3\\%)$ , demonstrating the challenges of our proposed benchmark. Our contributions are summarized below:",
+        "bbox": [
+            507,
+            500,
+            884,
+            595
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce FINDVER, the first comprehensive context-grounded claim verification benchmark for financial domains, presenting new challenges for state-of-the-art LLMs.",
+            "- We conduct an extensive evaluation that encompasses a wide range of LLMs, including those specialized in finance and math. We also evaluate both long-context and RAG settings to comprehensively assess the capabilities and limitations of existing LLMs in our task.",
+            "- Our experimental results reveal a noticeable performance gap compared to human experts. This highlights the limitations of current LLMs in complex real-world applications and the need for continued advancements."
+        ],
+        "bbox": [
+            507,
+            602,
+            884,
+            852
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            864,
+            665,
+            879
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Claim Verification Benchmark Claim verification is a well-established research area with two",
+        "bbox": [
+            507,
+            889,
+            884,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "14740",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "main settings. The first is the open-domain setting, which involves using an external retriever to find the most relevant information from a large corpus to verify claims (Vlachos and Riedel, 2014; Thorne et al., 2018; Aly et al., 2021; Wadden et al., 2022; Rangapur et al., 2024; Ma et al., 2024). The second setting is context-grounded claim verification, which requires models to verify claims based on the provided document context (Chen et al., 2020; Kamoi et al., 2023; Lu et al., 2023; Glockner et al., 2024). This work focuses on the second setting, as it allows us to eliminate variability and dependency on the retriever's performance, thereby focusing on the evaluation of LLM performance on an accurately verifying claims within a given context. However, as illustrated in Table 1, existing context-grounded claim verification benchmarks have four notable limitations: they typically 1) focus on general domains, overlooking the specific challenges and intricacies present in specialized fields, 2) focus solely on entailment classification and do not evaluate the reasoning processes of models, 3) do not involve claims that require intensive reasoning and complicated document comprehension. These limitations hinder their effectiveness for evaluating LLMs in real-world practice.",
+        "bbox": [
+            112,
+            84,
+            492,
+            502
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Financial Evaluation Benchmark NLP techniques have been applied to various financial tasks, such as named entity recognition (Salinas Alvarado et al., 2015; Shah et al., 2023), sentiment analysis (Malo et al., 2013; Maia et al., 2018), stock movement prediction (Soun et al., 2022; Xu and Cohen, 2018; Wu et al., 2018), and summarization (Zhou et al., 2021; Mukherjee et al., 2022; Liu et al., 2022). More recently, there has been an increasing focus on tasks involving financial documents (e.g., annual reports and regulatory filings), which are crucial for providing insights into a company's performance and strategies. Several QA benchmarks have been proposed to evaluate models' performance in answering questions based on financial documents, with a particular focus on numerical reasoning (Chen et al., 2021; Zhu et al., 2021; Zhao et al., 2022; Chen et al., 2022b; Koncel-Kedziorski et al., 2024; Zhao et al., 2024b,a). Despite these advancements, there remains a significant gap in the exploration of claim verification tasks within the financial domain. While the recent FIN-FACT benchmark (Rangapur et al., 2024) addresses explainable multimodal financial fact-checking, it primarily focuses on open-domain sce",
+        "bbox": [
+            112,
+            519,
+            490,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "narios. Verifying claims derived from financial documents is crucial, as inaccuracies can significantly influence investment decisions and market perceptions. To bridge this gap, we introduce FINDVER, the first context-grounded claim verification benchmark, specifically designed for real-world financial document comprehension.",
+        "bbox": [
+            507,
+            84,
+            885,
+            197
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 FINDVER Benchmark",
+        "text_level": 1,
+        "bbox": [
+            507,
+            210,
+            741,
+            225
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "FINDVER provides a robust evaluation benchmark for reasoning-intensive and explainable claim verification over long and hybrid-content financial documents. We present an overview of the FINDVER construction pipeline in Figure 2; and detail the task formulation, data construction, and quality validation process in the following subsections.",
+        "bbox": [
+            507,
+            237,
+            885,
+            350
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Task Formulation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            363,
+            697,
+            376
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We formally define the task of FINDVER within the context of LLMs as follows: Consider a single financial document  $d$ , containing textual data  $P$  and tabular data  $T$ , associated with a claim  $c$  that needs verification. The expert-annotated data we collect supports the following two tasks:",
+        "bbox": [
+            507,
+            384,
+            882,
+            481
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Entailment Classification The model is required to determine the entailment label  $\\ell \\in \\mathcal{L} = \\{\\text{\"entailed\", \"refuted\"}\\}$ , based on the document context:",
+        "bbox": [
+            507,
+            491,
+            885,
+            555
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell = \\arg \\max  _ {\\ell \\in \\mathcal {L}} P _ {\\mathbf {L L M}} (\\ell \\mid P, T, c) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            574,
+            574,
+            882,
+            596
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Reasoning-process Explanation Generation The model is required to generate a natural language explanation  $e$ :",
+        "bbox": [
+            507,
+            606,
+            882,
+            653
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ne = \\arg \\max _ {e} P _ {\\mathbf {L L M}} (e \\mid P, T, c) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            671,
+            884,
+            694
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "which articulates the reasoning process behind the validity of the provided claim  $c$ , based solely on the provided textual content  $P$  and tabular content  $T$  within the financial document.",
+        "bbox": [
+            507,
+            702,
+            882,
+            766
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Notably, some claim verification systems, particularly those developed prior to the era of LLMs and for previous datasets that did not require explanation generation (Chen et al., 2020; Yin et al., 2021; Koreeda et al., 2021), might not explicitly perform explanation generation. Instead, they directly output the final label. For such systems, FINDVER can also be used for evaluation by focusing on the entailment classification task.",
+        "bbox": [
+            507,
+            776,
+            885,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "14741",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1d6122a63969cf6dca72bcb4870ef402a5118c041f589dfcaa5c2ba513ec29da.jpg",
+        "image_caption": [
+            "Figure 2: An overview of FINDVER construction pipeline."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            152,
+            80,
+            848,
+            225
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 FINDVER Subset Design",
+        "text_level": 1,
+        "bbox": [
+            112,
+            274,
+            354,
+            288
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "FINDVER is designed to mirror the real-world challenges encountered in the financial domain. Therefore, we ensure that the included annotators are financial experts with professional experience in comprehending and processing financial documents. Table 7 in Appendix presents the detailed annotator biographies for FINDVER annotation.",
+        "bbox": [
+            112,
+            297,
+            487,
+            409
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To identify the common reasoning-intensive scenarios in claim verification based on financial documents, we engaged with domain experts and conducted a preliminary study. This helped us determine three key types of scenarios that frequently arise in real-world settings. Accordingly, we have created three corresponding subsets of FINDVER.",
+        "bbox": [
+            112,
+            411,
+            489,
+            523
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) FDV-IE (information extraction), which involves extracting information from both textual and tabular content within a long-context document.",
+            "(2) FDV-MATH (numerical reasoning), which necessitates performing calculations or statistical analysis based on data within the document.",
+            "(3) FDV-Know (knowledge-intensive reasoning), which requires integrating external domain-specific knowledge or regulations for claim verification."
+        ],
+        "bbox": [
+            112,
+            524,
+            487,
+            675
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Source Document Collection",
+        "text_level": 1,
+        "bbox": [
+            112,
+            693,
+            383,
+            707
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Similar to Zhao et al. (2023a), we use the quarterly (Form 10-Q) and annual reports (Form 10-K) of companies as the source documents, which are publicly available in the open-source database<sup>1</sup> of the U.S. Securities and Exchange Commission. We collect a total of 523 documents that were first released between January 1 to April 30, 2024, which is after the cutoff date of most pretraining corpora for training foundation models. This helps to alleviate issues related to data memorization to some extent. After collecting the raw HTML-format documents,",
+        "bbox": [
+            112,
+            715,
+            489,
+            892
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "we utilize the SEC  $\\mathrm{API}^2$ , a commercial platform API for extracting financial document content, to process the collected documents, obtaining documents with both textual and tabular data.",
+        "bbox": [
+            507,
+            273,
+            884,
+            337
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Claim Annotation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            363,
+            697,
+            376
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Entailed Claim Annotation To address the potential bias concerning the position of evidence within the documents, we initiate the process by randomly sampling multiple document contexts from the given document. Annotators are then tasked with creating \"entailed\" claims based on the textual and tabular data within these contexts. The annotators are instructed to simulate real-world document comprehension scenarios, ensuring the annotated claims are representative of practical financial analysis and align with the scenarios defined by the corresponding subsets. Annotators are then tasked with carefully locating all evidence (i.e., indices of relevant paragraphs and tables) within the entire document that support the claims, which are used for the subsequent data validation.",
+        "bbox": [
+            505,
+            391,
+            884,
+            650
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Refuted Claim Annotation Following established practices in the field (Wadden et al., 2020; Chen et al., 2020; Lu et al., 2023), and since directly obtaining \"refuted\" types is difficult, we instead perturb the original \"entailed\" claims into \"refuted\" claim through expert annotation. Specifically, expert annotators first create an \"entailed\" claim using the same procedure detailed in the \"Entailed Claim Annotation\" paragraph. The annotators are then instructed to perturb the \"entailed\" claim to introduce factual errors that are directly contradicted by the annotated evidence, and rewrite the annotated reasoning-process explanation.",
+        "bbox": [
+            505,
+            671,
+            885,
+            882
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "'https://www.sec.gov/edgar/search/",
+        "bbox": [
+            134,
+            904,
+            443,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://sec-api.io/",
+        "bbox": [
+            529,
+            904,
+            712,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "14742",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/6fb1c2983f98981ffd8514e50a8820c87649644b34a33d9d3340b122083a3e1f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Annotation Quality</td><td>%S ≥ 4</td></tr><tr><td>Claim</td><td></td></tr><tr><td>Fluency</td><td>92</td></tr><tr><td>Meaningfulness</td><td>90</td></tr><tr><td>Alignment with real-world scenarios</td><td>94</td></tr><tr><td>Evidence</td><td></td></tr><tr><td>Relevancy</td><td>89</td></tr><tr><td>Completeness</td><td>85</td></tr><tr><td>Reasoning-process Explanation</td><td></td></tr><tr><td>Fluency</td><td>95</td></tr><tr><td>Correctness</td><td>92</td></tr><tr><td>Comprehensiveness</td><td>90</td></tr><tr><td>Entailment Label</td><td></td></tr><tr><td>Correctness</td><td>94</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            485,
+            303
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Human evaluation over 100 samples from the FINDVER testmini set. Two internal evaluators were asked to rate the samples on a scale of 1 to 5 individually. We report percent of samples that have an average score  $\\geq 4$  to indicate the annotation quality of FINDVER.",
+        "bbox": [
+            112,
+            311,
+            489,
+            384
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.5 Explanation Annotation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            418,
+            349,
+            434
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After finishing the claim annotation, we pass it to another annotator for explanation annotation. The annotators are required to first read the claim carefully and annotate a detailed explanation of the reasoning process. Such reasoning-process explanations allow for a granular and informative evaluation of model outputs, helping future work identify reasoning errors and provide more accurate error feedback. We compare the entailment label annotated in this step with those in the claim annotation step. A third annotator is introduced if the two annotation versions are different. In practice, we achieve an inter-annotator agreement of  $90.3\\%$  for entailment label annotation.",
+        "bbox": [
+            112,
+            449,
+            489,
+            674
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During our pilot annotation phase, we observed variability in the format of reasoning-process explanation annotated by different annotators, which made the dataset less standardized. To ensure consistency and clarity in our benchmark, we developed a predefined template for annotators to follow. Specifically, annotators are required to commence with the extraction of relevant information phase, where they need to list all claim-relevant information in a numbered list. Subsequently, they are required to annotate the reasoning over the extracted information segment in a step-by-step manner. For each step, they should elucidate the associated reasoning. Finally, they annotate the entailment label feature.",
+        "bbox": [
+            112,
+            680,
+            489,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.6 Data Quality Validation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            742,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To ensure the high quality of our annotated data, for every annotated example, a qualified annotator is assigned to validate several key aspects: (1) the claim and reasoning-process explanation should be grammatically correct and free of spelling errors; (2) the claim should be closely related to financial domains and meaningful in real-world scenarios; (3) the annotated evidence should be relevant to the claim and complete enough to verify it; (4) the entailment label of the claim should be supported by the annotated evidence; and (5) the reasoning-process explanation should correctly interpret the extracted evidence and apply appropriate reasoning steps to correctly verify the claim. The validators are asked to revise examples that do not meet these standards. In practice, 347 out of 2,100 initial examples were revised by the validators. We also report the human evaluation scores over 100 sampled examples. As illustrated in Table 2, FINDVER has a high annotation quality.",
+        "bbox": [
+            505,
+            105,
+            884,
+            429
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.7 Dataset Preparation and Release",
+        "text_level": 1,
+        "bbox": [
+            507,
+            439,
+            811,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3 provides an overview of the key statistics for our benchmark. The dataset is randomly split into two subsets: testmini and test. The testmini set is intended for model development and validation. It contains 600 examples, with 200 examples from each subset. The test set comprises the remaining 1,500 examples and is designed for standard evaluation. To prevent data contamination (Jacovi et al., 2023; Shi et al., 2024; Deng et al., 2024), the ground-truth-related annotation features for the test set will not be publicly released. Instead, we provide an online evaluation platform where researchers can assess their models and participate in a public leaderboard.",
+        "bbox": [
+            505,
+            461,
+            884,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            700,
+            700,
+            717
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We next present the experimental setup, covering the evaluated LLMs, long-context and RAG setups, implementation details, and the measurement of human-level performance.",
+        "bbox": [
+            505,
+            727,
+            882,
+            790
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental LLMs",
+        "text_level": 1,
+        "bbox": [
+            507,
+            803,
+            721,
+            818
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We examine the performance of LLMs across two distinct categories on FINDVER: (1) Proprietary LLMs, including GPT-4* (OpenAI, 2023a,b, 2024), Gemini-1.5-* (Gemini, 2024), and Claude-3 (Anthropic, 2024); and (2) Open-source LLMs, including Gemma (Team et al., 2024), Llama",
+        "bbox": [
+            505,
+            825,
+            884,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "14743",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/f759e5dc082ea94ddc83e33d9106e25196b9bd1f71fde233002ff61ff2eb1e73.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Property</td><td>FDV-IE</td><td>FDV-MATH</td><td>FDV-Know</td></tr><tr><td>Real-world scenarios in financial domains</td><td>information extraction</td><td>numerical reasoning</td><td>knowledge-intensive reasoning</td></tr><tr><td># Document</td><td>221</td><td>225</td><td>217</td></tr><tr><td>Doc Length (i.e., word count) (Median/Avg/Max)</td><td>42K / 41K / 71K</td><td>43K / 41K / 71K</td><td>43K / 41K / 71K</td></tr><tr><td># Tables per document (Median/Avg)</td><td>62 / 78.9</td><td>62 / 79.1</td><td>62 / 79.0</td></tr><tr><td>Claim length (Median/Avg)</td><td>47 / 47.2</td><td>24 / 25.1</td><td>36 / 37.1</td></tr><tr><td># Text evidence per claim (Median/Avg)</td><td>2 / 1.8</td><td>1 / 1.3</td><td>3 / 2.6</td></tr><tr><td># Table evidence per claim (Median/Avg)</td><td>1 / 1.0</td><td>1 / 0.9</td><td>1 / 1.2</td></tr><tr><td>% Claims w. table evidence</td><td>66.3%</td><td>71.1%</td><td>70.8%</td></tr><tr><td>Explanation length (Median/Avg)</td><td>70 / 73.1</td><td>74 / 76.2</td><td>96 / 100.7</td></tr><tr><td>Benchmark size (# Claims)</td><td></td><td></td><td></td></tr><tr><td>testmini size</td><td>200</td><td>200</td><td>200</td></tr><tr><td>test size</td><td>500</td><td>500</td><td>500</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            880,
+            291
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Basic statistics of the FINDVER benchmark.",
+        "bbox": [
+            315,
+            303,
+            678,
+            317
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Adopted Chain-of-Thought Prompt",
+        "text_level": 1,
+        "bbox": [
+            139,
+            344,
+            359,
+            357
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "[System Input]",
+        "text_level": 1,
+        "bbox": [
+            141,
+            369,
+            240,
+            381
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As a financial expert, your task is to assess the truthfulness of the given claim by determining whether it is entailed or refuted based on the provided financial document. Follow these steps:",
+        "bbox": [
+            141,
+            382,
+            460,
+            428
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Carefully read the given context and the claim.",
+            "2. Analyze the document, focusing on the relevant financial data or facts that related to the claim.",
+            "3. Document each step of your reasoning process to ensure your assessment is clear and thorough.",
+            "4. Conclude your analysis with a final determination. In your last sentence, clearly state your conclusion in the following format: \"Therefore, the claim is {entailment_label}.\" Replace {entailment_label} with either 'entailed' (if the claim is supported by the document) or 'refuted' (if the claim contradicts or partially contradicts the document)."
+        ],
+        "bbox": [
+            141,
+            429,
+            460,
+            571
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "[User Input]",
+        "text_level": 1,
+        "bbox": [
+            141,
+            594,
+            226,
+            606
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Financial Report:",
+        "bbox": [
+            141,
+            607,
+            250,
+            618
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "{Financial Report}",
+        "bbox": [
+            141,
+            619,
+            255,
+            631
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Claim to verify:",
+        "bbox": [
+            141,
+            642,
+            240,
+            653
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "{Claim}",
+        "bbox": [
+            141,
+            655,
+            191,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Follow the instructions and think step by step to verify the claim.",
+        "bbox": [
+            139,
+            677,
+            460,
+            702
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 3: The Chain-of-Thought prompt used.",
+        "bbox": [
+            142,
+            731,
+            457,
+            746
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2&3 (Touvron et al., 2023a; Meta, 2024), Yi1.5 (AI et al., 2024), Qwen-2 (qwe, 2024), Mistral & Mixtral (Jiang et al., 2023a, 2024), InternLM2 (Team, 2024), C4AI (Aryabumi et al., 2024), GLM (Du et al., 2022), and Phi-3 (Abdin et al., 2024). Table 8 in Appendix presents the details of evaluated models (i.e., organizations, release time, max context length, and model version).",
+        "bbox": [
+            112,
+            774,
+            489,
+            903
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The experiments for open-sourced LLMs were",
+        "bbox": [
+            132,
+            904,
+            485,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "conducted using the vLMM framework (Kwon et al., 2023). For all the experiments, we set temperature as 1.0 and maximum output length as 512. We adopt the Chain-of-Thought (CoT) prompting methods (Wei et al., 2022) for the FINDVER benchmark. Specifically, the model is instructed to first output a detailed reasoning process for verifying claims, and then provide the entailment label of the claim based on the generated reasoning process. Figure 3 presents the used prompts.",
+        "bbox": [
+            507,
+            343,
+            884,
+            504
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Long-Context and RAG Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            514,
+            810,
+            530
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As presented in Table 3, the documents within our benchmark are notably lengthy. To effectively handle this, we have implemented two real-world application settings that are widely recognized for their utility in dealing with extensive texts. For Long-context Setting, we input the entire financial document into the model. We limit our evaluation to those models that have a context window larger than 100,000 tokens, which exceeds the maximum length of the included financial document. For RAG Setting, we leverage the current best-performing embedding models (i.e., OpenAI's text-embedding-3-large) to retrieve the top-10 paragraphs or tables that are most relevant to the claims. These elements are then concatenated in their original order as found in the document before being fed into the model.",
+        "bbox": [
+            507,
+            536,
+            884,
+            809
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            820,
+            741,
+            835
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Input Tabular Data Compilation Building on previous research that assessed LLMs on tasks involving tabular data (Chen, 2023; Zhao et al., 2023b,c), we introduce our methodology for processing tables within documents. Our approach",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "14744",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/3373d35b3dfa3ca793db6bde18602044c4031bc2f9c68728e63774c5f2bad87f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Notes</td><td colspan=\"2\">FDV-IE</td><td colspan=\"2\">FDV-MATH</td><td colspan=\"2\">FDV-Know</td><td colspan=\"2\">Average</td></tr><tr><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td></tr><tr><td>Random Choice</td><td></td><td>50.0</td><td></td><td>50.0</td><td></td><td>50.0</td><td></td><td>50.0</td><td></td></tr><tr><td>Human Non-Expert</td><td></td><td>90.0</td><td></td><td>85.0</td><td></td><td>85.0</td><td></td><td>86.7</td><td></td></tr><tr><td>Human Expert</td><td></td><td>95.0</td><td></td><td>90.0</td><td></td><td>95.0</td><td></td><td>93.3</td><td></td></tr><tr><td colspan=\"10\">Open-source LLMs</td></tr><tr><td>InternLM2-Math-7b</td><td>Math</td><td>-</td><td>58.0</td><td>-</td><td>53.5</td><td>-</td><td>54.5</td><td>-</td><td>55.3</td></tr><tr><td>InternLM2-7B</td><td></td><td>-</td><td>59.5</td><td>-</td><td>54.5</td><td>-</td><td>56.0</td><td>-</td><td>56.7</td></tr><tr><td>Gemma-7B</td><td></td><td>-</td><td>59.5</td><td>-</td><td>55.5</td><td>-</td><td>55.0</td><td>-</td><td>56.7</td></tr><tr><td>GLM-4-9b</td><td></td><td>58.5</td><td>61.0</td><td>54.5</td><td>54.5</td><td>55.0</td><td>56.5</td><td>56.0</td><td>57.3</td></tr><tr><td>Llama-2-7B</td><td></td><td>-</td><td>60.0</td><td>-</td><td>56.5</td><td>-</td><td>56.5</td><td>-</td><td>57.7</td></tr><tr><td>Mistral-7B-v3</td><td></td><td>-</td><td>59.5</td><td>-</td><td>56.5</td><td>-</td><td>57.0</td><td>-</td><td>57.7</td></tr><tr><td>Phi-3-medium-4k</td><td></td><td>-</td><td>61.5</td><td>-</td><td>54.0</td><td>-</td><td>58.0</td><td>-</td><td>57.8</td></tr><tr><td>Llama-2-70B</td><td></td><td>-</td><td>61.5</td><td>-</td><td>54.5</td><td>-</td><td>58.0</td><td>-</td><td>58.0</td></tr><tr><td>Phi-3-medium-128k</td><td></td><td>58.0</td><td>61.5</td><td>54.0</td><td>55.5</td><td>56.5</td><td>57.5</td><td>56.2</td><td>58.2</td></tr><tr><td>Meta-Llama-3-8B</td><td></td><td>-</td><td>62.5</td><td>-</td><td>55.0</td><td>-</td><td>59.5</td><td>-</td><td>59.0</td></tr><tr><td>Yi-1.5-34B</td><td></td><td>-</td><td>62.5</td><td>-</td><td>58.0</td><td>-</td><td>58.0</td><td>-</td><td>59.5</td></tr><tr><td>Meta-Llama-3-70B</td><td>MoE</td><td>-</td><td>65.5</td><td>-</td><td>61.5</td><td>-</td><td>61.5</td><td>-</td><td>62.8</td></tr><tr><td>C4AI Command R+</td><td></td><td>-</td><td>67.5</td><td>-</td><td>60.0</td><td>-</td><td>64.5</td><td>-</td><td>64.0</td></tr><tr><td>Qwen2-72B</td><td></td><td>67.0</td><td>68.0</td><td>62.5</td><td>61.5</td><td>60.5</td><td>65.0</td><td>63.3</td><td>64.8</td></tr><tr><td>Mixtral-8x22B</td><td></td><td>-</td><td>70.0</td><td>-</td><td>62.0</td><td>-</td><td>67.0</td><td>-</td><td>66.3</td></tr><tr><td colspan=\"10\">Proprietary LLMs</td></tr><tr><td>Gemini-1.5-Flash</td><td></td><td>71.0</td><td>70.5</td><td>62.5</td><td>60.5</td><td>65.0</td><td>65.5</td><td>66.2</td><td>65.5</td></tr><tr><td>GPT-3.5-turbo</td><td></td><td>-</td><td>79.0</td><td>-</td><td>64.0</td><td>-</td><td>70.5</td><td>-</td><td>71.2</td></tr><tr><td>GPT-4o</td><td></td><td>80.0</td><td>78.5</td><td>70.5</td><td>68.0</td><td>76.5</td><td>74.5</td><td>75.7</td><td>73.7</td></tr><tr><td>Claude-3.5-Sonnet</td><td></td><td>83.5</td><td>80.5</td><td>71.0</td><td>69.0</td><td>77.0</td><td>75.5</td><td>77.2</td><td>75.0</td></tr></table>",
+        "bbox": [
+            122,
+            80,
+            872,
+            529
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Accuracy of entailment classification on the testmini set of FINDVER. We report results for LLMs with CoT prompting under the long-context (LongC) and RAG settings. Numbers underscored indicate that models under the long-context setting achieve better results than under the RAG setting.",
+        "bbox": [
+            112,
+            539,
+            882,
+            583
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "involves distinguishing headers and cells in different columns using a vertical bar (l) and separating rows with new lines. This format allows us to input flattened table data directly into LLMs. In our initial experiments, we found that LLMs such as GPT-* and Llama-3 can effectively interpret this table representation. However, we suggest that future studies should investigate more sophisticated methods for encoding tabular data to enhance comprehension by LLMs.",
+        "bbox": [
+            112,
+            607,
+            489,
+            770
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Model Response Processing Following previous work (Lu et al., 2024), we adopt LLM for processing model response. Specifically, we utilize GPT-4o-mini to extract labels from the LLM output, which can be either \"entailed\", \"refuted\" or \"none\". The \"none\" label typically indicates that the LLM output contains nonsensical symbols or unintelligible text rather than meaningful content. In cases where the output is labeled as \"none\", we",
+        "bbox": [
+            110,
+            776,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "assign the final label by making a random guess.",
+        "bbox": [
+            507,
+            608,
+            870,
+            624
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Human-level Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            634,
+            764,
+            648
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To provide a rough but informative estimate of human-level performance by non-experts and experts on FINDVER, we randomly sampled 5 documents  $\\times 4$  claims / document  $= 20$  claims from each validation subset, totaling 60 claims. We enroll two experts (i.e., professionals with CFA license) and two non-experts (i.e., undergraduate students majored in computer science) to individually verify the claims by providing the NL explanations. Table 4 presents the human-level performance.",
+        "bbox": [
+            505,
+            655,
+            884,
+            816
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Experiment Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            827,
+            714,
+            843
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Main Findings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            852,
+            670,
+            868
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4 and Table 5 display the primary results for FINDVER. We reveals a significant accuracy gap between human experts and LLMs. Notably,",
+        "bbox": [
+            507,
+            873,
+            884,
+            922
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "14745",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/332e5d66f77efbe6b84906a5d108a225faac6811ec3fb1ae95be1db976d56ff0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>IE</td><td>MATH</td><td>KNOW</td><td>Avg</td></tr><tr><td>Human Non-Expert</td><td>90.0</td><td>85.0</td><td>85.0</td><td>86.7</td></tr><tr><td>Human Expert</td><td>95.0</td><td>90.0</td><td>95.0</td><td>93.3</td></tr><tr><td colspan=\"5\">Open-source LLMs</td></tr><tr><td>InternLM2-Math-7B</td><td>57.0</td><td>53.8</td><td>54.6</td><td>55.1</td></tr><tr><td>InternLM2-7B</td><td>59.6</td><td>53.4</td><td>55.4</td><td>56.1</td></tr><tr><td>Gemma-7B</td><td>59.8</td><td>54.4</td><td>55.0</td><td>56.4</td></tr><tr><td>GLM-4-9B</td><td>61.4</td><td>54.2</td><td>57.4</td><td>57.7</td></tr><tr><td>Llama-2-7B</td><td>61.0</td><td>57.2</td><td>57.2</td><td>58.5</td></tr><tr><td>Mistral-7B-v3</td><td>59.8</td><td>56.0</td><td>56.4</td><td>57.4</td></tr><tr><td>Phi-3-medium-4k</td><td>61.8</td><td>54.0</td><td>57.2</td><td>57.7</td></tr><tr><td>Llama-2-70B</td><td>60.6</td><td>53.8</td><td>58.0</td><td>57.5</td></tr><tr><td>Phi-3-medium-128k</td><td>61.2</td><td>55.4</td><td>58.2</td><td>58.3</td></tr><tr><td>Meta-Llama-3-8B</td><td>63.4</td><td>55.0</td><td>60.2</td><td>59.5</td></tr><tr><td>Yi-1.5-34B</td><td>62.8</td><td>57.2</td><td>56.8</td><td>58.9</td></tr><tr><td>Meta-Llama-3-70B</td><td>65.0</td><td>61.2</td><td>60.4</td><td>62.2</td></tr><tr><td>C4AI Command R+</td><td>67.4</td><td>59.0</td><td>65.4</td><td>63.9</td></tr><tr><td>Qwen2-72B</td><td>69.0</td><td>62.2</td><td>65.2</td><td>65.5</td></tr><tr><td>Mixtral-8x22B</td><td>71.0</td><td>62.8</td><td>68.2</td><td>67.3</td></tr><tr><td colspan=\"5\">Proprietary LLMs</td></tr><tr><td>Gemini-1.5-Flash</td><td>71.2</td><td>61.0</td><td>65.8</td><td>66.0</td></tr><tr><td>GPT-3.5-turbo</td><td>79.6</td><td>65.2</td><td>70.8</td><td>71.9</td></tr><tr><td>GPT-4o</td><td>78.6</td><td>69.4</td><td>73.8</td><td>73.9</td></tr><tr><td>Claude-3.5-Sonnet</td><td>81.0</td><td>68.2</td><td>74.0</td><td>74.4</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            485,
+            416
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Claude-3.5-Sonnet, the highest-performing LLM, achieves an accuracy rate of only  $77.2\\%$ , in stark contrast to the  $93.3\\%$  accuracy of financial experts. This discrepancy highlights the complexity and challenges of our benchmark.",
+        "bbox": [
+            112,
+            510,
+            487,
+            590
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "For the less competitive LLMs, such as Qwen2-72B, GLM-4-9B, and Phi-3-medium-128k, they exhibit improved performance under the RAG setting. In contrast, the currently more competitive LLMs, such as GPT-4o and Claude-3.5-Sonnet, generally perform better under the long-context setting compared to the RAG setting. This indicates the potential of developing long-context techniques to manage tasks involving extensive documents in specialized domains.",
+        "bbox": [
+            112,
+            593,
+            489,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 Chain-of-Thought Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            769,
+            374,
+            785
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To better understand the effectiveness of CoT prompting methods for our tasks, we select the commonly-used proprietary and open-source LLMs, GPT-4o and Qwen2-72B, for our experiments. In the w/o CoT setting, we instruct the LLMs to directly output the entailment label of the claim using the provided document context (Figure 4). As illustrated in Table 6, both LLMs' per",
+        "bbox": [
+            112,
+            791,
+            490,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9713aa09a204cbf751b49bf68651efb363464e8122bfeb61414b27d2f3d1ae02.jpg",
+        "image_caption": [
+            "Figure 4: The Direct Output prompt used."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            514,
+            82,
+            868,
+            324
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/ae32b0824ded7982267046c5c31cc27d6e69cc85ab867ddd68b38743441b8cf3.jpg",
+        "table_caption": [
+            "Table 5: Accuracy of entailment classification on the FINDVER test set. We report results for LLMs with CoT prompting under the RAG setting. Due to computation constraint, we did not evaluate the long-context setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">w/o CoT</td><td colspan=\"2\">w/ CoT</td></tr><tr><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td></tr><tr><td>Qwen2-72B</td><td>57.7 (-5.6)</td><td>59.5 (-5.3)</td><td>63.3</td><td>64.8</td></tr><tr><td>GPT-4o</td><td>70.1 (-7.1)</td><td>69.5 (-5.5)</td><td>77.2</td><td>75.0</td></tr></table>",
+        "bbox": [
+            510,
+            366,
+            878,
+            441
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Accuracy of entailment for GPT-4o and Qwen2-72B with and without CoT Prompting methods on the FINDVER testmini set.",
+        "bbox": [
+            507,
+            451,
+            882,
+            494
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "formance degrades in the w/o CoT setting. These results highlight the importance of CoT reasoning in enhancing performance for our task.",
+        "bbox": [
+            507,
+            523,
+            882,
+            571
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3 Error Analysis of Reasoning Process",
+        "text_level": 1,
+        "bbox": [
+            507,
+            590,
+            843,
+            606
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The Claude-3.5-Sonnet model achieves a top accuracy of  $77.2\\%$  under the long context setting. To better understand the model's limitations, we perform a detailed error analysis with human evaluators. We randomly select 25 instances from each of the three subsets where the Claude-3.5-Sonnet model fails to perform correctly. Our analysis has identified four primary categories of errors: (1) Extraction error: The model fails to correctly retrieve relevant information from the context, resulting in inaccurate verification. (2) Numerical reasoning error: The model encounters difficulties with correct mathematical reasoning. (3) Domain knowledge deficiency: The model lacks sufficient knowledge in finance-related areas, which hampers its ability to reason accurately. (4) Computation error: While the model's reasoning is correct, it makes computational mistakes during intermediate or final steps, resulting in incorrect verification.",
+        "bbox": [
+            505,
+            615,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "14746",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            247,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper presents FINDVER, a comprehensive benchmark designed to evaluate LLMs in claim verification over long and hybrid-content financial documents. Through extensive experiments involving 19 LLMs under long-context and RAG settings, we have demonstrated that even the top-performing models exhibit a significant performance gap compared to financial experts. Our detailed findings and insights reveal the strengths and limitations of current LLMs in this new task. We believe that FINDVER provides a valuable benchmark for future research on LLMs' ability to handle complex claim verification tasks within the expert domain.",
+        "bbox": [
+            112,
+            109,
+            492,
+            319
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            331,
+            220,
+            344
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose FINDVER and conduct comprehensive analysis of different LLMs' capabilities on our task. However, there are still some limitations: First, our evaluation does not include recently released finance-specific LLMs (Wu et al., 2023; Yang et al., 2023; Xie et al., 2023, 2024), as these models are not yet compatible with the vLLM framework used for inference. Due to computational resource constraints, we do not tune LLMs on a large-scale finance-domain data ourselves. However, we believe that training on finance data can help improve LLMs' capabilities in FINDVER. Moreover, we only conduct human error analysis on the generated reasoning process of models. We believe future work could explore the development of LLM-based automated evaluation systems (Liu et al., 2023; Zheng et al., 2023) for automatically detecting reasoning errors within the generated explanation.",
+        "bbox": [
+            112,
+            356,
+            489,
+            663
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            690,
+            213,
+            703
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "2024. Qwen2 technical report.",
+        "text_level": 1,
+        "bbox": [
+            114,
+            713,
+            324,
+            727
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, Alon Benhaim, Misha Bilenko, Johan Bjorck, Sebastien Bubeck, Martin Cai, Caio Cesar Teodoro Mendes, Weizhu Chen, Vishrav Chaudhary, Parul Chopra, Allie Del Giorno, Gustavo de Rosa, Matthew Dixon, Ronen Eldan, Dan Iter, Amit Garg, Abhishek Goswami, Suriya Gunasekar, Emman Haider, Junheng Hao, Russell J. Hewett, Jamie Huynh, Mojan Javaheripi, Xin Jin, Piero Kauffmann, Nikos Karampatziakis, Dongwoo Kim, Mahoud Khademi, Lev Kurilenko, James R. Lee, Yin Tat Lee, Yuanzhi Li, Chen Liang, Weishung Liu, Eric",
+        "bbox": [
+            115,
+            736,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Lin, Zeqi Lin, Piyush Madan, Arindam Mitra, Hardik Modi, Anh Nguyen, Brandon Norick, Barun Patra, Daniel Perez-Becker, Thomas Portet, Reid Pryzant, Heyang Qin, Marko Radmilac, Corby Rosset, Sambudha Roy, Olatunj Ruwase, Olli Saarikivi, Amin Saied, Adil Salim, Michael Santacroce, Shital Shah, Ning Shang, Hiteshi Sharma, Xia Song, Masahiro Tanaka, Xin Wang, Rachel Ward, Guanhua Wang, Philipp Witte, Michael Wyatt, Can Xu, Jiahang Xu, Sonali Yadav, Fan Yang, Ziyi Yang, Donghan Yu, Chengruidong Zhang, Cyril Zhang, Jianwen Zhang, Li Lyna Zhang, Yi Zhang, Yue Zhang, Yunan Zhang, and Xiren Zhou. 2024. Phi-3 technical report: A highly capable language model locally on your phone.",
+        "bbox": [
+            526,
+            85,
+            884,
+            282
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "01. AI, :, Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, Kaidong Yu, Peng Liu, Qiang Liu, Shawn Yue, Senbin Yang, Shiming Yang, Tao Yu, Wen Xie, Wenhao Huang, Xiaohui Hu, Xiaoyi Ren, Xinyao Niu, Pengcheng Nie, Yuchi Xu, Yudong Liu, Yue Wang, Yuxuan Cai, Zhenyu Gu, Zhiyuan Liu, and Zonghong Dai. 2024. Yi: Open foundation models by 01.ai.",
+        "bbox": [
+            509,
+            291,
+            885,
+            411
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. 2022. PubHealthTab: A public health table-based dataset for evidence-based fact checking. In *Findings of the Association for Computational Linguistics: NAACL* 2022, pages 1-16, Seattle, United States. Association for Computational Linguistics.",
+        "bbox": [
+            509,
+            420,
+            885,
+            500
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Rami Aly, Zhijiang Guo, Michael Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. 2021. Feverous: Fact extraction and verification over unstructured and structured information. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1.",
+        "bbox": [
+            509,
+            508,
+            885,
+            602
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Anthropic. 2024. Introducing the next generation of claude.",
+        "bbox": [
+            509,
+            612,
+            882,
+            639
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Viraat Aryabumi, John Dang, Dwarak Talupuru, Saurabh Dash, David Cairuz, Hangyu Lin, Bharat Venkitesh, Madeline Smith, Jon Ander Campos, Yi Chern Tan, Kelly Marchisio, Max Bartolo, Sebastian Ruder, Acyr Locatelli, Julia Kreutzer, Nick Frosst, Aidan Gomez, Phil Blunsom, Marzieh Fadaee, Ahmet Üstün, and Sara Hooker. 2024. Aya 23: Open weight releases to further multilingual progress.",
+        "bbox": [
+            510,
+            649,
+            884,
+            755
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Pepa Atanasova, Oana-Maria Camburu, Christina Lioma, Thomas Lukasiewicz, Jakob Grue Simonsen, and Isabelle Augenstein. 2023. Faithfulness tests for natural language explanations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 283-294, Toronto, Canada. Association for Computational Linguistics.",
+        "bbox": [
+            509,
+            764,
+            885,
+            871
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. Generating fact checking explanations. In Proceedings of the 58th",
+        "bbox": [
+            509,
+            879,
+            882,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "14747",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Annual Meeting of the Association for Computational Linguistics, pages 7352-7364, Online. Association for Computational Linguistics.",
+            "Jifan Chen, Aniruddh Sriram, Eunsol Choi, and Greg Durrett. 2022a. Generating literal and implied subquestions to fact-check complex claims. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3495-3516, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Wenhu Chen. 2023. Large language models are few(1)-shot table reasoners. In Findings of the Association for Computational Linguistics: EACL 2023, pages 1120-1130, Dubrovnik, Croatia. Association for Computational Linguistics.",
+            "Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2020. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations.",
+            "Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. 2021. FinQA: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697-3711, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. 2022b. ConvFinQA: Exploring the chain of numerical reasoning in conversational finance question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6279-6292, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Chunyuan Deng, Yilun Zhao, Yuzhao Heng, Yitong Li, Jiannan Cao, Xiangru Tang, and Arman Cohan. 2024. Unveiling the spectrum of data contamination in language model: A survey from detection to remediation. In Findings of the Association for Computational Linguistics ACL 2024, pages 16078-16092, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics.",
+            "Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. GLM: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, Dublin, Ireland. Association for Computational Linguistics.",
+            "Gemini. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.",
+            "Max Glockner, Ieva Staliūnaitè, James Thorne, Gisela Vallejo, Andreas Vlachos, and Iryna Gurevych. 2024."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "AmbiFC: Fact-checking ambiguous claims with evidence. Transactions of the Association for Computational Linguistics, 12:1-18.",
+            "Vivek Gupta, Maitrey Mehta, Pegah Nokhiz, and Vivek Srikumar. 2020. INFOTABS: Inference on tables as semi-structured data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2309–2324, Online. Association for Computational Linguistics.",
+            "Alon Jacovi, Avi Caciularu, Omer Goldman, and Yoav Goldberg. 2023. Stop uploading test data in plain text: Practical strategies for mitigating data contamination by evaluation benchmarks. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5075-5084, Singapore. Association for Computational Linguistics.",
+            "Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023a. Mistral 7b. arXiv preprint arXiv:2310.06825.",
+            "Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024. Mixtral of experts.",
+            "Albert Qiaochu Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L'elio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023b. Mistral 7b. ArXiv, abs/2310.06825.",
+            "Yichen Jiang, Shikha Bordia, Zheng Zhong, Charles Dognin, Maneesh Singh, and Mohit Bansal. 2020. HoVer: A dataset for many-hop fact extraction and claim verification. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3441-3460, Online. Association for Computational Linguistics.",
+            "Ryo Kamoi, Tanya Goyal, Juan Diego Rodriguez, and Greg Durrett. 2023. WiCE: Real-world entailment for claims in Wikipedia. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7561-7583, Singapore. Association for Computational Linguistics.",
+            "Rik Koncel-Kedziorski, Michael Krumdick, Viet Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. 2024. Bizbench: A quantitative reasoning benchmark for business and finance."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "14748",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yuta Koreeda, Manning, and Christopher. 2021. ContractNLI: A dataset for document-level natural language inference for contracts. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 1907-1919, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.",
+            "Shuaiqi Liu, Jiannong Cao, Ruosong Yang, and Zhiyuan Wen. 2022. Long text and multi-table summarization: Dataset and method. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 1995-2010, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: NLG evaluation using gpt-4 with better human alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 2511–2522, Singapore. Association for Computational Linguistics.",
+            "Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, KaiWei Chang, Michel Galley, and Jianfeng Gao. 2024. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In International Conference on Learning Representations (ICLR).",
+            "Xinyuan Lu, Liangming Pan, Qian Liu, Preslav Nakov, and Min-Yen Kan. 2023. SCITAB: A challenging benchmark for compositional reasoning and claim verification on scientific tables. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7787-7813, Singapore. Association for Computational Linguistics.",
+            "Huanhuan Ma, Weizhi Xu, Yifan Wei, Liuji Chen, Liang Wang, Qiang Liu, Shu Wu, and Liang Wang. 2024. Ex-fever: A dataset for multi-hop explainable fact verification.",
+            "Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. 2018. Ww'18 open challenge: Financial opinion mining and question answering. In Companion Proceedings of the The Web Conference 2018, WWW '18, page 1941-1942, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee.",
+            "Pekka Malo, Ankur Sinha, Pyry Takala, Pekka Korhonen, and Jyrki Wallenius. 2013. Good debt or bad debt: Detecting semantic orientations in economic texts."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "AI Meta. 2024. Introducing meta llama 3: The most capable openly available llm to date. Meta AI.",
+            "Rajdeep Mukherjee, Abhinav Bohra, Akash Banerjee, Soumya Sharma, Manjunath Hegde, Afreen Shaikh, Shivani Shrivastava, Koustuv Dasgupta, Niloy Ganguly, Saptarshi Ghosh, and Pawan Goyal. 2022. ECTSum: A new benchmark dataset for bullet point summarization of long earnings call transcripts. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10893-10906, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "OpenAI. 2022. Chatgpt: Optimizing language models for dialogue.",
+            "OpenAI. 2023a. Gpt-4 technical report. ArXiv, abs/2303.08774.",
+            "OpenAI. 2023b. GPT-4V(ision) system card.",
+            "OpenAI. 2024. Hello gpt-4o.",
+            "Aman Rangapur, Haoran Wang, Ling Jian, and Kai Shu. 2024. Fin-fact: A benchmark dataset for multimodal financial fact checking and explanation generation.",
+            "Daniel Russo, Serra Sinem Tekiroglu, and Marco Guerini. 2023. Benchmarking the generation of fact checking explanations. Transactions of the Association for Computational Linguistics, 11:1250-1264.",
+            "Julio Cesar Salinas Alvarado, Karin Verspoor, and Timothy Baldwin. 2015. Domain adaption of named entity recognition to support credit risk assessment. In Proceedings of the Australasian Language Technology Association Workshop 2015, pages 84-90, Parramatta, Australia.",
+            "Agam Shah, Ruchit Vithani, Abhinav Gullapalli, and Sudheer Chava. 2023. Finer: Financial named entity recognition dataset and weak-supervision model.",
+            "Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. 2024. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations.",
+            "Yejun Soun, Jaemin Yoo, Minyong Cho, Jihyeong Jeon, and U Kang. 2022. Accurate stock movement prediction with self-supervised learning from sparse noisy tweets. In 2022 IEEE International Conference on Big Data (Big Data), pages 1691-1700.",
+            "Liyan Tang, Philippe Laban, and Greg Durrett. 2024. Minicheck: Efficient fact-checking of llms on grounding documents.",
+            "Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Leonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "14749",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Machel Reid, Maciej Mikuţa, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. 2024. Gemma: Open models based on gemini research and technology.",
+        "bbox": [
+            132,
+            85,
+            489,
+            502
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "InternLM Team. 2024. Internlm2 technical report.",
+        "bbox": [
+            114,
+            514,
+            457,
+            529
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. 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 809-819, New Orleans, Louisiana. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            541,
+            489,
+            659
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinez, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan",
+        "bbox": [
+            114,
+            671,
+            489,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023a. Llama 2: Open foundation and finetuned chat models.",
+        "bbox": [
+            526,
+            85,
+            884,
+            137
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hugo Touvron, Louis Martin, Kevin R. Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M. Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony S. Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel M. Kloumann, A. V. Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, R. Subramanian, Xia Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zhengxu Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open foundation and fine-tuned chat models.",
+        "bbox": [
+            510,
+            148,
+            884,
+            449
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Andreas Vlachos and Sebastian Riedel. 2014. Fact checking: Task definition and dataset construction. In Proceedings of the ACL 2014 Workshop on Language Technologies and Computational Social Science, pages 18-22, Baltimore, MD, USA. Association for Computational Linguistics.",
+        "bbox": [
+            510,
+            458,
+            884,
+            539
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or fiction: Verifying scientific claims. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7534-7550, Online. Association for Computational Linguistics.",
+        "bbox": [
+            510,
+            546,
+            882,
+            640
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "David Wadden, Kyle Lo, Bailey Kuehl, Arman Cohan, Iz Beltagy, Lucy Lu Wang, and Hannaneh Hajishirzi. 2022. SciFact-open: Towards open-domain scientific claim verification. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 4719-4734, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+        "bbox": [
+            510,
+            650,
+            882,
+            741
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems.",
+        "bbox": [
+            510,
+            752,
+            884,
+            819
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Huizhe Wu, Wei Zhang, Weiwei Shen, and Jun Wang. 2018. Hybrid deep sequential modeling for social text-driven stock prediction. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM '18, page 1627-1630, New York, NY, USA. Association for Computing Machinery.",
+        "bbox": [
+            510,
+            827,
+            882,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "14750",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. 2023. Bloomberggpt: A large language model for finance. ArXiv, abs/2303.17564.",
+            "Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, and Jimin Huang. 2023. Pixiu: A large language model, instruction data and evaluation benchmark for finance.",
+            "Qianqian Xie, Dong Li, Mengxi Xiao, Zihao Jiang, Ruoyu Xiang, Xiao Zhang, Zhengyu Chen, Yueru He, Weiguang Han, Yuzhe Yang, Shunian Chen, Yifei Zhang, Lihang Shen, Daniel Kim, Zhiwei Liu, Zheheng Luo, Yangyang Yu, Yupeng Cao, Zhiyang Deng, Zhiyuan Yao, Haohang Li, Duanyu Feng, Yongfu Dai, VijayaSai Somasundaram, Peng Lu, Yilun Zhao, Yitao Long, Guojun Xiong, Kaleb Smith, Honghai Yu, Yanzhao Lai, Min Peng, Jianyun Nie, Jordan W. Suchow, Xiao-Yang Liu, Benyou Wang, Alejandro Lopez-Lira, Jimin Huang, and Sophia Ananiadou. 2024. Open-finllms: Open multimodal large language models for financial applications.",
+            "Yumo Xu and Shay B. Cohen. 2018. Stock movement prediction from tweets and historical prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1970-1979, Melbourne, Australia. Association for Computational Linguistics.",
+            "Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. 2023. Fingpt: Open-source financial large language models. FinLLM Symposium at IJCAI 2023.",
+            "Wenpeng Yin, Dragomir Radev, and Caiming Xiong. 2021. DocNLI: A large-scale dataset for document-level natural language inference. In *Findings of the Association for Computational Linguistics: ACLIJCNLP* 2021, pages 4913-4922, Online. Association for Computational Linguistics.",
+            "Yilun Zhao, Yunxiang Li, Chenying Li, and Rui Zhang. 2022. MultiHiertt: Numerical reasoning over multi hierarchical tabular and textual data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588-6600, Dublin, Ireland. Association for Computational Linguistics.",
+            "Yilun Zhao, Hongjun Liu, Yitao Long, Rui Zhang, Chen Zhao, and Arman Cohan. 2024a. Financemath: Knowledge-intensive math reasoning in finance domains. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12841-12858, Bangkok, Thailand. Association for Computational Linguistics.",
+            "Yilun Zhao, Yitao Long, Hongjun Liu, Ryo Kamoi, Linyong Nan, Lyuhao Chen, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. 2024b. DocMath-eval: Evaluating math reasoning capabilities of LLMs in understanding long and specialized"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "documents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16103-16120, Bangkok, Thailand. Association for Computational Linguistics.",
+            "Yilun Zhao, Yitao Long, Hongjun Liu, Linyong Nan, Lyuhao Chen, Ryo Kamoi, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. 2023a. Docmath-eval: Evaluating numerical reasoning capabilities of llms in understanding long documents with tabular data.",
+            "Yilun Zhao, Zhenting Qi, Linyong Nan, Boyu Mi, Yixin Liu, Weijin Zou, Simeng Han, Ruizhe Chen, Xiangru Tang, Yumo Xu, Dragomir Radev, and Arman Cohan. 2023b. QTSumm: Query-focused summarization over tabular data. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1157-1172, Singapore. Association for Computational Linguistics.",
+            "Yilun Zhao, Haowei Zhang, Shengyun Si, Linyong Nan, Xiangru Tang, and Arman Cohan. 2023c. Investigating table-to-text generation capabilities of large language models in real-world information seeking scenarios. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 160-175, Singapore. Association for Computational Linguistics.",
+            "Yilun Zhao, Chen Zhao, Linyong Nan, Zhenting Qi, Wenlin Zhang, Xiangru Tang, Boyu Mi, and Dragomir Radev. 2023d. RobuT: A systematic study of table QA robustness against human-annotated adversarial perturbations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6064–6081, Toronto, Canada. Association for Computational Linguistics.",
+            "Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track.",
+            "Zhihan Zhou, Liqian Ma, and Han Liu. 2021. Trade the event: Corporate events detection for news-based event-driven trading. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2114-2124, Online. Association for Computational Linguistics.",
+            "Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and TatSeng Chua. 2021. TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance. 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 3277-3287, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            910
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "14751",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/01cc874d54df8b9951fd1cbcf5be31cbbbd1e64026fed6a0dc5b10c1e1300347.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>ID</td><td>Finance Industry Experience</td><td>English Proficiency</td><td>Annotation Sets</td><td>Evaluator?</td></tr><tr><td>1</td><td>1 working and 1 internship at US</td><td>Native speaker</td><td>FDV-KNOW</td><td>✓</td></tr><tr><td>2</td><td>&gt;= 2 internship at US</td><td>&gt; 15 years</td><td>FDV-MATH</td><td>✓</td></tr><tr><td>3</td><td>1 working at Singapore and 2 internship at US</td><td>Native speaker</td><td>FDV-KNOW</td><td>✓</td></tr><tr><td>4</td><td>2 working and &gt;= 1 internship at US</td><td>Native speaker</td><td>FDV-KNOW</td><td>✗</td></tr><tr><td>5</td><td>1 internship at US, 2 internship at China</td><td>10 years</td><td>FDV-IE</td><td>✗</td></tr><tr><td>6</td><td>1 internships at HK, China</td><td>15 years</td><td>FDV-IE, FDV-MATH</td><td>✓</td></tr><tr><td>7</td><td>1 internships at China</td><td>10 years</td><td>FDV-IE, FDV-MATH</td><td>✗</td></tr></table>",
+        "bbox": [
+            115,
+            146,
+            878,
+            265
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/5ff4777ff2de5d6b4d65b65fd3c33352a452021782e44af3238913af5a123ed2.jpg",
+        "table_caption": [
+            "Table 7: Details of annotators involved in dataset construction. FINDVER is annotated by financial professionals with extensive experience in comprehending financial documents, ensuring it accurately reflects the real-world challenges in the financial domain."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Organization</td><td>Release Time</td><td>Max Length</td><td>Source</td></tr><tr><td>GPT-4o (OpenAI, 2023a)</td><td>OpenAI</td><td>2023-03</td><td>128k</td><td>https://platform.openai.com/</td></tr><tr><td>GPT-3.5-turbo (OpenAI, 2022)</td><td>OpenAI</td><td>2022-11</td><td>16k</td><td>https://platform.openai.com/</td></tr><tr><td>Gemini-1.5-* (Gemini, 2024)</td><td>Google</td><td>2024-02</td><td>128k</td><td>https://ai.google.dev/</td></tr><tr><td>Claude-3.5 (Anthropic, 2024)</td><td>Anthropic</td><td>2024-03</td><td>200k</td><td>https://www.anthropic.com/api</td></tr><tr><td>Gemma (Team et al., 2024)</td><td>Google</td><td>2024-02</td><td>8k</td><td>google/gemma-7b-it</td></tr><tr><td>Llama-2 (Touvron et al., 2023a)</td><td>Meta</td><td>2023-02</td><td>4k</td><td>meta-llama/Llama-2-*-chat-hf</td></tr><tr><td>Llama-3 (Meta, 2024)</td><td>Meta</td><td>2024-04</td><td>8k</td><td>meta-llama/Meta-Llama-3-*-Instruct</td></tr><tr><td>Yi-1.5 (AI et al., 2024)</td><td>01-ai</td><td>2024-05</td><td>32k</td><td>01-ai/Yi-1.5-*-Chat</td></tr><tr><td>Qwen-2 (qwe, 2024)</td><td>Qwen</td><td>2024-06</td><td>128k</td><td>Qwen/Qwen2-*-Instruct</td></tr><tr><td>Mistral (Jiang et al., 2024, 2023a)</td><td>Mistral AI</td><td>2024-05</td><td>32k</td><td>mistralai/Mistral-7B-Instruct-v0.3</td></tr><tr><td>Mixtral (Jiang et al., 2024, 2023a)</td><td>Mistral AI</td><td>2024-04</td><td>64k</td><td>mistralai/Mixtral-8x22B-v0.1</td></tr><tr><td>InternLM2 (Team, 2024)</td><td>internlm</td><td>2024-01</td><td>200k</td><td>internlm/internlm2-chat-*</td></tr><tr><td>GLM (Du et al., 2022)</td><td>THUDM</td><td>2024-06</td><td>128k</td><td>THUDM/glm-4-9b-chat</td></tr><tr><td>Phi-3 (Abdin et al., 2024)</td><td>microsoft</td><td>2024-04</td><td>128k</td><td>microsoft/Phi-3-*-instruct</td></tr></table>",
+        "bbox": [
+            137,
+            455,
+            858,
+            813
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Details of the organization, release time, maximum context length, and model source (i.e., url for proprietary models and Huggingface model name for open-source models) for the LLMs evaluated in FINDVER.",
+        "bbox": [
+            112,
+            822,
+            882,
+            851
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "14752",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2024/FinDVer_ Explainable Claim Verification over Long and Hybrid-content Financial Documents/3adc16a8-3414-4d00-95ba-3c3dbbf67df2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb9e0f89ba76cbe7769425a202ec899087b9b89e6d4db749cbd4a831ba7eb38d
+size 1159137

2024/FinDVer_ Explainable Claim Verification over Long and Hybrid-content Financial Documents/full.md

@@ -0,0 +1,340 @@
+# FINDVER: Explainable Claim Verification over Long and Hybrid-Content Financial Documents
+
+Yilun Zhao Yitao Long Yuru Jiang Chengye Wang Weiyuan Chen Hongjun Liu Yiming Zhang Xiangru Tang Chen Zhao Arman Cohan
+
+Yale NLP - FINDVER Team
+
+# Abstract
+
+We introduce FINDVER, a comprehensive benchmark specifically designed to evaluate the explainable claim verification capabilities of LLMs in the context of understanding and analyzing long, hybrid-content financial documents. FINDVER is divided into three subsets: information extraction, numerical reasoning, and knowledge-intensive reasoning—each addressing common scenarios encountered in real-world financial contexts. We assess a broad spectrum of LLMs under long-context and RAG settings. Our results show that even the current best-performing system, Claude-3.5-Sonnet, significantly lags behind human experts. Our detailed findings and insights highlight the strengths and limitations of existing LLMs in this new task. We believe FINDVER can serve as a valuable benchmark for evaluating LLM capabilities in claim verification over complex, expert-domain documents.
+
+$\mathbb{O}$  github.com/yilunzhao/FinDVer
+
+# 1 Introduction
+
+In today's information explosion era, the responsibility of verifying the truthfulness of the item is often passed on to the audience. unverified claims about a company's financial performance frequently circulate in online media, potentially misleading investors. Therefore, it is crucial to verify these claims using the companies' original financial documents (i.e., earnings reports and regulatory filings). Recent advancements in Large Language Models (LLMs) have attracted significant attention due to their capabilities in solving a broad range of tasks (Touvron et al., 2023b; Jiang et al., 2023b; OpenAI, 2023a). However, it remains particularly difficult for applying them to document-grounded claim verification in real-world financial domains due to the following two reasons:
+
+First, financial documents are typically long, intricate and dense, and they include both quantita
+
+![](images/c7ea5953a21de17fd7ca8addbe3db246487ff65e8d9360c84f983e220cf93462.jpg)  
+Figure 1: An example from the numerical reasoning subset of the FINDVER benchmark. To verify the claim, the LLM is required to first locate claim-relevant data points within long and hybrid-content financial documents, and then apply numerical reasoning over the extracted data points for claim verification.
+
+tive tables and qualitative text (Chen et al., 2021; Zhu et al., 2021; Zhao et al., 2022, 2023d; Koncel-Kedziorski et al., 2024). Extracting and analyzing claim-relevant data from these documents requires complicated document comprehension abilities and professional knowledge in financial domains. Moreover, the type of reasoning involved encompasses various unique aspects that are less studied, necessitating a dedicated approach to evaluation and application.
+
+Second, in the financial domain, where decisions often involve significant stakes, it is often critical to provide clear and comprehensible rationales for any claim verification decisions (Atanasova et al., 2020, 2023). However, existing context-grounded claim verification benchmarks (Chen et al., 2020; Kamoi et al., 2023; Lu et al., 2023; Glockner et al., 2024) primarily focus on the task of entailment classification and do not evaluate the reasoning process. This hinders the practical application and evaluation of LLMs in real-world scenarios.
+
+In response to the aforementioned pressing need,
+
+<table><tr><td>Dataset</td><td>Input Context</td><td>Annotation / Data Creation</td><td># Label</td><td>w. Expla-
+nation?</td><td>Reasoning-
+Intensive?</td></tr><tr><td>PubMedTab (Akhtar et al., 2022)</td><td>Wikipedia table</td><td>Crowd-sourced</td><td>4</td><td>X</td><td>X</td></tr><tr><td>TABFACT (Chen et al., 2020)</td><td>Wikipedia table</td><td>Crowd-sourced</td><td>2</td><td>X</td><td>✓</td></tr><tr><td>INFOTABS (Gupta et al., 2020)</td><td>Wikipedia table</td><td>Crowd-sourced</td><td>3</td><td>X</td><td>✓</td></tr><tr><td>SCITAB (Lu et al., 2023)</td><td>Scientific table</td><td>Expert &amp; InstructGPT</td><td>3</td><td>X</td><td>✓</td></tr><tr><td>HOVER (Jiang et al., 2020)</td><td>Wikipedia articles</td><td>Crowd-sourced</td><td>2</td><td>X</td><td>X</td></tr><tr><td>DOCNLI (Yin et al., 2021)</td><td>News article</td><td>From summarization datasets</td><td>2</td><td>X</td><td>X</td></tr><tr><td>ContractNLI (Koreeda et al., 2021)</td><td>Contract</td><td>Expert &amp; Crowd-sourced</td><td>2</td><td>X</td><td>X</td></tr><tr><td>LLM-AGGREFACT (Tang et al., 2024)</td><td>Doc from various domains</td><td>From existing benchmarks</td><td>2</td><td>X</td><td>X</td></tr><tr><td>WICE (Kamoi et al., 2023)</td><td>Wikipedia article</td><td>Crowd-sourced</td><td>3</td><td>X</td><td>X</td></tr><tr><td>AMBIFC (Glockner et al., 2024)</td><td>Wikipedia article</td><td>Crowd-sourced</td><td>3</td><td>X</td><td>X</td></tr><tr><td>CLAIMDECOMP (Chen et al., 2022a)</td><td>Political article</td><td>Expert</td><td>6</td><td>X</td><td>X</td></tr><tr><td>SCIFACT (Wadden et al., 2020)</td><td>Scientific paper abstracts</td><td>Expert</td><td>2</td><td>X</td><td>✓</td></tr><tr><td>LIAR++ (Russo et al., 2023)</td><td>Political article</td><td>From fact-check website</td><td>2</td><td>✓</td><td>X</td></tr><tr><td>FullFact (Russo et al., 2023)</td><td>Web page</td><td>From fact-check website</td><td>2</td><td>✓</td><td>X</td></tr><tr><td>PUBHEALTH (Akhtar et al., 2022)</td><td>Health Web page</td><td>From fact check website</td><td>4</td><td>✓</td><td>X</td></tr><tr><td>FINDVER (ours)</td><td>Long financial doc with tables</td><td>Expert</td><td>2</td><td>✓</td><td>✓</td></tr></table>
+
+Table 1: Comparison between FINDVER and existing context-grounded claim verification datasets. FINDVER is distinguished by four unique characteristics: (1) Expert Annotation: It is annotated by financial experts to ensure high data quality; (2) Complex Document Comprehension: It requires interpreting a mix of textual and tabular data within a long-context financial document; (3) Examination on Reasoning-Process Explanation: It enhances claim verifications with detailed explanations about the reasoning process, increasing the benchmark's practical value; and (4) Diverse Reasoning for Real-world Scenarios: It incorporates various reasoning challenges, such as extracting complicated information, performing numerical calculations, applying external professional knowledge, and conducting comparative analyses. Accordingly, we divide the benchmark into three focused subsets, each tailored to mirror distinct real-world financial analysis scenarios.
+
+we present FINDVER, a comprehensive and domain expert-annotated explainable claim verification benchmark that first explores in the context of financial documents. The LLMs are tasked with generating explanations of their reasoning to verify claims labeled as "entailed" or "refuted", based on the information in the provided document, which contains both textual and tabular data. To identify the common reasoning-intensive scenarios in claim verification based on financial documents, we engage with domain experts and conducted a preliminary study. This helped us determine three key types of scenarios that frequently arise in real-world settings: information extraction, numerical reasoning, and knowledge-intensive reasoning. For each scenario, we construct an evaluation set. Each example in our dataset is annotated with detailed supporting evidence and step-by-step reasoning-process explanations.
+
+We evaluate a wide spectrum of open- and closed-source LLMs, specifically, 19 models from 10 organizations. The documents in our benchmark are exceedingly long; therefore, we employ two widely adopted real-world application settings—retrieval augmented generation (RAG) and long-context—in this study. The experiment
+
+tal results indicate that even the existing best-performing LLM (i.e., Claude-3.5-Sonnet) still significantly lags behind human experts  $(77.2\%$  versus  $93.3\%)$ , demonstrating the challenges of our proposed benchmark. Our contributions are summarized below:
+
+- We introduce FINDVER, the first comprehensive context-grounded claim verification benchmark for financial domains, presenting new challenges for state-of-the-art LLMs.  
+- We conduct an extensive evaluation that encompasses a wide range of LLMs, including those specialized in finance and math. We also evaluate both long-context and RAG settings to comprehensively assess the capabilities and limitations of existing LLMs in our task.  
+- Our experimental results reveal a noticeable performance gap compared to human experts. This highlights the limitations of current LLMs in complex real-world applications and the need for continued advancements.
+
+# 2 Related Work
+
+Claim Verification Benchmark Claim verification is a well-established research area with two
+
+main settings. The first is the open-domain setting, which involves using an external retriever to find the most relevant information from a large corpus to verify claims (Vlachos and Riedel, 2014; Thorne et al., 2018; Aly et al., 2021; Wadden et al., 2022; Rangapur et al., 2024; Ma et al., 2024). The second setting is context-grounded claim verification, which requires models to verify claims based on the provided document context (Chen et al., 2020; Kamoi et al., 2023; Lu et al., 2023; Glockner et al., 2024). This work focuses on the second setting, as it allows us to eliminate variability and dependency on the retriever's performance, thereby focusing on the evaluation of LLM performance on an accurately verifying claims within a given context. However, as illustrated in Table 1, existing context-grounded claim verification benchmarks have four notable limitations: they typically 1) focus on general domains, overlooking the specific challenges and intricacies present in specialized fields, 2) focus solely on entailment classification and do not evaluate the reasoning processes of models, 3) do not involve claims that require intensive reasoning and complicated document comprehension. These limitations hinder their effectiveness for evaluating LLMs in real-world practice.
+
+Financial Evaluation Benchmark NLP techniques have been applied to various financial tasks, such as named entity recognition (Salinas Alvarado et al., 2015; Shah et al., 2023), sentiment analysis (Malo et al., 2013; Maia et al., 2018), stock movement prediction (Soun et al., 2022; Xu and Cohen, 2018; Wu et al., 2018), and summarization (Zhou et al., 2021; Mukherjee et al., 2022; Liu et al., 2022). More recently, there has been an increasing focus on tasks involving financial documents (e.g., annual reports and regulatory filings), which are crucial for providing insights into a company's performance and strategies. Several QA benchmarks have been proposed to evaluate models' performance in answering questions based on financial documents, with a particular focus on numerical reasoning (Chen et al., 2021; Zhu et al., 2021; Zhao et al., 2022; Chen et al., 2022b; Koncel-Kedziorski et al., 2024; Zhao et al., 2024b,a). Despite these advancements, there remains a significant gap in the exploration of claim verification tasks within the financial domain. While the recent FIN-FACT benchmark (Rangapur et al., 2024) addresses explainable multimodal financial fact-checking, it primarily focuses on open-domain sce
+
+narios. Verifying claims derived from financial documents is crucial, as inaccuracies can significantly influence investment decisions and market perceptions. To bridge this gap, we introduce FINDVER, the first context-grounded claim verification benchmark, specifically designed for real-world financial document comprehension.
+
+# 3 FINDVER Benchmark
+
+FINDVER provides a robust evaluation benchmark for reasoning-intensive and explainable claim verification over long and hybrid-content financial documents. We present an overview of the FINDVER construction pipeline in Figure 2; and detail the task formulation, data construction, and quality validation process in the following subsections.
+
+# 3.1 Task Formulation
+
+We formally define the task of FINDVER within the context of LLMs as follows: Consider a single financial document  $d$ , containing textual data  $P$  and tabular data  $T$ , associated with a claim  $c$  that needs verification. The expert-annotated data we collect supports the following two tasks:
+
+Entailment Classification The model is required to determine the entailment label  $\ell \in \mathcal{L} = \{\text{"entailed", "refuted"}\}$ , based on the document context:
+
+$$
+\ell = \arg \max  _ {\ell \in \mathcal {L}} P _ {\mathbf {L L M}} (\ell \mid P, T, c) \tag {1}
+$$
+
+Reasoning-process Explanation Generation The model is required to generate a natural language explanation  $e$ :
+
+$$
+e = \arg \max _ {e} P _ {\mathbf {L L M}} (e \mid P, T, c) \tag {2}
+$$
+
+which articulates the reasoning process behind the validity of the provided claim  $c$ , based solely on the provided textual content  $P$  and tabular content  $T$  within the financial document.
+
+Notably, some claim verification systems, particularly those developed prior to the era of LLMs and for previous datasets that did not require explanation generation (Chen et al., 2020; Yin et al., 2021; Koreeda et al., 2021), might not explicitly perform explanation generation. Instead, they directly output the final label. For such systems, FINDVER can also be used for evaluation by focusing on the entailment classification task.
+
+![](images/1d6122a63969cf6dca72bcb4870ef402a5118c041f589dfcaa5c2ba513ec29da.jpg)  
+Figure 2: An overview of FINDVER construction pipeline.
+
+# 3.2 FINDVER Subset Design
+
+FINDVER is designed to mirror the real-world challenges encountered in the financial domain. Therefore, we ensure that the included annotators are financial experts with professional experience in comprehending and processing financial documents. Table 7 in Appendix presents the detailed annotator biographies for FINDVER annotation.
+
+To identify the common reasoning-intensive scenarios in claim verification based on financial documents, we engaged with domain experts and conducted a preliminary study. This helped us determine three key types of scenarios that frequently arise in real-world settings. Accordingly, we have created three corresponding subsets of FINDVER.
+
+(1) FDV-IE (information extraction), which involves extracting information from both textual and tabular content within a long-context document.  
+(2) FDV-MATH (numerical reasoning), which necessitates performing calculations or statistical analysis based on data within the document.  
+(3) FDV-Know (knowledge-intensive reasoning), which requires integrating external domain-specific knowledge or regulations for claim verification.
+
+# 3.3 Source Document Collection
+
+Similar to Zhao et al. (2023a), we use the quarterly (Form 10-Q) and annual reports (Form 10-K) of companies as the source documents, which are publicly available in the open-source database<sup>1</sup> of the U.S. Securities and Exchange Commission. We collect a total of 523 documents that were first released between January 1 to April 30, 2024, which is after the cutoff date of most pretraining corpora for training foundation models. This helps to alleviate issues related to data memorization to some extent. After collecting the raw HTML-format documents,
+
+we utilize the SEC  $\mathrm{API}^2$ , a commercial platform API for extracting financial document content, to process the collected documents, obtaining documents with both textual and tabular data.
+
+# 3.4 Claim Annotation
+
+Entailed Claim Annotation To address the potential bias concerning the position of evidence within the documents, we initiate the process by randomly sampling multiple document contexts from the given document. Annotators are then tasked with creating "entailed" claims based on the textual and tabular data within these contexts. The annotators are instructed to simulate real-world document comprehension scenarios, ensuring the annotated claims are representative of practical financial analysis and align with the scenarios defined by the corresponding subsets. Annotators are then tasked with carefully locating all evidence (i.e., indices of relevant paragraphs and tables) within the entire document that support the claims, which are used for the subsequent data validation.
+
+Refuted Claim Annotation Following established practices in the field (Wadden et al., 2020; Chen et al., 2020; Lu et al., 2023), and since directly obtaining "refuted" types is difficult, we instead perturb the original "entailed" claims into "refuted" claim through expert annotation. Specifically, expert annotators first create an "entailed" claim using the same procedure detailed in the "Entailed Claim Annotation" paragraph. The annotators are then instructed to perturb the "entailed" claim to introduce factual errors that are directly contradicted by the annotated evidence, and rewrite the annotated reasoning-process explanation.
+
+<table><tr><td>Annotation Quality</td><td>%S ≥ 4</td></tr><tr><td>Claim</td><td></td></tr><tr><td>Fluency</td><td>92</td></tr><tr><td>Meaningfulness</td><td>90</td></tr><tr><td>Alignment with real-world scenarios</td><td>94</td></tr><tr><td>Evidence</td><td></td></tr><tr><td>Relevancy</td><td>89</td></tr><tr><td>Completeness</td><td>85</td></tr><tr><td>Reasoning-process Explanation</td><td></td></tr><tr><td>Fluency</td><td>95</td></tr><tr><td>Correctness</td><td>92</td></tr><tr><td>Comprehensiveness</td><td>90</td></tr><tr><td>Entailment Label</td><td></td></tr><tr><td>Correctness</td><td>94</td></tr></table>
+
+Table 2: Human evaluation over 100 samples from the FINDVER testmini set. Two internal evaluators were asked to rate the samples on a scale of 1 to 5 individually. We report percent of samples that have an average score  $\geq 4$  to indicate the annotation quality of FINDVER.
+
+# 3.5 Explanation Annotation
+
+After finishing the claim annotation, we pass it to another annotator for explanation annotation. The annotators are required to first read the claim carefully and annotate a detailed explanation of the reasoning process. Such reasoning-process explanations allow for a granular and informative evaluation of model outputs, helping future work identify reasoning errors and provide more accurate error feedback. We compare the entailment label annotated in this step with those in the claim annotation step. A third annotator is introduced if the two annotation versions are different. In practice, we achieve an inter-annotator agreement of  $90.3\%$  for entailment label annotation.
+
+During our pilot annotation phase, we observed variability in the format of reasoning-process explanation annotated by different annotators, which made the dataset less standardized. To ensure consistency and clarity in our benchmark, we developed a predefined template for annotators to follow. Specifically, annotators are required to commence with the extraction of relevant information phase, where they need to list all claim-relevant information in a numbered list. Subsequently, they are required to annotate the reasoning over the extracted information segment in a step-by-step manner. For each step, they should elucidate the associated reasoning. Finally, they annotate the entailment label feature.
+
+# 3.6 Data Quality Validation
+
+To ensure the high quality of our annotated data, for every annotated example, a qualified annotator is assigned to validate several key aspects: (1) the claim and reasoning-process explanation should be grammatically correct and free of spelling errors; (2) the claim should be closely related to financial domains and meaningful in real-world scenarios; (3) the annotated evidence should be relevant to the claim and complete enough to verify it; (4) the entailment label of the claim should be supported by the annotated evidence; and (5) the reasoning-process explanation should correctly interpret the extracted evidence and apply appropriate reasoning steps to correctly verify the claim. The validators are asked to revise examples that do not meet these standards. In practice, 347 out of 2,100 initial examples were revised by the validators. We also report the human evaluation scores over 100 sampled examples. As illustrated in Table 2, FINDVER has a high annotation quality.
+
+# 3.7 Dataset Preparation and Release
+
+Table 3 provides an overview of the key statistics for our benchmark. The dataset is randomly split into two subsets: testmini and test. The testmini set is intended for model development and validation. It contains 600 examples, with 200 examples from each subset. The test set comprises the remaining 1,500 examples and is designed for standard evaluation. To prevent data contamination (Jacovi et al., 2023; Shi et al., 2024; Deng et al., 2024), the ground-truth-related annotation features for the test set will not be publicly released. Instead, we provide an online evaluation platform where researchers can assess their models and participate in a public leaderboard.
+
+# 4 Experiment Setup
+
+We next present the experimental setup, covering the evaluated LLMs, long-context and RAG setups, implementation details, and the measurement of human-level performance.
+
+# 4.1 Experimental LLMs
+
+We examine the performance of LLMs across two distinct categories on FINDVER: (1) Proprietary LLMs, including GPT-4* (OpenAI, 2023a,b, 2024), Gemini-1.5-* (Gemini, 2024), and Claude-3 (Anthropic, 2024); and (2) Open-source LLMs, including Gemma (Team et al., 2024), Llama
+
+<table><tr><td>Property</td><td>FDV-IE</td><td>FDV-MATH</td><td>FDV-Know</td></tr><tr><td>Real-world scenarios in financial domains</td><td>information extraction</td><td>numerical reasoning</td><td>knowledge-intensive reasoning</td></tr><tr><td># Document</td><td>221</td><td>225</td><td>217</td></tr><tr><td>Doc Length (i.e., word count) (Median/Avg/Max)</td><td>42K / 41K / 71K</td><td>43K / 41K / 71K</td><td>43K / 41K / 71K</td></tr><tr><td># Tables per document (Median/Avg)</td><td>62 / 78.9</td><td>62 / 79.1</td><td>62 / 79.0</td></tr><tr><td>Claim length (Median/Avg)</td><td>47 / 47.2</td><td>24 / 25.1</td><td>36 / 37.1</td></tr><tr><td># Text evidence per claim (Median/Avg)</td><td>2 / 1.8</td><td>1 / 1.3</td><td>3 / 2.6</td></tr><tr><td># Table evidence per claim (Median/Avg)</td><td>1 / 1.0</td><td>1 / 0.9</td><td>1 / 1.2</td></tr><tr><td>% Claims w. table evidence</td><td>66.3%</td><td>71.1%</td><td>70.8%</td></tr><tr><td>Explanation length (Median/Avg)</td><td>70 / 73.1</td><td>74 / 76.2</td><td>96 / 100.7</td></tr><tr><td>Benchmark size (# Claims)</td><td></td><td></td><td></td></tr><tr><td>testmini size</td><td>200</td><td>200</td><td>200</td></tr><tr><td>test size</td><td>500</td><td>500</td><td>500</td></tr></table>
+
+Table 3: Basic statistics of the FINDVER benchmark.
+
+# Adopted Chain-of-Thought Prompt
+
+# [System Input]
+
+As a financial expert, your task is to assess the truthfulness of the given claim by determining whether it is entailed or refuted based on the provided financial document. Follow these steps:
+
+1. Carefully read the given context and the claim.  
+2. Analyze the document, focusing on the relevant financial data or facts that related to the claim.  
+3. Document each step of your reasoning process to ensure your assessment is clear and thorough.  
+4. Conclude your analysis with a final determination. In your last sentence, clearly state your conclusion in the following format: "Therefore, the claim is {entailment_label}." Replace {entailment_label} with either 'entailed' (if the claim is supported by the document) or 'refuted' (if the claim contradicts or partially contradicts the document).
+
+# [User Input]
+
+Financial Report:
+
+{Financial Report}
+
+Claim to verify:
+
+{Claim}
+
+Follow the instructions and think step by step to verify the claim.
+
+Figure 3: The Chain-of-Thought prompt used.
+
+2&3 (Touvron et al., 2023a; Meta, 2024), Yi1.5 (AI et al., 2024), Qwen-2 (qwe, 2024), Mistral & Mixtral (Jiang et al., 2023a, 2024), InternLM2 (Team, 2024), C4AI (Aryabumi et al., 2024), GLM (Du et al., 2022), and Phi-3 (Abdin et al., 2024). Table 8 in Appendix presents the details of evaluated models (i.e., organizations, release time, max context length, and model version).
+
+The experiments for open-sourced LLMs were
+
+conducted using the vLMM framework (Kwon et al., 2023). For all the experiments, we set temperature as 1.0 and maximum output length as 512. We adopt the Chain-of-Thought (CoT) prompting methods (Wei et al., 2022) for the FINDVER benchmark. Specifically, the model is instructed to first output a detailed reasoning process for verifying claims, and then provide the entailment label of the claim based on the generated reasoning process. Figure 3 presents the used prompts.
+
+# 4.2 Long-Context and RAG Settings
+
+As presented in Table 3, the documents within our benchmark are notably lengthy. To effectively handle this, we have implemented two real-world application settings that are widely recognized for their utility in dealing with extensive texts. For Long-context Setting, we input the entire financial document into the model. We limit our evaluation to those models that have a context window larger than 100,000 tokens, which exceeds the maximum length of the included financial document. For RAG Setting, we leverage the current best-performing embedding models (i.e., OpenAI's text-embedding-3-large) to retrieve the top-10 paragraphs or tables that are most relevant to the claims. These elements are then concatenated in their original order as found in the document before being fed into the model.
+
+# 4.3 Implementation Details
+
+Input Tabular Data Compilation Building on previous research that assessed LLMs on tasks involving tabular data (Chen, 2023; Zhao et al., 2023b,c), we introduce our methodology for processing tables within documents. Our approach
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Notes</td><td colspan="2">FDV-IE</td><td colspan="2">FDV-MATH</td><td colspan="2">FDV-Know</td><td colspan="2">Average</td></tr><tr><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td></tr><tr><td>Random Choice</td><td></td><td>50.0</td><td></td><td>50.0</td><td></td><td>50.0</td><td></td><td>50.0</td><td></td></tr><tr><td>Human Non-Expert</td><td></td><td>90.0</td><td></td><td>85.0</td><td></td><td>85.0</td><td></td><td>86.7</td><td></td></tr><tr><td>Human Expert</td><td></td><td>95.0</td><td></td><td>90.0</td><td></td><td>95.0</td><td></td><td>93.3</td><td></td></tr><tr><td colspan="10">Open-source LLMs</td></tr><tr><td>InternLM2-Math-7b</td><td>Math</td><td>-</td><td>58.0</td><td>-</td><td>53.5</td><td>-</td><td>54.5</td><td>-</td><td>55.3</td></tr><tr><td>InternLM2-7B</td><td></td><td>-</td><td>59.5</td><td>-</td><td>54.5</td><td>-</td><td>56.0</td><td>-</td><td>56.7</td></tr><tr><td>Gemma-7B</td><td></td><td>-</td><td>59.5</td><td>-</td><td>55.5</td><td>-</td><td>55.0</td><td>-</td><td>56.7</td></tr><tr><td>GLM-4-9b</td><td></td><td>58.5</td><td>61.0</td><td>54.5</td><td>54.5</td><td>55.0</td><td>56.5</td><td>56.0</td><td>57.3</td></tr><tr><td>Llama-2-7B</td><td></td><td>-</td><td>60.0</td><td>-</td><td>56.5</td><td>-</td><td>56.5</td><td>-</td><td>57.7</td></tr><tr><td>Mistral-7B-v3</td><td></td><td>-</td><td>59.5</td><td>-</td><td>56.5</td><td>-</td><td>57.0</td><td>-</td><td>57.7</td></tr><tr><td>Phi-3-medium-4k</td><td></td><td>-</td><td>61.5</td><td>-</td><td>54.0</td><td>-</td><td>58.0</td><td>-</td><td>57.8</td></tr><tr><td>Llama-2-70B</td><td></td><td>-</td><td>61.5</td><td>-</td><td>54.5</td><td>-</td><td>58.0</td><td>-</td><td>58.0</td></tr><tr><td>Phi-3-medium-128k</td><td></td><td>58.0</td><td>61.5</td><td>54.0</td><td>55.5</td><td>56.5</td><td>57.5</td><td>56.2</td><td>58.2</td></tr><tr><td>Meta-Llama-3-8B</td><td></td><td>-</td><td>62.5</td><td>-</td><td>55.0</td><td>-</td><td>59.5</td><td>-</td><td>59.0</td></tr><tr><td>Yi-1.5-34B</td><td></td><td>-</td><td>62.5</td><td>-</td><td>58.0</td><td>-</td><td>58.0</td><td>-</td><td>59.5</td></tr><tr><td>Meta-Llama-3-70B</td><td>MoE</td><td>-</td><td>65.5</td><td>-</td><td>61.5</td><td>-</td><td>61.5</td><td>-</td><td>62.8</td></tr><tr><td>C4AI Command R+</td><td></td><td>-</td><td>67.5</td><td>-</td><td>60.0</td><td>-</td><td>64.5</td><td>-</td><td>64.0</td></tr><tr><td>Qwen2-72B</td><td></td><td>67.0</td><td>68.0</td><td>62.5</td><td>61.5</td><td>60.5</td><td>65.0</td><td>63.3</td><td>64.8</td></tr><tr><td>Mixtral-8x22B</td><td></td><td>-</td><td>70.0</td><td>-</td><td>62.0</td><td>-</td><td>67.0</td><td>-</td><td>66.3</td></tr><tr><td colspan="10">Proprietary LLMs</td></tr><tr><td>Gemini-1.5-Flash</td><td></td><td>71.0</td><td>70.5</td><td>62.5</td><td>60.5</td><td>65.0</td><td>65.5</td><td>66.2</td><td>65.5</td></tr><tr><td>GPT-3.5-turbo</td><td></td><td>-</td><td>79.0</td><td>-</td><td>64.0</td><td>-</td><td>70.5</td><td>-</td><td>71.2</td></tr><tr><td>GPT-4o</td><td></td><td>80.0</td><td>78.5</td><td>70.5</td><td>68.0</td><td>76.5</td><td>74.5</td><td>75.7</td><td>73.7</td></tr><tr><td>Claude-3.5-Sonnet</td><td></td><td>83.5</td><td>80.5</td><td>71.0</td><td>69.0</td><td>77.0</td><td>75.5</td><td>77.2</td><td>75.0</td></tr></table>
+
+Table 4: Accuracy of entailment classification on the testmini set of FINDVER. We report results for LLMs with CoT prompting under the long-context (LongC) and RAG settings. Numbers underscored indicate that models under the long-context setting achieve better results than under the RAG setting.
+
+involves distinguishing headers and cells in different columns using a vertical bar (l) and separating rows with new lines. This format allows us to input flattened table data directly into LLMs. In our initial experiments, we found that LLMs such as GPT-* and Llama-3 can effectively interpret this table representation. However, we suggest that future studies should investigate more sophisticated methods for encoding tabular data to enhance comprehension by LLMs.
+
+Model Response Processing Following previous work (Lu et al., 2024), we adopt LLM for processing model response. Specifically, we utilize GPT-4o-mini to extract labels from the LLM output, which can be either "entailed", "refuted" or "none". The "none" label typically indicates that the LLM output contains nonsensical symbols or unintelligible text rather than meaningful content. In cases where the output is labeled as "none", we
+
+assign the final label by making a random guess.
+
+# 4.4 Human-level Performance
+
+To provide a rough but informative estimate of human-level performance by non-experts and experts on FINDVER, we randomly sampled 5 documents  $\times 4$  claims / document  $= 20$  claims from each validation subset, totaling 60 claims. We enroll two experts (i.e., professionals with CFA license) and two non-experts (i.e., undergraduate students majored in computer science) to individually verify the claims by providing the NL explanations. Table 4 presents the human-level performance.
+
+# 5 Experiment Results
+
+# 5.1 Main Findings
+
+Table 4 and Table 5 display the primary results for FINDVER. We reveals a significant accuracy gap between human experts and LLMs. Notably,
+
+<table><tr><td>Model</td><td>IE</td><td>MATH</td><td>KNOW</td><td>Avg</td></tr><tr><td>Human Non-Expert</td><td>90.0</td><td>85.0</td><td>85.0</td><td>86.7</td></tr><tr><td>Human Expert</td><td>95.0</td><td>90.0</td><td>95.0</td><td>93.3</td></tr><tr><td colspan="5">Open-source LLMs</td></tr><tr><td>InternLM2-Math-7B</td><td>57.0</td><td>53.8</td><td>54.6</td><td>55.1</td></tr><tr><td>InternLM2-7B</td><td>59.6</td><td>53.4</td><td>55.4</td><td>56.1</td></tr><tr><td>Gemma-7B</td><td>59.8</td><td>54.4</td><td>55.0</td><td>56.4</td></tr><tr><td>GLM-4-9B</td><td>61.4</td><td>54.2</td><td>57.4</td><td>57.7</td></tr><tr><td>Llama-2-7B</td><td>61.0</td><td>57.2</td><td>57.2</td><td>58.5</td></tr><tr><td>Mistral-7B-v3</td><td>59.8</td><td>56.0</td><td>56.4</td><td>57.4</td></tr><tr><td>Phi-3-medium-4k</td><td>61.8</td><td>54.0</td><td>57.2</td><td>57.7</td></tr><tr><td>Llama-2-70B</td><td>60.6</td><td>53.8</td><td>58.0</td><td>57.5</td></tr><tr><td>Phi-3-medium-128k</td><td>61.2</td><td>55.4</td><td>58.2</td><td>58.3</td></tr><tr><td>Meta-Llama-3-8B</td><td>63.4</td><td>55.0</td><td>60.2</td><td>59.5</td></tr><tr><td>Yi-1.5-34B</td><td>62.8</td><td>57.2</td><td>56.8</td><td>58.9</td></tr><tr><td>Meta-Llama-3-70B</td><td>65.0</td><td>61.2</td><td>60.4</td><td>62.2</td></tr><tr><td>C4AI Command R+</td><td>67.4</td><td>59.0</td><td>65.4</td><td>63.9</td></tr><tr><td>Qwen2-72B</td><td>69.0</td><td>62.2</td><td>65.2</td><td>65.5</td></tr><tr><td>Mixtral-8x22B</td><td>71.0</td><td>62.8</td><td>68.2</td><td>67.3</td></tr><tr><td colspan="5">Proprietary LLMs</td></tr><tr><td>Gemini-1.5-Flash</td><td>71.2</td><td>61.0</td><td>65.8</td><td>66.0</td></tr><tr><td>GPT-3.5-turbo</td><td>79.6</td><td>65.2</td><td>70.8</td><td>71.9</td></tr><tr><td>GPT-4o</td><td>78.6</td><td>69.4</td><td>73.8</td><td>73.9</td></tr><tr><td>Claude-3.5-Sonnet</td><td>81.0</td><td>68.2</td><td>74.0</td><td>74.4</td></tr></table>
+
+Claude-3.5-Sonnet, the highest-performing LLM, achieves an accuracy rate of only  $77.2\%$ , in stark contrast to the  $93.3\%$  accuracy of financial experts. This discrepancy highlights the complexity and challenges of our benchmark.
+
+For the less competitive LLMs, such as Qwen2-72B, GLM-4-9B, and Phi-3-medium-128k, they exhibit improved performance under the RAG setting. In contrast, the currently more competitive LLMs, such as GPT-4o and Claude-3.5-Sonnet, generally perform better under the long-context setting compared to the RAG setting. This indicates the potential of developing long-context techniques to manage tasks involving extensive documents in specialized domains.
+
+# 5.2 Chain-of-Thought Analysis
+
+To better understand the effectiveness of CoT prompting methods for our tasks, we select the commonly-used proprietary and open-source LLMs, GPT-4o and Qwen2-72B, for our experiments. In the w/o CoT setting, we instruct the LLMs to directly output the entailment label of the claim using the provided document context (Figure 4). As illustrated in Table 6, both LLMs' per
+
+![](images/9713aa09a204cbf751b49bf68651efb363464e8122bfeb61414b27d2f3d1ae02.jpg)  
+Figure 4: The Direct Output prompt used.
+
+Table 5: Accuracy of entailment classification on the FINDVER test set. We report results for LLMs with CoT prompting under the RAG setting. Due to computation constraint, we did not evaluate the long-context setting.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">w/o CoT</td><td colspan="2">w/ CoT</td></tr><tr><td>LongC</td><td>RAG</td><td>LongC</td><td>RAG</td></tr><tr><td>Qwen2-72B</td><td>57.7 (-5.6)</td><td>59.5 (-5.3)</td><td>63.3</td><td>64.8</td></tr><tr><td>GPT-4o</td><td>70.1 (-7.1)</td><td>69.5 (-5.5)</td><td>77.2</td><td>75.0</td></tr></table>
+
+Table 6: Accuracy of entailment for GPT-4o and Qwen2-72B with and without CoT Prompting methods on the FINDVER testmini set.
+
+formance degrades in the w/o CoT setting. These results highlight the importance of CoT reasoning in enhancing performance for our task.
+
+# 5.3 Error Analysis of Reasoning Process
+
+The Claude-3.5-Sonnet model achieves a top accuracy of  $77.2\%$  under the long context setting. To better understand the model's limitations, we perform a detailed error analysis with human evaluators. We randomly select 25 instances from each of the three subsets where the Claude-3.5-Sonnet model fails to perform correctly. Our analysis has identified four primary categories of errors: (1) Extraction error: The model fails to correctly retrieve relevant information from the context, resulting in inaccurate verification. (2) Numerical reasoning error: The model encounters difficulties with correct mathematical reasoning. (3) Domain knowledge deficiency: The model lacks sufficient knowledge in finance-related areas, which hampers its ability to reason accurately. (4) Computation error: While the model's reasoning is correct, it makes computational mistakes during intermediate or final steps, resulting in incorrect verification.
+
+# 6 Conclusion
+
+This paper presents FINDVER, a comprehensive benchmark designed to evaluate LLMs in claim verification over long and hybrid-content financial documents. Through extensive experiments involving 19 LLMs under long-context and RAG settings, we have demonstrated that even the top-performing models exhibit a significant performance gap compared to financial experts. Our detailed findings and insights reveal the strengths and limitations of current LLMs in this new task. We believe that FINDVER provides a valuable benchmark for future research on LLMs' ability to handle complex claim verification tasks within the expert domain.
+
+# Limitations
+
+In this work, we propose FINDVER and conduct comprehensive analysis of different LLMs' capabilities on our task. However, there are still some limitations: First, our evaluation does not include recently released finance-specific LLMs (Wu et al., 2023; Yang et al., 2023; Xie et al., 2023, 2024), as these models are not yet compatible with the vLLM framework used for inference. Due to computational resource constraints, we do not tune LLMs on a large-scale finance-domain data ourselves. However, we believe that training on finance data can help improve LLMs' capabilities in FINDVER. Moreover, we only conduct human error analysis on the generated reasoning process of models. We believe future work could explore the development of LLM-based automated evaluation systems (Liu et al., 2023; Zheng et al., 2023) for automatically detecting reasoning errors within the generated explanation.
+
+# References
+
+# 2024. Qwen2 technical report.
+
+Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, Alon Benhaim, Misha Bilenko, Johan Bjorck, Sebastien Bubeck, Martin Cai, Caio Cesar Teodoro Mendes, Weizhu Chen, Vishrav Chaudhary, Parul Chopra, Allie Del Giorno, Gustavo de Rosa, Matthew Dixon, Ronen Eldan, Dan Iter, Amit Garg, Abhishek Goswami, Suriya Gunasekar, Emman Haider, Junheng Hao, Russell J. Hewett, Jamie Huynh, Mojan Javaheripi, Xin Jin, Piero Kauffmann, Nikos Karampatziakis, Dongwoo Kim, Mahoud Khademi, Lev Kurilenko, James R. Lee, Yin Tat Lee, Yuanzhi Li, Chen Liang, Weishung Liu, Eric
+
+Lin, Zeqi Lin, Piyush Madan, Arindam Mitra, Hardik Modi, Anh Nguyen, Brandon Norick, Barun Patra, Daniel Perez-Becker, Thomas Portet, Reid Pryzant, Heyang Qin, Marko Radmilac, Corby Rosset, Sambudha Roy, Olatunj Ruwase, Olli Saarikivi, Amin Saied, Adil Salim, Michael Santacroce, Shital Shah, Ning Shang, Hiteshi Sharma, Xia Song, Masahiro Tanaka, Xin Wang, Rachel Ward, Guanhua Wang, Philipp Witte, Michael Wyatt, Can Xu, Jiahang Xu, Sonali Yadav, Fan Yang, Ziyi Yang, Donghan Yu, Chengruidong Zhang, Cyril Zhang, Jianwen Zhang, Li Lyna Zhang, Yi Zhang, Yue Zhang, Yunan Zhang, and Xiren Zhou. 2024. Phi-3 technical report: A highly capable language model locally on your phone.
+
+01. AI, :, Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, Kaidong Yu, Peng Liu, Qiang Liu, Shawn Yue, Senbin Yang, Shiming Yang, Tao Yu, Wen Xie, Wenhao Huang, Xiaohui Hu, Xiaoyi Ren, Xinyao Niu, Pengcheng Nie, Yuchi Xu, Yudong Liu, Yue Wang, Yuxuan Cai, Zhenyu Gu, Zhiyuan Liu, and Zonghong Dai. 2024. Yi: Open foundation models by 01.ai.
+
+Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. 2022. PubHealthTab: A public health table-based dataset for evidence-based fact checking. In *Findings of the Association for Computational Linguistics: NAACL* 2022, pages 1-16, Seattle, United States. Association for Computational Linguistics.
+
+Rami Aly, Zhijiang Guo, Michael Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. 2021. Feverous: Fact extraction and verification over unstructured and structured information. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1.
+
+Anthropic. 2024. Introducing the next generation of claude.
+
+Viraat Aryabumi, John Dang, Dwarak Talupuru, Saurabh Dash, David Cairuz, Hangyu Lin, Bharat Venkitesh, Madeline Smith, Jon Ander Campos, Yi Chern Tan, Kelly Marchisio, Max Bartolo, Sebastian Ruder, Acyr Locatelli, Julia Kreutzer, Nick Frosst, Aidan Gomez, Phil Blunsom, Marzieh Fadaee, Ahmet Üstün, and Sara Hooker. 2024. Aya 23: Open weight releases to further multilingual progress.
+
+Pepa Atanasova, Oana-Maria Camburu, Christina Lioma, Thomas Lukasiewicz, Jakob Grue Simonsen, and Isabelle Augenstein. 2023. Faithfulness tests for natural language explanations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 283-294, Toronto, Canada. Association for Computational Linguistics.
+
+Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. Generating fact checking explanations. In Proceedings of the 58th
+
+Annual Meeting of the Association for Computational Linguistics, pages 7352-7364, Online. Association for Computational Linguistics.  
+Jifan Chen, Aniruddh Sriram, Eunsol Choi, and Greg Durrett. 2022a. Generating literal and implied subquestions to fact-check complex claims. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3495-3516, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Wenhu Chen. 2023. Large language models are few(1)-shot table reasoners. In Findings of the Association for Computational Linguistics: EACL 2023, pages 1120-1130, Dubrovnik, Croatia. Association for Computational Linguistics.  
+Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2020. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations.  
+Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. 2021. FinQA: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697-3711, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. 2022b. ConvFinQA: Exploring the chain of numerical reasoning in conversational finance question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6279-6292, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Chunyuan Deng, Yilun Zhao, Yuzhao Heng, Yitong Li, Jiannan Cao, Xiangru Tang, and Arman Cohan. 2024. Unveiling the spectrum of data contamination in language model: A survey from detection to remediation. In Findings of the Association for Computational Linguistics ACL 2024, pages 16078-16092, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics.  
+Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. GLM: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, Dublin, Ireland. Association for Computational Linguistics.  
+Gemini. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.  
+Max Glockner, Ieva Staliūnaitè, James Thorne, Gisela Vallejo, Andreas Vlachos, and Iryna Gurevych. 2024.
+
+AmbiFC: Fact-checking ambiguous claims with evidence. Transactions of the Association for Computational Linguistics, 12:1-18.  
+Vivek Gupta, Maitrey Mehta, Pegah Nokhiz, and Vivek Srikumar. 2020. INFOTABS: Inference on tables as semi-structured data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2309–2324, Online. Association for Computational Linguistics.  
+Alon Jacovi, Avi Caciularu, Omer Goldman, and Yoav Goldberg. 2023. Stop uploading test data in plain text: Practical strategies for mitigating data contamination by evaluation benchmarks. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5075-5084, Singapore. Association for Computational Linguistics.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023a. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024. Mixtral of experts.  
+Albert Qiaochu Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L'elio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023b. Mistral 7b. ArXiv, abs/2310.06825.  
+Yichen Jiang, Shikha Bordia, Zheng Zhong, Charles Dognin, Maneesh Singh, and Mohit Bansal. 2020. HoVer: A dataset for many-hop fact extraction and claim verification. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3441-3460, Online. Association for Computational Linguistics.  
+Ryo Kamoi, Tanya Goyal, Juan Diego Rodriguez, and Greg Durrett. 2023. WiCE: Real-world entailment for claims in Wikipedia. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7561-7583, Singapore. Association for Computational Linguistics.  
+Rik Koncel-Kedziorski, Michael Krumdick, Viet Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. 2024. Bizbench: A quantitative reasoning benchmark for business and finance.
+
+Yuta Koreeda, Manning, and Christopher. 2021. ContractNLI: A dataset for document-level natural language inference for contracts. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 1907-1919, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.  
+Shuaiqi Liu, Jiannong Cao, Ruosong Yang, and Zhiyuan Wen. 2022. Long text and multi-table summarization: Dataset and method. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 1995-2010, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: NLG evaluation using gpt-4 with better human alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 2511–2522, Singapore. Association for Computational Linguistics.  
+Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, KaiWei Chang, Michel Galley, and Jianfeng Gao. 2024. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In International Conference on Learning Representations (ICLR).  
+Xinyuan Lu, Liangming Pan, Qian Liu, Preslav Nakov, and Min-Yen Kan. 2023. SCITAB: A challenging benchmark for compositional reasoning and claim verification on scientific tables. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7787-7813, Singapore. Association for Computational Linguistics.  
+Huanhuan Ma, Weizhi Xu, Yifan Wei, Liuji Chen, Liang Wang, Qiang Liu, Shu Wu, and Liang Wang. 2024. Ex-fever: A dataset for multi-hop explainable fact verification.  
+Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. 2018. Ww'18 open challenge: Financial opinion mining and question answering. In Companion Proceedings of the The Web Conference 2018, WWW '18, page 1941-1942, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee.  
+Pekka Malo, Ankur Sinha, Pyry Takala, Pekka Korhonen, and Jyrki Wallenius. 2013. Good debt or bad debt: Detecting semantic orientations in economic texts.
+
+AI Meta. 2024. Introducing meta llama 3: The most capable openly available llm to date. Meta AI.  
+Rajdeep Mukherjee, Abhinav Bohra, Akash Banerjee, Soumya Sharma, Manjunath Hegde, Afreen Shaikh, Shivani Shrivastava, Koustuv Dasgupta, Niloy Ganguly, Saptarshi Ghosh, and Pawan Goyal. 2022. ECTSum: A new benchmark dataset for bullet point summarization of long earnings call transcripts. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10893-10906, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+OpenAI. 2022. Chatgpt: Optimizing language models for dialogue.  
+OpenAI. 2023a. Gpt-4 technical report. ArXiv, abs/2303.08774.  
+OpenAI. 2023b. GPT-4V(ision) system card.  
+OpenAI. 2024. Hello gpt-4o.  
+Aman Rangapur, Haoran Wang, Ling Jian, and Kai Shu. 2024. Fin-fact: A benchmark dataset for multimodal financial fact checking and explanation generation.  
+Daniel Russo, Serra Sinem Tekiroglu, and Marco Guerini. 2023. Benchmarking the generation of fact checking explanations. Transactions of the Association for Computational Linguistics, 11:1250-1264.  
+Julio Cesar Salinas Alvarado, Karin Verspoor, and Timothy Baldwin. 2015. Domain adaption of named entity recognition to support credit risk assessment. In Proceedings of the Australasian Language Technology Association Workshop 2015, pages 84-90, Parramatta, Australia.  
+Agam Shah, Ruchit Vithani, Abhinav Gullapalli, and Sudheer Chava. 2023. Finer: Financial named entity recognition dataset and weak-supervision model.  
+Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. 2024. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations.  
+Yejun Soun, Jaemin Yoo, Minyong Cho, Jihyeong Jeon, and U Kang. 2022. Accurate stock movement prediction with self-supervised learning from sparse noisy tweets. In 2022 IEEE International Conference on Big Data (Big Data), pages 1691-1700.  
+Liyan Tang, Philippe Laban, and Greg Durrett. 2024. Minicheck: Efficient fact-checking of llms on grounding documents.  
+Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Leonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam
+
+Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Machel Reid, Maciej Mikuţa, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. 2024. Gemma: Open models based on gemini research and technology.
+
+InternLM Team. 2024. Internlm2 technical report.
+
+James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. 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 809-819, New Orleans, Louisiana. Association for Computational Linguistics.
+
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinez, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan
+
+Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023a. Llama 2: Open foundation and finetuned chat models.
+
+Hugo Touvron, Louis Martin, Kevin R. Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M. Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony S. Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel M. Kloumann, A. V. Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, R. Subramanian, Xia Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zhengxu Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open foundation and fine-tuned chat models.
+
+Andreas Vlachos and Sebastian Riedel. 2014. Fact checking: Task definition and dataset construction. In Proceedings of the ACL 2014 Workshop on Language Technologies and Computational Social Science, pages 18-22, Baltimore, MD, USA. Association for Computational Linguistics.
+
+David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or fiction: Verifying scientific claims. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7534-7550, Online. Association for Computational Linguistics.
+
+David Wadden, Kyle Lo, Bailey Kuehl, Arman Cohan, Iz Beltagy, Lucy Lu Wang, and Hannaneh Hajishirzi. 2022. SciFact-open: Towards open-domain scientific claim verification. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 4719-4734, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems.
+
+Huizhe Wu, Wei Zhang, Weiwei Shen, and Jun Wang. 2018. Hybrid deep sequential modeling for social text-driven stock prediction. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM '18, page 1627-1630, New York, NY, USA. Association for Computing Machinery.
+
+Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. 2023. Bloomberggpt: A large language model for finance. ArXiv, abs/2303.17564.  
+Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, and Jimin Huang. 2023. Pixiu: A large language model, instruction data and evaluation benchmark for finance.  
+Qianqian Xie, Dong Li, Mengxi Xiao, Zihao Jiang, Ruoyu Xiang, Xiao Zhang, Zhengyu Chen, Yueru He, Weiguang Han, Yuzhe Yang, Shunian Chen, Yifei Zhang, Lihang Shen, Daniel Kim, Zhiwei Liu, Zheheng Luo, Yangyang Yu, Yupeng Cao, Zhiyang Deng, Zhiyuan Yao, Haohang Li, Duanyu Feng, Yongfu Dai, VijayaSai Somasundaram, Peng Lu, Yilun Zhao, Yitao Long, Guojun Xiong, Kaleb Smith, Honghai Yu, Yanzhao Lai, Min Peng, Jianyun Nie, Jordan W. Suchow, Xiao-Yang Liu, Benyou Wang, Alejandro Lopez-Lira, Jimin Huang, and Sophia Ananiadou. 2024. Open-finllms: Open multimodal large language models for financial applications.  
+Yumo Xu and Shay B. Cohen. 2018. Stock movement prediction from tweets and historical prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1970-1979, Melbourne, Australia. Association for Computational Linguistics.  
+Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. 2023. Fingpt: Open-source financial large language models. FinLLM Symposium at IJCAI 2023.  
+Wenpeng Yin, Dragomir Radev, and Caiming Xiong. 2021. DocNLI: A large-scale dataset for document-level natural language inference. In *Findings of the Association for Computational Linguistics: ACLIJCNLP* 2021, pages 4913-4922, Online. Association for Computational Linguistics.  
+Yilun Zhao, Yunxiang Li, Chenying Li, and Rui Zhang. 2022. MultiHiertt: Numerical reasoning over multi hierarchical tabular and textual data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588-6600, Dublin, Ireland. Association for Computational Linguistics.  
+Yilun Zhao, Hongjun Liu, Yitao Long, Rui Zhang, Chen Zhao, and Arman Cohan. 2024a. Financemath: Knowledge-intensive math reasoning in finance domains. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12841-12858, Bangkok, Thailand. Association for Computational Linguistics.  
+Yilun Zhao, Yitao Long, Hongjun Liu, Ryo Kamoi, Linyong Nan, Lyuhao Chen, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. 2024b. DocMath-eval: Evaluating math reasoning capabilities of LLMs in understanding long and specialized
+
+documents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16103-16120, Bangkok, Thailand. Association for Computational Linguistics.  
+Yilun Zhao, Yitao Long, Hongjun Liu, Linyong Nan, Lyuhao Chen, Ryo Kamoi, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. 2023a. Docmath-eval: Evaluating numerical reasoning capabilities of llms in understanding long documents with tabular data.  
+Yilun Zhao, Zhenting Qi, Linyong Nan, Boyu Mi, Yixin Liu, Weijin Zou, Simeng Han, Ruizhe Chen, Xiangru Tang, Yumo Xu, Dragomir Radev, and Arman Cohan. 2023b. QTSumm: Query-focused summarization over tabular data. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1157-1172, Singapore. Association for Computational Linguistics.  
+Yilun Zhao, Haowei Zhang, Shengyun Si, Linyong Nan, Xiangru Tang, and Arman Cohan. 2023c. Investigating table-to-text generation capabilities of large language models in real-world information seeking scenarios. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 160-175, Singapore. Association for Computational Linguistics.  
+Yilun Zhao, Chen Zhao, Linyong Nan, Zhenting Qi, Wenlin Zhang, Xiangru Tang, Boyu Mi, and Dragomir Radev. 2023d. RobuT: A systematic study of table QA robustness against human-annotated adversarial perturbations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6064–6081, Toronto, Canada. Association for Computational Linguistics.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track.  
+Zhihan Zhou, Liqian Ma, and Han Liu. 2021. Trade the event: Corporate events detection for news-based event-driven trading. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2114-2124, Online. Association for Computational Linguistics.  
+Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and TatSeng Chua. 2021. TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance. 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 3277-3287, Online. Association for Computational Linguistics.
+
+<table><tr><td>ID</td><td>Finance Industry Experience</td><td>English Proficiency</td><td>Annotation Sets</td><td>Evaluator?</td></tr><tr><td>1</td><td>1 working and 1 internship at US</td><td>Native speaker</td><td>FDV-KNOW</td><td>✓</td></tr><tr><td>2</td><td>&gt;= 2 internship at US</td><td>&gt; 15 years</td><td>FDV-MATH</td><td>✓</td></tr><tr><td>3</td><td>1 working at Singapore and 2 internship at US</td><td>Native speaker</td><td>FDV-KNOW</td><td>✓</td></tr><tr><td>4</td><td>2 working and &gt;= 1 internship at US</td><td>Native speaker</td><td>FDV-KNOW</td><td>✗</td></tr><tr><td>5</td><td>1 internship at US, 2 internship at China</td><td>10 years</td><td>FDV-IE</td><td>✗</td></tr><tr><td>6</td><td>1 internships at HK, China</td><td>15 years</td><td>FDV-IE, FDV-MATH</td><td>✓</td></tr><tr><td>7</td><td>1 internships at China</td><td>10 years</td><td>FDV-IE, FDV-MATH</td><td>✗</td></tr></table>
+
+Table 7: Details of annotators involved in dataset construction. FINDVER is annotated by financial professionals with extensive experience in comprehending financial documents, ensuring it accurately reflects the real-world challenges in the financial domain.  
+
+<table><tr><td>Model</td><td>Organization</td><td>Release Time</td><td>Max Length</td><td>Source</td></tr><tr><td>GPT-4o (OpenAI, 2023a)</td><td>OpenAI</td><td>2023-03</td><td>128k</td><td>https://platform.openai.com/</td></tr><tr><td>GPT-3.5-turbo (OpenAI, 2022)</td><td>OpenAI</td><td>2022-11</td><td>16k</td><td>https://platform.openai.com/</td></tr><tr><td>Gemini-1.5-* (Gemini, 2024)</td><td>Google</td><td>2024-02</td><td>128k</td><td>https://ai.google.dev/</td></tr><tr><td>Claude-3.5 (Anthropic, 2024)</td><td>Anthropic</td><td>2024-03</td><td>200k</td><td>https://www.anthropic.com/api</td></tr><tr><td>Gemma (Team et al., 2024)</td><td>Google</td><td>2024-02</td><td>8k</td><td>google/gemma-7b-it</td></tr><tr><td>Llama-2 (Touvron et al., 2023a)</td><td>Meta</td><td>2023-02</td><td>4k</td><td>meta-llama/Llama-2-*-chat-hf</td></tr><tr><td>Llama-3 (Meta, 2024)</td><td>Meta</td><td>2024-04</td><td>8k</td><td>meta-llama/Meta-Llama-3-*-Instruct</td></tr><tr><td>Yi-1.5 (AI et al., 2024)</td><td>01-ai</td><td>2024-05</td><td>32k</td><td>01-ai/Yi-1.5-*-Chat</td></tr><tr><td>Qwen-2 (qwe, 2024)</td><td>Qwen</td><td>2024-06</td><td>128k</td><td>Qwen/Qwen2-*-Instruct</td></tr><tr><td>Mistral (Jiang et al., 2024, 2023a)</td><td>Mistral AI</td><td>2024-05</td><td>32k</td><td>mistralai/Mistral-7B-Instruct-v0.3</td></tr><tr><td>Mixtral (Jiang et al., 2024, 2023a)</td><td>Mistral AI</td><td>2024-04</td><td>64k</td><td>mistralai/Mixtral-8x22B-v0.1</td></tr><tr><td>InternLM2 (Team, 2024)</td><td>internlm</td><td>2024-01</td><td>200k</td><td>internlm/internlm2-chat-*</td></tr><tr><td>GLM (Du et al., 2022)</td><td>THUDM</td><td>2024-06</td><td>128k</td><td>THUDM/glm-4-9b-chat</td></tr><tr><td>Phi-3 (Abdin et al., 2024)</td><td>microsoft</td><td>2024-04</td><td>128k</td><td>microsoft/Phi-3-*-instruct</td></tr></table>
+
+Table 8: Details of the organization, release time, maximum context length, and model source (i.e., url for proprietary models and Huggingface model name for open-source models) for the LLMs evaluated in FINDVER.

2024/FinDVer_ Explainable Claim Verification over Long and Hybrid-content Financial Documents/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6363964dac82d57bcb786fc780b24495ce64a3319ee932c0a86586a34a21e789
+size 797027

2024/Finding Blind Spots in Evaluator LLMs with Interpretable Checklists/e1a38a24-c9fd-4bd0-b41f-2477351b5a81_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d65b1ea8200fa8242bca778a7f4b4df1b75bec183aadd9dcb43fb989b4a80768
+size 1853639

2024/Finding Blind Spots in Evaluator LLMs with Interpretable Checklists/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97999ba8b7d289af6b9b1ccba21ab3e4eb00269edb10274c4613143d10e1de6f
+size 2705343

2024/Fine-Grained Detection of Solidarity for Women and Migrants in 155 Years of German Parliamentary Debates/ec6157f5-fff1-423b-b564-2664514ae3c7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca2d741453d06acb6a3f06c62bf9d4f33e9df517f0f46ee8acb03843bc03e42e
+size 1860657

2024/Fine-Grained Detection of Solidarity for Women and Migrants in 155 Years of German Parliamentary Debates/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:271700974c5b6fe0153bd957002f63046432c3c9ae3907c7002c7ba86de8f32c
+size 3183167

2024/Fine-Grained Prediction of Reading Comprehension from Eye Movements/6bd75cfc-d816-4947-81b0-c69bbdffc4eb_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e43422aaf26ee02859a222ceb1ec803fefba2633930a0c2fb596aa7e9e0be94
+size 5240313

2024/Fine-Grained Prediction of Reading Comprehension from Eye Movements/full.md

@@ -0,0 +1,424 @@
+# Fine-Grained Prediction of Reading Comprehension from Eye Movements
+
+Omer Shubi<sup>1</sup>, Yoav Meiri<sup>1</sup>, Cfir Avraham Hadar<sup>1</sup>, Yevgeni Berzak<sup>1,2</sup>
+
+$^{1}$ Faculty of Data and Decision Sciences,
+
+Technion - Israel Institute of Technology, Haifa, Israel
+
+$^{2}$ Department of Brain and Cognitive Sciences,
+
+Massachusetts Institute of Technology, Cambridge, USA
+
+{shubi,meiri.yoav,kfir-hadar}@campus.technion.ac.il,berzak@technion.ac.il
+
+# Abstract
+
+Can human reading comprehension be assessed from eye movements in reading? In this work, we address this longstanding question using large-scale eyetracking data. We focus on a cardinal and largely unaddressed variant of this question: predicting reading comprehension of a single participant for a single question from their eye movements over a single paragraph. We tackle this task using a battery of recent models from the literature, and three new multimodal language models. We evaluate the models in two different reading regimes: ordinary reading and information seeking, and examine their generalization to new textual items, new participants, and the combination of both. The evaluations suggest that the task is highly challenging, and highlight the importance of benchmarking against a strong text-only baseline. While in some cases eye movements provide improvements over such a baseline, they tend to be small. This could be due to limitations of current modelling approaches, limitations of the data, or because eye movement behavior does not sufficiently pertain to fine-grained aspects of reading comprehension processes. Our study provides an infrastructure for making further progress on this question.<sup>1</sup>
+
+# 1 Introduction
+
+Reading comprehension is an indispensable skill for successful participation in modern society. Consequently, many efforts and resources are invested in the development of reading comprehension assessments by educational institutions and commercial companies. The standard, and to date the only practical way to assess reading comprehension is through behavioral tasks, most commonly reading comprehension questions. However, despite its clear value and ubiquitous use, this approach is extremely time-consuming and costly, which severely
+
+limits the volume and public availability of reading comprehension tests. Further, this testing methodology relies on offline behavioral signals – the end responses to a few select reading comprehension questions, and has no ability to trace the rich online reading comprehension processes as they unfold over time.
+
+An alternative vision for assessing reading comprehension has been emerging in psycholinguistics and the psychology of reading. It posits that reading comprehension may be decoded in real-time directly from eye movements in reading. This vision is rooted in literature that suggests a tight correspondence between eye movements and real time language comprehension (Just and Carpenter, 1980; Rayner, 1998; Rayner et al., 2016, among others). With the rise of modern machine learning and NLP, multiple studies over the past decade attempted to use eye movement data to predict reading comprehension (Copeland et al., 2014; Ahn et al., 2020; Reich et al., 2022; Mézière et al., 2023b, among others). This line of work suggests that in some cases various aspects of reading comprehension can be predicted from eye movements with above-chance performance. However, despite the advances so far, predictive modeling of reading comprehension from gaze is still in its infancy.
+
+A number of factors have been hindering progress in this area. One is the paucity and small size of reading comprehension data paired with eye movements. Second, the task of reading comprehension prediction has thus far been predominantly formulated as prediction of aggregated scores across multiple questions rather than prediction of comprehension at the resolution of an individual question. Further, reading comprehension has been primarily studied when the reader has no specific goals with respect to the text beyond general comprehension, a regime that we refer to as ordinary reading. Many other reading regimes common in daily life, such as explicit information
+
+seeking, remain largely unaddressed. Finally, despite the dramatic progress in machine learning and NLP in recent years, effective joint modeling of text and eye movements remains a nascent and challenging domain of investigation.
+
+In this work, we take a step forward in advancing the state-of-the-art in eye movement-based prediction of reading comprehension by combining new models, new data, and systematic evaluations. Our primary contributions are the following:
+
+- Task: we introduce the challenging and largely unaddressed task of predicting the reading comprehension of a single reader with respect to a single reading comprehension question over one passage. This task is enabled by OneStop Eye Movements (Malmaud et al., 2020), the largest eyetracking for reading comprehension dataset to date with 486 multiple-choice questions and 19,440 question responses from 360 participants.  
+- Modeling: we develop three new models that combine text and eye movements based on the transformer encoder architecture: RoBERTa-QEye, MAG-QEye, and PostFusion-QEye. These models address both test format-agnostic and multiple-choice specific variants of the task.  
+- Reading Regimes: we study reading comprehension not only in ordinary reading but also in information seeking, a highly common but understudied reading scenario.  
+- Evaluation: we evaluate our models against a battery of existing models for prediction of reading comprehension from eye movements, and a strong text-only baseline. To this end, we use a detailed evaluation protocol targeting three different levels of model generalization: new participant, new textual item, and the combination of both.
+
+# 2 Related Work
+
+Our study contributes to an existing body of work on the prediction of reading comprehension from eye movements in reading. To address various aspects of this task, prior studies used a wide range of models, including linear models (Mézière et al., 2023b,a), kernel methods (Makowski et al., 2019), feed-forward networks (e.g. Copeland et al., 2014), CNNs (Ahn et al., 2020) and RNNs (e.g. Ahn et al.,
+
+2020; Reich et al., 2022). These were typically applied to the prediction of aggregated comprehension scores over multiple items. In this work, we evaluate multiple models from prior work on the single-item reading comprehension task.
+
+While transformer models (Vaswani et al., 2017), have been used for joint modeling of eye movements and text (e.g. Deng et al., 2023; Yang and Hollenstein, 2023), they have not been applied to the problem of reading comprehension prediction from eye movements. In this work we introduce three new transformer models which draw on multimodal transformers, in particular MAG (Rahman et al., 2020) which integrated text, speech and vision for sentiment analysis, and language vision models such as VisualBERT (Li et al., 2019) (see Zhu et al. (2023); Xu et al. (2023) for reviews).
+
+Most prior studies on reading comprehension prediction from eye movements relied solely on eye movement features (Copeland et al., 2014; Southwell et al., 2020; Ahn et al., 2020; Mézière et al., 2023b,a), while a few combined eye movements with properties of the underlying text (Martínez-Gómez and Aizawa, 2014; Makowski et al., 2019; Reich et al., 2022). In the current work, we take the latter, under-explored approach. The importance of combining eye movements with attributes of the text is motivated by a large literature in the psychology of reading which points to systematic effects of linguistic properties of the text on reading times (Rayner, 1998; Rayner et al., 2004; Kliegl et al., 2004; Demberg and Keller, 2008; Smith and Levy, 2013, among others), in particular in the context of reading comprehension (Just and Carpenter, 1980) and linguistic proficiency (Berzak et al., 2018; Berzak and Levy, 2023).
+
+While highly informative, existing work is critically limited by small data, especially with respect to the number of available questions and participants. For example, Copeland et al. (2014) have 9 text pages, 18 questions and 39 participants. SB-SAT (Ahn et al., 2020), the only publicly available eyetracking dataset for reading comprehension, has 22 text pages, 20 questions, and 95 participants. The small size of previously used datasets severely limits the potential of NLP and machine learning approaches for reading comprehension prediction. At the same time, the reading comprehension component of broad coverage eyetracking datasets such as MECO (Siegelman et al., 2022) and CELER (Berzak et al., 2022) comprises only simple comprehension questions that serve as attention checks,
+
+and as such are not well suited for studying reading comprehension. OneStop, used here, has a large number of items, participants and questions, enabling meaningfully addressing item-level prediction of comprehension.
+
+Prior work varies in experimental designs. In several studies, multiple questions are presented after reading a multi-screen text without the ability to return to the text (Makowski et al., 2019; Ahn et al., 2020; Reich et al., 2022). This design is advantageous in the separation of text reading and question answering, but can lead to loose relations between eye movements and question-answering behavior due to memory limitations. In other studies, such as Copeland et al. (2014), participants can switch back and forth between the text and the questions. This creates a complex mix of ordinary reading and information seeking components which are difficult to disentangle. In OneStop, a single question appears immediately after reading a single text page, setting a middle ground between the two primary existing approaches for question presentation, and alleviating their main disadvantages. At the same time, it includes a question preview manipulation which allows to systematically compare reading comprehension in ordinary reading and question guided information seeking.
+
+An additional limitation of prior work is the scope and nature of the evaluations. With the exception of Copeland et al. (2014), both training and evaluation were previously carried out over aggregated responses across multiple questions, and in some cases also across multiple texts. These approaches, which focus on measuring overall comprehension, do not enable testing direct links between eye movements and understanding specific aspects of the text. In several studies (Martínez-Gómez and Aizawa, 2014; Makowski et al., 2019; Ahn et al., 2020; Reich et al., 2022), an additional step was taken, binning comprehension scores into two binary categories, high versus low comprehension, thus further simplifying the task.
+
+A second important evaluation limitation in prior work is evaluations in which eyetracking data for both the test participants and items is used in the training set. To our knowledge, except for Makowski et al. (2019), no work has evaluated reading comprehension prediction when neither the participant nor the item appears in the training data. This evaluation regime is needed to fully characterize model generalization ability. Importantly, even in less challenging regimes and with aggre
+
+gated scores and binning, model performance in prior work is typically only modestly higher than chance level. More stringent evaluations without binning comprehension scores (Martínez-Gómez and Aizawa, 2014), or with held-out participants and/or items (Makowski et al., 2019; Reich et al., 2022) tend to exhibit chance level performance. These results suggest that generalization in reading comprehension prediction is highly challenging.
+
+# 3 Eyetracking Data
+
+We use OneStop, an extended version of the dataset collected by Malmaud et al. (2020) over the textual materials of OneStopQA (Berzak et al., 2020). OneStop is the largest English L1 eyetracking for reading corpus to date. The data was collected using an Eyelink  $1000+$  eyetracker at a sampling rate of  $1000\mathrm{Hz}$ . In this dataset, 360 adult native English participants read newswire articles from the Guardian, and answer a multiple-choice reading comprehension question about each paragraph. The dataset includes 30 articles divided into 162 paragraphs. The average paragraph length is 109 words. Each paragraph has 3 possible questions, corresponding to a total of 486 questions.
+
+<table><tr><td>Answer</td><td>Category</td><td>Degree of Comprehension</td><td colspan="2">Gathering</td><td colspan="2">Hunting</td></tr><tr><td>A</td><td>Correct</td><td>Full comprehension</td><td colspan="2">7,890 (81.2)</td><td colspan="2">8,450 (86.9)</td></tr><tr><td>B</td><td>Incorrect</td><td>Identified question-relevant information</td><td colspan="2">1000 (10.3)</td><td colspan="2">744 (7.7)</td></tr><tr><td>C</td><td>Incorrect</td><td>Some degree of attention to the text</td><td colspan="2">568 (5.8)</td><td colspan="2">374 (3.8)</td></tr><tr><td>D</td><td>Incorrect</td><td>No evidence for comprehension</td><td colspan="2">260 (2.7)</td><td colspan="2">152 (1.6)</td></tr></table>
+
+Table 1: Summary of the STARC annotation framework for answer types  $A - D$  ,their corresponding degree of comprehension,and number of trials in which each answer type was chosen in OneStop. Values in parentheses are percentages by reading regime.
+
+The articles are divided into three 10-article batches, where each participant is assigned to one batch. In each trial of the experiment, participants read a paragraph and then proceed to answer one of the three possible questions on a new screen, without the ability to return to the paragraph. 180 participants are in an ordinary reading (Gathering) regime where they do not see the question prior to reading the paragraph. The remaining 180 participants are in an information seeking regime (Hunting) where they are presented with the question (but not the answers) before reading the paragraph. The total number of trials is 19,440, split equally across the two reading regimes. This corresponds to 40 responses per question, 20 for each regime–paragraph combination. The total number of word
+
+![](images/37e5ea15acf00790e9d287b00624f80dc4f3f14c30c01f8216419b53917be5ce.jpg)  
+Figure 1: Left: an example of an eye movement trajectory over a paragraph, where red circles represent fixations, and blue arrows represent saccades. Right: a schematic depiction of word-level feature extraction, resulting in a vector  $E_{w_i}$ : an eye movements and linguistic word properties feature representation for each word.
+
+tokens over which eyetracking data was collected in OneStop is 3,827,216.
+
+The underlying textual materials and reading comprehension questions follow the STARC annotation framework (Berzak et al., 2020), where answer  $A$  is the correct answer, answer  $B$  is a miscomprehension of the information required to answer correctly,  $C$  refers to another part of the text that does not provide the answer to the question and  $D$  has no textual support. These answer types correspond to an ordering of the answers by degree of comprehension. Table 1 presents a summary of the framework along with answer choice statistics in the OneStop eyetracking data.
+
+# 4 Tasks
+
+# 4.1 Correct versus Incorrect Comprehension
+
+The primary task we address is item-level prediction of whether a participant will respond correctly to a single question about a paragraph from the participant's eye movements over the paragraph. For each paragraph  $p$  and a corresponding question  $q^p$ , the possible answers are  $\text{Ans}^{q^p} = \{a_1^{q^p}, a_2^{q^p}, a_3^{q^p}, a_4^{q^p}\}$ . Note that the correct answer  $A$  and the three distractors  $\{B, C, D\}$  are randomly mapped per trial to  $a_1$  through  $a_4$ . The set of  $p, q^p$ , and optionally  $\text{Ans}^{q^p}$ , defines a textual item  $W$ . Given a participant  $S$  tested on item  $W$ , where the participant's eye movements over the paragraph are  $\text{Eyes}_S^p$ , the complete trial information is  $\text{Trial}_S^W := \{W, \text{Eyes}_p^S\}$ . We make  $W$  optional to allow for models that use only eye movements without the text.
+
+The prediction problem can then be formulated as a binary classification task, we predict whether the participant will answer the question correctly. Formally, given a classifier  $h$ :
+
+$$
+h: \operatorname {T r i a l} _ {S} ^ {W} \mapsto \{0, 1 \} \tag {1}
+$$
+
+where 1 indicates a correct answer  $(A)$  and 0 indicates an incorrect answer  $(B / C / D)$ .
+
+Note that this task formulation abstracts away from the multiple-choice format. This allows assessing comprehension without depending on the format of the subsequent assessment task (e.g. answer choice, answer production), nor its details such as the number of answer choices and their specific content in the multiple-choice format. The combination of these task characteristics enables applying prior models from the literature, all of which predict a binary outcome without taking into account the answers, and some of which use only eye movements without the text.
+
+# 4.2 Specific Answer Choice
+
+We further address a task that takes advantage of the multiple-choice assessment format. In this task, given the answers, we predict which specific answer the participant will select:
+
+$$
+h: \operatorname {T r i a l} _ {S} ^ {W} \mapsto \left\{a _ {1}, a _ {2}, a _ {3}, a _ {4} \right\} \tag {2}
+$$
+
+# 5 Models
+
+We introduce three new models, RoBERTa-QEye, MAG-QEye and PostFusion-QEye, all of which combine text and eye movements information, and rely on the transformer language model encoder. Specifically, we use the RoBERTaLarge model (Liu et al., 2019). Each of these models uses a different strategy for combining text with eye movements. RoBERTa-QEye augments the textual input with additional eye movement features. MAG-QEye uses eye movement information to modify contextualized word representations at intermediate layers of the language model. PostFusion-QEye processes text and eye movements separately and then combines them via cross-attention mechanisms. We further adjust a number of prior models
+
+![](images/48eb7081f304aefc26bcba04098cec736b31276bee613772a17abc0008a83ed8.jpg)  
+Figure 2: Model architectures. (a) RoBERTa-QEye treats eye movements as additional input features. (b) MAG-QEye uses eye movement information to modify contextualized word representations. (c) PostFusion-QEye processes text and eye movements separately and combines them via cross-attention mechanisms. Model input:  $Eyes^{P}$  represents the participant's eye movements over the paragraph  $p$ ,  $q^{p}$  is a question and  $[Ans^{q^{p}}]$  are optional answer choices which are provided only in the multiple choice version of the task.
+
+![](images/54b912d9268d2689312ba4ad4194e6ebdadfe9802eefba4d058af9d610da8632.jpg)
+
+![](images/b8e15ea93ef883a9a176ade4f16b2f9e2969d4353d420a5e86da9dbb78a0a333.jpg)
+
+from the literature for the single-item reading comprehension prediction task.
+
+Eye Movement Feature Representations The eyetracking record is commonly represented as a scanpath consisting of fixations (periods in which the gaze position is stable) and saccades (rapid transitions between fixations). The examined models represent this information in three different ways, in increasing level of granularity:
+
+- Global: Summarizing fixation and saccade information across all the words in the input.  
+- Words: Summarizing fixation and saccade information for each word.  
+- Fixations: Accounting for each fixation and its preceding and following saccade.
+
+Our new models focus on the word and fixation level approaches, using a variety of eye movement measures from the psycholinguistic literature. As reading times are known to be affected by linguistic word properties such as predictability, frequency, and length (Rayner et al., 2004; Kliegl et al., 2004; Rayner et al., 2011), which are not directly encoded in word embeddings, we further add such properties to the eye movement representations to allow the models to learn eye movements-word property interactions. The strength of such interactions has been shown to be indicative of the readers' linguistic proficiency (Berzak et al., 2018; Berzak and Levy, 2023), which is directly related to reading
+
+comprehension. The eye movement and linguistic word property features used in all the models are listed in Appendix A. Note that two different feature sets are used for representing eye movements at the word and fixation levels. Figure 1 presents an example of an eye movement trajectory over a paragraph and a schematic visualization of the word-level feature extraction approach.
+
+# 5.1 RoBERTa-QEye
+
+RoBERTa-QEye incorporates eye movements as additional input sequences to RoBERTa by projecting them to the word embedding space. An overview of the architecture is presented in Figure 2a. The model is implemented in two variants, RoBERTa-QEye-Words which has a word-level feature representation and RoBERTa-QEye-Fixations, which uses a fixation-level representation. Both variants combine a textual input  $Z_{W}$  with eye movements input  $Z_{E_{P}}$ .
+
+The textual representation  $Z_W$  is the word embedding sequence [CLS;  $p$ ; SEP;  $q^p$ ;  $[Ans^{q^p}]$ ; SEP], where  $p$  is the paragraph,  $q^p$  is the question,  $[Ans^{q^p}]$  are optional answers, and SEP is a separator token. The eye movement representation for the paragraph  $Z_{EP} = [Z_{E_{w_1}},\dots,Z_{E_{w_n}}]$  consists of a representation for each fixation or word  $i$  as:
+
+$$
+Z _ {E _ {w _ {i}}} = \operatorname {F C} \left(E _ {w _ {i}}\right) + \operatorname {E m b} _ {\text {p o s}} (i) + \operatorname {E m b} _ {\text {e y e}} \tag {3}
+$$
+
+where  $E_{w_i}$  are the eye movement and word property features and FC is a fully connected layer
+
+projecting this feature representation to the word embedding space.  $\mathrm{Emb}_{\mathrm{pos}}(i)$  is the positional embedding of the  $i$ -th word or fixation, initialized to the model's original positional embedding, which ties the eye movement representation to its respective word index.  $\mathrm{Emb}_{\mathrm{eye}}$  is an additional learnable embedding marking the presence of eye movement information.  $Z_{E_P}$  is concatenated with the word embedding representation  $Z_W$ , separated by a special token  $\mathrm{SEP}_E$ , initialized as SEP. The combined sequence  $[Z_{E_P};\mathrm{SEP}_E;Z_W]$  is passed through the transformer encoder language model. The resulting CLS token is then provided to a multilayer perceptron for response prediction.
+
+# 5.2 MAG-QEye
+
+MAG-QEye, shown in Figure 2b, modifies the transformer encoder's hidden word representations based on eye movement information. It is an adaptation of the MAG architecture (Rahman et al., 2020) originally developed for multimodal sentiment analysis. The goal of this model is to emphasize or de-emphasize words based on their respective eye movement features. Formally, for a given model layer  $k$ , each hidden token representation in the paragraph  $Z_{W_i}^k$  is shifted by  $H_{W_i}$ :
+
+$$
+\bar {Z} _ {W _ {i}} ^ {k} = Z _ {W _ {i}} ^ {k} + \alpha H _ {W _ {i}} \tag {4}
+$$
+
+where  $H_{W_i}$  is a scaled version of eye movements  $E_{W_i}$  transformed into the word embedding space. The final resulting CLS token is passed through a multilayer perceptron classifier. Appendix B.1 provides a detailed description of the architecture.
+
+# 5.3 PostFusion-QEye
+
+PostFusion-QEye, outlined in Figure 2c, processes text and eye movements separately and combines their representations through two cross-attention mechanisms. The primary objective of these mechanisms is to transform both text and eye movement data into a unified space, which we refer to as the reading space while taking into account the reading comprehension prediction task.
+
+The input paragraph is passed through a language model to obtain contextualized embeddings  $Z_P$ . The eye movement input features are processed through two 1D convolution layers, resulting in the eye movement representation  $Z_{E_P}$ . Cross-attention is then applied between the paragraph embedding  $Z_P$  and  $Z_{E_P}$ , with eye movements as the query and text embeddings as the key
+
+and the value. This step modifies the paragraph words based on the eye movements. The output is provided along with  $Z_{E_P}$  to a fully connected layer, yielding  $Z_{E_P + P}$ , a projection of the two into a shared space. Another cross-attention layer is applied between  $Z_{E_P + P}$  as key and value and the question embedding  $Z_Q$  as query, weighting the shared representation by the relevance to the question. The output of this step is passed to a multi-layer perceptron classifier to predict the response.
+
+# 5.4 Multiple-Choice Variants
+
+For the specific-answer prediction task, we add to the model input the answer choices:  $[a_1^{q^p}, a_2^{q^p}, a_3^{q^p}, a_4^{q^p}]$ . The answer choices are provided to the model in a randomized order, as presented to the participants.
+
+# 5.5 Baseline Models
+
+We compare the proposed models to a number of eye movement models from prior work. We focus on models that were either designed for reading comprehension prediction or can be adjusted to the binary task with minimal modifications. As none of the prior models allow encoding of answers, we cannot apply them to the multiple-choice task.
+
+Logistic Regression (Mézière et al., 2023b) Based on Mézière et al. (2023b) who used linear regression for reading comprehension prediction. We use the same feature set which includes reading speed, and global averages of standard eye movement measures.
+
+CNN (Ahn et al., 2020) Similarly to Mézière et al. (2023b), this model is based only on eye movement information, without the underlying text. It uses the fixation sequence, represented by x and y coordinates on the screen, fixation durations, and pupil size, which are passed through a Convolutional Neural Network (CNN) to predict a binary comprehension outcome.
+
+BEyeLSTM (Reich et al., 2022) A model for predicting reading comprehension from eye movements which represents both the fixation sequence and text features, combining LSTMs with affine transformations. BEyeLSTM outperforms the CNN model of Ahn et al. (2020), on the high versus low comprehension task with SB-SAT.
+
+Eyettention (Deng et al., 2023) This model was originally developed for scanpath prediction. Eyettention is a word sequence encoder and a fixation
+
+sequence encoder that uses a pre-trained BERT (Devlin et al., 2019) and an LSTM (Hochreiter and Schmidhuber, 1997), with a cross-attention mechanism for the alignment of the input sequences. We adjust this model for prediction of reading comprehension by using global cross-attention instead of windowed attention, and represent the scanpath using the last hidden representation. Further details on this model are provided in Appendix B.
+
+# 5.6 No Eye Movements Baselines
+
+We further introduce two baselines with no eye movements. The first is a majority class baseline. The second is Text-only RoBERTa. This baseline is of special importance as it is able to take into account item difficulty as reflected in the item textual characteristics and the distribution of item responses in the training data. To our knowledge, no previous reading comprehension prediction method was benchmarked against this kind of baseline.
+
+# 6 Experimental Setup
+
+We evaluate the models in three evaluation regimes that test different aspects of model generalization.
+
+- New Participant: No eyetracking data is available for the given participant, but eyetracking data from other participants is available for the given item (paragraph).  
+- New Item: No eyetracking data is available for the item, but prior eyetracking data is available for the participant on other items.  
+- New Item & Participant: No prior eyetracking data is available for the participant nor for the item.
+
+We further report aggregated results across all three regimes.
+
+We perform model training, hyperparameter tuning, and evaluation separately for the ordinary reading and information seeking parts of the data, with 10-fold cross-validation. Figure 3 presents schematically one of the 10 data splits for a 10article 60-participant batch. A full data split for a reading regime (ordinary reading or information seeking) is the union of three such splits. In each split, approximately  $64\%$  of the data is allocated for training,  $17\%$  for validation, and  $19\%$  for testing. The test data is further divided into  $9\%$  in the New Participant,  $9\%$  New Item, and  $1\%$  New Item & Participant regimes. In total across the
+
+![](images/2fd5d649641eebae2f8abcc946edc930f5caa5485f0e59b51c484f4375fdd491.jpg)  
+Figure 3: A schematic depiction of a 10-article 60-participant batch split, divided into a train set, a validation set, and the three test sets. A full data split for a reading regime (ordinary reading or information seeking) consists of the union of three batch splits.
+
+10 splits, approximately  $90\%$  of the trials in the dataset appear in each of the New Participant and New Item evaluation regimes, and  $10\%$  in the New Item & Participant regime. Items are assigned to the train, validation and test portions of each split at the article level, such that no article is split across different data portions, ensuring generalization to items whose content is unrelated to items seen in training. See Appendix C for further information on the splits.
+
+Because the data is unbalanced across classes, we use balanced accuracy as the evaluation metric. As prior work has shown considerable differences in reading behavior between the ordinary reading and information seeking reading conditions (Hahn and Keller, 2023; Malmaud et al., 2020; Shubi and Berzak, 2023), we train and evaluate the models on each type of trials separately. We perform hyperparameter tuning for each split, and report balanced accuracy results on the aggregation of the predictions across the 10 test sets. We assume that at test time the evaluation regime of the trial is unknown. Model hyperparameter tuning is therefore based on the entire validation set of the split. As prior models from the literature were developed for different tasks and on different datasets, we run a hyperparameter search for each model over a search space that includes the original parameter settings. Hyperparameters are also optimized for the Text-only RoBERTa baseline. To address the
+
+<table><tr><td colspan="3">Binary Reading Comprehension</td><td colspan="4">Ordinary Reading (Gathering)</td><td colspan="4">Information Seeking (Hunting)</td></tr><tr><td>Model</td><td>Gaze Representation</td><td>Text Representation</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td></tr><tr><td>Majority</td><td>None</td><td>None</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td></tr><tr><td>Text-only RoBERTa</td><td>None</td><td>Emb</td><td>54.8</td><td>63.1</td><td>55.2</td><td>58.7</td><td>51.8</td><td>63.1</td><td>50.5</td><td>57.1</td></tr><tr><td>Log. Reg. (Mézière et al., 2023b)</td><td>Global</td><td>None</td><td>53.3</td><td>50.8</td><td>53.8</td><td>52.2</td><td>53.2</td><td>52.2</td><td>52.3</td><td>52.7</td></tr><tr><td>CNN (Ahn et al., 2020)</td><td>Fixations</td><td>None</td><td>51.0</td><td>51.0</td><td>51.9</td><td>51.1</td><td>51.4</td><td>51.3</td><td>49.2</td><td>51.2</td></tr><tr><td>BEyeLSTM (Reich et al., 2022)</td><td>Fixations</td><td>Ling. Feat.</td><td>50.6</td><td>55.7</td><td>51.1</td><td>53.0</td><td>50.5</td><td>55.1</td><td>55.1</td><td>53.0</td></tr><tr><td>Eyettention (Deng et al., 2023)</td><td>Fixations</td><td>Emb + Word Len.</td><td>54.8</td><td>60.4</td><td>57.1</td><td>57.6</td><td>50.5</td><td>56.4</td><td>52.3</td><td>53.4</td></tr><tr><td>RoBERTa-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>55.5</td><td>63.5</td><td>52.1</td><td>59.1</td><td>50.5</td><td>63.8</td><td>51.0</td><td>56.8</td></tr><tr><td>RoBERTa-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>53.3</td><td>61.3</td><td>57.1</td><td>57.3</td><td>50.3</td><td>60.3</td><td>50.8</td><td>55.1</td></tr><tr><td>MAG-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>54.8</td><td>64.1*</td><td>53.8</td><td>59.2</td><td>52.5</td><td>62.3</td><td>51.3</td><td>57.1</td></tr><tr><td>PostFusion-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>54.8</td><td>63.5</td><td>55.0</td><td>58.9</td><td>53.8*</td><td>62.7</td><td>53.8</td><td>58.0</td></tr></table>
+
+Table 2: Results on balanced accuracy for the main binary reading comprehension prediction task (correct vs incorrect comprehension). 'All' denotes results for the aggregation of all the trials across the three test regimes. 'Emb' stands for word embeddings, 'Ling.Feat.' for linguistic word properties. Statistically significant improvements over the Text-only RoBERTa baseline, using a paired bootstrap test, chosen based on considerations described in (Dror et al., 2018), are marked with  $*$  at  $p < 0.05$ .  
+
+<table><tr><td colspan="3">Multiple-Choice Reading Comprehension</td><td colspan="4">Ordinary Reading (Gathering)</td><td colspan="4">Information Seeking (Hunting)</td></tr><tr><td>Model</td><td>Gaze Representation</td><td>Text Representation</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td></tr><tr><td>Majority</td><td>None</td><td>None</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td></tr><tr><td>Text-only RoBERTa</td><td>None</td><td>Emb</td><td>25.3</td><td>33.0</td><td>25.2</td><td>29.0</td><td>25.0</td><td>31.7</td><td>24.8</td><td>28.2</td></tr><tr><td>RoBERTa-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>28.2*</td><td>31.5</td><td>32.1**</td><td>29.9</td><td>28.9***</td><td>30.1</td><td>27.1</td><td>29.3</td></tr><tr><td>RoBERTa-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>29.2*</td><td>32.9</td><td>28.1</td><td>30.9</td><td>30.3***</td><td>31.0</td><td>29.5</td><td>30.5***</td></tr><tr><td>MAG-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>27.9***</td><td>32.5</td><td>30.4***</td><td>30.2**</td><td>26.8</td><td>30.0</td><td>29.0</td><td>28.4</td></tr><tr><td>PostFusion-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>29.4**</td><td>31.7</td><td>32.9*</td><td>30.6*</td><td>27.5*</td><td>27.9</td><td>26.7</td><td>27.6</td></tr></table>
+
+Table 3: Results on balanced accuracy for the multiple-choice specific answer prediction task. Statistically significant improvements over the Text-only RoBERTa baseline, using a paired bootstrap test, are marked with ‘*’ at  $p < {0.05}$  , '***' at  $p < {0.01}$  and ’***’ at  $p < {0.001}$  . We note that in some cases, higher balanced accuracy scores correspond to lower p-values due to higher variability in the predictions of the minority classes.
+
+unbalanced nature of the data, shown in Table 1, we sample the same number of trials from each answer class during training. Additional details on feature normalization, model training, hyperparameter search, and number of model parameters are provided in Appendix D.
+
+# 7 Results
+
+# 7.1 Correct vs Incorrect Comprehension
+
+In Table 2, we present trial-level reading comprehension prediction results for ordinary reading and information seeking. The best results are achieved by different models under the different evaluation regimes. MAG-QEye achieves the highest overall balanced accuracy in ordinary reading with a score of 59.2, while PostFusion-QEye performs best in information seeking, with a score of 58.0. In all the evaluation regimes, the best performing model outperforms the Text-only RoBERTa baseline. In all but the New Item & Participant evaluation regime, the best performing model is one of our proposed models. Text-only RoBERTa turns out to be a key benchmark, whereby most models are below this baseline especially in the New Participant regime.
+
+We note several key trends in the results. First, results in the New Participant regime tend to be higher than in the New Item regime, highlighting the importance and the challenge of generalization to new items. The strong performance of the RoBERTa text-only baseline in the New Participant regime suggests that much of the gains in this regime do not stem from eye movement information, but rather from item properties and statistics. This highlights the importance of benchmarking against such a baseline for assessing the contribution of eye movement information. It further underscores the importance of explicit representation of the text; the Logistic Regression, CNN and BEyeLSTM models, which do not include such a representation, perform poorly in the New Participant regime. Finally, for any given model, the ordinary reading regime tends to yield higher accuracies compared information seeking. We hypothesize that this difference could be related to higher variability in reading strategies in information seeking across participants (Shubi and Berzak, 2023). We leave a detailed investigation of this hypothesis to future work.
+
+# 7.2 Multiple-Choice Task
+
+In Table 3 we use our models, MAG-QEye and PostFusion-QEye, and the two RoBERTa-QEye variants to predict participants' specific answer response among the four provided answers. As mentioned above, prior models from the literature are not applicable for this task. We find that all the models outperform the Text-only RoBERTa baseline in the two regimes that involve new items, but not in the New Participant regime. The best performing model in the overall evaluations is RoBERTa-QEye-Fixations. The general trends regarding higher performance in the New Participant regime compared to the New Item regime, as well as the stronger within-model performance in ordinary reading compared to information seeking, extend to this evaluation.
+
+# 7.3 Additional Experiments
+
+We perform two additional sets of experiments of preliminary nature. In Appendix E we provide ablation experiments on the effect of linguistic word properties on model performance. In Appendix F we further examine different variants of the textual backbone of the models. Finally, we provide validation set results in Appendix G.
+
+# 8 Summary and Discussion
+
+This paper presents a systematic evaluation of the ability to predict reading comprehension from eye movements in reading at the level of a single question over a single paragraph. We address this task using a range of existing and new models, applied to large scale data across several task variants and evaluation regimes. Our experiments indicate that the task at hand is highly challenging, and further highlight the importance of text-only baselines for assessing the added value of eye movements information. However, we do find that small improvements over a strong text-only baseline are achievable with the proposed and some of the past modeling approaches.
+
+Given the presented results, the extent to which specific aspects of reading comprehension can be reliably decoded from eye movements signal remains an open question. It is possible that eye movements simply do not contain sufficient information for decoding comprehension at high accuracy rates for the examined level of granularity. Alternatively, it may be the case that current modeling techniques do not represent or process eye
+
+movements data effectively enough for this task. Another factor whose role in task difficulty needs to be investigated in more detail is the imbalanced nature of the data, where only a relatively small fraction of the responses are incorrect.
+
+Additional work on eye movement data analysis, new model architectures, feature representations and training regimes is needed for making further progress on this task. Additionally, new datasets with other task variants and other populations such as children and L2 readers are required to study the problem in a more comprehensive manner. We envision that the models, tasks, evaluation protocols, and data presented here will serve as a stepping stone for such work, as well as a broader scientific investigation of the relations between eye movements and reading comprehension.
+
+# 9 Ethical Considerations
+
+The eyetracking data used in this work was collected by Malmaud et al. (2020) under an institutional IRB protocol. All the participants provided written consent prior to participating in the eyetracking study. The data is anonymized. Analyses of the relations between eye movements and reading comprehension, and predictive models of comprehension are among the primary use cases for which the data was collected.
+
+Automatic reading comprehension assessments from eye movements can potentially address shortcomings of standard assessment methodologies by reducing test development and test taking costs, and enhancing test availability. However, they also introduce potential risks for biased and inaccurate assessments that may put various populations and individuals at a disadvantage. These include non-native speakers, older participants, participants with cognitive impairments, disabilities, eye conditions and others. Much higher model performance than the current state-of-the-art and a thorough examination of potential biases due to factors unrelated to reading comprehension are needed before considering deploying such assessments.
+
+It has previously been shown that eye movements can be used for user identification (e.g. Bednarik et al., 2005; Jäger et al., 2020). We do not perform user identification in this study. We further emphasize that future reading comprehension assessment systems are to be used only with explicit consent from potential users to have their eye movements collected and analyzed for this purpose.
+
+# 10 Limitations
+
+Our work has a number of limitations which are related to the experimental design of OneStop. First, the textual data consists of articles with 4-7 paragraphs. Each question is over the content of a single paragraph. Longer and shorter texts, as well as questions that require integration of information from several paragraphs, are not covered. The experimental design does not allow participants to go back and forth between the question and passage, which is common in question answering tasks. Further, participant expectations for upcoming reading comprehension questions, as well as the setting of an in-lab experiment may result in reading patterns that deviate from reading in everyday settings (Huettig and Ferreira, 2022) and could impact the predictive performance of the model.
+
+While our work examines the feasibility of automated assessment of reading comprehension from eye movements, the accuracy of the models presented is still very far from being relevant for deployment in real world scenarios. Our results are further limited to the equipment at hand. Our approach has only been tested using a state-of-the-art eyetracker (Eyelink 1000 Plus) at a sampling rate of  $1000\mathrm{Hz}$ . This allows extracting gaze position and duration at a very high temporal resolution and character-level precision. While studies such as Ishimaru et al. (2017) and Chen et al. (2023) have demonstrated predictive modeling capabilities using lower spatial and temporal resolution eye tracking systems, additional work is required to test the feasibility of reading comprehension prediction using such equipment.
+
+Although we use the largest eyetracking for reading comprehension dataset to date, OneStop was collected from adult L1 English speakers, with no cognitive impairments, and in the large majority of cases no eye conditions. We acknowledge that this pool of participants excludes multiple populations, including children, elderly, participants with cognitive and physical impairments and others. Future data collection and analysis work is required to test the generalization capabilities and potential biases of the models in other populations.
+
+In this work we assume the availability of both suitable eyetracking data and a pretrained language model for the language at hand. Although language models for lower-resource languages (e.g. Chriqui and Yahav, 2022; Vamvas et al., 2023) and multilingual models (e.g. Lai et al., 2023) have been made
+
+available, many languages still lack such models. Similarly, to the best of our knowledge, no eyetracking data with a substantial reading comprehension component is currently available for languages other than English. This limits the generality of the results. More eyetracking data collection and language model development work is required to include additional languages.
+
+# Acknowledgments
+
+This work was supported by ISF grant 1499/22.
+
+# References
+
+Seoyoung Ahn, Conor Kelton, Aruna Balasubramanian, and Greg Zelinsky. 2020. Towards predicting reading comprehension from gaze behavior. In ACM Symposium on Eye Tracking Research and Applications, ETRA '20 Short Papers, New York, NY, USA. Association for Computing Machinery.  
+Roman Bednarik, Tomi Kinnunen, Andrei Mihaila, and Pasi Franti. 2005. Eye-movements as a biometric. In Image Analysis: 14th Scandinavian Conference, SCIA 2005, Joensuu, Finland, June 19-22, 2005. Proceedings 14, pages 780-789. Springer.  
+Yevgeni Berzak, Boris Katz, and Roger Levy. 2018. Assessing Language Proficiency from Eye Movements in Reading. 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 1986-1996, Stroudsburg, PA, USA. Association for Computational Linguistics.  
+Yevgeni Berzak and Roger Levy. 2023. Eye movement traces of linguistic knowledge in native and non-native reading. Open Mind, 7:179-196.  
+Yevgeni Berzak, Jonathan Malmaud, and Roger Levy. 2020. STARC: Structured annotations for reading comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5726-5735. Association for Computational Linguistics.  
+Yevgeni Berzak, Chie Nakamura, Amelia Smith, Emily Weng, Boris Katz, Suzanne Flynn, and Roger Levy. 2022. CELER: A 365-participant corpus of eye movements in L1 and L2 English reading. Open Mind, 6:1-10.  
+Xiuge Chen, Namrata Srivastava, Rajiv Jain, Jennifer Healey, and Tilman Dingler. 2023. Characteristics of Deep and Skim Reading on Smartphones vs. Desktop: A Comparative Study. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-14, Hamburg Germany. ACM.
+
+Avihay Chriqui and Inbal Yahav. 2022. HeBERT and HebEMO: A Hebrew BERT Model and a Tool for Polarity Analysis and Emotion Recognition. INFORMS Journal on Data Science, 1(1):81-95.  
+Leana Copeland, Tom Gedeon, and Balapuwaduge Mendis. 2014. Predicting reading comprehension scores from eye movements using artificial neural networks and fuzzy output error. Artificial Intelligence Research, 3.  
+Vera Demberg and Frank Keller. 2008. Data from eyetracking corpora as evidence for theories of syntactic processing complexity. Cognition, 109(2):193-210.  
+Shuwen Deng, David R. Reich, Paul Prasse, Patrick Haller, Tobias Scheffer, and Lena A. Jäger. 2023. Eyettention: An attention-based dual-sequence model for predicting human scanpaths during reading. In Proceedings of the ACM on Human-Computer Interaction, pages 1-24. Association for Computing Machinery.  
+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.  
+Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. The hitchhiker's guide to testing statistical significance in natural language processing. In Proceedings of the 56th annual meeting of the association for computational linguistics (volume 1: Long papers), pages 1383-1392.  
+William Falcon and The PyTorch Lightning team. 2019. PyTorch Lightning.  
+Michael Hahn and Frank Keller. 2023. Modeling task effects in human reading with neural network-based attention. Cognition, 230:105289.  
+John Hale. 2001. A probabilistic earley parser as a psycholinguistic model. In Second meeting of the north american chapter of the association for computational linguistics.  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation, 9(8):1735-1780. Conference Name: Neural Computation.  
+Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020. spaCy: Industrialstrength Natural Language Processing in Python.  
+Falk Huettig and Fernanda Ferreira. 2022. The Myth of Normal Reading. Perspectives on Psychological Science, page 17456916221127226. Publisher: SAGE Publications Inc.
+
+Shoya Ishimaru, Kensuke Hoshika, Kai Kunze, Koichi Kise, and Andreas Dengel. 2017. Towards reading trackers in the wild: detecting reading activities by EOG glasses and deep neural networks. In Proceedings of the 2017 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2017 ACM International Symposium on Wearable Computers, UbiComp '17, pages 704-711, New York, NY, USA. Association for Computing Machinery.  
+Lena A Jäger, Silvia Makowski, Paul Prasse, Sascha Liehr, Maximilian Seidler, and Tobias Scheffer. 2020. Deep eyedentification: Biometric identification using micro-movements of the eye. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2019, Würzburg, Germany, September 16-20, 2019, Proceedings, Part II, pages 299-314. Springer.  
+Marcel Adam Just and Patricia A. Carpenter. 1980. A theory of reading: From eye fixations to comprehension. Psychological Review, 87(4):329.  
+Reinhold Kliegl, Ellen Grabner, Martin Rolfs, and Ralf Engbert. 2004. Length, frequency, and predictability effects of words on eye movements in reading. European Journal of Cognitive Psychology - EUR J COGN PSYCHOL, 16:262-284.  
+Viet Lai, Nghia Ngo, Amir Pouran Ben Veyseh, Hieu Man, Franck Dernoncourt, Trung Bui, and Thien Nguyen. 2023. ChatGPT beyond English: Towards a comprehensive evaluation of large language models in multilingual learning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13171-13189, Singapore. Association for Computational Linguistics.  
+Roger Levy. 2008. Expectation-based syntactic comprehension. Cognition, 106(3):1126-1177.  
+Lianian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. 2019. VisualBERT: A Simple and Performant Baseline for Vision and Language. ArXiv:1908.03557 [cs].  
+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:1907.11692 [cs]. ArXiv: 1907.11692.  
+Ilya Loshchilov and Frank Hutter. 2018. Decoupled Weight Decay Regularization. In International Conference on Learning Representations.  
+Silvia Makowski, Lena A Jäger, Ahmed Abdelwahab, Niels Landwehr, and Tobias Scheffer. 2019. A discriminative model for identifying readers and assessing text comprehension from eye movements. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2018, Dublin, Ireland, September 10-14, 2018, Proceedings, Part I 18, pages 209-225. Springer.
+
+Jonathan Malmaud, Roger Levy, and Yevgeni Berzak. 2020. Bridging Information-Seeking Human Gaze and Machine Reading Comprehension. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 142–152, Stroudsburg, PA, USA. Association for Computational Linguistics.  
+Pascual Martínez-Gómez and Akiko Aizawa. 2014. Recognition of understanding level and language skill using measurements of reading behavior. In Proceedings of the 19th International Conference on Intelligent User Interfaces, IUI '14, page 95-104, New York, NY, USA. Association for Computing Machinery.  
+Marius Mosbach, Maksym Andriushchenko, and Dietrich Klakow. 2021. On the stability of fine-tuning  $\{\text{bert}\}$ : Misconceptions, explanations, and strong baselines. In International Conference on Learning Representations.  
+Diane C. Mézière, Lili Yu, Erik D. Reichle, Genevieve McArthur, and Titus von der Malsburg. 2023a. Scannpath regularity as an index of reading comprehension. Scientific Studies of Reading.  
+Diane C. Mézière, Lili Yu, Erik D. Reichle, Titus von der Malsburg, and Genevieve McArthur. 2023b. Using eye-tracking measures to predict reading comprehension. Reading Research Quarterly, 58(3):425-449.  
+Nicki Skafte Detlefsen, Jiri Borovec, Justus Schock, Ananya Harsh, Teddy Koker, Luca Di Liello, Daniel Stancl, Changsheng Quan, Maxim Grechkin, and William Falcon. 2022. TorchMetrics - Measuring Reproducibility in PyTorch.  
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems 32, pages 8024–8035. Curran Associates, Inc.  
+Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard Duchesnay. 2011. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12(85):2825-2830.  
+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.  
+Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, AmirAli Bagher Zadeh, Chengfeng Mao, Louis-Philippe
+
+Morency, and Ehsan Hoque. 2020. Integrating Multimodal Information in Large Pretrained Transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2359-2369, Online. Association for Computational Linguistics.  
+Keith Rayner. 1998. Eye movements in reading and information processing: 20 years of research. *Psychological Bulletin*, 124(3):372-422.  
+Keith Rayner, Jane Ashby, Alexander Pollatsek, and Erik D Reichle. 2004. The effects of frequency and predictability on eye fixations in reading: implications for the ez reader model. Journal of Experimental Psychology: Human Perception and Performance, 30(4):720.  
+Keith Rayner, Elizabeth R Schotter, Michael EJ Masson, Mary C Potter, and Rebecca Treiman. 2016. So much to read, so little time: How do we read, and can speed reading help? Psychological Science in the Public Interest, 17(1):4-34.  
+Keith Rayner, Timothy J Slattery, Denis Drieghe, and Simon P Liversedge. 2011. Eye movements and word skipping during reading: Effects of word length and predictability. Journal of Experimental Psychology: Human Perception and Performance, 37(2):514.  
+David R. Reich, Paul Prasse, Chiara Tschirner, Patrick Haller, Frank Goldhammer, and Lena A. Jäger. 2022. Inferring native and non-native human reading comprehension and subjective text difficulty from scanpaths in reading. In Symposium on Eye Tracking Research and Applications, ETRA '22. Association for Computing Machinery.  
+Omer Shubi and Yevgeni Berzak. 2023. Eye movements in information-seeking reading. In Proceedings of the Annual Meeting of the Cognitive Science Society.  
+Noam Siegelman, Sascha Schroeder, Cengiz Acartürk, Hee-Don Ahn, Svetlana Alexeeva, Simona Amenta, Raymond Bertram, Rolando Bonandrini, Marc Brysbaert, Daria Chernova, et al. 2022. Expanding horizons of cross-linguistic research on reading: The Multilingual Eye-movement Corpus (MECO). Behavior Research Methods, 54(6):2843-2863.  
+Nathaniel J Smith and Roger Levy. 2013. The effect of word predictability on reading time is logarithmic. Cognition, 128(3):302-319.  
+Rosy Southwell, Julie Gregg, Robert Bixler, and Sidney K D'Mello. 2020. What eye movements reveal about later comprehension of long connected texts. Cognitive Science, 44(10):e12905.  
+Robyn Speer. 2022. rspeer/wordfreq: v3.0.  
+Jannis Vamvas, Johannes Graen, and Rico Sennrich. 2023. Swissbert: The multilingual language model for switzerland. In Proceedings of the 8th edition of the Swiss Text Analytics Conference, pages 54-69.
+
+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 Neural Information Processing Systems, volume 30. Curran Associates, Inc.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi 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 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.  
+Peng Xu, Xiatian Zhu, and David A. Clifton. 2023. Multimodal Learning With Transformers: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132. Conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence.  
+Duo Yang and Nora Hollenstein. 2023. PLM-AS: Pre-trained Language Models Augmented with Scanpaths for Sentiment Classification. Proceedings of the Northern Lights Deep Learning Workshop, 4.  
+Linan Zhu, Zhechao Zhu, Chenwei Zhang, Yifei Xu, and Xiangjie Kong. 2023. Multimodal sentiment analysis based on fusion methods: A survey. Information Fusion, 95:306-325.
+
+<table><tr><td>Feature Name</td><td>Description</td></tr><tr><td colspan="2">Word-Level Eye Movement Features</td></tr><tr><td>IA_DWELL_TIME</td><td>The sum of the duration across all fixations that fell in the current interest area</td></tr><tr><td>IA_DWELL_TIME_%</td><td>Percentage of trial time spent on the current interest area (IA_DWELL_TIME / TRIAL_DWELL_TIME).</td></tr><tr><td>IA FIXATION_%</td><td>Percentage of all fixations in a trial falling in the current interest area.</td></tr><tr><td>IA FIXATION_COUNT</td><td>Total number of fixations falling in the interest area.</td></tr><tr><td>IA_REGRESSION_IN_COUNT</td><td>Number of times interest area was entered from a higher IA_ID (from the right in English).</td></tr><tr><td>IA_REGRESSION_OUT_FULL_COUNT</td><td>Number of times interest area was exited to a lower IA_ID (to the left in English).</td></tr><tr><td>IA_run_COUNT</td><td>Number of times the Interest Area was entered and left (runs).</td></tr><tr><td>IA_FIRST_fix_PROGRESSIVE</td><td>Checks whether the first fixation in the interest area is a first-pass fixation.</td></tr><tr><td>IA_FIRST_fixATION_duration</td><td>Duration of the first fixation event that was within the current interest area</td></tr><tr><td>IA_FIRST_fixATION_VISITED_IA_COUNT</td><td>This reports the number of different interest areas visited so far before the first fixation is made to the current interest area.</td></tr><tr><td>IA_FIRST_run_DWELL_TIME</td><td>Dwell time of the first run (i.e., the sum of the duration of all fixations in the first run of fixations within the current interest area).</td></tr><tr><td>IA_FIRST_run-fixATION_COUNT</td><td>Number of all fixations in a trial falling in the first run of the current interest area.</td></tr><tr><td>IA_SKIP</td><td>An interest area is considered skipped (i.e., IA_SKIP = 1) if no fixation occurred in first-pass reading.</td></tr><tr><td>IA_TOP</td><td>Y coordinate of the top of the interest area.</td></tr><tr><td>IA_LEFT</td><td>X coordinate of the left-most part of the interest area.</td></tr><tr><td>normalizedWord_ID</td><td>Position in the paragraph of the word interest area, normalized from zero to one.</td></tr><tr><td>IA_REGRESSION_PATH.Duration</td><td>The summed fixation duration from when the current interest area is first fixated until the eyes enter an interest area with a higher IA_ID.</td></tr><tr><td>IA_REGRESSION_OUT_COUNT</td><td>Number of times interest area was exited to a lower IA_ID (to the left in English) before a higher IA_ID was fixated in the trial.</td></tr><tr><td>IA_SELECTIVE_REGRESSION_PATH.Duration</td><td>Duration of fixations and refixations of the current interest area before the eyes enter an interest area with a higher ID.</td></tr><tr><td>IA LAST FIXATION-duration</td><td>Duration of the last fixation event that was within the current interest area.</td></tr><tr><td>IA LAST_run_DWELL_TIME</td><td>Dwell time of the last run (i.e., the sum of the duration of all fixations in the last run of fixations within the current interest area).</td></tr><tr><td>PARAGRAPH_RT</td><td>Reading time of the entire paragraph.</td></tr><tr><td>total SKIP</td><td>Binary indicator whether the word was fixated on.</td></tr><tr><td colspan="2">Fixation-level Eye Movement Features</td></tr><tr><td>CURRENT FIX_INDEX</td><td>The position of the current fixation in the trial.</td></tr><tr><td>CURRENT FIX_DURATION</td><td>Duration of the current fixation.</td></tr><tr><td>CURRENT FIX_PUPIL</td><td>Average pupil size during the current fixation.</td></tr><tr><td>CURRENT FIX_X</td><td>X coordinate of the current fixation.</td></tr><tr><td>CURRENT FIX_Y</td><td>Y coordinate of the current fixation.</td></tr><tr><td>NEXT FIXAngle, PREVIOUS FIXAngle</td><td>Angle between the horizontal plane and the line connecting the current fixation and the next/previous fixation.</td></tr><tr><td>NEXT FIXDistance, PREVIOUS FIXDistance</td><td>Distance between the current fixation and the next/previous fixation in degrees of visual angle.</td></tr><tr><td>NEXT_SAC_AMPLITURE</td><td>Amplitude of the following saccade in degrees of visual angle.</td></tr><tr><td>NEXT_SACAngle</td><td>Angle between the horizontal plane and the direction of the next saccade.</td></tr><tr><td>NEXT_SAC_AVG_VELOCITY</td><td>Average velocity of the next saccade.</td></tr><tr><td>NEXT_SAC_DURATION</td><td>Duration of the next saccade in milliseconds.</td></tr><tr><td>NEXT_SAC_PEAK_VELOCITY</td><td>Peak values of gaze velocity (in visual degrees per second) of the next saccade.</td></tr></table>
+
+Table 4: Word-level and fixation-level eye movement features, defined and extracted by SR Data Viewer.  
+
+<table><tr><td>Feature Name</td><td>Description</td></tr><tr><td>Surprisal</td><td>(Hale, 2001; Levy, 2008), formulated as - log2(p(word|context)) for each word given the preceding textual content of the paragraph as context, probabilities extracted from the GPT-2-small language model (Radford et al., 2019; Wolf et al., 2020).</td></tr><tr><td>Wordfreq_Frequency</td><td>Frequency of the word based on the Wordfreq package (Speer, 2022), formulated as - log2(p(word)).</td></tr><tr><td>Length</td><td>Length of the word in characters.</td></tr><tr><td>start_of_line</td><td>Binary indicator of whether the word appeared at the beginning of a line.</td></tr><tr><td>end_of_line</td><td>Binary indicator of whether the word appeared at the end of a line.</td></tr><tr><td>Is_Content_Word</td><td>Binary indicator of whether the word is a content word.</td></tr><tr><td>n_Lefts</td><td>A content word is defined as a word that has a part-of-speech tag of either PROPN, NOUN, VERB, ADV, or ADJ.</td></tr><tr><td>n_Rights</td><td>The number of leftward immediate children of the word in the syntactic dependency parse.</td></tr><tr><td>Distance2Head</td><td>The number of rightward immediate children of the word in the syntactic dependency parse.</td></tr></table>
+
+Table 5: Linguistic word properties and their descriptions. POS tags and parse trees were obtained using SpaCy (Honnibal et al., 2020).
+
+# B Adaptations of Prior Models
+
+# B.1 MAG
+
+We replace the vision and acoustic input with word-level eye movement features. To align them with the tokenized text, we duplicate the word-level features for each subword token. Additionally, for a fair comparison with other models, we replace BERT with RoBERTa $_{\text{LARGE}}$  as the textual backbone model.
+
+Formally, each token embedding  $Z_{i}$  is displaced by  $H_{i}$ .
+
+$$
+\bar {Z} _ {i} = Z _ {i} + \alpha H _ {i} \tag {5}
+$$
+
+$H_{i}$  is a scaled and transformed version of the eye movements  $E_{i}$ ,
+
+$$
+H _ {i} = g _ {i} \cdot \left(W _ {e} E _ {i}\right) + b _ {H} \tag {6}
+$$
+
+where the scaling is defined by,
+
+$$
+g _ {i} = \operatorname {R e L U} \left(W _ {g} \left[ Z _ {i}; A _ {i} \right] + b _ {g}\right) \tag {7}
+$$
+
+The amount of displacement is defined by
+
+$$
+\alpha = \min  \left(\frac {\left| \left| Z _ {i} \right| \right| _ {2}}{\left| \left| H _ {i} \right| \right| _ {2}} \beta , 1\right) \tag {8}
+$$
+
+where  $\beta$  is a hyper-parameter, and  $W_{e}, W_{g}, b_{H}, b_{g}$  are learned.
+
+Finally, the contextualized CLS token is used for classification.
+
+# B.2 Eyettention
+
+We adjust the prediction objective of the model from next fixation to trial-level classification. To this end, we use global cross attention between the word sequence and the scanpath sequence instead of fixed window cross attention, as suggested in Deng et al. (2023). We then represent the whole scanpath using the last hidden representation of the scanpath LSTM. We further replace BERT, with RoBERTa $_{\text{LARGE}}$  for consistency with the other models.
+
+# B.3 BEyeLSTM
+
+First, we employ SpaCy tokenization based on paragraph-level input rather than word-level input, resulting in a more precise tokenization. Second, the textual materials used here include a more fine-grained set of part-of-speech tags and named entities, which results in a larger final feature set. Lastly, we omit the "words in fixed context on unigrams" feature, as it presupposes that all the participants read the same texts, which is not the case in OneStop.
+
+# B.4 CNN
+
+Ahn et al. (2020) resort to artificially subdividing SB-SAT texts into smaller segments in order generate a sufficient number of training examples to make the dataset usable for their task of predicting low versus high comprehension over multiple items. This heuristic is problematic in general, and not applicable to the single item task addressed here. In the current work we use the entire fixation sequence as the input to the model.
+
+# C Cross Validation Splits
+
+Each split guarantees an equal number of participants from each OneStopQA batch in each portion of the split, and is approximately stratified by answer type. Recall that each participant is presented with a specific combination of a paragraph and one of its three associated questions. Due to the stratification by answer type, it is not guaranteed that the appearances of any given paragraph will be balanced across the three possible questions in any of the split portions. Note that across the 10 test sets, not all participant - item combinations are covered in the test sets, as this would require 100 data splits.
+
+# D Feature Standardization and Hyperparameter Tuning
+
+We apply standardization for each feature in  $E_P$ , where the statistics are computed on the train set and applied to the validation and test sets, separately for each split. Feature normalization is performed using Scikit-learn (Pedregosa et al., 2011).
+
+For all the neural models, we use the AdamW optimizer (Loshchilov and Hutter, 2018) with a batch size of 16, a linear warmup ratio of 0.1, and a weight decay of 0.1, following best practice recommendations from Liu et al. (2019) and Mosbach et al. (2021). The search space for learning rates is  $\{0.00001, 0.00003, 0.0001\}$  and for dropout  $\{0.1, 0.3, 0.5\}$ .
+
+- For Logistic Regression, we search over regularization parameter C values of  $\{0.1, 5, 10, 50, 100\}$  with and without an L2 penalty.  
+- For the CNN we include a learning rate of 0.001 as in Ahn et al. (2020).  
+- Following (Reich et al., 2022), for BEyeLSTM the search space for learning rates is  $\{0.001, 0.003, 0.01\}$ , embedding dimensions of  $\{4,8\}$  and hidden dimension of  $\{64,128\}$ .  
+- For Eyettention we also include a learning rate of 0.001 and dropout of 0.2, as in Deng et al. (2023).  
+- For MAG-QEye, the search space for the injection layer index is  $\{0,11,23\}$ . We set the MAG-internal dropout to 0.5, and the scalar parameter to 1e-3, as suggested by (Rahman et al., 2020).  
+- In PostFusion-QEye, the 1D convolution layers have a kernel size of three, stride 1, and padding 1.
+
+All neural networks are trained using the Pytorch Lighting library (Falcon and The PyTorch Lightning team, 2019; Paszke et al., 2019) and evaluated using torch-metrics (Nicki Skafte Detlefsen et al., 2022) on a NVIDIA A100-40GB and A40-48GB GPUs. We adapt Huggingface's RoBERTa implementation (Wolf et al., 2020). The baselines described in Section 5.5 are reimplemented in this framework as well. A single training epoch took approximately 5 minutes. We train for a maximum of ten epochs, stopping after three epochs without improvement on the validation set.
+
+The number of model parameters is 355M for the RoBERTa $_{\text{LARGE}}$  backbone, and an additional 1.1M for MAG-QEye and RoBERTa-QEye, and 9M for PostFusion-QEye.
+
+# E The Role of Linguistic Word Property Features
+
+Our proposed models tend to outperform the Text-only RoBERTa baseline, especially in the two evaluation regimes that involve new items. Note however, that in addition to eye movements, these models also include linguistic word properties, which may provide information on the textual item that is not fully encoded in word embeddings. Some of them (e.g. word length, frequency and surprisal) are also known to be predictive of reading times.
+
+What is the effect of these features on model performance? To examine this question, we carry out two ablation experiments. In the first experiment, we ablate the linguistic word property features. In the second experiment we ablate the eye movement features. The latter ablation is not possible with fixation based models, because even with the eye movement features removed, these models still have information about the gaze trajectory through the order and word identity of the fixations. We therefore perform these experiments only with the word based models RoBERTa-QEye-Words and MAG-QEye.
+
+Table 6 in Appendix E presents the ablation results for the binary task. In the first experiment, removal of linguistic word properties does not substantially affect model performance. This outcome does not match our expectation regarding the potential benefits of allowing models to learn eye movement – linguistic word property interactions. In the second experiment, overall, we again do not observe performance degradation when ablating the eye movement features. While this experiment is not sufficient for drawing general conclusions regarding the value of eye movement information for our task, it suggests that in our two instances of word-based models, eye movements do not seem to provide substantial performance gains above and beyond features that can be readily extracted from the text. We leave a more extensive investigation regarding the impact of linguistic features on model performance to future work.
+
+<table><tr><td>Binary Reading Comprehension</td><td colspan="4">Gathering Trials</td><td colspan="4">Hunting Trials</td></tr><tr><td>Model</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td></tr><tr><td>Text-only RoBERTa</td><td>54.8</td><td>63.1</td><td>55.2</td><td>58.7</td><td>51.8</td><td>63.1</td><td>50.5</td><td>57.1</td></tr><tr><td>MAG-QEye</td><td>54.8</td><td>64.1*</td><td>53.8</td><td>59.2</td><td>52.5</td><td>62.3</td><td>51.3</td><td>57.1</td></tr><tr><td>MAG-QEye w/o Ling. Feat</td><td>55.9</td><td>63.8</td><td>55.5</td><td>59.6</td><td>52.3</td><td>63.3</td><td>54.8</td><td>57.7</td></tr><tr><td>MAG-QEye w/o Eyes</td><td>54.2</td><td>63.7</td><td>56.7</td><td>58.8</td><td>51.9</td><td>63.3</td><td>53.8</td><td>57.4</td></tr><tr><td>RoBERTa-QEye-Words</td><td>55.5</td><td>63.5</td><td>52.1</td><td>59.1</td><td>50.5</td><td>63.8</td><td>51.0</td><td>56.8</td></tr><tr><td>RoBERTa-QEye-Words w/o Ling. Feat</td><td>55.4</td><td>63.3</td><td>56.3</td><td>59.2</td><td>51.1</td><td>62.7</td><td>50.7</td><td>56.6</td></tr><tr><td>RoBERTa-QEye-Words w/o Eyes</td><td>56.7*</td><td>63.7</td><td>57.5</td><td>60.0**</td><td>49.3</td><td>63.2</td><td>51.2</td><td>56.0</td></tr></table>
+
+Table 6: The effect of ablating word-level eye movement features (Table 4) and linguistic word properties (Table 5) on balanced accuracy for binary classification of the word based models MAG-QEye and RoBERTa-QEye-Words. Statistically significant improvements over Text-only RoBERTa, using a paired bootstrap test, are marked with \* at  $p <   0.05$ $^{**}$  at  $p <   0.01$  and \*\*\* at  $p <   0.001$
+
+# F Textual Backbone Variants
+
+Our models use RoBERTa as a textual backbone model, and the parameters of this backbone are subjected to change during model training. Other choices for this model component are possible. For example, one can pre-train the model on multiple choice question answering, freeze the textual backbone parameters during model training, or choose a different textual backbone model altogether. Preliminary experiments with MAG-QEye in Appendix F Table 7 do not show a consistent effect of these choices on model performance in the main prediction task. We leave a comprehensive investigation of textual backbone model choice and training to future work.
+
+<table><tr><td>Binary Reading Comprehension</td><td colspan="4">Gathering Trials</td><td colspan="4">Hunting Trials</td></tr><tr><td>MAG-QEye Backbone</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td></tr><tr><td>RoBERTa Large</td><td>54.8</td><td>64.1</td><td>53.8</td><td>59.2</td><td>52.5</td><td>62.3</td><td>51.3</td><td>57.1</td></tr><tr><td>RoBERTa Large Frozen</td><td>54.3</td><td>61.4</td><td>51.4</td><td>57.5</td><td>51.9</td><td>60.0</td><td>53.3</td><td>55.8</td></tr><tr><td>RoBERTa Large Trained for QA on RACE</td><td>54.8</td><td>64.6</td><td>52.7</td><td>59.3</td><td>48.3</td><td>62.7</td><td>44.9</td><td>54.9</td></tr><tr><td>RoBERTa Base</td><td>52.8</td><td>64.0</td><td>56.9</td><td>58.3</td><td>50.8</td><td>63.5*</td><td>51.6</td><td>56.9</td></tr></table>
+
+Table 7: Balanced accuracy performance comparison of different backbone architectures and training strategies for MAG-QEye. Statistically significant improvements compared to an unfrozen RoBERTa Large backbone are marked with ‘*’ at  $p < {0.05}$  ,‘**’ at  $p < {0.01}$  and ‘***’ at  $p < {0.001}$  using a paired bootstrap test.
+
+# G Validation Set Results
+
+<table><tr><td colspan="3">Binary Reading Comprehension</td><td colspan="4">Ordinary Reading (Gathering)</td><td colspan="4">Information Seeking (Hunting)</td></tr><tr><td>Model</td><td>Gaze Representation</td><td>Text Representation</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td></tr><tr><td>Majority</td><td>None</td><td>None</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td><td>50.0</td></tr><tr><td>Text-only RoBERTa</td><td>None</td><td>Emb</td><td>59.8</td><td>65.8</td><td>57.9</td><td>62.5</td><td>57.1</td><td>65.1</td><td>56.8</td><td>60.8</td></tr><tr><td>Log. Reg. (Mézière et al., 2023b)</td><td>Global</td><td>None</td><td>53.4</td><td>51.1</td><td>53.9</td><td>52.3</td><td>51.8</td><td>53.0</td><td>51.9</td><td>52.4</td></tr><tr><td>CNN (Ahn et al., 2020)</td><td>Fixations</td><td>None</td><td>53.3</td><td>53.7</td><td>53.4</td><td>53.5</td><td>55.1</td><td>54.5</td><td>55.0</td><td>54.8</td></tr><tr><td>BEyeLSTM (Reich et al., 2022)</td><td>Fixations</td><td>Ling. Feat.</td><td>55.0</td><td>58.5</td><td>55.7</td><td>56.7</td><td>57.3</td><td>58.6</td><td>58.3</td><td>58.0</td></tr><tr><td>Eyettention (Deng et al., 2023)</td><td>Fixations</td><td>Emb + Word Len.</td><td>58.5</td><td>62.4</td><td>57.9</td><td>60.3</td><td>57.0</td><td>59.5</td><td>56.9</td><td>58.2</td></tr><tr><td>RoBERTa-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>57.0</td><td>65.5</td><td>60.5</td><td>61.2</td><td>55.3</td><td>64.7</td><td>52.2</td><td>59.6</td></tr><tr><td>RoBERTa-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>57.0</td><td>63.5</td><td>60.4</td><td>60.3</td><td>54.6</td><td>62.4</td><td>56.5</td><td>58.4</td></tr><tr><td>MAG-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>60.4</td><td>65.8</td><td>58.9</td><td>62.9</td><td>57.3</td><td>66.0</td><td>59.5</td><td>61.6</td></tr><tr><td>PostFusion-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>60.1</td><td>65.2</td><td>60.4</td><td>62.5</td><td>58.3</td><td>65.8</td><td>59.3</td><td>61.9*</td></tr></table>
+
+Table 8: Balanced accuracy for the binary reading comprehension prediction task (correct vs incorrect comprehension).  
+
+<table><tr><td colspan="3">Multiple-Choice Reading Comprehension</td><td colspan="4">Ordinary Reading (Gathering)</td><td colspan="4">Information Seeking (Hunting)</td></tr><tr><td>Model</td><td>Gaze Representation</td><td>Text Representation</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td><td>New Item</td><td>New Participant</td><td>New Item &amp; Participant</td><td>All</td></tr><tr><td>Majority</td><td>None</td><td>None</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td><td>25.0</td></tr><tr><td>Text-only RoBERTa</td><td>None</td><td>Emb</td><td>25.7</td><td>35.7</td><td>25.6</td><td>30.4</td><td>25.0</td><td>34.4</td><td>25.5</td><td>29.5</td></tr><tr><td>RoBERTa-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>34.0***</td><td>34.4</td><td>37.4**</td><td>34.3***</td><td>33.3***</td><td>34.3</td><td>32.9</td><td>33.7*</td></tr><tr><td>RoBERTa-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>33.6***</td><td>34.7</td><td>37.9***</td><td>34.3***</td><td>34.0***</td><td>34.4</td><td>37.4</td><td>34.3***</td></tr><tr><td>MAG-QEye</td><td>Words</td><td>Emb + Ling. Feat.</td><td>33.8***</td><td>36.1</td><td>34.3**</td><td>34.9**</td><td>34.8***</td><td>33.6</td><td>32.9</td><td>34.1***</td></tr><tr><td>PostFusion-QEye</td><td>Fixations</td><td>Emb + Ling. Feat.</td><td>33.2***</td><td>35.1</td><td>33.5*</td><td>34.1**</td><td>34.0**</td><td>31.8</td><td>35.4</td><td>33.0</td></tr></table>
+
+Table 9: Balanced accuracy for the multiple-choice specific answer prediction task.

2024/Fine-Grained Prediction of Reading Comprehension from Eye Movements/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe7df23e5bd3cbaf1f532dc2a69a9ed359f0084525595c4ecff08f202f2e0752
+size 845153

2024/Fine-Tuning Large Language Models to Translate_ Will a Touch of Noisy Data in Misaligned Languages Suffice_/8e3a98d3-e6ed-4c7a-afff-0b0d7bc4b4fb_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1696dae110b989a6feae60026ab0e8bda82559a7a4cadd6fcd2af20612da0a8b
+size 1273730

2024/Fine-Tuning Large Language Models to Translate_ Will a Touch of Noisy Data in Misaligned Languages Suffice_/full.md

@@ -0,0 +1,445 @@
+# Fine-Tuning Large Language Models to Translate: Will a Touch of Noisy Data in Misaligned Languages Suffice?
+
+Dawei Zhu
+
+Barry Haddow
+
+Pinzhen Chen
+
+Xiaoyu Shen $^{3*}$
+
+Miaoran Zhang
+
+Dietrich Klakow<sup>1</sup>
+
+$^{1}$ Saarland University, Saarland Informatics Campus
+
+$^{3}$ Digital Twin Institute, Eastern Institute of Technology, Ningbo
+
+{dzhu,mzhang}@1sv.uni-saarland.de
+
+2University of Edinburgh
+
+f Technology, Ningbo
+
+pinzhen.chen@ed.ac.uk
+
+# Abstract
+
+Traditionally, success in multilingual machine translation can be attributed to three key factors in training data: large volume, diverse translation directions, and high quality. In the current practice of fine-tuning large language models (LLMs) for translation, we revisit the importance of these factors. We find that LLMs display strong translation capability after being fine-tuned on as few as 32 parallel sentences and that fine-tuning on a single translation direction enables translation in multiple directions. However, the choice of direction is critical: fine-tuning LLMs with only English on the target side can lead to task misinterpretation, which hinders translation into non-English languages. Problems also arise when noisy synthetic data is placed on the target side, especially when the target language is well-represented in LLM pre-training. Yet interestingly, synthesized data in an under-represented language has a less pronounced effect. Our findings suggest that when adapting LLMs to translation, the requirement on data quantity can be eased but careful considerations are still crucial to prevent an LLM from exploiting unintended data biases. $^{1}$
+
+# 1 Introduction
+
+Large language models (LLMs) have reached new heights in various NLP tasks (Radford et al., 2019; Brown et al., 2020; Touvron et al., 2023; Jiang et al., 2023). Supervised fine-tuning (SFT, Ouyang et al., 2022, alternatively, instruction tuning or simply fine-tuning in some literature) further prepares these models for better generalization and reliability in downstream tasks by training on task input-output data combined with instructions in natural languages (Sanh et al., 2022; Wei et al., 2022; Mishra et al., 2022). In this research direction, various works have studied the "scaling up" of SFT
+
+data size, number of languages, etc (Chung et al., 2024; Muennighoff et al., 2023). On the other hand, recent papers also embraced the philosophy of "less is more" by achieving strong results with a small set of high-quality training instances, claiming a "superficial alignment hypothesis" (Zhou et al., 2023) with similar findings by others.
+
+This work investigates the role of SFT data in aligning LLMs to machine translation (MT), a cross-lingual generation task with high demands in practical domains. Prior research has found fine-tuning to improve translation performance (Zhang et al., 2023c) and more recent works also integrated continued pre-training with more data to provide further improvement (Xu et al., 2024a; Alves et al., 2024). For encoder-decoder models, Wu et al. (2024a) used little data to enable an English-centric model to translate between any two languages. Nonetheless, the feasibility of "less is more" in LLM translation fine-tuning is rather under-explored. In translation prompting, researchers have suggested that a model's translation capability can be attributed to the bilingual signals exposed during pre-training (Briakou et al., 2023) and task recognition in LLM layers (Sia et al., 2024), hinting that the translation capability has been picked up during pre-training. A natural question follows: Can we put reduced effort into data?
+
+From a data efficiency perspective, we squeeze the translation SFT data to a mere size of 32 or the translation direction to 1 for multilingual translation, for which we believe LLMs already possess a strong pre-trained foundation in multilingual understanding and generation. Beyond quantity and language diversity, we perform SFT on synthesized data via machine translation, which is a common data augmentation practice for under-served languages. To summarize, our analysis is grounded in the task of MT, with "scaling down" in mind. In multiple dimensions—data size (\$3.2), translation direction (\$3.3 and $3.4), and data synthesis
+
+$(\S 3.5)$  — our findings verify, complement, and refine the existing superficial alignment hypothesis for fine-tuning LLMs for translation tasks:
+
+1. 32 data instances successfully enable an LLM to translate in 11 directions. More data still helps but the return diminishes.  
+2. Data in a single translation direction can effectively align an LLM to translate to and from multiple directions. Yet, it is crucial to pick the right direction—we recommend not placing English on the target side.  
+3. When fine-tuning on lower-quality synthetic data, LLMs are affected if the data is placed on the target side, but they show greater resilience against such flaws in low-resource languages, which are less represented during pre-training.
+
+# 2 Preliminaries
+
+# 2.1 Supervised fine-tuning
+
+In this work, we perform SFT to prepare pre-trained LLMs for MT. Let  $S$  denote a source input and  $T = [t_1,t_2,\dots,t_{|T|}]$  denote a target-side reference. We start with placing the input into a prompt template by applying  $\mathcal{I}(\cdot)$  to  $S$ . For each training instance, the instruction template is randomly selected from a pre-defined pool. We fine-tune an LLM parameterized by  $\theta$  by optimizing the log-likelihood:
+
+$$
+\begin{array}{l} \mathcal {L} _ {S F T} (\mathcal {I} (S), T; \theta) = - \log P (T | \mathcal {I} (S); \theta) \\ = - \log \prod_ {k = 1} ^ {| T |} P \left(t _ {k} \mid t _ {<   k}, \mathcal {I} (S); \theta\right) \\ = - \sum_ {k = 1} ^ {| T |} \log P \left(t _ {k} \mid t _ {<   k}, \mathcal {I} (S); \theta\right) \\ \end{array}
+$$
+
+# 2.2 Superficial alignment hypothesis
+
+Zhou et al. (2023) claim that a model's knowledge and capabilities are acquired almost entirely during pre-training, and the effect of alignment tuning might be "superficial", in that it teaches the model the format for interacting with users. This idea is further supported by recent works (Lin et al., 2024; Ghosh et al., 2024). However, to what extent this applies to multilingual translation in LLMs is little known. To bridge this gap, we conduct a series of controlled experiments on fine-tuning LLMs for translation, complementing previous research across three dimensions. First, we study the parallel
+
+data efficiency in the era of LLMs, aiming to determine the minimum data needed for effective model alignment to the translation task. Next, we explore the scope of alignment by probing whether aligning one translation direction influences other directions. Finally, we investigate how synthesized fine-tuning data quality impacts the LLMs' behaviour in generating translations.
+
+# 3 Experiments and Results
+
+# 3.1 Experimental setup
+
+Training. By default, we take the test sets from WMT17 to WMT20 as our parallel training data (Bojar et al., 2017, 2018; Barrault et al., 2019, 2020); we also use the development sets in WMT21 (Akhbardeh et al., 2021) for training if a language pair of interest is not available in earlier years. The specific training data configurations will be detailed in the subsequent sections. The test sets from WMT21 are used for validation. Detailed data statistics can be found in Appendix F.1. The LLM we use for SFT is the base version of Llama-2 7B (Touvron et al., 2023). When performing SFT, we use a learning rate of 5e-6, an effective batch size of 64, and a linear learning rate scheduling with a warmup ratio of 0.1. We select the model checkpoint based on COMET scores on the validation sets. To form the model input for SFT, we feed the source sentence into the Alpaca prompt template (Taori et al., 2023), supplementing it with a translation instruction that is randomly selected from a pool of 31 diverse instructions. Refer to Table 4 in the appendix for a complete list of templates.
+
+Evaluation. We primarily evaluate the models on the WMT22 test sets (Kocmi et al., 2022) covering 11 translation directions: en←cs, en←de, en←jp, en←ru, en←zh, and en→hr. Languages in these 11 directions are explicitly included in Llama-2's pre-training corpus. In Section 3.4, we extend our evaluation to translation directions involving medium and low resource languages: Icelandic and Hausa (i.e., en←is, en←ha), which comes from WMT21's test set. At inference time, a fixed translation instruction is applied (Table 4 row 1). We
+
+![](images/da086094ec00749048aa286285f533d9e15a64a364aa16b1df29f825afb8f630.jpg)  
+Figure 1: Performance comparison between instruction-tuned baselines and Llama-2 fine-tuned with different training data sizes. Average COMET (left) and BLEU (right) scores across 11 translation directions are presented. For training data sizes of 1 and 3, ICL is applied, marked with an asterisk “*”; otherwise, we perform SFT. With only 32 training examples for SFT, Llama-2 outperforms general-purpose, instruction-tuned baselines. Base.: instruction-tuned baseline models. See individual performance for the 11 translation directions in Appendix A.
+
+use beam search with a beam size of 4 for generation, as our preliminary results indicate that it offers better translation quality than sampling-based generation, an observation consistent with recent works (Jiao et al., 2023; Zeng et al., 2024). The maximum generation length is set to 256 tokens. We used a reference-based COMET22 checkpoint $^{4}$  (Rei et al., 2020) and BLEU (Papineni et al., 2002) as the evaluation metrics. See Appendix F.3 for detailed software configurations.
+
+# 3.2 How much SFT data enables LLMs to translate?
+
+Recent works in machine translation suggest that pre-trained LLMs require significantly less parallel data for fine-tuning (via SFT), compared to training conventional translation models from scratch. However, the SFT process in these works still operates with an order of  $10^{5}$  parallel samples (Jiao et al., 2023; Zhang et al., 2023c; Zeng et al., 2024; Xu et al., 2024a, i.a.), without a clear justification for selecting this specific data size and source. This raises a pivotal question, inspired by the recently proposed "superficial alignment hypothesis" (Zhou et al., 2023): Is SFT mainly a method for superficially aligning LLMs for translation tasks? If so, what is the actual minimal amount of data required to achieve effective "alignment"?
+
+Setup. We fine-tune Llama-2 7B using different numbers of training samples and evaluate the multilingual translation performance of the resulting models. We collect training data covering 10 translation directions:  $\mathrm{en} \leftrightarrow \{\mathrm{cs}, \mathrm{de}, \mathrm{jp}, \mathrm{ru}, \mathrm{zh}\}$ . The training data sourced from WMT17-20 contains a
+
+total of 74,623 parallel examples. Note that the training samples across translation directions are not evenly distributed. To create training sets of varying sizes, we subsample the original data into subsets that are powers of 2, starting from  $16(2^{4})$  and ending with  $4096(2^{12})$ ; larger subsets always contain smaller ones. To ensure balanced language representation in our subsets, we distribute samples as evenly as possible among the language pairs. $^5$
+
+We refer to the fine-tuned model as SFT-MT. Considering LLMs can also perform translation through prompting, we compare SFT-MT with 1- and 3-shot in-context learning (ICL), denoted as ICL-MT. For ICL, we randomly select demonstrations from the training set in the test direction for each test sentence. We do not consider Llama-2's zero-shot performance because, although it sometimes produces acceptable translations in the beginning, it often continues generating, which makes it difficult to accurately estimate its performance. Lastly, since LLMs fine-tuned on diverse tasks also serve as strong translation systems (Zhu et al., 2024), we compare our models with open-source general-purpose instruction-tuned LLMs, which we denote as IT-LLM. These include Vicuna-v1.5-7b (Chiang et al., 2023), Mistral-7b-Instruct (Jiang et al., 2023), and Llama-2-7b-chat (Touvron et al., 2023).
+
+Results. Figure 1 illustrates the effect of varying training sizes on translation performance. In both 1- and 3-shot cases, ICL-MT underperforms IT-LLM baselines like Llama-2-7b-chat despite sharing the
+
+![](images/6408bb49373507ea01e8ad38412677613174d1f06bfb6445948c8b0608179803.jpg)  
+Figure 2: Normalized COMET score (as a % of performance from fine-tuning on an equivalent sized dataset of all 10 directions) resulted from varying combinations of train and test translation directions. In most cases, Llama-2 fine-tuned on a single translation direction can effectively translate across other directions, achieving performance comparable to models trained on all directions, with a few exceptions when trained on  $\mathrm{X}\rightarrow \mathrm{en}$  but tested on  $\mathrm{en}\rightarrow \mathrm{X}$ . Performance measured in BLEU score is provided in Appendix B.
+
+same foundation model, indicating that a few incontext demonstrations may not effectively align Llama-2 for translation.
+
+However, performance significantly improves when Llama-2 is fine-tuned with just 16 samples. With further increases in the training size to 32 samples, Llama-2 performs on par with or surpasses all three IT-LLM baselines in both COMET and BLEU metrics. This suggests that a handful of high-quality parallel data can effectively specialize the model into a performant translation system. Increasing parallel data further boosts performance, though with diminishing returns: the COMET score rises by an average of 2 points when expanding from 32 to 1024 samples, but only by 0.5 points when increasing further from 1024 to 75K samples (full training set). Given that it is unlikely that these 32 training samples "teach" Llama-2 new translation skills, this shows strong evidence that superficial alignment applies to MT. We observe a similar trend in Mistral-7B and Llama-2-13B. Refer to Appendix A for their performance across varying data sizes. In summary, effective translation alignment begins with minimal training data, revealing less is good alignment and more is better with diminishing gains.
+
+# 3.3 Do we need to include all directions?
+
+In the preceding section, we follow the traditional practice in multilingual MT by including multiple
+
+translation directions during training. However, the observation that only a few dozen examples make Llama-2 translate well leads us to reconsider the necessity of including samples from all directions of interest. Specifically, will training on just a single translation direction be sufficient to help LLMs perform multilingual translation?
+
+Setup. We explore six training configurations, each focusing on a single translation direction: de→en, zh→en, en→de, en→zh, fr→de, and de→fr. These configurations include cases where English appears on the source side, the target side, as well as settings with English excluded, to investigate if specific languages have a different impact on the overall performance. The training size is set to 1024 for SFT. Evaluations are conducted across the same 11 test directions as used in the previous section. Additionally, we explore similar settings in ICL, where we present demonstrations with translation directions that do not match those used in evaluations, to determine if the mechanisms of both SFT and ICL exhibit similarities. Lastly, we conduct a joint evaluation, progressively expanding both the training size and the range of covered translation directions to understand the combined effect of these factors.
+
+SFT results. Figure 2 demonstrates the normalized performance of Llama-2 when fine-tuned in various single directions. Remarkably, training
+
+<table><tr><td colspan="5">Evaluation on de→en</td></tr><tr><td rowspan="2">demo lang</td><td colspan="2">1-shot</td><td colspan="2">3-shot</td></tr><tr><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td></tr><tr><td>de→en</td><td>73.47</td><td>19.7</td><td>75.04</td><td>22.4</td></tr><tr><td>en→de</td><td>55.96</td><td>7.3</td><td>44.39</td><td>3.5</td></tr><tr><td>de→fr</td><td>66.35</td><td>12.1</td><td>64.61</td><td>17.6</td></tr><tr><td>fr→de</td><td>58.06</td><td>7.8</td><td>57.13</td><td>10.5</td></tr><tr><td>zh→en</td><td>56.66</td><td>10.7</td><td>54.82</td><td>7.1</td></tr><tr><td>en→zh</td><td>51.30</td><td>7.8</td><td>56.87</td><td>1.8</td></tr></table>
+
+<table><tr><td colspan="5">Evaluation on en→de</td></tr><tr><td rowspan="2">demo lang</td><td colspan="2">1-shot</td><td colspan="2">3-shot</td></tr><tr><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td></tr><tr><td>en→de</td><td>67.37</td><td>10.5</td><td>69.80</td><td>14.3</td></tr><tr><td>de→en</td><td>57.83</td><td>8.7</td><td>45.54</td><td>5.0</td></tr><tr><td>en→zh</td><td>59.76</td><td>9.5</td><td>59.53</td><td>8.4</td></tr><tr><td>zh→en</td><td>47.31</td><td>4.5</td><td>49.24</td><td>5.0</td></tr><tr><td>fr→de</td><td>59.36</td><td>8.6</td><td>66.01</td><td>12.9</td></tr><tr><td>de→fr</td><td>60.70</td><td>11.0</td><td>61.76</td><td>11.3</td></tr></table>
+
+Table 1: ICL-MT performance with aligned vs. misaligned demonstrations, evaluated on de→en and en→de. 1-shot/3-shot: using 1 or 3 demonstrations randomly sampled from the training set. Misaligned demonstrations consistently cause a substantial performance drop.
+
+with just one direction enables Llama-2 to translate between multiple languages. For instance, after fine-tuning on de→en or zh→en, the model can translate from all considered languages to English, scoring at least 98.6% of the original COMET scores for training on all directions. Similarly, the model fine-tuned on en→de, en→zh, fr→de or de→fr also demonstrates only a slight performance decline when translating from English.
+
+Notable declines are observed in two scenarios: (1) trained to translate to English and evaluated on translating to non-English; and (2) trained to translate to non-English and evaluated on translating to English.7 Of these two scenarios, scenario 1 exhibits a much larger performance drop. The fact that both scenarios involve a mismatch between using English and non-English suggests that Llama-2, as an English-centric LLM, may process English differently compared to other languages. When fine-tuned for English generation, the model may misinterpret the task as only generating in English. Generalization among non-English languages is much easier than generalization between English and non-English languages, as evidenced by the negligible performance drop when fine-tuning and testing on two vastly different language pairs such as de→fr and en→zh. Overall, the findings suggest that SFT in one translation direction effectively enables the many directions, though avoiding misinterpretation is crucial.
+
+ICL results. We also provide results of performing ICL with misaligned translation directions between demonstration and test in Table 1. It can be seen that misaligned demonstrations significantly degrade translation performance, with 3-shot be
+
+![](images/318344c24bfe026248fd56df5c2b34e0b7b70ec210ff850d37dd0a57886f1e18.jpg)  
+Figure 3: Average performance (in COMET) across 11 test directions for models trained with varying data sizes and directions. Both factors positively impact performance. +=: training directions added on top of previous directions; two directions are added at each time. For example, "+" covers 10 directions: en  $\leftrightarrow$  {de, zh, cs, jp, ru}. Performance on individual test directions is provided in Appendix C.
+
+often worse than 1-shot. We observe that the model may output Chinese characters, emojis, time, etc., but no clear error patterns are observed. This contrasts sharply with findings from SFT: while SFT can recognize the format of translation, ICL requires language-aligned demonstrations.
+
+Joint evaluation. Figure 3 presents a joint evaluation of size and translation direction. For small training sizes, covering diverse translation directions in training proves to be beneficial. However, the benefits of such diversity level off as the training size increases. With a training size of 1024, models trained exclusively on two directions, en←de, perform on par with those trained on all directions.
+
+![](images/ddf34c93314f7d5321a4ba9c961c4180767880d423ac007c285008f0bc91a454.jpg)  
+Figure 4: Model performance (in COMET) across 15 translation directions under different training configurations. Training models on unseen languages (en  $\leftrightarrow$  is, en  $\leftrightarrow$  ha) results in slight improvements in translating these languages compared to models trained on en  $\leftrightarrow$  de. The differences in performance when translating between seen languages are minimal across all training configurations. Performance measured in BLEU score is provided in Appendix D.
+
+![](images/4fab667bb18a52606b66778a9233885387effd3c423faf8231b9562049db3729.jpg)
+
+# 3.4 Can alignment be achieved for unseen languages?
+
+Previous sections focus on translation directions involving languages explicitly included in Llama-2's pre-training corpus. We now extend our investigation to languages that do not have an identified presence of over  $0.005\%$  in the pre-training data (c.f. Touvron et al., 2023, p22), referred to as unseen languages. Here we seek answers to two questions: (1) Can we effectively make Llama-2 translate both from and to unseen languages by fine-tuning it with a small amount of data? (2) How well can this finetuned model translate from and to languages seen in Llama?
+
+Setup. We consider three training configurations: en  $\leftrightarrow$  is, en  $\leftrightarrow$  ha, and en  $\leftrightarrow$  de, with Icelandic (is) and Hausa (ha) being unseen languages. en  $\leftrightarrow$  de serves as a control to assess Llama-2's initial translation capabilities into unseen languages without specific fine-tuning. The training size is fixed at 1024 (512 samples for each direction). The test directions include the 11 directions as before, plus en  $\leftrightarrow$  is and en  $\leftrightarrow$  ha coming from the WMT21 test.
+
+Results. The results are presented in Figure 4. It can be seen that fine-tuning on Icelandic and Hausa enhances a model's translation quality on these languages compared to the control setup, yet the gains are modest. We observe that Llama-2 manages to produce tokens in these languages, however, the translations often largely deviate from the original meanings. This suggests that it is difficult to teach models new translation directions via SFT with limited data. Interestingly, we find fine-tuning on Icelandic or Hausa does not hinder Llama-2's
+
+ability to translate from and to all seen languages, maintaining performance levels comparable to the control scenario with en  $\leftrightarrow$  de. Based on these results, we propose a complement to the superficial alignment hypothesis in MT: LLMs may learn the essence of the translation task without requiring input-output mappings in languages it "understands" well.
+
+# 3.5 Can we use synthesized data?
+
+We have observed that LLMs quickly recognize the translation task with minimal high-quality, manually curated data, but what if the quality of the training data is subpar? This situation may occur, for example when parallel data is web-crawled or machine-generated. Can LLMs still adapt to the translation task or will they overfit to the imperfections in lower-quality data, leading to degraded translation performance?
+
+Setup. We replace either the source or target sentences in the original training set with lower-quality synthesized ones. We try two types of data synthesis: one by translating entire sentences on the other side and another by concatenating word-to-word translations. Pleasingly, these correspond to back-translation (Sennrich et al., 2016) using translation engines or bilingual word dictionaries which are practical at different levels of resource availability. Specifically, we use the OPUS-MT suite (Tiedemann and Thottingal, 2020) to translate from English to a target non-English language. $^{8}$
+
+![](images/694dc8adab57d40007c2f28fa51d8b375e3ca0e1073e2cb04050593d04e555ab.jpg)
+
+![](images/1763a3ec47dcb0dc8a71d783acc85707eaaf8b984f1905250ecf1168ace37031.jpg)
+
+![](images/918d296fa5707d1e538d790af669b2b4ab41ab7a7d7fcd69ffb01315f03e4475.jpg)  
+Figure 5: Model performance in COMET score varying training sizes, directions, and noise types. Top (Bottom): score averaged across all en  $\rightarrow$  X (X  $\rightarrow$  en) test directions. Training sizes considered are 32 and 1024. Generally, introducing noise on the target side tends to degrade model performance more, with the extent of impact also depending on the particular language involved. Performance measured in BLEU score is provided in Appendix E.
+
+<table><tr><td>Source</td><td>Ref./Data config.</td><td>Model output</td></tr><tr><td rowspan="5">Das finde ich ehrlich gesagt sehr ärgerlich.</td><td>reference</td><td>That really bothers me, I must say.</td></tr><tr><td>literal</td><td>The find I honest said very annoying.</td></tr><tr><td>en→de clean</td><td>I find that really annoying.</td></tr><tr><td>en→de sent. noise</td><td>I find that honestly very annoying.</td></tr><tr><td>en→de word noise</td><td>The find I honestly said very annoying.</td></tr><tr><td rowspan="5">以免再次发生这样的事情</td><td>reference</td><td>So that such a thing won’t happen again.</td></tr><tr><td>literal</td><td>in order to avoid again happen such thing.</td></tr><tr><td>en→de clean</td><td>Let’s not let it happen again.</td></tr><tr><td>en→de sent. noise</td><td>In order not to happen again.</td></tr><tr><td>en→de word noise</td><td>Avoid again happen this way.</td></tr></table>
+
+Table 2: Examples of testing Llama-2 trained on en→de with 1024 clean and noisy target sentences. The test directions are de→en (Top) and zh→en (Bottom). The reference translation is provided by the WMT22 test set. Word-to-word references were created by the authors in consultation with native speakers. Word-level noise makes Llama-2 degenerate into a literal translator.
+
+For word-level translation, we translate each space-delimited source word by feeding it into the MT model one at a time. Naturally, the synthesized versions introduce translation errors, adding "noise" to the training process. We investigate the impact of such noise in four translation directions:  $\mathrm{en} \rightarrow \mathrm{de}'$ ,  $\mathrm{de}' \rightarrow \mathrm{en}$ ,  $\mathrm{en} \rightarrow \mathrm{ha}'$ , and  $\mathrm{ha}' \rightarrow \mathrm{en}$ , where the prime (') notation denotes the side that is created using translation (noised). We consider two training sizes: 32 and 1024. In this section, our evaluation focuses on the 11 translation directions described in Section 3.1. Note that although Hausa is included in the current training setup, translation directions involving Hausa are excluded from our evaluation—because performance is sub-par for
+
+unseen languages as demonstrated in Section 3.4.
+
+Results. According to Figure 5, it can be seen that both types of data synthesis generally cause a drop in performance. However, The degree of degradation significantly varies depending on whether the noise appears on the source or target side of the translation as well as the language. Specifically, when noise is introduced to the target side, models fine-tuned on  $\mathrm{en} \rightarrow \mathrm{de}'$  and  $\mathrm{en} \rightarrow \mathrm{ha}'$  translations exhibit a sharp decline in performance. The impact of word noise is more severe than that of sentence noise. In the case of  $\mathrm{en} \rightarrow \mathrm{de}'$ , word-level synthesis causes the model to largely degenerate, leading to literal translations across many test
+
+cases across translation directions. An example of this behaviour is presented in Table 2. In contrast, the performance drop caused by word noise is less pronounced with  $\mathrm{en} \rightarrow \mathrm{ha}'$ , particularly when evaluated on  $\mathrm{en} \rightarrow \mathrm{X}$ .
+
+Conversely, when noise is introduced on the source side, the negative impact is much smaller, and the disparity in performance degradation between the two types of noise diminishes. Even more strikingly, when evaluated on  $\mathrm{en}\rightarrow \mathrm{X}$ , having noise at the source side often outperforms the clean settings. Notably, in Section 3.3, we show that finetuning models purely on  $\mathrm{X}\rightarrow \mathrm{en}$  risks task misinterpretation, leading to low performance on  $\mathrm{en}\rightarrow \mathrm{X}$ . However, adding noise appears to mitigate this issue, resulting in improvements in both COMET and BLEU scores, especially for the  $\mathrm{ha}^{\prime}\rightarrow \mathrm{en}$  case.
+
+Summarizing the observations, Llama-2 is much more robust against the noise introduced in Hausa, likely because it has limited familiarity with the language, making it more difficult to detect and imitate imperfections present in the training data. As a result, Llama-2 tends to just recognize the essence of the translation task instead of overfitting to the biases present in low-quality data. In contrast, with German, Llama-2's understanding leads to a misinterpretation of the training objectives, such as fitting the word-level noise with a directive for literal translations. Overall, LLMs may quickly fit translation imperfections in the training data, especially for seen languages; the resulting performance drop may be observable with just 32 training samples.
+
+# 4 Related Work
+
+# 4.1 What does LLM SFT bring us?
+
+Foundational language models become more robust and follow instructions better after being fine-tuned on task-oriented supervised data formulated as natural language text (Mishra et al., 2022; Sanh et al., 2022; Wei et al., 2022). We observe diverging trends in research on instruction tuning nowadays: (1) Many works attempt to scale up instruction data in terms of the number of tasks, languages, data size, and thus implicitly increasing training updates (Chung et al., 2024; Muennighoff et al., 2023; Wu et al., 2024c; Li et al., 2023; Üstün et al., 2024; Zhang et al., 2024). (2) Another stream of papers, argue that instruction tuning mainly alters a base model's response style but not content or knowledge—data quality and diversity outweigh
+
+quantity (Zhou et al., 2023; Mitchell et al., 2024; Lin et al., 2024; Chen et al., 2024a). This work is a continued exploration of the latter, focusing on the machine translation task. We verify the effect of size variations and include two new factors—language directions and quality—aiming to provide practical and cost-effective guidance on this matter.
+
+Specifically, language transfer has been demonstrated in smaller pre-trained models before LLMs (Wu and Dredze, 2019; Artetxe et al., 2020). For (sufficiently) multilingual models, training on certain languages might still benefit other languages at the test time (Choenni et al., 2023). In LLM instruction tuning, recent papers revealed crosslingual transfer and improved robustness in unseen languages via multilingual instruction tuning with a small data sample (Chen et al., 2024c; Kew et al., 2023; Shaham et al., 2024). Furthermore, it has been claimed that even monolingual instruction tuning is sufficient to elicit multilingual responses in the correct languages with a key ingredient being the right learning rate (Chirkova and Nikoulina, 2024a,b). In relation to our experiments, language transfer to unseen languages might account for improved performance in language directions that are not directly fine-tuned.
+
+# 4.2 How can we use LLMs for translation?
+
+In the field of machine translation, earlier works provided analysis of general-purpose prompting (Vilar et al., 2023; Agrawal et al., 2023; Zhang et al., 2023a) followed by a blossom of strategies focusing on specific aspects of the translation process (Sarti et al., 2023; Ghazvininejad et al., 2023; He et al., 2024; Moslem et al., 2023; Chen et al., 2024b; Raunak et al., 2023). Nonetheless, as shown in our experimental results, few-shot prompting is not on par with using instruction-tuned models, illustrating the importance of further understanding the role of instruction tuning in translation tasks.
+
+In terms of fine-tuning LLMs for translation, previous works have explored a wide range of subtasks: disambiguation, low-resource, document-level, and adaptive translation, etc (Li et al., 2024; Zhang et al., 2023b; Alves et al., 2023; Iyer et al., 2023; Mao and Yu, 2024; Wu et al., 2024b). These works focus on improving translation performance and specific applications. Stap et al. (2024) show that while fine-tuning improves translation quality, it can degrade certain key LLMs' advantages, such as the contextualization ability on document-level input. Some recent research aims to enhance
+
+the translation capabilities of LLMs by incorporating human preference data (Jiao et al., 2023; Zeng et al., 2024; Zhu et al., 2024) or by extending the pre-training phase before fine-tuning (Xu et al., 2024a,b; Alves et al., 2024), yet these approaches require significantly more data or computing resources. The aim of this paper is not to pursue the state of the art but to investigate the opportunities of extending instruction-tuned LLMs' translation capabilities in desirable compute-efficient scenarios. It is still worth noting that our investigation is orthogonal to previous works which employ relatively large monolingual and parallel data for continued pre-training.
+
+# 5 Conclusion and Future Work
+
+In this work, we conduct an in-depth analysis of fine-tuning LLMs for translation. We demonstrate that LLMs is capable of translating in multiple directions after being fine-tuned with minimal low-quality training data in a single direction. While this suggests pre-trained LLMs inherently possess multilingual translation capabilities which only need to be unlocked by aligning with the correct task format, we discover pitfalls and lessons in aligning LLMs; while LLMs make efforts to adjust to the translation task, they are good at imitating other patterns such as the noise in the parallel data. Future work could explore robust training methods that align LLMs with translation while minimizing the risk of overfitting to low-quality data.
+
+# Limitations
+
+This work offers a range of insights into fine-tuning LLMs for translation. However, our study is not exhaustive and is subject to the following limitations.
+
+Model size and diversity. Throughout our systematic study, we fine-tuned Llama-2 7B, Llama-2 12B, and Mistral 7B. These are strong and feasible options when the work is carried out. It is important to verify the generalizability of our findings to models with different capabilities or of different sizes.
+
+Non-English centric MT. Our evaluation is English-centric, which is the condition of most LLM pre-training. Findings will be more comprehensive if future work can extend it to translation directions not involving English.
+
+State-of-the-art performance. Our research primarily explores how SFT enables LLM to trans
+
+late to uncover data-efficient strategies in SFT and identify associated pitfalls. Recent studies have demonstrated that translation capabilities can be further enhanced through techniques such as continual pre-training (Xu et al., 2024a; Alves et al., 2024) and preference learning (Xu et al., 2024b; Zhu et al., 2024). However, these methods require significantly more training resources, which may pose challenges when applied to large models.
+
+Fine-tuning methods. Throughout this work, we perform SFT with full-parameter updates. It is worthwhile to explore parameter-efficient methods which bring in heavier regularization to understand whether they exhibit patterns similar to those observed in our work.
+
+# Ethical considerations
+
+Our work's sole aim is to study the influence of data factors in applying supervised fine-tuning to large language models. We expect minimal social risks to be associated with our efforts.
+
+# Acknowledgments
+
+We sincerely thank the reviewers of this work for their constructive and insightful feedback.
+
+Pinzhen Chen and Barry Haddow received funding from UK Research and Innovation (UKRI) under the UK government's Horizon Europe funding guarantee [grant number 10052546]. Miaoran Zhang received funding from the DFG (German Research Foundation) under project 232722074, SFB 1102. We thank EIT and IDT High Performance Computing Center for providing computational resources for this project.
+
+# References
+
+Sweta Agrawal, Chunting Zhou, Mike Lewis, Luke Zettlemoyer, and Marjan Ghazvininejad. 2023. Incontext examples selection for machine translation. In Findings of the Association for Computational Linguistics: ACL 2023.  
+Farhad Akhbardeh, Arkady Arkhangorodsky, Magdalena Biesialska, Ondrej Bojar, Rajen Chatterjee, Vishrav Chaudhary, Marta R. Costa-jussa, Cristina Espana-Bonet, Angela Fan, Christian Federmann, Markus Freitag, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Leonie Harter, Kenneth Heafield, Christopher Homan, Matthias Huck, Kwabena Amponsah-Kaakyire, Jungo Kasai, Daniel Khashabi, Kevin Knight, Tom Kocmi, Philipp Koehn, Nicholas Lourie, Christof Monz, Makoto Morishita, Masaaki Nagata, Ajay Nagesh, Toshiaki
+
+Nakazawa, Matteo Negri, Santanu Pal, Allahsera Auguste Tapo, Marco Turchi, Valentin Vydrin, and Marcos Zampieri. 2021. Findings of the 2021 conference on machine translation (WMT21). In Proceedings of the Sixth Conference on Machine Translation.  
+Duarte M. Alves, Nuno M. Guerreiro, João Alves, José Pombal, Ricardo Rei, José de Souza, Pierre Colombo, and Andre Martins. 2023. Steering large language models for machine translation with finetuning and in-context learning. In Findings of the Association for Computational Linguistics: EMNLP 2023.  
+Duarte M. Alves, José Pombal, Nuno M Guerreiro, Pedro H Martins, João Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, et al. 2024. Tower: An open multilingual large language model for translation-related tasks. arXiv preprint.  
+Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.  
+Loic Barrault, Magdalena Biesialska, Ondrej Bojar, Marta R. Costa-jussa, Christian Federmann, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Matthias Huck, Eric Joanis, Tom Kocmi, Philipp Koehn, Chi-ku Lo, Nikola Ljubesic, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Santanu Pal, Matt Post, and Marcos Zampieri. 2020. Findings of the 2020 conference on machine translation (WMT20). In Proceedings of the Fifth Conference on Machine Translation.  
+Loic Barrault, Ondrej Bojar, Marta R. Costa-jussa, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, Shervin Malmasi, Christof Monz, Mathias Muller, Santanu Pal, Matt Post, and Marcos Zampieri. 2019. Findings of the 2019 conference on machine translation (WMT19). In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1).  
+Ondrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Shujian Huang, Matthias Huck, Philipp Koehn, Qun Liu, Varvara Logacheva, Christof Monz, Matteo Negri, Matt Post, Raphael Rubino, Lucia Specia, and Marco Turchi. 2017. Findings of the 2017 conference on machine translation (WMT17). In Proceedings of the Second Conference on Machine Translation.  
+Ondrej Bojar, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, and Christof Monz. 2018. Findings of the 2018 conference on machine translation (WMT18). In Proceedings of the Third Conference on Machine Translation: Shared Task Papers.  
+Eleftheria Briakou, Colin Cherry, and George Foster. 2023. Searching for needles in a haystack: On the
+
+role of incidental bilingualism in PaLM's translation capability. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+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. In Advances in Neural Information Processing Systems.  
+Lichang Chen, Shiyang Li, Jun Yan, Hai Wang, Kalpa Gunaratna, Vikas Yadav, Zheng Tang, Vijay Srinivasan, Tianyi Zhou, Heng Huang, et al. 2024a. Alpagasus: Training a better Alpaca model with fewer data. In The Twelfth International Conference on Learning Representations.  
+Pinzhen Chen, Zhicheng Guo, Barry Haddow, and Kenneth Heafield. 2024b. Iterative translation refinement with large language models. In Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1).  
+Pinzhen Chen, Shaoxiong Ji, Nikolay Bogoychev, Andrey Kutuzov, Barry Haddow, and Kenneth Heafield. 2024c. Monolingual or multilingual instruction tuning: Which makes a better Alpaca. In Findings of the Association for Computational Linguistics: EACL 2024.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing GPT-4 with  $90\%$  ChatGPT quality. lmsys.org.  
+Nadezhda Chirkova and Vassilina Nikoulina. 2024a. Key ingredients for effective zero-shot cross-lingual knowledge transfer in generative tasks. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers).  
+Nadezhda Chirkova and Vassilina Nikoulina. 2024b. Zero-shot cross-lingual transfer in instruction tuning of large language models. In Proceedings of the 17th International Natural Language Generation Conference.  
+Rochelle Choenni, Dan Garrette, and Ekaterina Shutova. 2023. How do languages influence each other? studying cross-lingual data sharing during LM fine-tuning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research.
+
+Marjan Ghazvininejad, Hila Gonen, and Luke Zettle-moyer. 2023. Dictionary-based phrase-level prompting of large language models for machine translation. arXiv preprint.  
+Sreyan Ghosh, Chandra Kiran Reddy Evuru, Sonal Kumar, Ramaneswaran S, Deepali Aneja, Zeyu Jin, Ramani Duraiswami, and Dinesh Manocha. 2024. A closer look at the limitations of instruction tuning. In Proceedings of the 41st International Conference on Machine Learning.  
+Zhiwei He, Tian Liang, Wenxiang Jiao, Zhuosheng Zhang, Yujiu Yang, Rui Wang, Zhaopeng Tu, Shuming Shi, and Xing Wang. 2024. Exploring human-like translation strategy with large language models. Transactions of the Association for Computational Linguistics.  
+Vivek Iyer, Pinzhen Chen, and Alexandra Birch. 2023. Towards effective disambiguation for machine translation with large language models. In Proceedings of the Eighth Conference on Machine Translation.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7B. arXiv preprint.  
+Wenxiang Jiao, Jen-tse Huang, Wenxuan Wang, Zhiwei He, Tian Liang, Xing Wang, Shuming Shi, and Zhaopeng Tu. 2023. ParroT: Translating during chat using large language models tuned with human translation and feedback. In Findings of the Association for Computational Linguistics: EMNLP 2023.  
+Tannon Kew, Florian Schottmann, and Rico Sennrich. 2023. Turning english-centric LLMs into polyglots: How much multilinguality is needed? arXiv preprint.  
+Tom Kocmi, Rachel Bawden, Ondrej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Thamme Gowda, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Rebecca Knowles, Philipp Koehn, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Michal Novak, Martin Popel, and Maja Popovic. 2022. Findings of the 2022 conference on machine translation (WMT22). In Proceedings of the Seventh Conference on Machine Translation (WMT).  
+Haonan Li, Fajri Koto, Minghao Wu, Alham Fikri Aji, and Timothy Baldwin. 2023. Bactrian-X: Multilingual replicable instruction-following models with low-rank adaptation. arXiv preprint.  
+Jiahuan Li, Hao Zhou, Shujian Huang, Shanbo Cheng, and Jiajun Chen. 2024. Eliciting the translation ability of large language models via multilingual finetuning with translation instructions. Transactions of the Association for Computational Linguistics.  
+Bill Yuchen Lin, Abhilasha Ravichander, Ximing Lu, Nouha Dziri, Melanie Sclar, Khyathi Chandu, Chandra Bhagavatula, and Yejin Choi. 2024. Urial: Aligning untuned LLMs with just the 'write' amount of
+
+in-context learning. In The Twelfth International Conference on Learning Representations.  
+Zhuoyuan Mao and Yen Yu. 2024. Tuning LLMs with contrastive alignment instructions for machine translation in unseen, low-resource languages. In Proceedings of the Seventh Workshop on Technologies for Machine Translation of Low-Resource Languages (LoResMT 2024).  
+Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Eric Mitchell, Rafael Rafailov, Archit Sharma, Chelsea Finn, and Christopher D Manning. 2024. An emulator for fine-tuning large language models using small language models. In *The Twelfth International Conference on Learning Representations*.  
+Yasmin Moslem, Rejwanul Haque, John D. Kelleher, and Andy Way. 2023. Adaptive machine translation with large language models. In Proceedings of the 24th Annual Conference of the European Association for Machine Translation.  
+Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng Xin Yong, Hailey Schoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak, Samuel Albanie, Zaid Alyafeai, Albert Webson, Edward Raff, and Colin Raffel. 2023. Crosslingual generalization through multitask finetuning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems 35.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics.  
+Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog.  
+Vikas Raunak, Amr Sharaf, Hany Hassan Awadallah, and Arul Menezes. 2023. Leveraging GPT-4 for automatic translation post-editing. In Findings of the Association for Computational Linguistics: EMNLP 2023.
+
+Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2022. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations.  
+Gabriele Sarti, Phu Mon Htut, Xing Niu, Benjamin Hsu, Anna Currey, Georgiana Dinu, and Maria Nadejde. 2023. RAMP: Retrieval and attribute-marking enhanced prompting for attribute-controlled translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers).  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Uri Shaham, Jonathan Herzig, Roee Aharoni, Idan Szpektor, Reut Tsarfaty, and Matan Eyal. 2024. Multilingual instruction tuning with just a pinch of multilinguality. In Findings of the Association for Computational Linguistics ACL 2024.  
+Suzanna Sia, David Mueller, and Kevin Duh. 2024. Where does in-context translation happen in large language models. arXiv preprint.  
+David Stap, Eva Hasler, Bill Byrne, Christof Monz, and Ke Tran. 2024. The fine-tuning paradox: Boosting translation quality without sacrificing LLM abilities. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford Alpaca: An instruction-following LLaMA model. GitHub repository.  
+Jörg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT - building open translation services for the world. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint.  
+Ahmet Üstün, Viraat Aryabumi, Zheng-Xin Yong, WeiYin Ko, Daniel D'souza, Gbemileke Onilude, Neel Bhandari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, et al. 2024. Aya model: An instruction finetuned
+
+open-access multilingual language model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+David Vilar, Markus Freitag, Colin Cherry, Jiaming Luo, Viresh Ratnakar, and George Foster. 2023. Prompting PaLM for translation: Assessing strategies and performance. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V Le. 2022. Finetuned language models are zero-shot learners. In International Conference on Learning Representations.  
+Di Wu, Shaomu Tan, Yan Meng, David Stap, and Christof Monz. 2024a. How far can 100 samples go? unlocking zero-shot translation with tiny multiparallel data. In *Findings of the Association for Computational Linguistics* ACL 2024.  
+Minghao Wu, Thuy-Trang Vu, Lizhen Qu, George Foster, and Gholamreza Haffari. 2024b. Adapting large language models for document-level machine translation. arXiv preprint.  
+Minghao Wu, Abdul Waheed, Chiyu Zhang, Muhammad Abdul-Mageed, and Alham Aji. 2024c. LaMiniLM: A diverse herd of distilled models from large-scale instructions. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics.  
+Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. 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).  
+Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2024a. A paradigm shift in machine translation: Boosting translation performance of large language models. In The Twelfth International Conference on Learning Representations.  
+Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024b. Contrastive preference optimization: Pushing the boundaries of LLM performance in machine translation. In Proceedings of the 41st International Conference on Machine Learning.  
+Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2024. Teaching large language models to translate with comparison. In Proceedings of the AAAI Conference on Artificial Intelligence.  
+Biao Zhang, Barry Haddow, and Alexandra Birch. 2023a. Prompting large language model for machine translation: a case study. In Proceedings of the 40th International Conference on Machine Learning.
+
+Biao Zhang, Zhongtao Liu, Colin Cherry, and Orhan Firat. 2024. When scaling meets LLM finetuning: The effect of data, model and finetuning method. In *The Twelfth International Conference on Learning Representations*.  
+Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, et al. 2023b. BayLing: Bridging cross-lingual alignment and instruction following through interactive translation for large language models. arXiv preprint.  
+Xuan Zhang, Navid Rajabi, Kevin Duh, and Philipp Koehn. 2023c. Machine translation with large language models: Prompting, few-shot learning, and fine-tuning with QLoRA. In Proceedings of the Eighth Conference on Machine Translation.  
+Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. 2023. LIMA: Less is more for alignment. In Thirty-seventh Conference on Neural Information Processing Systems.  
+Dawei Zhu, Sony Trenous, Xiaoyu Shen, Dietrich Klakow, Bill Byrne, and Eva Hasler. 2024. A preference-driven paradigm for enhanced translation with large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.
+
+# A Model Performance with Varying Training Sample Sizes
+
+In Figure 6 and Figure 7, we present the performance for instruction-tuned baselines and our models on different evaluation directions. For most directions, using only 32 training samples can achieve competitive performance and beat all three instruction-tuned baselines. There are several exceptional cases, including en→zh and en→ja, in which the COMET score of SFT with a limited number of samples (32 or 64) is worse than 1-shot in-context learning.
+
+While we primarily report the results with Llama-2 7B in our experiments, we hypothesize that state-of-the-art LLMs are largely homogeneous in terms of language distribution and inherent translation capability making our findings applicable to other LLMs. To support this hypothesis, we conduct fine-tuning experiments with Mistral 7B and Llama-2 13B using varying data sizes: 32, 1024, and 70K. As shown in Figure 8, the general trend is quite similar to the Llama-2 7B case: fine-tuning with 32 examples results in competitive performance, matching or surpassing general-purpose instruction-tuned models. Furthermore, increasing the number of training examples leads to diminishing returns.
+
+# B Model Performance with Varying Training Directions
+
+Figure 9 shows normalized BLEU scores for different combinations of train and test translation directions. Similar to the COMET scores in Figure 2, we observe that when training the model on a single direction, its translation ability across other non-targeted directions is also elicited to a certain degree. It is worth noting that when the training direction is  $\mathrm{X}\rightarrow \mathrm{en}$ , the performance on directions  $\mathrm{en}\rightarrow \mathrm{X}$  is significantly worse than training on all directions.
+
+# C Combined Effect of Training Size and Direction
+
+Figure 12 illustrates the model performance across varying training sizes and translation directions, evaluated on en  $\rightarrow$  cs, de, zh. Similarly, Figure 13 presents the results on en  $\rightarrow$  cs, de, zh, and en  $\rightarrow$  hr. Consistently across all plots, we observe a positive impact on performance with an increasing number of training directions, particularly with smaller training sizes.
+
+# D Model Performance with Unseen Languages
+
+In Figure 10, we find similar patterns as the COMET score, where fine-tuning on unseen languages can elicit the model's ability to translate from and to all seen languages. However, the translation performance on unseen languages themselves remains subpar, suggesting that SFT primarily reveals the knowledge LLMs have possessed during pre-training.
+
+# E Model Performance with Noisy Data
+
+Figure 11 shows the BLEU score of different translation directions with two noise types. We can find that models are more sensitive to word-level noise than sentence-level noise. Also, the performance degradation is more noticeable when injecting noise into the source translation side. In comparison to the results of size 1024, using 32 training examples still achieves comparable or even better performance in the noisy condition.
+
+# F Technical Details
+
+# F.1 Datasets
+
+Our parallel data is derived from the development and test sets of WMT17 through WMT22. Detailed dataset statistics are available in Table 3. For most experiments, we use the test sets from WMT17 to WMT20 for training. The test set from WMT22 is used specifically for testing. An exception is noted in Section section 3.4, where models are trained using the en  $\leftrightarrow$  ha and en  $\leftrightarrow$  is language pairs from WMT21's development set. Subsequently, these models are evaluated using the corresponding test sets from WMT21.
+
+# F.2 Translation instructions
+
+The collection of translation instruction templates used in this work can be found in Table 4.
+
+# F.3 Evaluation packages
+
+To obtain COMET scores, we use Unbabel/wmt22-comet-da and for BLEU scores, we use sacreBLEU<sup>10</sup> (Post, 2018). The signature from the sacreBLEU package is nrefs:1, case:mixed, eff:no, tok:13a, smooth:exp, version:2.0.0 for all language pairs, except for tokenization for en→zh and en→jp, where we use tok:zh and tok:jp-mecab, respectively.
+
+<table><tr><td rowspan="2">Direction</td><td colspan="5">Training</td><td>Validation*</td><td>Test</td></tr><tr><td>WMT17</td><td>WMT18</td><td>WMT19</td><td>WMT20</td><td>WMT21dev</td><td>WMT21</td><td>WMT22</td></tr><tr><td>en-cs</td><td>3005</td><td>2983</td><td>1997</td><td>1418</td><td>0</td><td>1002</td><td>2037</td></tr><tr><td>en-de</td><td>3004</td><td>2998</td><td>1997</td><td>1418</td><td>0</td><td>1002</td><td>2037</td></tr><tr><td>en-hr</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td><td>1671</td></tr><tr><td>en-ja</td><td>0</td><td>0</td><td>0</td><td>1000</td><td>0</td><td>0</td><td>2037</td></tr><tr><td>en-ru</td><td>3001</td><td>3000</td><td>1997</td><td>2002</td><td>0</td><td>1002</td><td>2037</td></tr><tr><td>en-zh</td><td>2001</td><td>3981</td><td>1997</td><td>1418</td><td>0</td><td>1002</td><td>2037</td></tr><tr><td>cs-en</td><td>3005</td><td>2983</td><td>0</td><td>664</td><td>0</td><td>1000</td><td>1448</td></tr><tr><td>de-en</td><td>3004</td><td>2998</td><td>2000</td><td>785</td><td>0</td><td>1000</td><td>1984</td></tr><tr><td>ja-en</td><td>0</td><td>0</td><td>0</td><td>993</td><td>0</td><td>1005</td><td>2008</td></tr><tr><td>ru-en</td><td>3001</td><td>3000</td><td>2000</td><td>991</td><td>0</td><td>1000</td><td>2016</td></tr><tr><td>zh-en</td><td>2001</td><td>3981</td><td>2000</td><td>2000</td><td>0</td><td>1948</td><td>1875</td></tr><tr><td>en-ha</td><td>0</td><td>0</td><td>0</td><td>0</td><td>2000</td><td>1000</td><td>0</td></tr><tr><td>ha-en</td><td>0</td><td>0</td><td>0</td><td>0</td><td>2000</td><td>997</td><td>0</td></tr><tr><td>en-is</td><td>0</td><td>0</td><td>0</td><td>0</td><td>2004</td><td>1000</td><td>0</td></tr><tr><td>is-en</td><td>0</td><td>0</td><td>0</td><td>0</td><td>2004</td><td>1000</td><td>0</td></tr><tr><td>de-fr</td><td>0</td><td>0</td><td>1701</td><td>1619</td><td>0</td><td>⊗</td><td>1984</td></tr><tr><td>fr-de</td><td>0</td><td>0</td><td>1701</td><td>1619</td><td>0</td><td>⊗</td><td>2006</td></tr></table>
+
+Table 3: Data statistics. *Generally, WMT21 test is used for validation purposes; exceptions are en↔ha and en↔is, which are used for testing.  $^\otimes$  Although WMT21 includes data for de↔fr, these language pairs are excluded from experiments.
+
+# F.4 Hardware specifications and runtime
+
+Our experiments are conducted on a computing node with either 8 NVIDIA A100-40GB GPUs or 8 H100-80GB GPUs. DeepSpeed<sup>11</sup> with zero-stage 1 and mixed precision bfloat16 is used for performing SFT. Given the limited dataset size, typically fewer than 1024 samples, each SFT experiment can be completed within a mere 15 minutes using four H100 GPUs. However, given the necessity to evaluate the models across more than ten translation directions, the evaluation process may require up to four hours when performed on a single A100-40GB GPU.
+
+<table><tr><td>Instruction pool</td></tr><tr><td>Please provide the [TGT] translation for the following text</td></tr><tr><td>Convert the subsequent sentences from [SRC] into [TGT]:</td></tr><tr><td>Render the listed sentences in [TGT] from their original [SRC] form:</td></tr><tr><td>Transform the upcoming sentences from [SRC] language to [TGT] language:</td></tr><tr><td>Translate the given text from [SRC] to [TGT]:</td></tr><tr><td>Turn the following sentences from their [SRC] version to the [TGT] version:</td></tr><tr><td>Adapt the upcoming text from [SRC] to [TGT]:</td></tr><tr><td>Transpose the next sentences from the [SRC] format to the [TGT] format.</td></tr><tr><td>Reinterpret the ensuing text from [SRC] to [TGT] language.</td></tr><tr><td>Modify the forthcoming sentences, converting them from [SRC] to [TGT] .</td></tr><tr><td>What is the meaning of these sentences when translated to [TGT] ?</td></tr><tr><td>In the context of [TGT], what do the upcoming text signify? The text is:</td></tr><tr><td>How would you express the meaning of the following sentences in [TGT] ?</td></tr><tr><td>What is the significance of the mentioned sentences in [TGT] ?</td></tr><tr><td>In [TGT], what do the following text convey?</td></tr><tr><td>When translated to [TGT], what message do these sentences carry?</td></tr><tr><td>What is the intended meaning of the ensuing sentences in [TGT] ?</td></tr><tr><td>How should the following sentences be comprehended in [TGT] ?</td></tr><tr><td>In terms of [TGT], what do the next sentences imply?</td></tr><tr><td>Kindly furnish the [TGT] translation of the subsequent sentences.</td></tr><tr><td>Could you supply the [TGT] translation for the upcoming sentences?</td></tr><tr><td>Please offer the [TGT] rendition for the following statements.</td></tr><tr><td>I&#x27;d appreciate it if you could present the [TGT] translation for the following text:</td></tr><tr><td>Can you deliver the [TGT] translation for the mentioned sentences?</td></tr><tr><td>Please share the [TGT] version of the given sentences.</td></tr><tr><td>It would be helpful if you could provide the [TGT] translation of the ensuing sentences.</td></tr><tr><td>Kindly submit the [TGT] interpretation for the next sentences.</td></tr><tr><td>Please make available the [TGT] translation for the listed sentences.</td></tr><tr><td>Can you reveal the [TGT] translation of the forthcoming sentences?</td></tr><tr><td>Translate from [SRC] to [TGT]:</td></tr></table>
+
+Table 4: A collection of 31 translation prompts. Each instruction is randomly selected to form a training sample. At inference time, the first instruction is always selected. The placeholders [SRC] and [TGT] represent the source and target languages, respectively, and will be replaced with the appropriate languages depending on the specific example at hand.
+
+![](images/1f6854289348ad7d3f319bba49bedbbaa44260ca91008cb5fd316439b5eec137.jpg)
+
+![](images/85a701f86c78c40af99e7f68eb2d4bd86fdfbb0cffd6450ae9b22173428d34d0.jpg)
+
+![](images/054537b31c8e499d09b6f6efb0706132b2e3f3e9c44ae5ba9524bb62d0535543.jpg)
+
+![](images/ea3487d907d610e5219e61d6e24d539ccb7a4f7ff1f9444fffb3810113dea4e0.jpg)
+
+![](images/cf0a8cb8deb6dcf63674d6f8448bd778b586db171a0c21401a31969fdeb326ff.jpg)
+
+![](images/e1deb841cc85c26f1d5b850f42208df0017df3bd765604f277467787037c02ef.jpg)
+
+![](images/50b00f7833b438ffceafde6a3cdcdea3c1c07ce9516e23fda8a60be8d9e2e6c8.jpg)
+
+![](images/70b29120a14e0dd60993d3f8b14983282918105263c3b20a60801f36618cfa0a.jpg)
+
+![](images/38ce4f265fafa868be83744aa690da944ad27e637c53c24d352e47a60670b73a.jpg)
+
+![](images/02ee053228d5e68b539523ee23c67b95394e0f1e9cd4be231ad74b8aa89d98e0.jpg)  
+Llama-2-7b SFT-MT
+
+Figure 6: COMET scores between instruction-tuned baselines and our models at different training data sizes, evaluated on individual translation directions. ICL is used for training sizes at or below 3, indicated with "\("; otherwise, we perform SFT. With only 32 examples for SFT, Llama-2 outperforms general-purpose, instruction-tuned baselines. Base.: instruction-tuned baseline models.  
+![](images/4d84c9799fa59ef58ece962f47532032e308d3a5a8db789cd38058babf467384.jpg)  
+Llama-2-7b ICL-MT  
+★ vicuna-7b-v1.5  
+Mistral-7B-v0.1  
+Llama-2-7b-chat-hf
+
+![](images/46b23e97d2596df738347327c2f9e015761faa2ec964e282d61742f61ccd2280.jpg)
+
+![](images/9cec5a04bf3422c83cf76ec0224d4303b7b5881ab184e4bd1fad7ec08db9278e.jpg)
+
+![](images/65a5fac67721800fdcb2ecc18717caca4eb3397b8fece027ecf249ca6690aabc.jpg)
+
+![](images/78869169bec5b94f11e921e441b081ca4ac6b9510bfad9798001bb5b1b4ef938.jpg)
+
+![](images/34a20933646fc3d124f2ed77dcf8363548c93ba4f9dfbf92cd39c101d99b67f7.jpg)
+
+![](images/346918e668b42dda5d03603912f01164ebf6501b8d283b0bb731a079c58d252b.jpg)
+
+![](images/7c35712c79f7461929390f7e9815fe492bf486ae192f21b7122bc2340ec9232d.jpg)
+
+![](images/de0ea1412d69a56f2995d5da95b6dbe9b26aaf083460368bfd918e5160e7b2d2.jpg)
+
+![](images/0d4e67f311bc9971858d14aacff3e41075c31205aedd1aef905027ab71a3eb17.jpg)
+
+![](images/8aeba136c4d0e9c3c5ba2e604afbf56eefe83a862f398c95f3390f8e65e1de6b.jpg)  
+Llama-2-7b SFT-MT
+
+Figure 7: BLEU scores between instruction-tuned baselines and our models at different training data sizes, evaluated on individual translation directions. ICL is used for training sizes at or below 3, indicated with \*\*; otherwise, we perform SFT. With only 32 examples for SFT, Llama-2 outperforms general-purpose, instruction-tuned baselines. Base.: instruction-tuned baseline models.  
+![](images/1ec431dd1d7eb9cf3e4a1c01c7fc2e5173a0a55223aecffe721be6b291297a1b.jpg)  
+Llama-2-7b ICL-MT  
+★ vicuna-7b-v1.5  
+Mistral-7B-v0.1  
+Llama-2-7b-chat-hf
+
+![](images/38f38d42959478a6b1382e9ada2737d6d25aeace6957b5f2913cb5c38ffad003.jpg)  
+Figure 8: Performance comparison between instruction-tuned baselines and fine-tuned models with different training data sizes. "Instruct" refers to the instruction-tuned baselines, specifically Mistral-7B-Instruct-v0.1 and Llama-2-13b-chat. "32/1024/74623" represents models fine-tuned on 32, 1024, and 74623 examples, using pre-trained only models: Mistral-7B-v0.1 and Llama-2-13b.
+
+![](images/97cd45c2ec5f450601d9de5b470455661bc97d517146870eea94ea250ab66907.jpg)  
+Figure 9: Model performance  $(\%)$  in BLEU score resulted from varying combinations of train and test translation directions. The scores are normalized according to Llama-2 fine-tuned on all 10 training directions.
+
+![](images/f211bcd1ffa8940c1d674d1ea47b65911b567149b50ef063afa5f4349fc9a9a2.jpg)  
+Figure 10: Model performance evaluated across 15 translation directions. While models trained on unseen languages (en  $\leftrightarrow$  is, en  $\leftrightarrow$  ha) exhibit moderate improvements in translating these languages, they demonstrate accurate translations from and to seen languages.
+
+![](images/fa3e88bdca195feb9a4f1cbc388b0f43b39925a58318b9c9687927a7fc41fbba.jpg)
+
+![](images/e4741fa5f6220857f788e6a8b46164fc17d279e01751f2a482ae826c4abaa968.jpg)  
+Figure 11: Model performance in BLEU score varying training sizes, directions, and noise types. Top (Bottom): score averaged across all en  $\rightarrow$  X (X  $\rightarrow$  en) test directions. Training sizes considered are 32 and 1024.
+
+![](images/4b740f60ab6b9450b2f06df75104dbd9042ec4e25395706ff556de79ce48104d.jpg)  
+Test Direction: cs→en
+
+![](images/dd57915c56090ff375c2c73ad754fd4111c93a4662a5426ff55ee6be2d137d0f.jpg)  
+Test Direction: en  $\rightarrow$  cs
+
+![](images/7a919e50e34091f7ebab3294c74f2f89f3f2df60852a9e7bbf635327cd30df80.jpg)  
+Test Direction: de  $\rightarrow$  en
+
+![](images/3b51d4623b49a38d60630f98be9677a39a83acc43ce4644c8d029518efd7b5e1.jpg)  
+Test Direction: en  $\rightarrow$  de
+
+![](images/fab34831f28ad7e702aebced9e11ec32cea82d1cb289d6fbc8f3c942020c22fe.jpg)  
+Test Direction: zh→en
+
+![](images/5fdc9fcc5740b2a418a573d63d31cc0d7d19e0bed17c7c6c3d1a673c645de473.jpg)  
+Test Direction: en→zh  
+Figure 12: Model performance (in COMET) on individual directions for models trained with varying data sizes and directions. Both factors positively impact performance. +=: training directions added on top of previous directions; two directions (from and to English) at a time. For example, “+=ru” covers 10 directions: en  $\leftrightarrow$  {de, zh, cs, jp, ru}.
+
+![](images/755b35d5fa15c734c7959be0a1285a43cfdb5d9dfcd8ffae0800fe26b06442e1.jpg)  
+Test Direction: ja  $\rightarrow$  en
+
+![](images/5ad893b7b83e200e23e65f784482b9ac36ab02f883021de083cc1a34f5afa8c9.jpg)  
+Test Direction: en  $\rightarrow$  ja
+
+![](images/1ebf9b5aba032d33c5092f88830e861cf106b37d64b9ec06845d48d5fa74fb4e.jpg)  
+Test Direction: ru  $\rightarrow$  en
+
+![](images/12ac4f9f2debcd23bd3f0cb70b36265283a75112963aa0976dfc78a0b31f97e7.jpg)  
+Test Direction: en  $\rightarrow$  ru
+
+![](images/84e2f7576253b2f638d81ad9d4d373f784d496fde7c49ebb3841ab710e622e42.jpg)  
+Test Direction: en  $\rightarrow$  hr  
+Figure 13: Model performance (in COMET) on individual directions for models trained with varying data sizes and directions. Both factors positively impact performance. +=: training directions added on top of previous directions; two directions (from and to English) at a time. For example, “+=ru” covers 10 directions: en  $\leftrightarrow$  {de, zh, cs, jp, ru}.

2024/Fine-Tuning Large Language Models to Translate_ Will a Touch of Noisy Data in Misaligned Languages Suffice_/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbc055f54fe034e32159b3175a456d7193cc3329073b3650f68c64d176c18270
+size 1761465

2024/Fine-Tuning and Prompt Optimization_ Two Great Steps that Work Better Together/b6b2cc5a-f4ec-482a-8da6-09a5acae609f_content_list.json

@@ -0,0 +1,1567 @@
+[
+    {
+        "type": "text",
+        "text": "Fine-Tuning and Prompt Optimization: Two Great Steps that Work Better Together",
+        "text_level": 1,
+        "bbox": [
+            270,
+            89,
+            724,
+            133
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dilara Soylu Christopher Potts Omar Khattab",
+        "bbox": [
+            282,
+            151,
+            715,
+            168
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Stanford University",
+        "bbox": [
+            416,
+            181,
+            579,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{soylu, cgpotts, okhattab}@stanford.edu",
+        "bbox": [
+            309,
+            211,
+            685,
+            227
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Natural Language Processing (NLP) systems are increasingly taking the form of sophisticated modular pipelines, e.g., Retrieval Augmented Generation (RAG), where each module may involve a distinct Language Model (LM) and an associated prompt template. These compound systems often lack intermediate labels or gradient flow to optimize each module, making their end-to-end optimization challenging. Here we seek strategies to optimize both the module-level LM weights and the associated prompt templates of such systems to maximize a downstream task metric. We propose for the first time combining the weight and prompt optimization strategies to optimize a modular LM pipeline by alternating between the two to get the same LM to teach itself. In experiments with multi-hop QA, mathematical reasoning, and feature-based classification using mistral-7b, llama-2-7b, and llama-3-8b, these BetterTogether strategies optimizing the weights and prompts of a pipeline together outperform directly optimizing weights alone and prompts alone by up to  $60\\%$  and  $6\\%$ , respectively, on average across LMs and tasks. Our BetterTogether optimizer is released in DSPy at http://dspy.ai.",
+        "bbox": [
+            144,
+            282,
+            460,
+            665
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            671,
+            258,
+            686
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While the capabilities of language models (LMs) continue to grow, recent work has shown the potential of building more powerful Natural Language Processing (NLP) systems by composing multiple skills of LMs into pipelines. Examples of this include systems for retrieval-augmented generation (Guu et al., 2020; Lewis et al., 2020; Ma et al., 2023; Jiang et al., 2023b), multi-hop reasoning (Qi et al., 2021; Khattab et al., 2021), information extraction (Pourreza and Rafiei, 2023; D'Oosterlinck et al., 2024), and other sophisticated pipelines (Zelikman et al., 2022; Dohan et al., 2022; Khattab et al., 2022; Beurer-Kellner et al., 2023; Schlag et al., 2023; Viswanathan et al., 2023).",
+        "bbox": [
+            112,
+            696,
+            490,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Such LM Programs offer much more control for designing NLP systems, as they break down problems into modular, more manageable sub-tasks that can be assigned to LMs. If we could teach these LMs to accurately conduct their easier sub-tasks and to communicate effectively within multi-stage pipelines, this could greatly expand the scope of reliable NLP systems we can build.",
+        "bbox": [
+            507,
+            261,
+            884,
+            388
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To this end, Khattab et al. (2024) introduced the DSPy framework for defining and optimizing LM Programs. In it, a program is defined as a function  $\\Phi$  that composes a set of stages, which we will refer to as language modules  $M = \\langle M_1,\\dots ,M_{|M|}\\rangle$ , into a pipeline. Each language module  $M_{i}$  specifies a fuzzy natural-language transformation (e.g., generating a summary of a supplied document) that needs to be learned. To do so, each module learns a particular prompt (template)  $\\pi$  to make a call to a particular LM with weights  $\\theta$ . The optimization problem is then defined as maximizing the expected performance (per a downstream metric  $\\mu$ ) of the program  $\\Phi$  over a set of inputs by updating each module's  $\\pi$  and  $\\theta$ .",
+        "bbox": [
+            507,
+            390,
+            884,
+            630
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing work (Khattab et al., 2024; Opsahl-Ong et al., 2024) has studied optimizing the discrete string prompt of each module and has considered simple approaches for fine-tuning each module's LM weights. In this empirical study, we investigate updating each module's LM weights and prompt template together to maximize a downstream metric on the final output of the program. Doing this is challenging as  $\\Phi$  is not generally differentiable and its modules  $M_{i}$  generally lack labeled outputs, while exhibiting sophisticated dependencies. Moreover, in realistic settings, the training set is usually very small and only a small number of LM calls are possible for training and inference.",
+        "bbox": [
+            507,
+            631,
+            885,
+            856
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this challenge, we propose to alternate between fine-tuning LM weights and optimizing prompt templates and evaluate approximate optimization strategies in which we bootstrap",
+        "bbox": [
+            507,
+            857,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10696",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 10696-10710 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            152,
+            945,
+            843,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "training labels for all pipeline modules. In experiments with multi-hop QA (HotPotQA), mathematical reasoning (GSM8K), and feature-based classification (Iris), we show that these tandem strategies are highly effective across three different LMs, leading to  $5 - 78\\%$  gains for HotPotQA, 2.5-  $10\\%$  gains for GSM8K, and  $3.5 - 88\\%$  gains for Iris against prompts only and weights only strategies, averaged across mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct.",
+        "bbox": [
+            112,
+            84,
+            492,
+            247
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Problem Statement",
+        "text_level": 1,
+        "bbox": [
+            112,
+            261,
+            317,
+            275
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We are given an LM program  $\\Phi$ , which operates like a blackbox function  $\\Phi: \\mathcal{X} \\to \\mathcal{Y}$ , in which  $\\mathcal{X}$  and  $\\mathcal{Y}$  are typically in natural language (e.g., questions and their program-generated answers, respectively). For example, we may have a program  $\\Phi$  for answering complex questions with short factoid answers. In the course of its execution,  $\\Phi$  makes one or more calls to each of its  $|M| \\geq 1$  language modules,  $M = \\langle M_1, \\ldots, M_{|M|} \\rangle$ .",
+        "bbox": [
+            112,
+            288,
+            489,
+            435
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For example, the program may implement a multi-hop, retrieval-augmented pipeline for question answering. This common pipeline (Qi et al., 2021; Khattab et al., 2021; Press et al., 2023; Khattab et al., 2022) breaks down the input into subquestions that are used to iteratively find relevant passages (e.g., from a corpus like Wikipedia) until the question can be faithfully answered. In general terms, each module  $M_{i}:\\mathcal{X}_{i}\\to \\mathcal{Y}_{i}$  is a declarative LM invocation that defines, in inherently fuzzy natural-language terms, an input  $\\mathcal{X}_i$  domain (like a user-supplied question and a set of retrieved passages) and an output  $\\mathcal{Y}_i$  co-domain (like a search query to find additional relevant passages).",
+        "bbox": [
+            112,
+            437,
+            489,
+            661
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We seek to implement each language module as some specific, well-tuned strategy for invoking an underlying language model LM. Concretely, we assume that a module  $M_{i}$  will be fully implemented by specifying (1) the string prompt  $\\pi_{i}$  in which the module inputs  $\\mathcal{X}_i$  are plugged in to decode the module outputs  $\\mathcal{V}_i$  and (2) the floating-point weights  $\\theta_{i}$  assigned to the parameters of LM in the course of this module. We refer to the version of  $\\Phi$  in which the prompts and LM weights are assigned explicitly to  $\\Pi$  and  $\\Theta$ , respectively, as  $\\Phi_{\\langle \\Theta ,\\Pi \\rangle}$ .",
+        "bbox": [
+            112,
+            663,
+            487,
+            840
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given nothing but a small training set  $X = \\{(x_{1},m_{1}),\\ldots ,(x_{|X|},m_{|X|})\\}$  of inputs  $x_{i}\\in \\mathcal{X}$  and optional metadata like output labels or other hints  $m_{i}\\in \\mathcal{M}$  that can be used for determining the correctness of a given program run, and a metric",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "$\\mu : \\mathcal{Y} \\times \\mathcal{M} \\to \\mathbb{R}$ , our goal is to optimize  $\\Phi$ , that is, configure its modules' prompts and LM weights to maximize the following objective.",
+        "bbox": [
+            507,
+            84,
+            884,
+            134
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\underset {\\Theta , \\Pi} {\\arg \\max } \\frac {1}{| X |} \\sum_ {(x, m) \\in X} \\mu \\left(\\Phi_ {\\langle \\Theta , \\Pi \\rangle} (x), m\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            547,
+            155,
+            842,
+            195
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Researchers tuning LM pipelines are in effect seeking to achieve this objective. It is also a very large subspace of the optimization problem in the DSPy framework<sup>1</sup> for LM programs. Unfortunately, this problem is intractable: the search space is large and we don't have gradients or intermediate output labels to optimize each module, so we seek approximate strategies for such optimization.",
+        "bbox": [
+            507,
+            203,
+            884,
+            332
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 BetterTogether: Alternating Weight and Prompt Optimization Steps for LM Programs",
+        "text_level": 1,
+        "bbox": [
+            507,
+            344,
+            880,
+            394
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We now introduce the BetterTogether algorithm, which alternates the weight and prompt optimization steps for LM programs. We hypothesize that, when a large LM is used to teach itself how to tackle the task defined by an LM program, finetuning LM weights and prompts are both essential to achieve the highest quality. In particular, we expect that (1) prompt optimization before finetuning can lead to more successful datapoints for fine-tuning, and, (2) prompt optimization after finetuning can make adjustments to the behavior of the LM program, leading to higher quality outputs. Considering that fine-tuning is often perceived as a more powerful tool, this can be surprising, especially when both approaches are ultimately applied over the same set of training inputs  $X$ .",
+        "bbox": [
+            505,
+            401,
+            882,
+            659
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1 BetterTogether: Optimizing LM programs by alternating prompt and weight optimization steps, instantiated in Algorithm 2",
+        "text_level": 1,
+        "bbox": [
+            507,
+            671,
+            884,
+            720
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/979be1f94b68dcd0b1879cceace3c7680d398e0d48f7156c247dad1cfb1e87ae.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Input: Program Φ(Θ,Π) = ΦΘ ⊙ ΦΠ, \n    with module weights Θ = [θ1, ..., θ|Φ|] \n    and module prompts Π = [π1, ..., π|Φ|]</td></tr><tr><td colspan=\"2\">Training Set X and Metric μ</td></tr><tr><td colspan=\"2\">1: function BETTERTOGETHER(Φ(Θ,Π), X, μ)</td></tr><tr><td colspan=\"2\">2: Π&#x27; ← OPTIMIZEPROMPTS(Φ(Θ,Π), X, μ)</td></tr><tr><td colspan=\"2\">3: Θ&#x27; ← FINETUNEWEIGHTS(Φ(Θ,Π&#x27;), X, μ)</td></tr><tr><td colspan=\"2\">4: Π&#x27;&#x27; ← OPTIMIZEPROMPTS(Φ(Θ&#x27;,Π), X, μ)</td></tr><tr><td colspan=\"2\">5: return Φ(Θ&#x27;,Π&#x27;)</td></tr><tr><td colspan=\"2\">6: end function</td></tr></table>",
+        "bbox": [
+            510,
+            722,
+            880,
+            851
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Accordingly, the general optimization framework for our algorithm is defined in Algorithm 1.",
+        "bbox": [
+            507,
+            866,
+            884,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "http://dspy.ai",
+        "bbox": [
+            531,
+            906,
+            645,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10697",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given a program  $\\Phi$ , the algorithm begins by optimizing  $\\Phi$ 's prompts, then fine-tuning its set of LM weights with the data bootstrapped using the optimized prompts, and finally optimizing its prompts again using the fine-tuned weights. In principle, each of these steps could be treated as optional. This will define the different possible combinations of BetterTogether that we will seek to evaluate in Section 4. Specifically, we are interested in the quality of (1) the vanilla program  $\\Phi$  with simple user-supplied instructions as the prompts and no fine-tuning of LM, (2) optimizing the prompts only, (3) optimizing the weights only, (4) optimizing the prompts twice, i.e., using the prompt-optimized  $\\Phi$  as a starting point for a second round of prompt optimization, (5) optimizing the weights twice, (6) optimizing the prompts then the weights, (7) vice versa, and (8) optimizing the prompts, weights, then prompts. Overall, we expect the final three to consistently outperform the first five.",
+        "bbox": [
+            110,
+            84,
+            492,
+            405
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For Algorithm 1 to be complete, we need to instantiate Lines 1-3 with specific approaches for prompt optimization and LM fine-tuning. For this, we choose the Bootstrap- $*$  family of algorithms from Khattab et al. (2024), which work by executing an initial version of the program on input examples  $(x_{i},m_{i})\\in X$  and recording the inputs/outputs observed at each module when the final output is \"correct\", i.e.,  $\\mu (\\Phi (x_i),m_i)\\geq \\lambda$  for some threshold  $\\lambda$  (e.g., 1.0 for binary accuracy). This is important to note: in line with our problem formulation, our prompt and weight optimization regimes are not simply training on hand-labeled data but on self-generated program traces.",
+        "bbox": [
+            110,
+            406,
+            489,
+            631
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Instantiations for Lines 1-3 of Algorithm 1 are shown in Algorithm 2. For prompt optimization, we use BootstrapFewshotRS (BFRS) of DSPy, which self-generates potential few-shot examples of every module and applies a form of random search (RS) to select the specific generated few-shot examples that are used for prompting. Overall, BFRS first divides  $X$  into a training split  $T$  and a validation split  $V$  (Line 2). It then executes the provided  $\\Phi_{\\langle \\Theta, \\Pi \\rangle}$  on the training inputs, collecting input-output pairs for every module in  $\\Phi$  for each  $x_i \\in T$ . This is called a trace  $\\tau$ , and we keep only the traces assigned high scores by  $\\mu$  (Line 4). Given all of these traces, BFRS samples multiple different subsets of a few traces  $\\tau'$  (Line 6), each of them containing a potential few-shot example for each module in  $\\Phi$ , and ultimately selects the subset that, when used to construct few-shot prompts",
+        "bbox": [
+            112,
+            631,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 2 Instantiating Algorithm 1's prompt & weight optimizers with bootstrapping algorithms"
+        ],
+        "code_body": "Input: Training Set  $X$  and Metric  $\\mu$   \n1: function BOOTSTRAPFEWSHOTRS(  $\\Phi_{\\langle \\Theta ,\\Pi \\rangle}$ $X,\\mu)$   \n2:  $T,V\\gets$  SPLITINTOTRAINANDVALIDATION(X)  \n3:  $\\tau \\leftarrow$  BOOTSTRAPTRACES(  $\\Phi_{\\langle \\Theta ,\\Pi \\rangle}$ $T)$   \n4:  $\\tau \\leftarrow$  FILTERTRACES(  $\\tau ,\\mu)$   \n5: Initialize attempts list  $\\mathcal{A}\\gets \\{\\}$   \n6: for  $\\tau^{\\prime}\\in$  SAMPLEFEWSHOTSUBSETS(τ) do  \n7:  $\\Pi^{\\prime}\\gets$  CONSTRUCTFEWSHOTPROMPTS(τ')  \n8:  $\\sigma \\leftarrow \\frac{1}{|V|}\\sum_{\\langle x_i,m_i\\rangle \\in V}\\mu (\\Phi_{\\langle \\Theta ,\\Pi '}\\rangle (x_i),m_i)$   \n9: Extend A with  $(\\sigma ,\\Pi^{\\prime})$   \n10: end for  \n11: return  $\\Pi_{\\mathrm{max}}$  , A's highest-scoring prompts sequence  \n12: end function  \n13:  \n14: function BOOTSTRAPFINETUNE(  $\\Phi_{\\langle \\Theta ,\\Pi \\rangle}$ $X,\\mu)$   \n15:  $\\tau \\leftarrow$  BOOTSTRAPTRACES(  $\\Phi_{\\langle \\Theta ,\\Pi \\rangle}$ $X)$   \n16:  $\\tau \\leftarrow$  FILTERTRACES(  $\\tau ,\\mu)$   \n17:  $\\Theta^{\\prime}\\gets$  TRAINLM(τ)  \n18: return  $\\Theta^{\\prime}$   \n19: end function  \n20:  \n21: Set OPTIMIZEPROMPTS as BOOTSTRAPFEWSHOTRS  \n22: Set FINETUNEWEIGHTS as BOOTSTRAPFINETUNE",
+        "bbox": [
+            509,
+            117,
+            882,
+            401
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(Line 7), achieves the highest score (Line 8). This simple search strategy is known to consistently lead to large quality improvements in prompting LM programs (Khattab et al., 2024; Opsahl-Ong et al., 2024), often outperforming manually or automatically optimizing prompt instructions or writing examples by hand.",
+        "bbox": [
+            507,
+            455,
+            884,
+            568
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For fine-tuning, we extend BootstrapFinetune (BFT) of DSPy, which, given a program  $\\Phi$ , self-generates a large number examples for every module and combines them into one dataset to fine-tune the LM weights with an implicit multi-task objective, where the sub-tasks are the modules' roles. Existing work has only considered BFT in a very narrow setting for LM programs: on HotPotQA, Khattab et al. (2024) train a T5-Large model using traces from a few-shot Llama2-13b program, without considering getting an LM to teach itself via BFT nor considering a role for BFRS in the fine-tuned program. In this work, we focus on allowing models to teach themselves and self-improve. We propose for the first time combining the strategies of BFRS and BFT via alternation to get the same LM to teach itself better than either prompt or weight optimization in isolation. One could test similar ideas in scenarios where a larger model does the bootstrapping for a smaller LM. This may lead to even better results but is outside our scope.",
+        "bbox": [
+            507,
+            583,
+            884,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10698",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/d5e97a34a24523253004041c5985c15094606844a48da6c8c59fc400cb4e411c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Strategy</td><td colspan=\"3\">mistral-7b-instruct-v0.2</td><td colspan=\"3\">llama-2-7b-chat</td><td colspan=\"3\">llama-3-8b-instruct</td></tr><tr><td>HotPotQA</td><td>GSM8K</td><td>Iris</td><td>HotPotQA</td><td>GSM8K</td><td>Iris</td><td>HotPotQA</td><td>GSM8K</td><td>Iris</td></tr><tr><td colspan=\"10\">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>17.2</td><td>40.3</td><td>26.0</td><td>13.2</td><td>24.0</td><td>0.0</td><td>31.6</td><td>72.7</td><td>48.0</td></tr><tr><td>Prompt Optimization (Π)</td><td>33.8</td><td>46.4</td><td>57.3</td><td>33.3</td><td>26.0</td><td>56.7</td><td>46.9</td><td>77.9</td><td>79.3</td></tr><tr><td>Weight Optimization (Θ)</td><td>22.9</td><td>40.7</td><td>29.3</td><td>12.2</td><td>24.0</td><td>-</td><td>34.8</td><td>75.1</td><td>37.3</td></tr><tr><td>Π → Π</td><td>33.8</td><td>47.7</td><td>59.3</td><td>32.6</td><td>24.7</td><td>64.0</td><td>46.5</td><td>77.6</td><td>82.0</td></tr><tr><td>Θ → Θ</td><td>24.0</td><td>42.8</td><td>38.0</td><td>13.0</td><td>24.1</td><td>-</td><td>34.4</td><td>44.1</td><td>39.3</td></tr><tr><td colspan=\"10\">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>36.3</td><td>47.3</td><td>30.7</td><td>32.7</td><td>27.3</td><td>26.7</td><td>42.8</td><td>77.6</td><td>44.0</td></tr><tr><td>Θ → Π</td><td>33.0</td><td>48.3</td><td>66.7</td><td>34.2</td><td>26.6</td><td>-</td><td>43.6</td><td>78.9</td><td>78.7</td></tr><tr><td>Π → Θ → Π</td><td>37.6</td><td>46.8</td><td>52.7</td><td>34.8</td><td>26.3</td><td>65.3</td><td>46.7</td><td>77.0</td><td>79.3</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            878,
+            286
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Main Results. Percentage accuracies of baseline and BetterTogether strategies on HotPotQA, GSM8K, and Iris evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are average performance of 3 runs on a held-out test set using different random seeds. Bold font is used to mark the highest score in a given column. Strategies where weight optimization is the first step use the vanilla (zero-shot) strategy to generate the initial fine-tuning dataset. If a model generates very few or no correct outputs under the vanilla strategy on the training set used to bootstrap the fine-tuning data, there will not be a sufficient dataset for fine-tuning. These settings are marked with “-”.",
+        "bbox": [
+            112,
+            296,
+            884,
+            395
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Evaluation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            420,
+            364,
+            437
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We now seek to evaluate our hypothesis on the importance of optimizing both LM weights and prompts of LM programs. We conduct our evaluation across three datasets that span different tasks (and thus LM programs) each. In particular, we use HotPotQA (Yang et al., 2018) for multi-hop reasoning, GSM8K (Cobbe et al., 2021) for arithmetic reasoning, and Iris (Fisher, 1988) for classification. We run our experiments using three models, mistral-7b-instruct-v0.2 (Jiang et al., 2023a), llama-2-7b-chat (Touvron et al., 2023), llama-3-8b-instruct (MetaAI, 2024), keeping the model used for prompt optimization, bootstrapping training traces, and fine-tuning the same in a given experiment for all modules. We initialize all the modules of an input program  $\\Phi$  to use the same LM weights (e.g. mistral-7b-instruct-v0.2), but distinct prompt templates specialized for their particular module-level task, such as generating a search query or answering a question. We implement all of our programs and optimizers as extensions to the DSPy framework.",
+        "bbox": [
+            112,
+            451,
+            489,
+            804
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each dataset, we split the data into training, development, and test sets. We shuffle the training set every time before we perform a prompt or weight optimization, controlled by random seed. For prompt optimization, we sub-sample non-overlapping sets of training and validation examples from the initial training set for each task",
+        "bbox": [
+            112,
+            809,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "and use the BFRS prompt optimizer to optimize the module level prompts of a given  $\\Phi$ , leaving its underlying LM weights untouched. For weight optimization, we use all the available training examples as potential candidates to generate the data for finetuning  $\\Phi$ 's LM weights and pass them to the BFT weight optimizer, which: (1) runs a given  $\\Phi$  on all the training examples, (2) keeps the traces where the final output of  $\\Phi$  was correct and filters out the rest, (3) gets module level prompt-completion pairs for each trace, (4) creates new pairs by replacing the prompt generated by  $\\Phi$  with a vanilla prompt, (5) combines all the module level prompt-completion pairs into one dataset, (6) fine-tunes  $\\Phi$ 's LM weights on this dataset, and finally, (7) returns an updated  $\\Phi$  where the LM weights of all the modules are set to the fine-tuned LM. We use the Low Rank Adaptation (LoRA; Hu et al. 2022) method to fine-tune our LMs.",
+        "bbox": [
+            507,
+            420,
+            884,
+            725
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The full text for our programs and the vanilla prompts are shared in Appendix A. The license information for all LMs and datasets used as well as our implementation details such as hyperparameters and software are reported in Appendix B and Appendix C, respectively.",
+        "bbox": [
+            507,
+            728,
+            884,
+            826
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Multi-hop Reasoning HotPotQA (in the \"full-wiki\" setting) is a question answering task in which systems must find two Wikipedia page abstracts out of a corpus of 5 million via search and use them to answer a factoid question. Therefore it can be",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10699",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "implemented as a program that has three LM modules: the first two for generating search queries (i.e., hops) and the last one for generating an answer. Each module uses Chain-of-Thought (CoT; Wei et al. 2022) to generate its outputs, producing a reasoning string before the search query or the answer. Search queries are passed to a frozen CoBERTv2 (Santhanam et al., 2022) retriever. Accuracy is measured using the exact match score of the answer with the ground truth answer for the given question, after normalizing case, stripping surrounding whitespace characters, and removing punctuation. We use the following splits for HotPotQA: 1000 training examples and 500 development examples drawn from the original training set, along with 1500 test examples drawn from the original validation set, since the original test set is not public. We sub-sample 100 and 250 non-overlapping examples for the prompt optimization training and validation sets, respectively.",
+        "bbox": [
+            112,
+            84,
+            492,
+            407
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Arithmetic Reasoning GSM8K is a benchmark consisting of grade school math problems. We implement it as an LM program with a single module using CoT prompting, where the LM generates a reasoning string followed by an answer. We measure accuracy by extracting the last number from the first line of the model's response and comparing it to the ground truth response. For GSM8K, we use a training set of 1000 examples and a development set of 500 examples, both sampled from the original training set. We use all the 1319 test examples available for our test set. We use 100 and 250 non-overlapping examples for training and validation sets used during prompt optimization.",
+        "bbox": [
+            112,
+            414,
+            489,
+            640
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Classification Iris is a classic classification task, where the goal is to classify species of Iris flowers. We use a single-module CoT DSPy program for Iris, with the goal of assessing whether it being a feature-based classification task gives a large advantage to methods based entirely on gradient descent (fine-tuning). This tests the extrapolation of our hypothesis to a different setting from the other two tasks. We measure accuracy using the exact match score of the answer and the correct answer given a question after normalizing both, as is the case for HotPotQA. The Iris dataset has a total of 150 examples, from which we create training, development, and test splits of equal size. We sub-sample 15 and 35 non-overlapping examples to be used as the training and validation sets for prompt optimization, respectively.",
+        "bbox": [
+            112,
+            646,
+            490,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Results & Discussion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            83,
+            726,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Held-out test set performance of the strategies described in Section 3 is shared in Table 1. Reported values are averaged across three runs with unique random seeds. Results for the individual runs are reported separately in Appendix D.",
+        "bbox": [
+            507,
+            110,
+            884,
+            187
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In 7 out of the 9 dataset and LM pairs, we observe that the best-performing strategies are the strategies that utilize prompt (Π) and weight (Θ) optimization steps together, although there is no clear winner among the benchmarked BetterTogether strategies that optimize both. Overall, optimizing prompts is essential on all the tasks, but optimizing prompts and weights together leads to strong gains over the best setting that only optimizes one of the two.",
+        "bbox": [
+            507,
+            190,
+            884,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In summary, we have proposed to alternate between optimizing prompts and fine-tuning LM weights and explored a few strategies for doing so. In experiments with multi-hop QA (HotPotQA), mathematical reasoning (GSM8K), and feature-based classification (Iris), we show that our strategies are highly effective for getting an LM to teach itself to perform an LM program via bootstrapping, leading to  $5 - 78\\%$  gains for HotPotQA,  $2.5 - 10\\%$  gains for GSM8K, and  $3.5 - 88\\%$  gains for Iris.",
+        "bbox": [
+            507,
+            351,
+            885,
+            512
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            524,
+            645,
+            539
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While this paper presents strong evidence from nine cases, spanning three tasks (and their corresponding LM programs) and three LMs, it is possible that other tasks, programs, or LMs will change the pattern in unforeseen ways. In particular, we have only experimented with weight optimization in the form of LoRA fine-tuning of pre-trained models. It is in principle possible that some other fine-tuning strategy would be so powerful and cost-effective as to remove the need for prompt optimization.",
+        "bbox": [
+            507,
+            550,
+            882,
+            711
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In addition, though we expect our findings to inform many researchers and practitioners interested in optimizing LM programs, and encourage them to explore optimizing prompts and fine-tuning LM weights together, we do not yet understand why both are important. The role of prompt optimization and the role of fine-tuning in multi-stage LM programs are both new, and the relative lack of understanding of these roles in the emerging literature could pose risks in unanticipated interactions between these components, compared with standard gradient descent for neural networks, which has been studied for decades.",
+        "bbox": [
+            507,
+            712,
+            884,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10700",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            278,
+            101
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "D.S. is supported by Ravi Family Graduate Fellowship. This work was partially supported by IBM as a founding member of the Stanford Institute for Human-Centered Artificial Intelligence (HAI), and by the HAI Hoffman-Yee Grant \"Dendritic Computation for Knowledge Systems\".",
+        "bbox": [
+            112,
+            109,
+            490,
+            206
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            233,
+            213,
+            248
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. 2023. Prompting is programming: A query language for large language models. Proc. ACM Program. Lang., 7(PLDI).",
+            "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.",
+            "David Dohan, Winnie Xu, Aitor Lewkowycz, Jacob Austin, David Bieber, Raphael Gontijo Lopes, Yuhuai Wu, Henryk Michalewski, Rif A. Saurous, Jascha Sohl-dickstein, Kevin Murphy, and Charles Sutton. 2022. Language model cascades. Preprint, arXiv:2207.10342.",
+            "Karel D'Oosterlinck, Omar Khattab, François Remy, Thomas Demeester, Chris Develder, and Christopher Potts. 2024. In-context learning for extreme multi-label classification. Preprint, arXiv:2401.12178.",
+            "Ronald A. Fisher. 1988. Iris. UCI Machine Learning Repository.",
+            "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Retrieval augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 3929-3938. PMLR.",
+            "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.",
+            "HuggingFace. 2023. Text generation inference.",
+            "Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023a. Mistral 7b. Preprint, arXiv:2310.06825."
+        ],
+        "bbox": [
+            115,
+            256,
+            490,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhengbao Jiang, Frank Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023b. Active retrieval augmented generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7969-7992, Singapore. Association for Computational Linguistics.",
+            "Omar Khattab, Christopher Potts, and Matei A. Zaharia. 2021. Baleen: Robust multi-hop reasoning at scale via condensed retrieval. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 27670-27682.",
+            "Omar Khattab, Keshav Santhanam, Xiang Lisa Li, David Hall, Percy Liang, Christopher Potts, and Matei Zaharia. 2022. Demonstrate-search-predict: Composing retrieval and language models for knowledge-intensive nlp. ArXiv preprint, abs/2212.14024.",
+            "Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan A, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2024. DSPy: Compiling declarative language model calls into state-of-the-art pipelines. In The Twelfth International Conference on Learning Representations.",
+            "Patrick S. H. Lewis, Ethan Perez, Aleksandra Pik-tus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.",
+            "Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting in retrieval-augmented large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5303-5315, Singapore. Association for Computational Linguistics.",
+            "Dirk Merkel. 2014. Docker: lightweight linux containers for consistent development and deployment. Linux J., 2014(239).",
+            "MetaAI. 2024. Meta llama 3.",
+            "Krista Opsahl-Ong, Michael J Ryan, Josh Purtell, David Broman, Christopher Potts, Matei Zaharia, and Omar Khattab. 2024. Optimizing instructions and demonstrations for multi-stage language model programs.",
+            "Mohammadreza Pourreza and Davood Rafiei. 2023. DIN-SQL: decomposed in-context learning of text-to-sql with self-correction. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10701",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 5687-5711, Singapore. Association for Computational Linguistics.",
+            "Peng Qi, Haejun Lee, Tg Sido, and Christopher Manning. 2021. Answering open-domain questions of varying reasoning steps from text. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3599-3614, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. 2022. ColBERTv2: Effective and efficient retrieval via lightweight late interaction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3715-3734, Seattle, United States. Association for Computational Linguistics.",
+            "Imanol Schlag, Sainbayar Sukhbaatar, Asli Celikyilmaz, Wen tau Yih, Jason Weston, Jürgen Schmidhuber, and Xian Li. 2023. Large language model programs. Preprint, arXiv:2305.05364.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.",
+            "Vijay Viswanathan, Chenyang Zhao, Amanda Bertsch, Tongshuang Wu, and Graham Neubig. 2023. Prompt2Model: Generating deployable models from natural language instructions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 413-421, Singapore. Association for Computational Linguistics.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le,"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.",
+            "Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics.",
+            "Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. 2022. Star: Bootstrapping reasoning with reasoning. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10702",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Appendices",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            218,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A Programs",
+        "text_level": 1,
+        "bbox": [
+            114,
+            109,
+            238,
+            124
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The DSPy programs for HotPotQA, GSM8K, and Iris are shared in Python Snippets A.1, A.2, A.3, respectively. Tables A.1, A.2, A.3 show the vanilla (zero-shot) programs generated with these programs, along with sample model completions.",
+        "bbox": [
+            110,
+            134,
+            885,
+            183
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "class HotPotQAProgram(dspyModule): def __init__(self, passages_per_hop  $= 3$  .. super().__init__() self.retrieve  $=$  dspy.Retrieve  $(k =$  passages_per_hop) self_generate_query  $=$  [dspyChainOfThought(\"context，question-> search_query\")for_in range(2)] self_generate_answer  $=$  dspyChainOfThought(\"context，question->answer\") def forward(self,question): context  $=$  [] for hop in range(2): search_query  $=$  self_generate_query[hop](context  $\\equiv$  context，question= question).search_query passages  $=$  self.retrieve.search_query).passages context  $=$  disp.utils.deduplicate(context+passages) return self_generate_answer(context  $\\equiv$  context，question  $\\equiv$  question).copy( context  $\\equiv$  context)",
+        "guess_lang": "txt",
+        "bbox": [
+            105,
+            186,
+            868,
+            425
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Python Snippets A.1: DSPy program for HotPotQA, consisting of three Chain-of-Thought (CoT) LM modules, along with a frozen retriever. Each LM module in the program uses its own prompt and LM weights under the hood.",
+        "bbox": [
+            112,
+            429,
+            882,
+            458
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10703",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/56fd1b9747fedf1393a53e73a8151b41ddf8e5c623666bc2178772039113346e.jpg",
+        "table_caption": [
+            "Table A.1: Vanilla (zero-shot) prompts for the HotPotQA program given in Python Snippets A.1 with three LM modules, each with corresponding prompts and LM weights. (1) the generate_query[0] module prompts the corresponding LM to generate a search query given a question along with an empty context; (2) the generate_query[1] module prompts the corresponding model to generate a search query given the context paragraphs retrieved using the search query outputted by the prior module; (3) the generate_answer module prompts the corresponding LM to answer the question given the context paragraphs retrieved by the earlier modules. The green text shows the model completions obtained by setting all the modules of the program to use mistralai/Mistral-7b-Instruct-v0.2 as their corresponding LM. The correct answer to the question is indeed Key deer as predicted."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Module Label</td><td>Prompt</td></tr><tr><td>generate_query[0]</td><td>Given the fields `context`, `question`, produce the fields `search_query`.</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Follow the following format.</td></tr><tr><td></td><td>Context: ${context}</td></tr><tr><td></td><td>Question: ${question}</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to ${produce the search_query}. We ...</td></tr><tr><td></td><td>Search Query: ${search_query}</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Context: N/A</td></tr><tr><td></td><td>Question: What animal subject to the Foster&#x27;s rule is the smallest North American deer?</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to find the smallest North American deer subject to the Foster&#x27;s rule.</td></tr><tr><td></td><td>1. The Foster&#x27;s rule is a mathematical formula used to estimate the body weight of large mammals based on their ear surface area.2. To apply the Foster&#x27;s rule, we need to find the ear surface area of the deer species in question.3. We don&#x27;t have the ear surface area of each North American deer species in the context, so we need to search for it.4. Therefore, our search query should include the keywords &quot;North American deer,&quot; &quot;ear surface area,&quot; and &quot;Foster&#x27;s rule.&quot;</td></tr><tr><td></td><td>Search Query: North American deer ear surface area Foster&#x27;s rule smallest species</td></tr></table>",
+        "bbox": [
+            115,
+            206,
+            892,
+            675
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10704",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Module Label Prompt",
+        "text_level": 1,
+        "bbox": [
+            178,
+            87,
+            359,
+            102
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "generate_query[1] Given the fields `context`, `question`, produce the fields `search_query`.",
+        "bbox": [
+            126,
+            114,
+            860,
+            128
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Follow the following format.",
+        "bbox": [
+            302,
+            162,
+            512,
+            174
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Context: ${context}",
+        "bbox": [
+            302,
+            186,
+            447,
+            198
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Question: ${question}",
+        "bbox": [
+            302,
+            210,
+            460,
+            222
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Reasoning: Let's think step by step in order to  $\\{$ produce the search_query\\}. We ...",
+        "bbox": [
+            302,
+            233,
+            763,
+            256
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Search Query: ${search_query}",
+        "bbox": [
+            302,
+            269,
+            522,
+            281
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Context:",
+        "text_level": 1,
+        "bbox": [
+            302,
+            317,
+            361,
+            326
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "[1] «Key deer | The Key deer (\"Odocoileus virginianus clavium\") is an endangered deer that lives only in the Florida Keys. It is a subspecies of the white-tailed deer (\"O. virginianus\"). It is the smallest North American deer.»",
+        "bbox": [
+            302,
+            329,
+            868,
+            374
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "[2] «North American least shrew | The North American least shrew (\"Cryptotis parva\") is one of the smallest mammals, growing to be only up to 3 inches long. The North American least shrew has a long pointed snout and a tail never more than twice the length of its hind foot. It has a dense fur coat that is either grayish-brown or reddish-brown with a white belly. Its fur becomes lighter in the summer and darker in the winter. Although similar in appearance to several species of rodents, all shrews are members of the order Soricomorpha and should not be mistaken for a member of the Rodentia order. The North American least shrew's eyes are small and its ears are completely concealed within its short fur, giving it very poor eyesight and hearing.",
+        "bbox": [
+            302,
+            376,
+            868,
+            505
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "[3] «Wynnea americana | Wynnea americana, commonly known as moose antlers or rabbit ears, is a species of fungus in the family Sarcoscyphaceae. This uncommon inedible species is recognizable by its spoon-shaped or rabbit-ear shaped fruit bodies that may reach up to 13 cm tall. It has dark brown and warty outer surfaces, while the fertile spore-bearing inner surface is orange to pinkish to reddish brown. The fruit bodies grow clustered together from large underground masses of compacted mycelia known as sclerotia. In eastern North America, where it is typically found growing in the soil underneath hardwood trees, it is found from New York to Michigan south to Mexico. The species has also been collected from Costa Rica, India, and Japan.»",
+        "bbox": [
+            302,
+            506,
+            868,
+            634
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Question: What animal subject to the Foster's rule is the smallest North American deer?",
+        "bbox": [
+            302,
+            648,
+            847,
+            670
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Reasoning: Let's think step by step in order to produce the search query. We need to find the smallest North American deer, which is mentioned in the context as the Key deer. Therefore, the search query should include the terms \"Key deer\" and \"smallest North American deer\".",
+        "bbox": [
+            302,
+            684,
+            868,
+            728
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Search Query: Key deer smallest North American deer",
+        "bbox": [
+            302,
+            743,
+            687,
+            753
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10705",
+        "bbox": [
+            478,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Module Label Prompt",
+        "text_level": 1,
+        "bbox": [
+            178,
+            87,
+            359,
+            102
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "generate_answer Given the fields `context`, `question`, produce the fields `answer`.",
+        "bbox": [
+            144,
+            114,
+            816,
+            128
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Follow the following format.",
+        "bbox": [
+            302,
+            162,
+            512,
+            174
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Context: ${context}",
+        "bbox": [
+            302,
+            186,
+            447,
+            198
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Question: ${question}",
+        "bbox": [
+            302,
+            209,
+            460,
+            221
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Reasoning: Let's think step by step in order to ${produce the answer}. We ...",
+        "bbox": [
+            302,
+            233,
+            853,
+            256
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Answer: ${answer}",
+        "bbox": [
+            302,
+            269,
+            431,
+            281
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Context:",
+        "text_level": 1,
+        "bbox": [
+            302,
+            317,
+            361,
+            326
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "[1] «Key deer | The Key deer (\"Odocoileus virginianus clavium\") is an endangered deer that lives only in the Florida Keys. It is a subspecies of the white-tailed deer (\"O. virginianus\"). It is the smallest North American deer.»",
+        "bbox": [
+            304,
+            329,
+            868,
+            374
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "[2] «North American least shrew | The North American least shrew (\"Cryptotis parva\") is one of the smallest mammals, growing to be only up to 3 inches long. The North American least shrew has a long pointed snout and a tail never more than twice the length of its hind foot. It has a dense fur coat that is either grayish-brown or reddish-brown with a white belly. Its fur becomes lighter in the summer and darker in the winter. Although similar in appearance to several species of rodents, all shrews are members of the order Soricomorpha and should not be mistaken for a member of the Rodentia order. The North American least shrew's eyes are small and its ears are completely concealed within its short fur, giving it very poor eyesight and hearing.",
+        "bbox": [
+            304,
+            376,
+            868,
+            505
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "[3] «Wynnea americana | Wynnea americana, commonly known as moose antlers or rabbit ears, is a species of fungus in the family Sarcoscyphaceae. This uncommon inedible species is recognizable by its spoon-shaped or rabbit-ear shaped fruit bodies that may reach up to 13 cm tall. It has dark brown and warty outer surfaces, while the fertile spore-bearing inner surface is orange to pinkish to reddish brown. The fruit bodies grow clustered together from large underground masses of compacted mycelia known as sclerotia. In eastern North America, where it is typically found growing in the soil underneath hardwood trees, it is found from New York to Michigan south to Mexico. The species has also been collected from Costa Rica, India, and Japan.»",
+        "bbox": [
+            304,
+            506,
+            868,
+            634
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "[4] «Pudu | The pudús (Mapudungun \"pūdú\" or \"pūdu\", Spanish: pudú, ]） are two species of South American deer from the genus Pudu, and are the world's smallest deer. The name is a loanword from Mapudungun, the language of the indigenous Mapuche people of southern Chile and south-western Argentina. The two species of pudús are the northern pudú (\"Pudu mephistophiles\") from Venezuela, Colombia, Ecuador, and Peru, and the southern pudú (\"Pudu puda\"; sometimes incorrectly modified to \"Pudu pudú\") from southern Chile and south-western Argentina. Pudu range in size from 32 to tall, and up to 85 cm long. As of 2009, the southern pudu is classified as near threatened, while the northern pudu is classified as vulnerable in the IUCN Red List.",
+        "bbox": [
+            304,
+            636,
+            868,
+            753
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Question: What animal subject to the Foster's rule is the smallest North American deer?",
+        "bbox": [
+            304,
+            766,
+            847,
+            789
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Reasoning: Let's think step by step in order to find the answer. We know that the smallest North American deer is mentioned in the context. We also know that the Foster's rule refers to a group of animals. In this case, the Foster's rule refers to endangered deer species that live only in the Florida Keys. Therefore, the Key deer is the answer.",
+        "bbox": [
+            302,
+            802,
+            868,
+            860
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Answer: Key deer",
+        "bbox": [
+            302,
+            873,
+            423,
+            883
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10706",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "class CoTProgram(dsyModule): def __init__(self): super().__init__(self.createAnswer = dspyChainOfThought(\"question -> answer\") def forward(self, question): return self.createAnswer(question=question)",
+        "guess_lang": "python",
+        "bbox": [
+            107,
+            85,
+            774,
+            172
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Python Snippets A.2: DSPy program for GSM8K.",
+        "text_level": 1,
+        "bbox": [
+            332,
+            175,
+            662,
+            192
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/f09037c6d8a9b62d97a29cd6119d52d83530e28f196da7c008c1c770a93bf895.jpg",
+        "table_caption": [
+            "Table A.2: Vanilla (zero-shot) prompts for the GSM8K program given in Python Snippets A.2 with one LM module. The generate_answer module prompts the corresponding LM to answer a grade school math problem. The green text shows the model completion obtained by setting all the modules of the program to use mistralai/Mistral-7b-Instruct-v0.2 as their corresponding LM. The model's response is evaluated as correct because the correct answer (1200) is contained in the first line of model's generation."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Module Label</td><td>Prompt</td></tr><tr><td>generate_answer</td><td>Given the fields `question`, produce the fields `answer`.</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Follow the following format.</td></tr><tr><td></td><td>Question: ${question}Reasoning: Let&#x27;s think step by step in order to ${produce the answer}. We...Answer: ${answer}</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Question: Micah can type 20 words per minute and Isaiah can type 40 words per minute. How many more words can Isaiah type than Micah in an hour?Reasoning: Let&#x27;s think step by step in order to find out how many more words Isaiah can type than Micah in an hour. We know that Micah can type 20 words per minute and Isaiah can type 40 words per minute. In one minute, Isaiah types twice as many words as Micah. In 60 minutes, Isaiah types 60 minutes * 40 words per minute = &lt;&lt;60*40=2400&gt;&gt;2400 words. Micah types 60 minutes * 20 words per minute = &lt;&lt;60*20=1200&gt;&gt;1200 words. The difference between the number of words Isaiah and Micah can type in an hour is 2400 words - 1200 words = &lt;&lt;2400-1200=1200&gt;&gt;1200 words.</td></tr><tr><td></td><td>Answer: Isaiah can type 1200 more words than Micah in an hour.</td></tr></table>",
+        "bbox": [
+            115,
+            300,
+            878,
+            640
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "class IrisSignature(dspy.Signature): \"Given the petal and sepal dimensions in cm, predict the iris species.\"   \npetal_length  $=$  dspy.InputField()   \npetal_width  $=$  dspy.InputField()   \nsepal_length  $=$  dspy.InputField()   \nsepal_width  $=$  dspy.InputField()   \nanswer  $=$  dspy.OutputField desc  $\\equiv$  'setosa, versicolor, or virginica')   \nclass IrisProgram(dspyModule): def __init__(self): self.createanswer  $=$  dspyChainOfThought(IrisSignature)   \ndef forward(self,petal_length,petal_width,sepal_length,sepal_width): return self.createanswer(petal_length  $\\equiv$  petal_length,petal_width= petal_width,sepal_length  $\\equiv$  sepal_length,sepal_width  $\\equiv$  sepal_width)",
+        "guess_lang": "python",
+        "bbox": [
+            105,
+            680,
+            810,
+            885
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Python Snippets A.3: DSPy program for Iris, provided to us by the DSPy team.",
+        "bbox": [
+            221,
+            890,
+            773,
+            904
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "10707",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/e8f91d76685d3fc4a14d633f0ef23c0837182ebae8f9da9049176c3e8a9ef398.jpg",
+        "table_caption": [
+            "Table A.3: Vanilla (zero-shot) prompts for the Iris program given in Python Snippets A.3 with one LM module. The generate_answer module prompts the corresponding LM to classify the Iris flowers into one of the following species given their features: Setosa, Versicolor, Virginica. The green text shows the model completion obtained by setting all the modules of the program to use mistralai/Mistral-7b-Instruct-v0.2 as their corresponding LM. The correct answer is Setosa, just as predicted."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Module Label</td><td>Prompt</td></tr><tr><td>generate_answer</td><td>Given the petal and sepal dimensions in cm, predict the iris species.</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Follow the following format.</td></tr><tr><td></td><td>Petal Length: ${petal_length}</td></tr><tr><td></td><td>Petal Width: ${petal_width}</td></tr><tr><td></td><td>Sepal Length: ${sepal_length}</td></tr><tr><td></td><td>Sepal Width: ${sepal_width}</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to ${produce the answer}. We</td></tr><tr><td></td><td>...</td></tr><tr><td></td><td>Answer: setosa, versicolor, or virginica</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Petal Length: 1.4</td></tr><tr><td></td><td>Petal Width: 0.2</td></tr><tr><td></td><td>Sepal Length: 5.0</td></tr><tr><td></td><td>Sepal Width: 3.6</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to predict the iris species. We will compare the given measurements with the average measurements of each iris species. 1. Setosa: The average petal length for setosa is 1.3 cm and the average petal width is 0.3 cm. The given petal length (1.4 cm) is slightly larger than the average, but the petal width (0.2 cm) is smaller than the average. However, the sepal dimensions (5.0 cm and 3.6 cm) are within the range of setosa. Based on these measurements, it is likely that the iris is setosa. 2. Versicolour: The average petal length for versicolour is 4.2 cm and the average petal width is 1.4 cm. The given petal length (1.4 cm) is smaller than the average, and the petal width (0.2 cm) is much smaller than the average. The sepal dimensions (5.0 cm and 3.6 cm) are also within the range of versicolour. However, the small petal dimensions suggest that it is less likely to be versicolour. 3. Virginica: The average petal length for virginica is 5.8 cm and the average petal width is 2.0 cm. The given petal length (1.4 cm) is much smaller than the average, and the petal width (0.2 cm) is much smaller than the average. The sepal dimensions (5.0 cm and 3.6 cm) are also within the range of virginica. However, the small petal dimensions suggest that it is less likely to be virginica. Based on the given measurements, it is most likely that the iris is setosa.</td></tr><tr><td></td><td>Answer: setosa.</td></tr></table>",
+        "bbox": [
+            117,
+            166,
+            878,
+            809
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Asset Information",
+        "text_level": 1,
+        "bbox": [
+            112,
+            841,
+            309,
+            857
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We share the associated licenses for the models and datasets we used below. For models, we list the specific HuggingFace model id we used to retrieve the respective weights.",
+        "bbox": [
+            112,
+            866,
+            882,
+            898
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "1. mistralai/Mistral-7b-Instruct-v0.2: Apache License 2.0",
+        "bbox": [
+            129,
+            904,
+            645,
+            920
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "10708",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "2. meta-llama/Llama-2-7b-chat-hf: Meta Llama 2 Community License at https://ai.meta.com/llama/licenses/",
+            "3. meta-llama/Meta-Llama-3-8B-Instruct: Meta Llama 3 Community License at https://llama.meta.com/llama3/legalize/",
+            "4. HotPotQA: Apache License 2.0",
+            "5. GSM8K: MIT License",
+            "6. Iris: Creative Commons Attribution 4.0 International (CC BY 4.0)"
+        ],
+        "bbox": [
+            127,
+            84,
+            885,
+            239
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "All the LMs used in this work are intended for use in English.",
+        "bbox": [
+            112,
+            250,
+            571,
+            265
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            277,
+            356,
+            293
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "In this section, we share the implementation details as it pertains to the LM inference and fine-tuning infrastructure, as well as compute requirements.",
+        "bbox": [
+            112,
+            303,
+            882,
+            335
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Inference For inference, we host our models in Docker (Merkel, 2014) instances through HuggingFace's text-generation-inference (HuggingFace, 2023) toolkit. We keep the sampling parameters the same across all experiments, using TopK sampling with a temperature of 0.1, and top_k of 0.97, until the model either generates a stopping string or a total of 1024 tokens including the tokens in the prompt.",
+        "bbox": [
+            112,
+            344,
+            884,
+            425
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Prompt Optimization For prompt optimization, we use the BootstrapFewShotRS (BFRS) optimizer from the DSPy library. In particular, we allow BFRS to randomly search 6 candidate programs using up to 3 few-shot examples for each module prompt.",
+        "bbox": [
+            112,
+            434,
+            882,
+            482
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Fine-tuning For fine-tuning, we use Low Rank Adaptation (LoRA; Hu et al. 2022) to train the query and key self-attention layers of our models, using a LoRA rank of 32, alpha of 64, with no dropout. We fine-tune all of our models for 5 epochs using bf10at16 precision, with a learning rate of 1e-5 and an effective batch size of 8.",
+        "bbox": [
+            112,
+            489,
+            882,
+            555
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Compute The BetterTogether strategies explored in this paper are naturally more expensive to run when compared to just prompt optimizing or fine-tuning. How these two steps compare to each other in terms of compute requirements or wall clock time depends on the particular settings used for each as well as the size of the dataset used. Total approximate GPU hours to produce Table 1 was  $\\approx 75$  hours, using A100 GPUs.",
+        "bbox": [
+            112,
+            564,
+            882,
+            643
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Extended Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            655,
+            304,
+            671
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "The results shared in Table 1 are the average of three runs. Tables D.1, D.2, and D.3 show the breakdown of the individual runs for HotPotQA, GSM8K, and Iris, respectively.",
+        "bbox": [
+            112,
+            681,
+            882,
+            713
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "10709",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/fed6960508c0cfb1a0bf095b8a8a1326570550215f72bbb2c293be048c3ae108.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Strategy</td><td colspan=\"4\">mistral-7b-instruct-v0.2</td><td colspan=\"4\">llama-2-7b-chat</td><td colspan=\"4\">llama-3-8b-instruct</td></tr><tr><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td></tr><tr><td colspan=\"13\">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>17.2</td><td>17.2</td><td>17.2</td><td>17.2</td><td>13.2</td><td>13.2</td><td>13.2</td><td>13.2</td><td>31.6</td><td>31.6</td><td>31.6</td><td>31.6</td></tr><tr><td>Prompt Optimization (Π)</td><td>32.7</td><td>34.7</td><td>34.0</td><td>33.8</td><td>33.3</td><td>33.3</td><td>33.4</td><td>33.3</td><td>45.7</td><td>47.4</td><td>47.5</td><td>46.9</td></tr><tr><td>Weight Optimization (Θ)</td><td>22.0</td><td>23.1</td><td>23.5</td><td>22.9</td><td>12.4</td><td>11.8</td><td>12.3</td><td>12.2</td><td>34.9</td><td>35.3</td><td>34.3</td><td>34.8</td></tr><tr><td>Π → Π</td><td>31.7</td><td>36.0</td><td>33.7</td><td>33.8</td><td>31.7</td><td>33.1</td><td>33.1</td><td>32.6</td><td>47.3</td><td>45.4</td><td>46.7</td><td>46.5</td></tr><tr><td>Θ → Θ</td><td>24.1</td><td>23.9</td><td>23.9</td><td>24.0</td><td>12.4</td><td>13.5</td><td>13.3</td><td>13.0</td><td>35.1</td><td>34.1</td><td>34.1</td><td>34.4</td></tr><tr><td colspan=\"13\">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>34.9</td><td>39.1</td><td>34.9</td><td>36.3</td><td>32.8</td><td>32.3</td><td>33.1</td><td>32.7</td><td>40.6</td><td>42.1</td><td>45.7</td><td>42.8</td></tr><tr><td>Θ → Π</td><td>29.3</td><td>33.8</td><td>35.8</td><td>33.0</td><td>36.0</td><td>33.4</td><td>33.1</td><td>34.2</td><td>44.5</td><td>40.9</td><td>45.3</td><td>43.6</td></tr><tr><td>Π → Θ → Π</td><td>34.9</td><td>40.7</td><td>37.2</td><td>37.6</td><td>34.7</td><td>34.5</td><td>35.3</td><td>34.8</td><td>46.5</td><td>47.1</td><td>46.4</td><td>46.7</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            878,
+            272
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/3febb67d53588251859e32283b9f6d2309db9c8ab020ca9e905107fe46f40a41.jpg",
+        "table_caption": [
+            "Table D.1: Results of HotPotQA Runs. Percentage accuracies of baseline and BetterTogether strategies on HotPotQA evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are performance of 3 runs on a held-out test set of 1500 examples, using different random seeds. Bold font in the average columns (Avg) is used to mark the highest score in a given column."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Strategy</td><td colspan=\"4\">mistral-7b-instruct-v0.2</td><td colspan=\"4\">llama-2-7b-chat</td><td colspan=\"4\">llama-3-8b-instruct</td></tr><tr><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td></tr><tr><td colspan=\"13\">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>40.3</td><td>40.3</td><td>40.3</td><td>40.3</td><td>24.0</td><td>24.0</td><td>24.0</td><td>24.0</td><td>72.7</td><td>72.7</td><td>72.7</td><td>72.7</td></tr><tr><td>Prompt Optimization (Π)</td><td>45.0</td><td>47.2</td><td>47.1</td><td>46.4</td><td>27.3</td><td>25.1</td><td>25.5</td><td>26.0</td><td>76.9</td><td>77.9</td><td>78.9</td><td>77.9</td></tr><tr><td>Weight Optimization (Θ)</td><td>40.8</td><td>40.0</td><td>41.2</td><td>40.7</td><td>23.7</td><td>24.2</td><td>24.0</td><td>24.0</td><td>75.7</td><td>74.8</td><td>74.8</td><td>75.1</td></tr><tr><td>Π → Π</td><td>46.3</td><td>47.2</td><td>49.6</td><td>47.7</td><td>28.4</td><td>24.0</td><td>21.8</td><td>24.7</td><td>76.5</td><td>80.1</td><td>76.1</td><td>77.6</td></tr><tr><td>Θ → Θ</td><td>42.9</td><td>41.8</td><td>43.8</td><td>42.8</td><td>24.0</td><td>24.3</td><td>24.0</td><td>24.1</td><td>52.2</td><td>36.6</td><td>43.4</td><td>44.0</td></tr><tr><td colspan=\"13\">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>46.4</td><td>47.3</td><td>48.2</td><td>47.3</td><td>27.8</td><td>28.1</td><td>25.9</td><td>27.3</td><td>77.6</td><td>75.4</td><td>79.8</td><td>77.6</td></tr><tr><td>Θ → Π</td><td>50.1</td><td>46.0</td><td>48.8</td><td>48.3</td><td>26.8</td><td>26.1</td><td>27.0</td><td>26.6</td><td>78.5</td><td>79.8</td><td>78.4</td><td>78.9</td></tr><tr><td>Π → Θ → Π</td><td>44.9</td><td>48.5</td><td>47.1</td><td>46.8</td><td>27.1</td><td>25.9</td><td>25.9</td><td>26.3</td><td>77.6</td><td>75.4</td><td>77.8</td><td>77.0</td></tr></table>",
+        "bbox": [
+            119,
+            348,
+            878,
+            539
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/eaebff68b94cd91c15c487f2e142876757b3e3d487f9bdbbf704a75043277635.jpg",
+        "table_caption": [
+            "Table D.2: Results of GSM8K Runs. Percentage accuracies of baseline and BetterTogether strategies on GSM8K evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are performance of 3 runs on a held-out test set of 1319 examples, using different random seeds. Bold font in the average columns (Avg) is used to mark the highest score in a given column."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Strategy</td><td colspan=\"4\">mistral-7b-instruct-v0.2</td><td colspan=\"4\">llama-2-7b-chat</td><td colspan=\"4\">llama-3-8b-instruct</td></tr><tr><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td></tr><tr><td colspan=\"13\">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>26.0</td><td>26.0</td><td>26.0</td><td>26.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>48.0</td><td>48.0</td><td>48.0</td><td>48.0</td></tr><tr><td>Prompt Optimization (Π)</td><td>52.0</td><td>54.0</td><td>66.0</td><td>57.3</td><td>44.0</td><td>68.0</td><td>58.0</td><td>56.7</td><td>62.0</td><td>96.0</td><td>80.0</td><td>79.3</td></tr><tr><td>Weight Optimization (Θ)</td><td>24.0</td><td>34.0</td><td>30.0</td><td>29.3</td><td>-</td><td>-</td><td>-</td><td>-</td><td>38.0</td><td>40.0</td><td>34.0</td><td>37.3</td></tr><tr><td>Π → Π</td><td>48.0</td><td>64.0</td><td>66.0</td><td>59.3</td><td>66.0</td><td>70.0</td><td>56.0</td><td>64.0</td><td>70.0</td><td>94.0</td><td>82.0</td><td>82.0</td></tr><tr><td>Θ → Θ</td><td>40.0</td><td>36.0</td><td>38.0</td><td>38.0</td><td>-</td><td>-</td><td>-</td><td>-</td><td>44.0</td><td>36.0</td><td>38.0</td><td>39.3</td></tr><tr><td colspan=\"13\">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>32.0</td><td>26.0</td><td>34.0</td><td>30.7</td><td>30.0</td><td>26.0</td><td>24.0</td><td>26.7</td><td>50.0</td><td>42.0</td><td>40.0</td><td>44.0</td></tr><tr><td>Θ → Π</td><td>80.0</td><td>54.0</td><td>66.0</td><td>66.7</td><td>-</td><td>-</td><td>-</td><td>-</td><td>78.0</td><td>78.0</td><td>80.0</td><td>78.7</td></tr><tr><td>Π → Θ → Π</td><td>52.0</td><td>44.0</td><td>62.0</td><td>52.7</td><td>62.0</td><td>70.0</td><td>64.0</td><td>65.3</td><td>74.0</td><td>80.0</td><td>84.0</td><td>79.3</td></tr></table>",
+        "bbox": [
+            119,
+            615,
+            878,
+            806
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Table D.3: Results of Iris Runs. Percentage accuracies of baseline and BetterTogether strategies on Iris evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are performance of 3 runs on a held-out test set of 50 examples, using different random seeds. Bold font in the average columns (Avg) is used to mark the highest score in a given column. Strategies where weight optimization is the first step use the vanilla (zero-shot) strategy to generate the initial fine-tuning dataset. If a model generates very few or no correct outputs under the vanilla strategy on the training set used to bootstrap the fine-tuning data, there will not be a sufficient dataset for fine-tuning. These settings are marked with “-”.",
+        "bbox": [
+            112,
+            816,
+            882,
+            917
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "10710",
+        "bbox": [
+            477,
+            928,
+            526,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2024/Fine-Tuning and Prompt Optimization_ Two Great Steps that Work Better Together/b6b2cc5a-f4ec-482a-8da6-09a5acae609f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d40230b78c0e044c39aeba29dca68dbdca080f3f611bffc5ed3fc5a7df7e77d5
+size 421838

2024/Fine-Tuning and Prompt Optimization_ Two Great Steps that Work Better Together/full.md

@@ -0,0 +1,308 @@
+# Fine-Tuning and Prompt Optimization: Two Great Steps that Work Better Together
+
+Dilara Soylu Christopher Potts Omar Khattab
+
+Stanford University
+
+{soylu, cgpotts, okhattab}@stanford.edu
+
+# Abstract
+
+Natural Language Processing (NLP) systems are increasingly taking the form of sophisticated modular pipelines, e.g., Retrieval Augmented Generation (RAG), where each module may involve a distinct Language Model (LM) and an associated prompt template. These compound systems often lack intermediate labels or gradient flow to optimize each module, making their end-to-end optimization challenging. Here we seek strategies to optimize both the module-level LM weights and the associated prompt templates of such systems to maximize a downstream task metric. We propose for the first time combining the weight and prompt optimization strategies to optimize a modular LM pipeline by alternating between the two to get the same LM to teach itself. In experiments with multi-hop QA, mathematical reasoning, and feature-based classification using mistral-7b, llama-2-7b, and llama-3-8b, these BetterTogether strategies optimizing the weights and prompts of a pipeline together outperform directly optimizing weights alone and prompts alone by up to  $60\%$  and  $6\%$ , respectively, on average across LMs and tasks. Our BetterTogether optimizer is released in DSPy at http://dspy.ai.
+
+# 1 Introduction
+
+While the capabilities of language models (LMs) continue to grow, recent work has shown the potential of building more powerful Natural Language Processing (NLP) systems by composing multiple skills of LMs into pipelines. Examples of this include systems for retrieval-augmented generation (Guu et al., 2020; Lewis et al., 2020; Ma et al., 2023; Jiang et al., 2023b), multi-hop reasoning (Qi et al., 2021; Khattab et al., 2021), information extraction (Pourreza and Rafiei, 2023; D'Oosterlinck et al., 2024), and other sophisticated pipelines (Zelikman et al., 2022; Dohan et al., 2022; Khattab et al., 2022; Beurer-Kellner et al., 2023; Schlag et al., 2023; Viswanathan et al., 2023).
+
+Such LM Programs offer much more control for designing NLP systems, as they break down problems into modular, more manageable sub-tasks that can be assigned to LMs. If we could teach these LMs to accurately conduct their easier sub-tasks and to communicate effectively within multi-stage pipelines, this could greatly expand the scope of reliable NLP systems we can build.
+
+To this end, Khattab et al. (2024) introduced the DSPy framework for defining and optimizing LM Programs. In it, a program is defined as a function  $\Phi$  that composes a set of stages, which we will refer to as language modules  $M = \langle M_1,\dots ,M_{|M|}\rangle$ , into a pipeline. Each language module  $M_{i}$  specifies a fuzzy natural-language transformation (e.g., generating a summary of a supplied document) that needs to be learned. To do so, each module learns a particular prompt (template)  $\pi$  to make a call to a particular LM with weights  $\theta$ . The optimization problem is then defined as maximizing the expected performance (per a downstream metric  $\mu$ ) of the program  $\Phi$  over a set of inputs by updating each module's  $\pi$  and  $\theta$ .
+
+Existing work (Khattab et al., 2024; Opsahl-Ong et al., 2024) has studied optimizing the discrete string prompt of each module and has considered simple approaches for fine-tuning each module's LM weights. In this empirical study, we investigate updating each module's LM weights and prompt template together to maximize a downstream metric on the final output of the program. Doing this is challenging as  $\Phi$  is not generally differentiable and its modules  $M_{i}$  generally lack labeled outputs, while exhibiting sophisticated dependencies. Moreover, in realistic settings, the training set is usually very small and only a small number of LM calls are possible for training and inference.
+
+To address this challenge, we propose to alternate between fine-tuning LM weights and optimizing prompt templates and evaluate approximate optimization strategies in which we bootstrap
+
+training labels for all pipeline modules. In experiments with multi-hop QA (HotPotQA), mathematical reasoning (GSM8K), and feature-based classification (Iris), we show that these tandem strategies are highly effective across three different LMs, leading to  $5 - 78\%$  gains for HotPotQA, 2.5-  $10\%$  gains for GSM8K, and  $3.5 - 88\%$  gains for Iris against prompts only and weights only strategies, averaged across mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct.
+
+# 2 Problem Statement
+
+We are given an LM program  $\Phi$ , which operates like a blackbox function  $\Phi: \mathcal{X} \to \mathcal{Y}$ , in which  $\mathcal{X}$  and  $\mathcal{Y}$  are typically in natural language (e.g., questions and their program-generated answers, respectively). For example, we may have a program  $\Phi$  for answering complex questions with short factoid answers. In the course of its execution,  $\Phi$  makes one or more calls to each of its  $|M| \geq 1$  language modules,  $M = \langle M_1, \ldots, M_{|M|} \rangle$ .
+
+For example, the program may implement a multi-hop, retrieval-augmented pipeline for question answering. This common pipeline (Qi et al., 2021; Khattab et al., 2021; Press et al., 2023; Khattab et al., 2022) breaks down the input into subquestions that are used to iteratively find relevant passages (e.g., from a corpus like Wikipedia) until the question can be faithfully answered. In general terms, each module  $M_{i}:\mathcal{X}_{i}\to \mathcal{Y}_{i}$  is a declarative LM invocation that defines, in inherently fuzzy natural-language terms, an input  $\mathcal{X}_i$  domain (like a user-supplied question and a set of retrieved passages) and an output  $\mathcal{Y}_i$  co-domain (like a search query to find additional relevant passages).
+
+We seek to implement each language module as some specific, well-tuned strategy for invoking an underlying language model LM. Concretely, we assume that a module  $M_{i}$  will be fully implemented by specifying (1) the string prompt  $\pi_{i}$  in which the module inputs  $\mathcal{X}_i$  are plugged in to decode the module outputs  $\mathcal{V}_i$  and (2) the floating-point weights  $\theta_{i}$  assigned to the parameters of LM in the course of this module. We refer to the version of  $\Phi$  in which the prompts and LM weights are assigned explicitly to  $\Pi$  and  $\Theta$ , respectively, as  $\Phi_{\langle \Theta ,\Pi \rangle}$ .
+
+Given nothing but a small training set  $X = \{(x_{1},m_{1}),\ldots ,(x_{|X|},m_{|X|})\}$  of inputs  $x_{i}\in \mathcal{X}$  and optional metadata like output labels or other hints  $m_{i}\in \mathcal{M}$  that can be used for determining the correctness of a given program run, and a metric
+
+$\mu : \mathcal{Y} \times \mathcal{M} \to \mathbb{R}$ , our goal is to optimize  $\Phi$ , that is, configure its modules' prompts and LM weights to maximize the following objective.
+
+$$
+\underset {\Theta , \Pi} {\arg \max } \frac {1}{| X |} \sum_ {(x, m) \in X} \mu \left(\Phi_ {\langle \Theta , \Pi \rangle} (x), m\right)
+$$
+
+Researchers tuning LM pipelines are in effect seeking to achieve this objective. It is also a very large subspace of the optimization problem in the DSPy framework<sup>1</sup> for LM programs. Unfortunately, this problem is intractable: the search space is large and we don't have gradients or intermediate output labels to optimize each module, so we seek approximate strategies for such optimization.
+
+# 3 BetterTogether: Alternating Weight and Prompt Optimization Steps for LM Programs
+
+We now introduce the BetterTogether algorithm, which alternates the weight and prompt optimization steps for LM programs. We hypothesize that, when a large LM is used to teach itself how to tackle the task defined by an LM program, finetuning LM weights and prompts are both essential to achieve the highest quality. In particular, we expect that (1) prompt optimization before finetuning can lead to more successful datapoints for fine-tuning, and, (2) prompt optimization after finetuning can make adjustments to the behavior of the LM program, leading to higher quality outputs. Considering that fine-tuning is often perceived as a more powerful tool, this can be surprising, especially when both approaches are ultimately applied over the same set of training inputs  $X$ .
+
+# Algorithm 1 BetterTogether: Optimizing LM programs by alternating prompt and weight optimization steps, instantiated in Algorithm 2
+
+<table><tr><td colspan="2">Input: Program Φ(Θ,Π) = ΦΘ ⊙ ΦΠ, 
+    with module weights Θ = [θ1, ..., θ|Φ|] 
+    and module prompts Π = [π1, ..., π|Φ|]</td></tr><tr><td colspan="2">Training Set X and Metric μ</td></tr><tr><td colspan="2">1: function BETTERTOGETHER(Φ(Θ,Π), X, μ)</td></tr><tr><td colspan="2">2: Π&#x27; ← OPTIMIZEPROMPTS(Φ(Θ,Π), X, μ)</td></tr><tr><td colspan="2">3: Θ&#x27; ← FINETUNEWEIGHTS(Φ(Θ,Π&#x27;), X, μ)</td></tr><tr><td colspan="2">4: Π&#x27;&#x27; ← OPTIMIZEPROMPTS(Φ(Θ&#x27;,Π), X, μ)</td></tr><tr><td colspan="2">5: return Φ(Θ&#x27;,Π&#x27;)</td></tr><tr><td colspan="2">6: end function</td></tr></table>
+
+Accordingly, the general optimization framework for our algorithm is defined in Algorithm 1.
+
+Given a program  $\Phi$ , the algorithm begins by optimizing  $\Phi$ 's prompts, then fine-tuning its set of LM weights with the data bootstrapped using the optimized prompts, and finally optimizing its prompts again using the fine-tuned weights. In principle, each of these steps could be treated as optional. This will define the different possible combinations of BetterTogether that we will seek to evaluate in Section 4. Specifically, we are interested in the quality of (1) the vanilla program  $\Phi$  with simple user-supplied instructions as the prompts and no fine-tuning of LM, (2) optimizing the prompts only, (3) optimizing the weights only, (4) optimizing the prompts twice, i.e., using the prompt-optimized  $\Phi$  as a starting point for a second round of prompt optimization, (5) optimizing the weights twice, (6) optimizing the prompts then the weights, (7) vice versa, and (8) optimizing the prompts, weights, then prompts. Overall, we expect the final three to consistently outperform the first five.
+
+For Algorithm 1 to be complete, we need to instantiate Lines 1-3 with specific approaches for prompt optimization and LM fine-tuning. For this, we choose the Bootstrap- $*$  family of algorithms from Khattab et al. (2024), which work by executing an initial version of the program on input examples  $(x_{i},m_{i})\in X$  and recording the inputs/outputs observed at each module when the final output is "correct", i.e.,  $\mu (\Phi (x_i),m_i)\geq \lambda$  for some threshold  $\lambda$  (e.g., 1.0 for binary accuracy). This is important to note: in line with our problem formulation, our prompt and weight optimization regimes are not simply training on hand-labeled data but on self-generated program traces.
+
+Instantiations for Lines 1-3 of Algorithm 1 are shown in Algorithm 2. For prompt optimization, we use BootstrapFewshotRS (BFRS) of DSPy, which self-generates potential few-shot examples of every module and applies a form of random search (RS) to select the specific generated few-shot examples that are used for prompting. Overall, BFRS first divides  $X$  into a training split  $T$  and a validation split  $V$  (Line 2). It then executes the provided  $\Phi_{\langle \Theta, \Pi \rangle}$  on the training inputs, collecting input-output pairs for every module in  $\Phi$  for each  $x_i \in T$ . This is called a trace  $\tau$ , and we keep only the traces assigned high scores by  $\mu$  (Line 4). Given all of these traces, BFRS samples multiple different subsets of a few traces  $\tau'$  (Line 6), each of them containing a potential few-shot example for each module in  $\Phi$ , and ultimately selects the subset that, when used to construct few-shot prompts
+
+Algorithm 2 Instantiating Algorithm 1's prompt & weight optimizers with bootstrapping algorithms  
+Input: Training Set  $X$  and Metric  $\mu$   
+1: function BOOTSTRAPFEWSHOTRS(  $\Phi_{\langle \Theta ,\Pi \rangle}$ $X,\mu)$   
+2:  $T,V\gets$  SPLITINTOTRAINANDVALIDATION(X)  
+3:  $\tau \leftarrow$  BOOTSTRAPTRACES(  $\Phi_{\langle \Theta ,\Pi \rangle}$ $T)$   
+4:  $\tau \leftarrow$  FILTERTRACES(  $\tau ,\mu)$   
+5: Initialize attempts list  $\mathcal{A}\gets \{\}$   
+6: for  $\tau^{\prime}\in$  SAMPLEFEWSHOTSUBSETS(τ) do  
+7:  $\Pi^{\prime}\gets$  CONSTRUCTFEWSHOTPROMPTS(τ')  
+8:  $\sigma \leftarrow \frac{1}{|V|}\sum_{\langle x_i,m_i\rangle \in V}\mu (\Phi_{\langle \Theta ,\Pi '}\rangle (x_i),m_i)$   
+9: Extend A with  $(\sigma ,\Pi^{\prime})$   
+10: end for  
+11: return  $\Pi_{\mathrm{max}}$  , A's highest-scoring prompts sequence  
+12: end function  
+13:  
+14: function BOOTSTRAPFINETUNE(  $\Phi_{\langle \Theta ,\Pi \rangle}$ $X,\mu)$   
+15:  $\tau \leftarrow$  BOOTSTRAPTRACES(  $\Phi_{\langle \Theta ,\Pi \rangle}$ $X)$   
+16:  $\tau \leftarrow$  FILTERTRACES(  $\tau ,\mu)$   
+17:  $\Theta^{\prime}\gets$  TRAINLM(τ)  
+18: return  $\Theta^{\prime}$   
+19: end function  
+20:  
+21: Set OPTIMIZEPROMPTS as BOOTSTRAPFEWSHOTRS  
+22: Set FINETUNEWEIGHTS as BOOTSTRAPFINETUNE
+
+(Line 7), achieves the highest score (Line 8). This simple search strategy is known to consistently lead to large quality improvements in prompting LM programs (Khattab et al., 2024; Opsahl-Ong et al., 2024), often outperforming manually or automatically optimizing prompt instructions or writing examples by hand.
+
+For fine-tuning, we extend BootstrapFinetune (BFT) of DSPy, which, given a program  $\Phi$ , self-generates a large number examples for every module and combines them into one dataset to fine-tune the LM weights with an implicit multi-task objective, where the sub-tasks are the modules' roles. Existing work has only considered BFT in a very narrow setting for LM programs: on HotPotQA, Khattab et al. (2024) train a T5-Large model using traces from a few-shot Llama2-13b program, without considering getting an LM to teach itself via BFT nor considering a role for BFRS in the fine-tuned program. In this work, we focus on allowing models to teach themselves and self-improve. We propose for the first time combining the strategies of BFRS and BFT via alternation to get the same LM to teach itself better than either prompt or weight optimization in isolation. One could test similar ideas in scenarios where a larger model does the bootstrapping for a smaller LM. This may lead to even better results but is outside our scope.
+
+<table><tr><td rowspan="2">Strategy</td><td colspan="3">mistral-7b-instruct-v0.2</td><td colspan="3">llama-2-7b-chat</td><td colspan="3">llama-3-8b-instruct</td></tr><tr><td>HotPotQA</td><td>GSM8K</td><td>Iris</td><td>HotPotQA</td><td>GSM8K</td><td>Iris</td><td>HotPotQA</td><td>GSM8K</td><td>Iris</td></tr><tr><td colspan="10">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>17.2</td><td>40.3</td><td>26.0</td><td>13.2</td><td>24.0</td><td>0.0</td><td>31.6</td><td>72.7</td><td>48.0</td></tr><tr><td>Prompt Optimization (Π)</td><td>33.8</td><td>46.4</td><td>57.3</td><td>33.3</td><td>26.0</td><td>56.7</td><td>46.9</td><td>77.9</td><td>79.3</td></tr><tr><td>Weight Optimization (Θ)</td><td>22.9</td><td>40.7</td><td>29.3</td><td>12.2</td><td>24.0</td><td>-</td><td>34.8</td><td>75.1</td><td>37.3</td></tr><tr><td>Π → Π</td><td>33.8</td><td>47.7</td><td>59.3</td><td>32.6</td><td>24.7</td><td>64.0</td><td>46.5</td><td>77.6</td><td>82.0</td></tr><tr><td>Θ → Θ</td><td>24.0</td><td>42.8</td><td>38.0</td><td>13.0</td><td>24.1</td><td>-</td><td>34.4</td><td>44.1</td><td>39.3</td></tr><tr><td colspan="10">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>36.3</td><td>47.3</td><td>30.7</td><td>32.7</td><td>27.3</td><td>26.7</td><td>42.8</td><td>77.6</td><td>44.0</td></tr><tr><td>Θ → Π</td><td>33.0</td><td>48.3</td><td>66.7</td><td>34.2</td><td>26.6</td><td>-</td><td>43.6</td><td>78.9</td><td>78.7</td></tr><tr><td>Π → Θ → Π</td><td>37.6</td><td>46.8</td><td>52.7</td><td>34.8</td><td>26.3</td><td>65.3</td><td>46.7</td><td>77.0</td><td>79.3</td></tr></table>
+
+Table 1: Main Results. Percentage accuracies of baseline and BetterTogether strategies on HotPotQA, GSM8K, and Iris evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are average performance of 3 runs on a held-out test set using different random seeds. Bold font is used to mark the highest score in a given column. Strategies where weight optimization is the first step use the vanilla (zero-shot) strategy to generate the initial fine-tuning dataset. If a model generates very few or no correct outputs under the vanilla strategy on the training set used to bootstrap the fine-tuning data, there will not be a sufficient dataset for fine-tuning. These settings are marked with “-”.
+
+# 4 Experimental Evaluation
+
+We now seek to evaluate our hypothesis on the importance of optimizing both LM weights and prompts of LM programs. We conduct our evaluation across three datasets that span different tasks (and thus LM programs) each. In particular, we use HotPotQA (Yang et al., 2018) for multi-hop reasoning, GSM8K (Cobbe et al., 2021) for arithmetic reasoning, and Iris (Fisher, 1988) for classification. We run our experiments using three models, mistral-7b-instruct-v0.2 (Jiang et al., 2023a), llama-2-7b-chat (Touvron et al., 2023), llama-3-8b-instruct (MetaAI, 2024), keeping the model used for prompt optimization, bootstrapping training traces, and fine-tuning the same in a given experiment for all modules. We initialize all the modules of an input program  $\Phi$  to use the same LM weights (e.g. mistral-7b-instruct-v0.2), but distinct prompt templates specialized for their particular module-level task, such as generating a search query or answering a question. We implement all of our programs and optimizers as extensions to the DSPy framework.
+
+For each dataset, we split the data into training, development, and test sets. We shuffle the training set every time before we perform a prompt or weight optimization, controlled by random seed. For prompt optimization, we sub-sample non-overlapping sets of training and validation examples from the initial training set for each task
+
+and use the BFRS prompt optimizer to optimize the module level prompts of a given  $\Phi$ , leaving its underlying LM weights untouched. For weight optimization, we use all the available training examples as potential candidates to generate the data for finetuning  $\Phi$ 's LM weights and pass them to the BFT weight optimizer, which: (1) runs a given  $\Phi$  on all the training examples, (2) keeps the traces where the final output of  $\Phi$  was correct and filters out the rest, (3) gets module level prompt-completion pairs for each trace, (4) creates new pairs by replacing the prompt generated by  $\Phi$  with a vanilla prompt, (5) combines all the module level prompt-completion pairs into one dataset, (6) fine-tunes  $\Phi$ 's LM weights on this dataset, and finally, (7) returns an updated  $\Phi$  where the LM weights of all the modules are set to the fine-tuned LM. We use the Low Rank Adaptation (LoRA; Hu et al. 2022) method to fine-tune our LMs.
+
+The full text for our programs and the vanilla prompts are shared in Appendix A. The license information for all LMs and datasets used as well as our implementation details such as hyperparameters and software are reported in Appendix B and Appendix C, respectively.
+
+Multi-hop Reasoning HotPotQA (in the "full-wiki" setting) is a question answering task in which systems must find two Wikipedia page abstracts out of a corpus of 5 million via search and use them to answer a factoid question. Therefore it can be
+
+implemented as a program that has three LM modules: the first two for generating search queries (i.e., hops) and the last one for generating an answer. Each module uses Chain-of-Thought (CoT; Wei et al. 2022) to generate its outputs, producing a reasoning string before the search query or the answer. Search queries are passed to a frozen CoBERTv2 (Santhanam et al., 2022) retriever. Accuracy is measured using the exact match score of the answer with the ground truth answer for the given question, after normalizing case, stripping surrounding whitespace characters, and removing punctuation. We use the following splits for HotPotQA: 1000 training examples and 500 development examples drawn from the original training set, along with 1500 test examples drawn from the original validation set, since the original test set is not public. We sub-sample 100 and 250 non-overlapping examples for the prompt optimization training and validation sets, respectively.
+
+Arithmetic Reasoning GSM8K is a benchmark consisting of grade school math problems. We implement it as an LM program with a single module using CoT prompting, where the LM generates a reasoning string followed by an answer. We measure accuracy by extracting the last number from the first line of the model's response and comparing it to the ground truth response. For GSM8K, we use a training set of 1000 examples and a development set of 500 examples, both sampled from the original training set. We use all the 1319 test examples available for our test set. We use 100 and 250 non-overlapping examples for training and validation sets used during prompt optimization.
+
+Classification Iris is a classic classification task, where the goal is to classify species of Iris flowers. We use a single-module CoT DSPy program for Iris, with the goal of assessing whether it being a feature-based classification task gives a large advantage to methods based entirely on gradient descent (fine-tuning). This tests the extrapolation of our hypothesis to a different setting from the other two tasks. We measure accuracy using the exact match score of the answer and the correct answer given a question after normalizing both, as is the case for HotPotQA. The Iris dataset has a total of 150 examples, from which we create training, development, and test splits of equal size. We sub-sample 15 and 35 non-overlapping examples to be used as the training and validation sets for prompt optimization, respectively.
+
+# 5 Results & Discussion
+
+Held-out test set performance of the strategies described in Section 3 is shared in Table 1. Reported values are averaged across three runs with unique random seeds. Results for the individual runs are reported separately in Appendix D.
+
+In 7 out of the 9 dataset and LM pairs, we observe that the best-performing strategies are the strategies that utilize prompt (Π) and weight (Θ) optimization steps together, although there is no clear winner among the benchmarked BetterTogether strategies that optimize both. Overall, optimizing prompts is essential on all the tasks, but optimizing prompts and weights together leads to strong gains over the best setting that only optimizes one of the two.
+
+In summary, we have proposed to alternate between optimizing prompts and fine-tuning LM weights and explored a few strategies for doing so. In experiments with multi-hop QA (HotPotQA), mathematical reasoning (GSM8K), and feature-based classification (Iris), we show that our strategies are highly effective for getting an LM to teach itself to perform an LM program via bootstrapping, leading to  $5 - 78\%$  gains for HotPotQA,  $2.5 - 10\%$  gains for GSM8K, and  $3.5 - 88\%$  gains for Iris.
+
+# 6 Limitations
+
+While this paper presents strong evidence from nine cases, spanning three tasks (and their corresponding LM programs) and three LMs, it is possible that other tasks, programs, or LMs will change the pattern in unforeseen ways. In particular, we have only experimented with weight optimization in the form of LoRA fine-tuning of pre-trained models. It is in principle possible that some other fine-tuning strategy would be so powerful and cost-effective as to remove the need for prompt optimization.
+
+In addition, though we expect our findings to inform many researchers and practitioners interested in optimizing LM programs, and encourage them to explore optimizing prompts and fine-tuning LM weights together, we do not yet understand why both are important. The role of prompt optimization and the role of fine-tuning in multi-stage LM programs are both new, and the relative lack of understanding of these roles in the emerging literature could pose risks in unanticipated interactions between these components, compared with standard gradient descent for neural networks, which has been studied for decades.
+
+# Acknowledgments
+
+D.S. is supported by Ravi Family Graduate Fellowship. This work was partially supported by IBM as a founding member of the Stanford Institute for Human-Centered Artificial Intelligence (HAI), and by the HAI Hoffman-Yee Grant "Dendritic Computation for Knowledge Systems".
+
+# References
+
+Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. 2023. Prompting is programming: A query language for large language models. Proc. ACM Program. Lang., 7(PLDI).  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.  
+David Dohan, Winnie Xu, Aitor Lewkowycz, Jacob Austin, David Bieber, Raphael Gontijo Lopes, Yuhuai Wu, Henryk Michalewski, Rif A. Saurous, Jascha Sohl-dickstein, Kevin Murphy, and Charles Sutton. 2022. Language model cascades. Preprint, arXiv:2207.10342.  
+Karel D'Oosterlinck, Omar Khattab, François Remy, Thomas Demeester, Chris Develder, and Christopher Potts. 2024. In-context learning for extreme multi-label classification. Preprint, arXiv:2401.12178.  
+Ronald A. Fisher. 1988. Iris. UCI Machine Learning Repository.  
+Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Retrieval augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 3929-3938. PMLR.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.  
+HuggingFace. 2023. Text generation inference.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023a. Mistral 7b. Preprint, arXiv:2310.06825.
+
+Zhengbao Jiang, Frank Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023b. Active retrieval augmented generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7969-7992, Singapore. Association for Computational Linguistics.  
+Omar Khattab, Christopher Potts, and Matei A. Zaharia. 2021. Baleen: Robust multi-hop reasoning at scale via condensed retrieval. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 27670-27682.  
+Omar Khattab, Keshav Santhanam, Xiang Lisa Li, David Hall, Percy Liang, Christopher Potts, and Matei Zaharia. 2022. Demonstrate-search-predict: Composing retrieval and language models for knowledge-intensive nlp. ArXiv preprint, abs/2212.14024.  
+Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan A, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2024. DSPy: Compiling declarative language model calls into state-of-the-art pipelines. In The Twelfth International Conference on Learning Representations.  
+Patrick S. H. Lewis, Ethan Perez, Aleksandra Pik-tus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.  
+Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting in retrieval-augmented large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5303-5315, Singapore. Association for Computational Linguistics.  
+Dirk Merkel. 2014. Docker: lightweight linux containers for consistent development and deployment. Linux J., 2014(239).  
+MetaAI. 2024. Meta llama 3.  
+Krista Opsahl-Ong, Michael J Ryan, Josh Purtell, David Broman, Christopher Potts, Matei Zaharia, and Omar Khattab. 2024. Optimizing instructions and demonstrations for multi-stage language model programs.  
+Mohammadreza Pourreza and Davood Rafiei. 2023. DIN-SQL: decomposed in-context learning of text-to-sql with self-correction. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.
+
+Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 5687-5711, Singapore. Association for Computational Linguistics.  
+Peng Qi, Haejun Lee, Tg Sido, and Christopher Manning. 2021. Answering open-domain questions of varying reasoning steps from text. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3599-3614, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. 2022. ColBERTv2: Effective and efficient retrieval via lightweight late interaction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3715-3734, Seattle, United States. Association for Computational Linguistics.  
+Imanol Schlag, Sainbayar Sukhbaatar, Asli Celikyilmaz, Wen tau Yih, Jason Weston, Jürgen Schmidhuber, and Xian Li. 2023. Large language model programs. Preprint, arXiv:2305.05364.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.  
+Vijay Viswanathan, Chenyang Zhao, Amanda Bertsch, Tongshuang Wu, and Graham Neubig. 2023. Prompt2Model: Generating deployable models from natural language instructions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 413-421, Singapore. Association for Computational Linguistics.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le,
+
+and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics.  
+Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. 2022. Star: Bootstrapping reasoning with reasoning. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.
+
+# Appendices
+
+# A Programs
+
+The DSPy programs for HotPotQA, GSM8K, and Iris are shared in Python Snippets A.1, A.2, A.3, respectively. Tables A.1, A.2, A.3 show the vanilla (zero-shot) programs generated with these programs, along with sample model completions.
+
+```txt
+class HotPotQAProgram(dspyModule): def __init__(self, passages_per_hop  $= 3$  .. super().__init__() self.retrieve  $=$  dspy.Retrieve  $(k =$  passages_per_hop) self_generate_query  $=$  [dspyChainOfThought("context，question-> search_query")for_in range(2)] self_generate_answer  $=$  dspyChainOfThought("context，question->answer") def forward(self,question): context  $=$  [] for hop in range(2): search_query  $=$  self_generate_query[hop](context  $\equiv$  context，question= question).search_query passages  $=$  self.retrieve.search_query).passages context  $=$  disp.utils.deduplicate(context+passages) return self_generate_answer(context  $\equiv$  context，question  $\equiv$  question).copy( context  $\equiv$  context)
+```
+
+Python Snippets A.1: DSPy program for HotPotQA, consisting of three Chain-of-Thought (CoT) LM modules, along with a frozen retriever. Each LM module in the program uses its own prompt and LM weights under the hood.
+
+Table A.1: Vanilla (zero-shot) prompts for the HotPotQA program given in Python Snippets A.1 with three LM modules, each with corresponding prompts and LM weights. (1) the generate_query[0] module prompts the corresponding LM to generate a search query given a question along with an empty context; (2) the generate_query[1] module prompts the corresponding model to generate a search query given the context paragraphs retrieved using the search query outputted by the prior module; (3) the generate_answer module prompts the corresponding LM to answer the question given the context paragraphs retrieved by the earlier modules. The green text shows the model completions obtained by setting all the modules of the program to use mistralai/Mistral-7b-Instruct-v0.2 as their corresponding LM. The correct answer to the question is indeed Key deer as predicted.  
+
+<table><tr><td>Module Label</td><td>Prompt</td></tr><tr><td>generate_query[0]</td><td>Given the fields `context`, `question`, produce the fields `search_query`.</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Follow the following format.</td></tr><tr><td></td><td>Context: ${context}</td></tr><tr><td></td><td>Question: ${question}</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to ${produce the search_query}. We ...</td></tr><tr><td></td><td>Search Query: ${search_query}</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Context: N/A</td></tr><tr><td></td><td>Question: What animal subject to the Foster&#x27;s rule is the smallest North American deer?</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to find the smallest North American deer subject to the Foster&#x27;s rule.</td></tr><tr><td></td><td>1. The Foster&#x27;s rule is a mathematical formula used to estimate the body weight of large mammals based on their ear surface area.2. To apply the Foster&#x27;s rule, we need to find the ear surface area of the deer species in question.3. We don&#x27;t have the ear surface area of each North American deer species in the context, so we need to search for it.4. Therefore, our search query should include the keywords &quot;North American deer,&quot; &quot;ear surface area,&quot; and &quot;Foster&#x27;s rule.&quot;</td></tr><tr><td></td><td>Search Query: North American deer ear surface area Foster&#x27;s rule smallest species</td></tr></table>
+
+# Module Label Prompt
+
+generate_query[1] Given the fields `context`, `question`, produce the fields `search_query`.
+
+Follow the following format.
+
+Context: ${context}
+
+Question: ${question}
+
+Reasoning: Let's think step by step in order to  $\{$ produce the search_query\}. We ...
+
+Search Query: ${search_query}
+
+# Context:
+
+[1] «Key deer | The Key deer ("Odocoileus virginianus clavium") is an endangered deer that lives only in the Florida Keys. It is a subspecies of the white-tailed deer ("O. virginianus"). It is the smallest North American deer.»
+
+[2] «North American least shrew | The North American least shrew ("Cryptotis parva") is one of the smallest mammals, growing to be only up to 3 inches long. The North American least shrew has a long pointed snout and a tail never more than twice the length of its hind foot. It has a dense fur coat that is either grayish-brown or reddish-brown with a white belly. Its fur becomes lighter in the summer and darker in the winter. Although similar in appearance to several species of rodents, all shrews are members of the order Soricomorpha and should not be mistaken for a member of the Rodentia order. The North American least shrew's eyes are small and its ears are completely concealed within its short fur, giving it very poor eyesight and hearing.
+
+[3] «Wynnea americana | Wynnea americana, commonly known as moose antlers or rabbit ears, is a species of fungus in the family Sarcoscyphaceae. This uncommon inedible species is recognizable by its spoon-shaped or rabbit-ear shaped fruit bodies that may reach up to 13 cm tall. It has dark brown and warty outer surfaces, while the fertile spore-bearing inner surface is orange to pinkish to reddish brown. The fruit bodies grow clustered together from large underground masses of compacted mycelia known as sclerotia. In eastern North America, where it is typically found growing in the soil underneath hardwood trees, it is found from New York to Michigan south to Mexico. The species has also been collected from Costa Rica, India, and Japan.»
+
+Question: What animal subject to the Foster's rule is the smallest North American deer?
+
+Reasoning: Let's think step by step in order to produce the search query. We need to find the smallest North American deer, which is mentioned in the context as the Key deer. Therefore, the search query should include the terms "Key deer" and "smallest North American deer".
+
+Search Query: Key deer smallest North American deer
+
+# Module Label Prompt
+
+generate_answer Given the fields `context`, `question`, produce the fields `answer`.
+
+Follow the following format.
+
+Context: ${context}
+
+Question: ${question}
+
+Reasoning: Let's think step by step in order to ${produce the answer}. We ...
+
+Answer: ${answer}
+
+# Context:
+
+[1] «Key deer | The Key deer ("Odocoileus virginianus clavium") is an endangered deer that lives only in the Florida Keys. It is a subspecies of the white-tailed deer ("O. virginianus"). It is the smallest North American deer.»
+
+[2] «North American least shrew | The North American least shrew ("Cryptotis parva") is one of the smallest mammals, growing to be only up to 3 inches long. The North American least shrew has a long pointed snout and a tail never more than twice the length of its hind foot. It has a dense fur coat that is either grayish-brown or reddish-brown with a white belly. Its fur becomes lighter in the summer and darker in the winter. Although similar in appearance to several species of rodents, all shrews are members of the order Soricomorpha and should not be mistaken for a member of the Rodentia order. The North American least shrew's eyes are small and its ears are completely concealed within its short fur, giving it very poor eyesight and hearing.
+
+[3] «Wynnea americana | Wynnea americana, commonly known as moose antlers or rabbit ears, is a species of fungus in the family Sarcoscyphaceae. This uncommon inedible species is recognizable by its spoon-shaped or rabbit-ear shaped fruit bodies that may reach up to 13 cm tall. It has dark brown and warty outer surfaces, while the fertile spore-bearing inner surface is orange to pinkish to reddish brown. The fruit bodies grow clustered together from large underground masses of compacted mycelia known as sclerotia. In eastern North America, where it is typically found growing in the soil underneath hardwood trees, it is found from New York to Michigan south to Mexico. The species has also been collected from Costa Rica, India, and Japan.»
+
+[4] «Pudu | The pudús (Mapudungun "pūdú" or "pūdu", Spanish: pudú, ]） are two species of South American deer from the genus Pudu, and are the world's smallest deer. The name is a loanword from Mapudungun, the language of the indigenous Mapuche people of southern Chile and south-western Argentina. The two species of pudús are the northern pudú ("Pudu mephistophiles") from Venezuela, Colombia, Ecuador, and Peru, and the southern pudú ("Pudu puda"; sometimes incorrectly modified to "Pudu pudú") from southern Chile and south-western Argentina. Pudu range in size from 32 to tall, and up to 85 cm long. As of 2009, the southern pudu is classified as near threatened, while the northern pudu is classified as vulnerable in the IUCN Red List.
+
+Question: What animal subject to the Foster's rule is the smallest North American deer?
+
+Reasoning: Let's think step by step in order to find the answer. We know that the smallest North American deer is mentioned in the context. We also know that the Foster's rule refers to a group of animals. In this case, the Foster's rule refers to endangered deer species that live only in the Florida Keys. Therefore, the Key deer is the answer.
+
+Answer: Key deer
+
+```python
+class CoTProgram(dsyModule): def __init__(self): super().__init__(self.createAnswer = dspyChainOfThought("question -> answer") def forward(self, question): return self.createAnswer(question=question)
+```
+
+# Python Snippets A.2: DSPy program for GSM8K.
+
+Table A.2: Vanilla (zero-shot) prompts for the GSM8K program given in Python Snippets A.2 with one LM module. The generate_answer module prompts the corresponding LM to answer a grade school math problem. The green text shows the model completion obtained by setting all the modules of the program to use mistralai/Mistral-7b-Instruct-v0.2 as their corresponding LM. The model's response is evaluated as correct because the correct answer (1200) is contained in the first line of model's generation.  
+
+<table><tr><td>Module Label</td><td>Prompt</td></tr><tr><td>generate_answer</td><td>Given the fields `question`, produce the fields `answer`.</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Follow the following format.</td></tr><tr><td></td><td>Question: ${question}Reasoning: Let&#x27;s think step by step in order to ${produce the answer}. We...Answer: ${answer}</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Question: Micah can type 20 words per minute and Isaiah can type 40 words per minute. How many more words can Isaiah type than Micah in an hour?Reasoning: Let&#x27;s think step by step in order to find out how many more words Isaiah can type than Micah in an hour. We know that Micah can type 20 words per minute and Isaiah can type 40 words per minute. In one minute, Isaiah types twice as many words as Micah. In 60 minutes, Isaiah types 60 minutes * 40 words per minute = &lt;&lt;60*40=2400&gt;&gt;2400 words. Micah types 60 minutes * 20 words per minute = &lt;&lt;60*20=1200&gt;&gt;1200 words. The difference between the number of words Isaiah and Micah can type in an hour is 2400 words - 1200 words = &lt;&lt;2400-1200=1200&gt;&gt;1200 words.</td></tr><tr><td></td><td>Answer: Isaiah can type 1200 more words than Micah in an hour.</td></tr></table>
+
+```python
+class IrisSignature(dspy.Signature): "Given the petal and sepal dimensions in cm, predict the iris species."   
+petal_length  $=$  dspy.InputField()   
+petal_width  $=$  dspy.InputField()   
+sepal_length  $=$  dspy.InputField()   
+sepal_width  $=$  dspy.InputField()   
+answer  $=$  dspy.OutputField desc  $\equiv$  'setosa, versicolor, or virginica')   
+class IrisProgram(dspyModule): def __init__(self): self.createanswer  $=$  dspyChainOfThought(IrisSignature)   
+def forward(self,petal_length,petal_width,sepal_length,sepal_width): return self.createanswer(petal_length  $\equiv$  petal_length,petal_width= petal_width,sepal_length  $\equiv$  sepal_length,sepal_width  $\equiv$  sepal_width)
+```
+
+Python Snippets A.3: DSPy program for Iris, provided to us by the DSPy team.
+
+Table A.3: Vanilla (zero-shot) prompts for the Iris program given in Python Snippets A.3 with one LM module. The generate_answer module prompts the corresponding LM to classify the Iris flowers into one of the following species given their features: Setosa, Versicolor, Virginica. The green text shows the model completion obtained by setting all the modules of the program to use mistralai/Mistral-7b-Instruct-v0.2 as their corresponding LM. The correct answer is Setosa, just as predicted.  
+
+<table><tr><td>Module Label</td><td>Prompt</td></tr><tr><td>generate_answer</td><td>Given the petal and sepal dimensions in cm, predict the iris species.</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Follow the following format.</td></tr><tr><td></td><td>Petal Length: ${petal_length}</td></tr><tr><td></td><td>Petal Width: ${petal_width}</td></tr><tr><td></td><td>Sepal Length: ${sepal_length}</td></tr><tr><td></td><td>Sepal Width: ${sepal_width}</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to ${produce the answer}. We</td></tr><tr><td></td><td>...</td></tr><tr><td></td><td>Answer: setosa, versicolor, or virginica</td></tr><tr><td></td><td>---</td></tr><tr><td></td><td>Petal Length: 1.4</td></tr><tr><td></td><td>Petal Width: 0.2</td></tr><tr><td></td><td>Sepal Length: 5.0</td></tr><tr><td></td><td>Sepal Width: 3.6</td></tr><tr><td></td><td>Reasoning: Let&#x27;s think step by step in order to predict the iris species. We will compare the given measurements with the average measurements of each iris species. 1. Setosa: The average petal length for setosa is 1.3 cm and the average petal width is 0.3 cm. The given petal length (1.4 cm) is slightly larger than the average, but the petal width (0.2 cm) is smaller than the average. However, the sepal dimensions (5.0 cm and 3.6 cm) are within the range of setosa. Based on these measurements, it is likely that the iris is setosa. 2. Versicolour: The average petal length for versicolour is 4.2 cm and the average petal width is 1.4 cm. The given petal length (1.4 cm) is smaller than the average, and the petal width (0.2 cm) is much smaller than the average. The sepal dimensions (5.0 cm and 3.6 cm) are also within the range of versicolour. However, the small petal dimensions suggest that it is less likely to be versicolour. 3. Virginica: The average petal length for virginica is 5.8 cm and the average petal width is 2.0 cm. The given petal length (1.4 cm) is much smaller than the average, and the petal width (0.2 cm) is much smaller than the average. The sepal dimensions (5.0 cm and 3.6 cm) are also within the range of virginica. However, the small petal dimensions suggest that it is less likely to be virginica. Based on the given measurements, it is most likely that the iris is setosa.</td></tr><tr><td></td><td>Answer: setosa.</td></tr></table>
+
+# B Asset Information
+
+We share the associated licenses for the models and datasets we used below. For models, we list the specific HuggingFace model id we used to retrieve the respective weights.
+
+1. mistralai/Mistral-7b-Instruct-v0.2: Apache License 2.0
+
+2. meta-llama/Llama-2-7b-chat-hf: Meta Llama 2 Community License at https://ai.meta.com/llama/licenses/  
+3. meta-llama/Meta-Llama-3-8B-Instruct: Meta Llama 3 Community License at https://llama.meta.com/llama3/legalize/  
+4. HotPotQA: Apache License 2.0  
+5. GSM8K: MIT License  
+6. Iris: Creative Commons Attribution 4.0 International (CC BY 4.0)
+
+All the LMs used in this work are intended for use in English.
+
+# C Implementation Details
+
+In this section, we share the implementation details as it pertains to the LM inference and fine-tuning infrastructure, as well as compute requirements.
+
+Inference For inference, we host our models in Docker (Merkel, 2014) instances through HuggingFace's text-generation-inference (HuggingFace, 2023) toolkit. We keep the sampling parameters the same across all experiments, using TopK sampling with a temperature of 0.1, and top_k of 0.97, until the model either generates a stopping string or a total of 1024 tokens including the tokens in the prompt.
+
+Prompt Optimization For prompt optimization, we use the BootstrapFewShotRS (BFRS) optimizer from the DSPy library. In particular, we allow BFRS to randomly search 6 candidate programs using up to 3 few-shot examples for each module prompt.
+
+Fine-tuning For fine-tuning, we use Low Rank Adaptation (LoRA; Hu et al. 2022) to train the query and key self-attention layers of our models, using a LoRA rank of 32, alpha of 64, with no dropout. We fine-tune all of our models for 5 epochs using bf10at16 precision, with a learning rate of 1e-5 and an effective batch size of 8.
+
+Compute The BetterTogether strategies explored in this paper are naturally more expensive to run when compared to just prompt optimizing or fine-tuning. How these two steps compare to each other in terms of compute requirements or wall clock time depends on the particular settings used for each as well as the size of the dataset used. Total approximate GPU hours to produce Table 1 was  $\approx 75$  hours, using A100 GPUs.
+
+# D Extended Results
+
+The results shared in Table 1 are the average of three runs. Tables D.1, D.2, and D.3 show the breakdown of the individual runs for HotPotQA, GSM8K, and Iris, respectively.
+
+<table><tr><td rowspan="2">Strategy</td><td colspan="4">mistral-7b-instruct-v0.2</td><td colspan="4">llama-2-7b-chat</td><td colspan="4">llama-3-8b-instruct</td></tr><tr><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td></tr><tr><td colspan="13">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>17.2</td><td>17.2</td><td>17.2</td><td>17.2</td><td>13.2</td><td>13.2</td><td>13.2</td><td>13.2</td><td>31.6</td><td>31.6</td><td>31.6</td><td>31.6</td></tr><tr><td>Prompt Optimization (Π)</td><td>32.7</td><td>34.7</td><td>34.0</td><td>33.8</td><td>33.3</td><td>33.3</td><td>33.4</td><td>33.3</td><td>45.7</td><td>47.4</td><td>47.5</td><td>46.9</td></tr><tr><td>Weight Optimization (Θ)</td><td>22.0</td><td>23.1</td><td>23.5</td><td>22.9</td><td>12.4</td><td>11.8</td><td>12.3</td><td>12.2</td><td>34.9</td><td>35.3</td><td>34.3</td><td>34.8</td></tr><tr><td>Π → Π</td><td>31.7</td><td>36.0</td><td>33.7</td><td>33.8</td><td>31.7</td><td>33.1</td><td>33.1</td><td>32.6</td><td>47.3</td><td>45.4</td><td>46.7</td><td>46.5</td></tr><tr><td>Θ → Θ</td><td>24.1</td><td>23.9</td><td>23.9</td><td>24.0</td><td>12.4</td><td>13.5</td><td>13.3</td><td>13.0</td><td>35.1</td><td>34.1</td><td>34.1</td><td>34.4</td></tr><tr><td colspan="13">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>34.9</td><td>39.1</td><td>34.9</td><td>36.3</td><td>32.8</td><td>32.3</td><td>33.1</td><td>32.7</td><td>40.6</td><td>42.1</td><td>45.7</td><td>42.8</td></tr><tr><td>Θ → Π</td><td>29.3</td><td>33.8</td><td>35.8</td><td>33.0</td><td>36.0</td><td>33.4</td><td>33.1</td><td>34.2</td><td>44.5</td><td>40.9</td><td>45.3</td><td>43.6</td></tr><tr><td>Π → Θ → Π</td><td>34.9</td><td>40.7</td><td>37.2</td><td>37.6</td><td>34.7</td><td>34.5</td><td>35.3</td><td>34.8</td><td>46.5</td><td>47.1</td><td>46.4</td><td>46.7</td></tr></table>
+
+Table D.1: Results of HotPotQA Runs. Percentage accuracies of baseline and BetterTogether strategies on HotPotQA evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are performance of 3 runs on a held-out test set of 1500 examples, using different random seeds. Bold font in the average columns (Avg) is used to mark the highest score in a given column.  
+
+<table><tr><td rowspan="2">Strategy</td><td colspan="4">mistral-7b-instruct-v0.2</td><td colspan="4">llama-2-7b-chat</td><td colspan="4">llama-3-8b-instruct</td></tr><tr><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td></tr><tr><td colspan="13">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>40.3</td><td>40.3</td><td>40.3</td><td>40.3</td><td>24.0</td><td>24.0</td><td>24.0</td><td>24.0</td><td>72.7</td><td>72.7</td><td>72.7</td><td>72.7</td></tr><tr><td>Prompt Optimization (Π)</td><td>45.0</td><td>47.2</td><td>47.1</td><td>46.4</td><td>27.3</td><td>25.1</td><td>25.5</td><td>26.0</td><td>76.9</td><td>77.9</td><td>78.9</td><td>77.9</td></tr><tr><td>Weight Optimization (Θ)</td><td>40.8</td><td>40.0</td><td>41.2</td><td>40.7</td><td>23.7</td><td>24.2</td><td>24.0</td><td>24.0</td><td>75.7</td><td>74.8</td><td>74.8</td><td>75.1</td></tr><tr><td>Π → Π</td><td>46.3</td><td>47.2</td><td>49.6</td><td>47.7</td><td>28.4</td><td>24.0</td><td>21.8</td><td>24.7</td><td>76.5</td><td>80.1</td><td>76.1</td><td>77.6</td></tr><tr><td>Θ → Θ</td><td>42.9</td><td>41.8</td><td>43.8</td><td>42.8</td><td>24.0</td><td>24.3</td><td>24.0</td><td>24.1</td><td>52.2</td><td>36.6</td><td>43.4</td><td>44.0</td></tr><tr><td colspan="13">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>46.4</td><td>47.3</td><td>48.2</td><td>47.3</td><td>27.8</td><td>28.1</td><td>25.9</td><td>27.3</td><td>77.6</td><td>75.4</td><td>79.8</td><td>77.6</td></tr><tr><td>Θ → Π</td><td>50.1</td><td>46.0</td><td>48.8</td><td>48.3</td><td>26.8</td><td>26.1</td><td>27.0</td><td>26.6</td><td>78.5</td><td>79.8</td><td>78.4</td><td>78.9</td></tr><tr><td>Π → Θ → Π</td><td>44.9</td><td>48.5</td><td>47.1</td><td>46.8</td><td>27.1</td><td>25.9</td><td>25.9</td><td>26.3</td><td>77.6</td><td>75.4</td><td>77.8</td><td>77.0</td></tr></table>
+
+Table D.2: Results of GSM8K Runs. Percentage accuracies of baseline and BetterTogether strategies on GSM8K evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are performance of 3 runs on a held-out test set of 1319 examples, using different random seeds. Bold font in the average columns (Avg) is used to mark the highest score in a given column.  
+
+<table><tr><td rowspan="2">Strategy</td><td colspan="4">mistral-7b-instruct-v0.2</td><td colspan="4">llama-2-7b-chat</td><td colspan="4">llama-3-8b-instruct</td></tr><tr><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td><td>Run 1</td><td>Run 2</td><td>Run 3</td><td>Avg</td></tr><tr><td colspan="13">Baseline Strategies</td></tr><tr><td>Vanilla Zero-shot</td><td>26.0</td><td>26.0</td><td>26.0</td><td>26.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>48.0</td><td>48.0</td><td>48.0</td><td>48.0</td></tr><tr><td>Prompt Optimization (Π)</td><td>52.0</td><td>54.0</td><td>66.0</td><td>57.3</td><td>44.0</td><td>68.0</td><td>58.0</td><td>56.7</td><td>62.0</td><td>96.0</td><td>80.0</td><td>79.3</td></tr><tr><td>Weight Optimization (Θ)</td><td>24.0</td><td>34.0</td><td>30.0</td><td>29.3</td><td>-</td><td>-</td><td>-</td><td>-</td><td>38.0</td><td>40.0</td><td>34.0</td><td>37.3</td></tr><tr><td>Π → Π</td><td>48.0</td><td>64.0</td><td>66.0</td><td>59.3</td><td>66.0</td><td>70.0</td><td>56.0</td><td>64.0</td><td>70.0</td><td>94.0</td><td>82.0</td><td>82.0</td></tr><tr><td>Θ → Θ</td><td>40.0</td><td>36.0</td><td>38.0</td><td>38.0</td><td>-</td><td>-</td><td>-</td><td>-</td><td>44.0</td><td>36.0</td><td>38.0</td><td>39.3</td></tr><tr><td colspan="13">BetterTogether Strategies</td></tr><tr><td>Π → Θ</td><td>32.0</td><td>26.0</td><td>34.0</td><td>30.7</td><td>30.0</td><td>26.0</td><td>24.0</td><td>26.7</td><td>50.0</td><td>42.0</td><td>40.0</td><td>44.0</td></tr><tr><td>Θ → Π</td><td>80.0</td><td>54.0</td><td>66.0</td><td>66.7</td><td>-</td><td>-</td><td>-</td><td>-</td><td>78.0</td><td>78.0</td><td>80.0</td><td>78.7</td></tr><tr><td>Π → Θ → Π</td><td>52.0</td><td>44.0</td><td>62.0</td><td>52.7</td><td>62.0</td><td>70.0</td><td>64.0</td><td>65.3</td><td>74.0</td><td>80.0</td><td>84.0</td><td>79.3</td></tr></table>
+
+Table D.3: Results of Iris Runs. Percentage accuracies of baseline and BetterTogether strategies on Iris evaluated on mistral-7b-instruct-v0.2, llama-2-7b-chat, and llama-3-8b-instruct. Reported are performance of 3 runs on a held-out test set of 50 examples, using different random seeds. Bold font in the average columns (Avg) is used to mark the highest score in a given column. Strategies where weight optimization is the first step use the vanilla (zero-shot) strategy to generate the initial fine-tuning dataset. If a model generates very few or no correct outputs under the vanilla strategy on the training set used to bootstrap the fine-tuning data, there will not be a sufficient dataset for fine-tuning. These settings are marked with “-”.

2024/Fine-Tuning and Prompt Optimization_ Two Great Steps that Work Better Together/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1972536090af57e8d615041b152a71f6765a2cc93214a421eb165d5c04400ed9
+size 724923

2024/Fine-Tuning or Retrieval_ Comparing Knowledge Injection in LLMs/76375409-778a-4580-941a-496a79ee2e4d_content_list.json

@@ -0,0 +1,2254 @@
+[
+    {
+        "type": "text",
+        "text": "Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs",
+        "text_level": 1,
+        "bbox": [
+            142,
+            89,
+            855,
+            110
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Oded Ovadia*, Meni Brief*, Moshik Mishaeli, and Oren Elisha  \nMicrosoft, Israel",
+        "bbox": [
+            228,
+            149,
+            769,
+            184
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) encapsulate a vast amount of factual information within their pre-trained weights, as evidenced by their ability to answer diverse questions across different domains. However, this knowledge is inherently limited, relying heavily on the characteristics of the training data. Consequently, using external datasets to incorporate new information or refine the capabilities of LLMs on previously seen information poses a significant challenge. In this study, we compare two common approaches: unsupervised fine-tuning and retrieval-augmented generation (RAG). We evaluate both approaches on a variety of knowledge-intensive tasks across different topics. Our findings reveal that while unsupervised fine-tuning offers some improvement, RAG consistently outperforms it, both for existing knowledge encountered during training and entirely new knowledge. Moreover, we find that LLMs struggle to learn new factual information through unsupervised fine-tuning, and that exposing them to numerous variations of the same fact during training could alleviate this problem.",
+        "bbox": [
+            144,
+            285,
+            460,
+            640
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            650,
+            258,
+            665
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) are able to capture vast amounts of factual information (Petroni et al., 2019; Cohen et al., 2023; Hu et al., 2023). LLMs exhibit a remarkable level of knowledge in various domains due to their massive pre-training datasets. However, there are two significant limitations to this knowledge. First, it is static and does not update with time. Second, it is non-specific and thus may lack nuanced expertise in particular domains. While these are two different problems, they are deeply related since their solution is the same: enhancing the model's knowledge.",
+        "bbox": [
+            112,
+            675,
+            489,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, the idea of adapting LLMs to particular domains and updating their knowledge has",
+        "bbox": [
+            112,
+            869,
+            489,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "become increasingly common (Yu et al., 2022). Various models have been suggested to improve factual knowledge and capabilities in diverse fields such as healthcare (Singhal et al., 2023a,b; Wu et al., 2023a), finance (Wu et al., 2023b; Yang et al., 2023), and law (Huang et al., 2023; Nguyen, 2023).",
+        "bbox": [
+            507,
+            261,
+            884,
+            357
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, we focus on the evaluation of a model's knowledge and its ability to memorize, understand, and retrieve factual data. We aim to understand the concept of knowledge injection (Wang et al., 2020; Chen et al., 2022; Liu et al., 2020; Lauscher et al., 2020). Given some knowledge base in the form of a text corpus, what is the best way to teach a pre-trained model this knowledge?",
+        "bbox": [
+            507,
+            357,
+            884,
+            486
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "One way to add knowledge to a pre-trained model is through fine-tuning. With fine-tuning, we continue the model's training process and adapt it using task-specific data. By exposing the model to a specific knowledge base, we expect the model weights to adapt accordingly. This process is meant to optimize the model for targeted applications, enhancing its performance and contextual relevance in specialized domains.",
+        "bbox": [
+            507,
+            487,
+            885,
+            630
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Another method to enhance a model's knowledge base is through the use of in-context learning (ICL) (Chen et al., 2021; Radford et al., 2019; Min et al., 2021; Lampinen et al., 2022). The main idea behind ICL is to improve the performance of pretrained LLMs on new tasks by modifying the input query to the model without directly changing the weights of the model. One form of ICL is retrieval augmented generation (RAG) (Lewis et al., 2020; Neelakantan et al., 2022). RAG uses information retrieval techniques to enable LLMs to obtain relevant information from a knowledge source and incorporate it into generated text.",
+        "bbox": [
+            507,
+            632,
+            885,
+            840
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This study aims to evaluate the knowledge injection capabilities of LLMs through a comparison of fine-tuning and RAG. To illustrate the rationale, let us use an analogy. Consider three college students taking a test on a specific topic. All had access",
+        "bbox": [
+            507,
+            841,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal contribution.",
+        "bbox": [
+            136,
+            906,
+            262,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "237",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 237-250 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            166,
+            945,
+            828,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "to class materials but didn't know the topic beforehand. The first student had the textbook only during the test, the second had pre-test access and studied, and the third lost access upon the test announcement. Who would probably perform better?",
+        "bbox": [
+            112,
+            84,
+            489,
+            165
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Background",
+        "text_level": 1,
+        "bbox": [
+            112,
+            175,
+            257,
+            192
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To assess knowledge injection, we must first understand what knowledge means for LLMs.",
+        "bbox": [
+            112,
+            200,
+            489,
+            231
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Knowledge and Language Models Defining knowledge is a complex philosophical task far beyond the scope of this research. However, we can examine what factual knowledge means in the context of language models. If a model knows a fact, it can accurately and consistently answer questions about it. Furthermore, it can reliably distinguish between true and false statements related to this fact. We can then extend this definition to a whole knowledge base, not just a single fact.",
+        "bbox": [
+            112,
+            233,
+            487,
+            393
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Mathematically, let  $\\mathcal{Q} = \\{q_n\\}_{n=1}^N$  be a set of  $N$  multiple choice factual questions, where each question has  $L$  possible answers and exactly one correct answer. Let  $\\mathcal{A} = \\{(a_n^1, \\ldots, a_n^L)\\}_{n=1}^N$  be the corresponding set of possible answers, and  $\\mathcal{C} = \\{c_n\\}_{n=1}^N$  be the correct answers.",
+        "bbox": [
+            112,
+            394,
+            487,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Let  $\\mathcal{M}$  be a language model. We denote by  $\\mathcal{M}(q_n)\\in \\{a_n^1,\\ldots ,a_n^L\\}$  the predicted answer of the model to the  $n$  -th question. We define the knowledge score  $\\mathcal{L}$  of  $\\mathcal{M}$  in relation to  $\\mathcal{Q}$  to be the standard accuracy score:",
+        "bbox": [
+            112,
+            491,
+            487,
+            571
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathcal {M}, \\mathcal {Q}} := \\frac {\\# \\left\\{q _ {n} \\mid \\mathcal {M} \\left(q _ {n}\\right) = c _ {n} \\right\\}}{N}. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            580,
+            487,
+            612
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We say that the model  $\\mathcal{M}$  possesses any knowledge regarding the set of questions  $\\mathcal{Q}$  if the following holds:",
+        "bbox": [
+            112,
+            619,
+            489,
+            667
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathcal {M}, \\mathcal {Q}} > \\frac {1}{L}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            250,
+            678,
+            487,
+            709
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In simpler terms, the model can consistently give correct answers, outperforming a simple random guessing baseline. Naturally, if the knowledge score  $\\mathcal{L}_{\\mathcal{M},\\mathcal{Q}}$  is higher for one model compared to another, then we assert that the former is more knowledgeable with regards to  $\\mathcal{Q}$  compared to the latter.",
+        "bbox": [
+            112,
+            712,
+            487,
+            821
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Previously Seen Knowledge One important distinction to make is between knowledge that the model has been exposed to before during pretraining as opposed to entirely new facts. Considering the size of modern LLM training sets, they cover a vast amount of information available",
+        "bbox": [
+            112,
+            824,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "through web-sourced text. As a result, even in niche domains, the goal of knowledge injection is not necessarily to teach the model entirely new facts but rather to \"refresh\" its memory by inducing a bias toward a particular domain.",
+        "bbox": [
+            507,
+            84,
+            882,
+            164
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Knowledge and Reasoning We emphasize that this knowledge evaluation framework for LLMs is imperfect. Importantly, it doesn't address other quality metrics influencing a model's response. Creating a purely knowledge-intensive dataset without involving some level of reasoning is challenging. Consequently, a model with robust reasoning abilities might excel on unfamiliar knowledge-intensive tasks by making \"educated guesses\" in a multiple-choice exam. Therefore, any evaluation of knowledge in LLMs should consider this, with results seen as part of a broader range of benchmarks for reasoning (Sakaguchi et al., 2021), reading comprehension (Dua et al., 2019), and general language abilities (Srivastava et al., 2022). However, this evaluation framework still strongly emphasizes factual information above all else.",
+        "bbox": [
+            507,
+            165,
+            884,
+            437
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Causes for Factual Errors There are many possible reasons for the failure of models to answer factual questions accurately. In (Wang et al., 2023), Wang et al. introduce a taxonomy of five main model-level causes:",
+        "bbox": [
+            507,
+            439,
+            882,
+            518
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Domain knowledge deficit: A language model may lack comprehensive expertise in a specific domain to which it has not been exposed. For example, a model trained exclusively on texts written by William Shakespeare would perform poorly when asked about the works of Mark Twain.",
+            "- Outdated Information: LLMs invariably have a cutoff date determined by their training dataset. Consequently, any events, discoveries, or changes occurring after the last training update will not be within the model's knowledge without access to external sources.",
+            "- Immemorization: Sometimes, a model is exposed to knowledge during its training process but does not retain it. This is especially true for rare facts that appear in the training dataset only scarcely (Kandpal et al., 2023).",
+            "- Forgetting: Language models often undergo additional training after the pre-training phase (fine-tuning). In some cases, this might lead to a phenomenon called catastrophic forgetting"
+        ],
+        "bbox": [
+            507,
+            532,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "238",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e217435ac9acb8705b9e934e1fb27f2944114bb525f7e6bbdbb56719e40f6d1.jpg",
+        "image_caption": [
+            "Figure 1: A visualization of the knowledge injection framework."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            867,
+            376
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(Kirkpatrick et al., 2017; Goodfellow et al., 2013; Chen et al., 2020; Luo et al., 2023), where models lose some of the knowledge they had prior to the fine-tuning process.",
+        "bbox": [
+            127,
+            429,
+            489,
+            494
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "- Reasoning Failure: In certain instances, a language model might possess relevant knowledge about a fact but fail to utilize it properly. This is particularly evident in complex multi-step reasoning tasks (Tan et al., 2023) or when posed with different questions about the same fact, resulting in disparate outcomes (Berglund et al., 2023).",
+        "bbox": [
+            114,
+            504,
+            489,
+            617
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We observe that most of these issues arise during the pre-training phase, with catastrophic forgetting being the notable exception. Hence, many LLMs will suffer from factual errors of this kind regardless of any post-training process.",
+        "bbox": [
+            112,
+            627,
+            489,
+            708
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Injecting Knowledge to Language Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            718,
+            440,
+            750
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following the background given in Section 2, it is clear that general pre-training is insufficient for many knowledge-intensive tasks. To solve this, an additional post-processing step is essential to augment the knowledge of a pre-trained model. This step is often referred to as knowledge injection (Wang et al., 2020; Chen et al., 2022; Liu et al., 2020; Lauscher et al., 2020).",
+        "bbox": [
+            112,
+            760,
+            489,
+            888
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we examine two widely used frameworks for knowledge injection: fine-tuning",
+        "bbox": [
+            112,
+            889,
+            487,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(FT) and retrieval augmented generation (RAG). We begin by formulating the knowledge injection problem, aiming to explain both methods using consistent terminology.",
+        "bbox": [
+            507,
+            429,
+            884,
+            494
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Problem formulation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            504,
+            721,
+            518
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In Equations (1) and (2), we presented a formulation for knowledge in language models through the lens of question-answering (Q&A). We now extend this formulation to the problem of knowledge injection using the same terminology.",
+        "bbox": [
+            507,
+            525,
+            884,
+            604
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a set of factual questions, there exists some text corpus containing information that is relevant to these questions. The central assumption of knowledge injection is that given full access to this corpus, it could serve as an auxiliary knowledge base and improve the model's performance on this set of questions.",
+        "bbox": [
+            507,
+            606,
+            884,
+            717
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Mathematically, let  $\\mathcal{M}$  be a pre-trained model, and let  $\\mathcal{Q}$  be a set of factual questions, as before. Now, assume we have a relevant auxiliary knowledge base  $\\mathcal{B}_{\\mathcal{Q}}$ . Our objective is to discover a transformation, denoted as  $\\mathcal{F}$ , that, when applied, would enhance the knowledge about  $\\mathcal{Q}$ :",
+        "bbox": [
+            507,
+            718,
+            885,
+            814
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {M} ^ {\\prime} := \\mathcal {F} (\\mathcal {M}, \\mathcal {B} _ {\\mathcal {Q}}) \\quad s. t. \\quad \\mathcal {L} _ {\\mathcal {M} ^ {\\prime}, \\mathcal {Q}} > \\mathcal {L} _ {\\mathcal {M}, \\mathcal {Q}}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            843,
+            882,
+            860
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, we aim to compare two choices for  $\\mathcal{F}$ : fine-tuning and RAG to see which option performs better in this problem.",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "239",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Fine-Tuning",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            257,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Fine-tuning is the process of adjusting a pre-trained model on a specific, often narrower, dataset or task to enhance its performance in that particular domain. Here, it is vital to distinguish between different types of fine-tuning. FT techniques are commonly classified into supervised, unsupervised, and reinforcement learning (RL) based methods. We proceed by briefly reviewing these methods and their relation to the problem of knowledge injection.",
+        "bbox": [
+            112,
+            108,
+            489,
+            267
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Supervised Fine-Tuning Supervised fine-tuning (SFT) requires sets of labeled input-output pairs. One of the most common SFT methods is instruction tuning (Wang et al., 2022; Mishra et al., 2021; Ouyang et al., 2022; Taori et al., 2023), which has emerged as one of the most powerful methods to improve model performance. With instruction tuning, the input is a natural language task description, and the output is an example of the desired behavior. Many current state-of-the-art LLMs have gone through instruction tuning after their pre-training phase.",
+        "bbox": [
+            115,
+            271,
+            489,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Instruction tuning has been shown to be very effective at improving the overall quality of the model, with a particular emphasis on its zero-shot and reasoning capabilities. However, despite these advantages, instruction tuning does not necessarily teach the model new knowledge (Ouyang et al., 2022; Chung et al., 2022; Mitra et al., 2023; Chia et al., 2023; Zhou et al., 2023). As such, instruction tuning alone is not a viable solution to the knowledge injection problem.",
+        "bbox": [
+            112,
+            466,
+            489,
+            626
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Reinforcement Learning Another form of FT relies on RL or RL-inspired optimization strategies to better align the model after its pre-training phase. A few prominent examples are reinforcement learning from human feedback (RLHF) (OpenAI, 2023; Touvron et al., 2023), direct preference optimization (DPO) (Rafailov et al., 2023), and proximal policy optimization (PPO) (Schulman et al., 2017; Tunstall et al., 2023).",
+        "bbox": [
+            112,
+            629,
+            489,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "These techniques have been shown to be very useful, especially when used in conjunction with instruction tuning. However, similarly to instruction tuning, these methods focus on the overall quality of the response and its expected behavior and not necessarily on its breadth of knowledge.",
+        "bbox": [
+            112,
+            775,
+            489,
+            871
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised Fine-Tuning The final FT strategy we discuss is unsupervised, meaning there are no available labels for the model to learn from.",
+        "bbox": [
+            112,
+            873,
+            489,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "One common unsupervised FT technique is often referred to as continual pre-training or unstructured FT.",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this method, the FT process is viewed as a direct continuation of the pre-training phase. We start with a saved checkpoint of the original LLM and train it in a causal auto-regressive manner, i.e., predicting the next token. One major difference in comparison to actual pre-training is the learning rate. Usually, one would need a much lower learning rate when continuing the pre-training of the model to avoid catastrophic forgetting (Kirkpatrick et al., 2017).",
+        "bbox": [
+            507,
+            137,
+            884,
+            298
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is well known that LLMs store vast amounts of knowledge during their pre-training phase (Zhou et al., 2023). So, it makes sense to continue this process in order to inject knowledge into the model. Hence, we use the unsupervised FT approach throughout this work and evaluate its efficacy in enhancing the model's capacity for learning new information.",
+        "bbox": [
+            507,
+            304,
+            884,
+            432
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Retrieval Augmented Generation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            461,
+            816,
+            476
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Retrieval augmented generation (RAG) (Lewis et al., 2020) is a technique that expands LLMs' capabilities, especially in knowledge-intensive tasks, by using external knowledge sources. While the original formulation involved additional training per task, it has since been demonstrated (Neelakantan et al., 2022) that a pre-trained embedding model can achieve improved performance with no additional training involved.",
+        "bbox": [
+            507,
+            492,
+            884,
+            636
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The idea is that given an auxiliary knowledge base and an input query, we use the RAG architecture to find documents within the knowledge base that resemble the input query. These documents are then added to the input query, thus giving the model further context about the subject of the query.",
+        "bbox": [
+            507,
+            642,
+            882,
+            739
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In practice, implementing the suggested architecture is quite straightforward: Given an auxiliary knowledge base  $\\mathcal{B}_{\\mathcal{Q}}$  and a pre-trained embedding model  $\\mathcal{M}_e$ , we create a dense vector representation (embedding) per document  $b \\in \\mathcal{B}_{\\mathcal{Q}}$  and store these in a vector store. Upon receiving a new query  $q$ , we use its embedding,  $\\mathcal{M}_e(q)$ , to retrieve  $q$ 's top- $K$  closest neighbors,  $\\mathbf{b}_q = \\{b_k\\}_1^K$ , according to dot-product ranking. We then update  $q$  to be  $\\tilde{q} = \\mathbf{b}_q \\| q$ , where  $\\|\\cdot\\|$  denotes string concatenation. Finally, we return  $\\mathcal{M}(\\tilde{q})$  as the model's output.",
+        "bbox": [
+            507,
+            744,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "240",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/2304b45868ce26a540e175586ec076f8c1201675fd618c60c596f6da1d1f8196.jpg",
+        "table_caption": [
+            "Table 1: Results for the MMLU datasets described in Section 4.1 in terms of log-likelihood accuracy (Equation (4))."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Task</td><td>Model</td><td>Base model</td><td>Base model + RAG</td><td>Fine-tuned</td><td>Fine-tuned + RAG</td></tr><tr><td rowspan=\"3\">Anatomy (0-shot)</td><td>Mistral 7B</td><td>0.556</td><td>0.681</td><td>0.570</td><td>0.659</td></tr><tr><td>Llama2 7B</td><td>0.393</td><td>0.489</td><td>0.430</td><td>0.489</td></tr><tr><td>Orca2 7B</td><td>0.607</td><td>0.637</td><td>0.600</td><td>0.637</td></tr><tr><td rowspan=\"3\">Anatomy (5-shot)</td><td>Mistral 7B</td><td>0.600</td><td>0.681</td><td>0.622</td><td>0.674</td></tr><tr><td>Llama2 7B</td><td>0.467</td><td>0.563</td><td>0.496</td><td>0.548</td></tr><tr><td>Orca2 7B</td><td>0.570</td><td>0.659</td><td>0.593</td><td>0.674</td></tr><tr><td rowspan=\"3\">Astronomy (0-shot)</td><td>Mistral 7B</td><td>0.625</td><td>0.678</td><td>0.651</td><td>0.697</td></tr><tr><td>Llama2 7B</td><td>0.401</td><td>0.467</td><td>0.487</td><td>0.520</td></tr><tr><td>Orca2 7B</td><td>0.645</td><td>0.750</td><td>0.651</td><td>0.750</td></tr><tr><td rowspan=\"3\">Astronomy (5-shot)</td><td>Mistral 7B</td><td>0.658</td><td>0.724</td><td>0.651</td><td>0.697</td></tr><tr><td>Llama2 7B</td><td>0.401</td><td>0.474</td><td>0.447</td><td>0.520</td></tr><tr><td>Orca2 7B</td><td>0.664</td><td>0.763</td><td>0.664</td><td>0.743</td></tr><tr><td rowspan=\"3\">College biology (0-shot)</td><td>Mistral 7B</td><td>0.681</td><td>0.757</td><td>0.701</td><td>0.764</td></tr><tr><td>Llama2 7B</td><td>0.438</td><td>0.493</td><td>0.458</td><td>0.465</td></tr><tr><td>Orca2 7B</td><td>0.583</td><td>0.639</td><td>0.604</td><td>0.632</td></tr><tr><td rowspan=\"3\">College biology (5-shot)</td><td>Mistral 7B</td><td>0.722</td><td>0.778</td><td>0.736</td><td>0.771</td></tr><tr><td>Llama2 7B</td><td>0.451</td><td>0.521</td><td>0.424</td><td>0.479</td></tr><tr><td>Orca2 7B</td><td>0.604</td><td>0.660</td><td>0.625</td><td>0.653</td></tr><tr><td rowspan=\"3\">College chemistry (0-shot)</td><td>Mistral 7B</td><td>0.470</td><td>0.500</td><td>0.490</td><td>0.500</td></tr><tr><td>Llama2 7B</td><td>0.310</td><td>0.380</td><td>0.390</td><td>0.390</td></tr><tr><td>Orca2 7B</td><td>0.370</td><td>0.440</td><td>0.370</td><td>0.390</td></tr><tr><td rowspan=\"3\">College chemistry (5-shot)</td><td>Mistral 7B</td><td>0.470</td><td>0.540</td><td>0.500</td><td>0.500</td></tr><tr><td>Llama2 7B</td><td>0.370</td><td>0.380</td><td>0.360</td><td>0.390</td></tr><tr><td>Orca2 7B</td><td>0.430</td><td>0.470</td><td>0.370</td><td>0.380</td></tr><tr><td rowspan=\"3\">Prehistory (0-shot)</td><td>Mistral 7B</td><td>0.713</td><td>0.750</td><td>0.719</td><td>0.731</td></tr><tr><td>Llama2 7B</td><td>0.448</td><td>0.481</td><td>0.457</td><td>0.478</td></tr><tr><td>Orca2 7B</td><td>0.642</td><td>0.679</td><td>0.673</td><td>0.673</td></tr><tr><td rowspan=\"3\">Prehistory (5-shot)</td><td>Mistral 7B</td><td>0.722</td><td>0.762</td><td>0.725</td><td>0.762</td></tr><tr><td>Llama2 7B</td><td>0.515</td><td>0.531</td><td>0.503</td><td>0.537</td></tr><tr><td>Orca2 7B</td><td>0.664</td><td>0.698</td><td>0.667</td><td>0.694</td></tr></table>",
+        "bbox": [
+            137,
+            118,
+            858,
+            592
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/899c7ed477b423962c9f2fd6978921cf4305b4a9bb9b8a054a1e8d3359acd03b.jpg",
+        "table_caption": [
+            "Table 2: Current events results. Models that were fine-tuned on the original dataset are labeled as  $FT-reg$ , while those trained on the dataset with multiple paraphrases are labeled as  $FT-par$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Base model</td><td>Base model + RAG</td><td>FT-reg</td><td>FT-par</td><td>FT-reg + RAG</td><td>FT-par + RAG</td></tr><tr><td>Mistral 7B</td><td>0.481</td><td>0.875</td><td>0.504</td><td>0.588</td><td>0.810</td><td>0.830</td></tr><tr><td>Llama2 7B</td><td>0.353</td><td>0.585</td><td>0.219</td><td>0.392</td><td>0.326</td><td>0.520</td></tr><tr><td>Orca2 7B</td><td>0.456</td><td>0.876</td><td>0.511</td><td>0.566</td><td>0.820</td><td>0.826</td></tr></table>",
+        "bbox": [
+            169,
+            661,
+            826,
+            727
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Knowledge Base Creation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            755,
+            371,
+            771
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Task Selection and Rationale",
+        "text_level": 1,
+        "bbox": [
+            112,
+            785,
+            386,
+            800
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "MMLU Benchmark To properly evaluate the capabilities of LLMs on knowledge-intensive tasks, we selected four distinct tasks from the Massively Multilingual Language Understanding Evaluation (MMLU) benchmark (Hendrycks et al., 2021) in the topics of anatomy, astronomy, college biology, college chemistry and prehistory. The chosen tasks",
+        "bbox": [
+            112,
+            808,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "were selected based on their emphasis on factual knowledge and the minimal reliance on reasoning. As a heuristic, we opted for tasks where the questions are short and involve no context. In practice we selected four STEM subjects as well as one humanities subject, to ensure the evaluation is not limited to certain fields. Note that prehistory involves questions spanning all non-modern history. This approach aims to enable us to test LLM proficiency in comprehending and manipulating information in",
+        "bbox": [
+            507,
+            756,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "241",
+        "bbox": [
+            485,
+            928,
+            512,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "isolation from its reasoning processes.",
+        "bbox": [
+            112,
+            85,
+            400,
+            99
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Current Events Task To further isolate LLMs' abilities to learn new knowledge, we created a task comprising multiple-choice questions about current events. This task includes multiple-choice questions about events that occurred after the cutoff of the various models' training data. Specifically, we focused on \"current events\" from the USA, in the time span of August-November 2023, that are included in the relevant Wikipedia indexes<sup>1</sup>. This method enables us to mostly guarantee that the models have not been exposed to these facts, thus allowing us to directly test knowledge injection capabilities.",
+        "bbox": [
+            112,
+            99,
+            489,
+            309
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Data Collection and Preprocessing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            321,
+            431,
+            336
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To effectively evaluate the LLMs' performance on these knowledge-intensive tasks, a comprehensive auxiliary dataset was collected by scraping relevant articles per topic from Wikipedia. The rationale behind selecting Wikipedia as the primary source of knowledge is its broad coverage of relevant topics and its reliability as a repository of crowd-verified knowledge. All articles pertinent to the tasks were retrieved via the official Wikipedia API<sup>2</sup> by identifying the relevant central page per topic.",
+        "bbox": [
+            112,
+            342,
+            489,
+            502
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Subsequently, a rigorous cleaning process was utilized to transform the data from raw subsections to clean chunks. This step was done with the \"wikiextractor\" tool (Attardi, 2015). The division into small, clean (e.g., remove HTML, URLs, etc.) chunks was aimed at enhancing the evaluation of the LLMs' understanding across various knowledge domains and aiding the LLMs in the fine-tuning process.",
+        "bbox": [
+            112,
+            504,
+            489,
+            646
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Current Events Task Creation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            659,
+            398,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "After collecting the relevant chunks from Wikipedia, we created a new multiple-choice dataset with the help of GPT-4 (OpenAI, 2023). First, we removed any small chunks. For each remaining chunk in the corpus, GPT-4 was instructed to create four highly specific, high-quality multiple-choice questions with only one correct answer. By specific, we mean that the question can be answered without knowledge of which context the question refers to and with minimal ambiguity. Next, GPT-4 was asked to select the two most specific of the four. This was followed",
+        "bbox": [
+            112,
+            680,
+            489,
+            872
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "by a manual evaluation and verification step. In total, this resulted in 910 new questions.",
+        "bbox": [
+            507,
+            84,
+            882,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Paraphrases Generation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            128,
+            747,
+            142
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "After creating the dataset, we utilized GPT-4 to generate augmentations of the dataset. We instructed GPT-4 to provide paraphrased versions of the input data that fully retain the information while being reworded. Each paraphrasing iteration was done with a different seed to ensure variety.",
+        "bbox": [
+            507,
+            148,
+            882,
+            243
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We selected 240 chunks at random for each task and created two paraphrases per chunk. These were set aside to be used as validation sets for hyperparameter tuning. For the current events dataset, we created ten paraphrases for each chunk used in the fine-tuning process described in Section 6.",
+        "bbox": [
+            507,
+            244,
+            884,
+            340
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments and Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            353,
+            759,
+            369
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Experimental Framework We used the popular LM-Evaluation-Harness (Gao et al., 2021) repository to evaluate the performance of LLMs on the selected knowledge-intensive tasks. LM-Evaluation-Harness is a robust benchmarking tool that currently serves as the industry standard for model evaluation and is the basis of the HuggingFace leaderboard<sup>3</sup>. Leveraging this platform ensured a standardized evaluation framework and allowed consistent comparison across models, methods, and datasets. More importantly, by using the industry standard for evaluation, we could avoid any differences stemming from prompt engineering and formatting issues and replicate the reported baseline results for each model.",
+        "bbox": [
+            507,
+            379,
+            882,
+            618
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Model Selection We chose three models for inference evaluation: Llama2-7B (Touvron et al., 2023), Mistral-7B (Jiang et al., 2023), and Orca2-7B (Mitra et al., 2023). The choice of these models was meant to represent the most popular open-source base models and an instruction-tuned model across various baseline capabilities. Additionally, we selected bge-large-en (Xiao et al., 2023) as the embedding model for the RAG component and used FAISS (Johnson et al., 2019) as its vectorstore. This embedding model is currently the SOTA of open-source embedding models, according to the HuggingFace MTEB leaderboard<sup>4</sup>.",
+        "bbox": [
+            507,
+            620,
+            882,
+            828
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Configuration Variations Our evaluation included multiple configurations, with a grid-search",
+        "bbox": [
+            507,
+            829,
+            882,
+            860
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "1https://en.wikipedia.org/wiki/Category: 2023_events_in_the_United_States_by_month 2https://www.mediawiki.org/wiki/API:Main_page",
+        "bbox": [
+            112,
+            881,
+            477,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://huggingface.co/spaces/HuggingFaceH4/open_11m_leaderboard",
+        "bbox": [
+            507,
+            870,
+            873,
+            895
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "4https://huggingface.co/spaces/mteb/ leaderboard",
+        "bbox": [
+            507,
+            896,
+            806,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "242",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/bfec4d2d932d68807210f49de4e130c4710eec410f36d4658b5a3e658c50cd6a.jpg",
+        "image_caption": [
+            "Mean Accuracy Gain  $(\\%)$  by Model and Method",
+            "Figure 2: The relative accuracy gain (as explained in Equation (5)) for each knowledge-injection method, averaged (columnwise) across all experiments in Table 1."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            95,
+            485,
+            236
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "over them, to allow for more comprehensive benchmarking.",
+        "bbox": [
+            112,
+            325,
+            489,
+            356
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Firstly, we compared the baseline and fine-tuned models and their performance with the RAG component. Secondly, we explored the optimal number of text chunks to add to the context in RAG. Specifically, different values of  $K \\in \\{0, \\dots, 5\\}$  were employed to analyze the impact on model performance. Finally, we explored 5-shot performance vs. 0-shot.",
+        "bbox": [
+            112,
+            357,
+            487,
+            483
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Training Setup We trained all of the models using the unsupervised training procedure described in Section 3.2. For each dataset, we divided the auxiliary knowledge base into equal chunks of size 256 by concatenating or splitting the original chunks based on their length. We also added two special tokens,  $\\langle \\mathrm{BOS} \\rangle$  and  $\\langle \\mathrm{EOS} \\rangle$ , to demarcate the original chunks' beginnings and ends to preserve the documents' structure.",
+        "bbox": [
+            112,
+            486,
+            487,
+            630
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The models were trained using learning rates between  $1 \\times 10^{-6}$  and  $5 \\times 10^{-5}$ , which were found through a hyperparameter search. All models were trained on 4 NVIDIA A-100 GPUs for a maximum of 5 epochs and a batch size of 64.",
+        "bbox": [
+            112,
+            631,
+            487,
+            709
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation method All evaluations were done by appending each of the multiple-choice options to the question, followed by passing the concatenation through the model to get a log probability score per option. The highest score was interpreted as the model's choice and used for accuracy calculation. More formally, this means that in Equation (1) we say that  $\\mathcal{M}(q_n) = c_n$  if:",
+        "bbox": [
+            112,
+            711,
+            487,
+            840
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc _ {n} = \\underset {l} {\\arg \\max } \\left\\{\\mathcal {M} \\left(q _ {n} \\| a _ {n} ^ {1}\\right), \\dots , \\mathcal {M} \\left(q _ {n} \\| a _ {n} ^ {L}\\right) \\right\\}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            848,
+            487,
+            887
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{M}(q_n\\| a_n^l) = \\log P_{\\mathcal{M}}(q_n\\| a_n^l)$ .",
+        "bbox": [
+            112,
+            888,
+            396,
+            904
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "MMLU Results For each task and model, we",
+        "bbox": [
+            131,
+            904,
+            485,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "compared four approaches: using just the base model, RAG, FT, and finally combining FT and RAG by using the fine-tuned model as the generator. Furthermore, we tested the MMLU tasks using both 0-shot and 5-shot scenarios. The full results are shown in Table 1. An aggregation of the relative accuracy gain, i.e.,",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left(\\mathcal {L} _ {\\mathcal {M} ^ {\\prime}, \\mathcal {Q}} - \\mathcal {L} _ {\\mathcal {M}, \\mathcal {Q}}\\right) / \\mathcal {L} _ {\\mathcal {M}, \\mathcal {Q}}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            205,
+            882,
+            222
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{M}$  is the base model and  $\\mathcal{M}'$  is the knowledge-injected model, is shown in Figure 2.",
+        "bbox": [
+            507,
+            228,
+            880,
+            261
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In all cases, RAG performed significantly better compared to the base models. Furthermore, using RAG with the base model as the generator was consistently better than only fine-tuning. In some cases, using the fine-tuned model instead of the base model as the generator in the RAG pipeline improved results even further. However, this is not consistent and thus demonstrates the inherent instability of fine-tuning. Additionally, we found that the 5-shot approach boosts the results by a small margin in most cases, with a similar trend being observed in all of the different approaches.",
+        "bbox": [
+            507,
+            261,
+            882,
+            455
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Current Events Results The evaluation on the current events task is shown in Table 2. RAG proves particularly effective due to the one-to-one correspondence between the questions and the auxiliary dataset (see Section 4.3). Fine-tuning is not competitive with RAG. However, fine-tuning with multiple paraphrases still provides a significant improvement over the baseline. We note that combining RAG with fine-tuning shows inferior performance compared to RAG alone.",
+        "bbox": [
+            507,
+            455,
+            882,
+            615
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "It is worth noting that although the questions are based on information the models were not exposed to during training, the results of the base models surpass  $\\frac{1}{L} = 0.25$ . This can partially be explained by the models using reasoning and/or pre-existing knowledge when answering questions that are not independent of the past information. Some examples of this can be found in Appendix D.",
+        "bbox": [
+            507,
+            615,
+            882,
+            744
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Fine-Tuning vs. RAG: In the results of both the MMLU and current events tasks, a significant advantage for RAG over fine-tuning is evident. While fine-tuning improved results compared to the base model in most cases, it was not competitive with the RAG approach.",
+        "bbox": [
+            507,
+            744,
+            882,
+            840
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Several factors might contribute to this behavior. Firstly, RAG not only adds knowledge to a model but also incorporates context relevant to the question, a feature lacking in fine-tuning. Additionally, fine-tuning may impact other capabilities",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "243",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "of the model due to a degree of catastrophic forgetting. Finally, it's plausible that unsupervised fine-tuned models might benefit from further alignment through supervised or RL-based fine-tuning, as evidenced by the vastly improved performance of Orca2 over the base Llama2.",
+        "bbox": [
+            112,
+            84,
+            489,
+            180
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 The Importance of Repetition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            192,
+            405,
+            209
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Unlike the other tasks, where the model has been exposed to aspects related to the topic during pretraining, current events includes new information. In this case, standard regular fine-tuning not only did not improve the performance of Llama2 but also significantly degraded it. To improve the fine-tuning results, we explored augmentation of the data using paraphrases.",
+        "bbox": [
+            112,
+            216,
+            489,
+            344
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Data Augmentation Data augmentation is a well-established method for enhancing the performance of language models and has been surveyed extensively (Shorten et al., 2021). Using generative models for augmentations has also been used successfully to improve classification models in the past (Sharma et al., 2022). An example of data augmentation using paraphrasing can be found in Appendix C.",
+        "bbox": [
+            112,
+            344,
+            489,
+            488
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Monotonic Improvement This approach resulted in notable improvements in our results, showing a direct correlation between the number of paraphrases utilized and the models' accuracy. Our experimentation revealed a compelling trend. For all models tested, the accuracy was a monotonically increasing function of the number of paraphrases used (visualized in Appendix A, Figure 4). This observation strongly suggests the positive impact of paraphrase augmentation, yielding information repetition, on the model's ability to comprehend and generalize new knowledge from limited data.",
+        "bbox": [
+            112,
+            489,
+            489,
+            681
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Learning New Information In Appendix A, Figure 3, we can see an interesting phenomenon observed throughout our experiments. After each epoch, i.e., completing another iteration over the entire dataset, the training loss drops significantly. This is consistent with what is known about LLMs memorizing the data during training and overfitting (Tirumala et al., 2022).",
+        "bbox": [
+            112,
+            683,
+            489,
+            810
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our hypothesis is as follows:",
+        "bbox": [
+            132,
+            812,
+            351,
+            827
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In order to teach pre-trained LLMs new knowledge, the knowledge must be repeated in numerous ways.",
+        "bbox": [
+            149,
+            834,
+            453,
+            882
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This is well known for LLM pre-training (Kandpal et al., 2023), and we see in this case that this",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "holds for fine-tuning as well. The rationale for this hypothesis is that mere memorization of sentences does not entail knowledge of their content, as was already shown in (Berglund et al., 2023). By providing the information in numerous forms (like the data augmentation process we used), the various relationships in the data (e.g.,  $a \\Rightarrow b$ ,  $b \\Rightarrow c$ ) stand a higher chance of appearing naturally. We believe this can potentially both increase  $\\mathcal{L}_{\\mathcal{M},\\mathcal{Q}}$  in general, as well as ameliorate Berglund et al.'s Reversal Curse. While promising, this result still warrants further research.",
+        "bbox": [
+            507,
+            84,
+            884,
+            275
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            290,
+            796,
+            306
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Large language models possess vast amounts of knowledge on various topics. In this work, we tested their capability to adapt to new knowledge: both specialized and completely unseen. This is among the first studies to compare two prominent approaches in this domain, namely fine-tuning and retrieval augmented generation. While fine-tuning can be useful for many use-cases, we found that RAG is a more reliable choice for knowledge injection.",
+        "bbox": [
+            507,
+            317,
+            884,
+            475
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Some aspects of this work still warrant further research. For example, we focused on unsupervised training as our primary fine-tuning method, as opposed to instruction-tuning or RL-based methods. Researching combinations of various techniques, with diverse auxiliary knowledge bases, may yield improved results. This approach, combined with our hypothesis from Section 6, could further enhance our understanding of knowledge injection via FT.",
+        "bbox": [
+            507,
+            479,
+            884,
+            637
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "While we believe that this work further enhances our understanding of knowledge in LLMs, there is a lot more work to be done in this field. Specifically, more research is required regarding the question of knowledge representation in LLMs, especially from a theoretical perspective.",
+        "bbox": [
+            507,
+            640,
+            882,
+            736
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Finally, further efforts are needed to measure knowledge in LLMs. While we employed an empirical approach as described in Equation (2), it is important to explore other definitions and perspectives on knowledge as well, and extend upon this work.",
+        "bbox": [
+            507,
+            737,
+            882,
+            832
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "8 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            846,
+            645,
+            860
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As in all machine learning applications, the choice of hyperparameters significantly impacts the results. We therefore strongly recommend optimiz",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "244",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ing all relevant hyperparameters for specific cases. We have supported our claims by running the experiments on three different models. However, generalization to other LLMs should be tested thoroughly. For example, GPT-4 achieves near perfect accuracy for some MMLU tasks (Nori et al., 2023), and thus further improvement is not applicable.",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Finally, while we chose various topics for the knowledge bases, all of our sources came from Wikipedia. Other datasets may yield different results, and must be evaluated carefully.",
+        "bbox": [
+            112,
+            198,
+            489,
+            262
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            291,
+            213,
+            307
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Giuseppe Attardi. 2015. Wikiextractor. https://github.com/attardi/wikiextractor.",
+            "Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, and Owain Evans. 2023. The reversal curse: Llms trained on\" a is b\" fail to learn\" b is a\". arXiv preprint arXiv:2309.12288.",
+            "Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan Yu. 2020. Recall and learn: Fine-tuning deep pretrained language models with less forgetting. arXiv preprint arXiv:2004.12651.",
+            "Xiang Chen, Ningyu Zhang, Xin Xie, Shumin Deng, Yunzhi Yao, Chuanqi Tan, Fei Huang, Luo Si, and Huajun Chen. 2022. Knowprompt: Knowledge-aware prompt-tuning with synergistic optimization for relation extraction. In Proceedings of the ACM Web conference 2022, pages 2778-2788.",
+            "Yanda Chen, Ruiqi Zhong, Sheng Zha, George Karypis, and He He. 2021. Meta-learning via language model in-context tuning. arXiv preprint arXiv:2110.07814.",
+            "Yew Ken Chia, Pengfei Hong, Lidong Bing, and Soujanya Poria. 2023. Instructeval: Towards holistic evaluation of instruction-tuned large language models. arXiv preprint arXiv:2306.04757.",
+            "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.",
+            "Roi Cohen, Mor Geva, Jonathan Berant, and Amir Globerson. 2023. Crawling the internal knowledge-base of language models. arXiv preprint arXiv:2301.12810.",
+            "Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs. arXiv preprint arXiv:1903.00161."
+        ],
+        "bbox": [
+            115,
+            316,
+            487,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2021. A framework for few-shot language model evaluation.",
+            "Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. 2013. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211.",
+            "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR).",
+            "Linmei Hu, Zeyi Liu, Ziwang Zhao, Lei Hou, Liqiang Nie, and Juanzi Li. 2023. A survey of knowledge enhanced pre-trained language models. IEEE Transactions on Knowledge and Data Engineering.",
+            "Quzhe Huang, Mingxu Tao, Zhenwei An, Chen Zhang, Cong Jiang, Zhibin Chen, Zirui Wu, and Yansong Feng. 2023. Lawyer llama technical report. arXiv preprint arXiv:2305.15062.",
+            "Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.",
+            "Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547.",
+            "Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In International Conference on Machine Learning, pages 15696-15707. PMLR.",
+            "James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526.",
+            "Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. 2022. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329.",
+            "Anne Lauscher, Olga Majewska, Leonardo FR Ribeiro, Iryna Gurevych, Nikolai Rozanov, and Goran Glavas. 2020. Common sense or world knowledge? investigating adapter-based knowledge injection into pretrained transformers. arXiv preprint arXiv:2005.11787."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "245",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474.",
+            "Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and Ping Wang. 2020. K-bert: Enabling language representation with knowledge graph. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2901-2908.",
+            "Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2023. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747.",
+            "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, and Hannaneh Hajishirzi. 2021. Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773.",
+            "Arindam Mitra, Luciano Del Corro, Shweti Mahajan, Andres Codas, Clarisse Simoes, Sahaj Agrawal, Xuxi Chen, Anastasia Razdaibiedina, Erik Jones, Kriti Aggarwal, et al. 2023. Orca 2: Teaching small language models how to reason. arXiv preprint arXiv:2311.11045.",
+            "Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas A. Tezak, Jong Wook Kim, Chris Hallacy, Johannes Heidecke, Pranav Shyam, Boris Power, Tyna Eloundou Nekoul, Girish Sastry, Gretchen Krueger, David P. Schnurr, Felipe Petroski Such, Kenny Sai-Kin Hsu, Madeleine Thompson, Tabarak Khan, Toki Sherbakov, Joanne Jang, Peter Welinder, and Lilian Weng. 2022. Text and code embeddings by contrastive pre-training. ArXiv, abs/2201.10005.",
+            "Ha-Thanh Nguyen. 2023. A brief report on lawgpt 1.0: A virtual legal assistant based on gpt-3. arXiv preprint arXiv:2302.05729.",
+            "Harsha Nori, Nicholas King, Scott Mayer McKinney, Dean Carignan, and Eric Horvitz. 2023. Capabilities of gpt-4 on medical challenge problems. ArXiv, abs/2303.13375.",
+            "OpenAI. 2023. Gpt-4 technical report. ArXiv, abs/2303.08774.",
+            "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. 2019. Language models as knowledge bases? arXiv preprint arXiv:1909.01066.",
+            "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.",
+            "Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290.",
+            "Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106.",
+            "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.",
+            "Saket Sharma, Aviral Joshi, Namrata Mukhija, Yiyun Zhao, Hanoz Bhathena, Prateek Singh, Sashank Santhanam, and Pritam Biswas. 2022. Systematic review of effect of data augmentation using paraphrasing on named entity recognition. In NeurIPS 2022 Workshop on Synthetic Data for Empowering ML Research.",
+            "Connor Shorten, Taghi M. Khoshgoftaar, and Borko Furht. 2021. Text data augmentation for deep learning. Journal of Big Data, 8.",
+            "Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. 2023a. Large language models encode clinical knowledge. Nature, 620(7972):172-180.",
+            "Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Le Hou, Kevin Clark, Stephen Pfohl, Heather Cole-Lewis, Darlene Neal, et al. 2023b. Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617.",
+            "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 preprint arXiv:2206.04615.",
+            "Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023. Can chatgpt replace traditional kbqa models? an in-depth analysis of the question answering performance of the gpt llm family. In International Semantic Web Conference, pages 348-367. Springer."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "246",
+        "bbox": [
+            485,
+            928,
+            515,
+            939
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. https://crfm.stanford.edu/2023/03/13/alpaca.html, 3(6):7.",
+            "Kushal Tirumala, Aram H. Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. 2022. Memorization without overfitting: Analyzing the training dynamics of large language models. *ArXiv*, abs/2205.10770.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.",
+            "Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von Werra, Clémentine Fourrier, Nathan Habib, et al. 2023. Zephyr: Direct distillation of lm alignment. arXiv preprint arXiv:2310.16944.",
+            "Cunxiang Wang, Xiaoze Liu, Yuanhao Yue, Xiangru Tang, Tianhang Zhang, Cheng Jiayang, Yunzhi Yao, Wenyang Gao, Xuming Hu, Zehan Qi, et al. 2023. Survey on factuality in large language models: Knowledge, retrieval and domain-specificity. arXiv preprint arXiv:2310.07521.",
+            "Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Guihong Cao, Daxin Jiang, Ming Zhou, et al. 2020. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808.",
+            "Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. 2022. Super-naturalinstructions: Generalization via declarative instructions on  $1600+$  nlp tasks. arXiv preprint arXiv:2204.07705.",
+            "Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2023a. Pmc-llama: Further finetuning llama on medical papers. arXiv preprint arXiv:2304.14454.",
+            "Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. 2023b. Bloomberggpt: A large language model for finance. arXiv preprint arXiv:2303.17564.",
+            "Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-pack: Packaged resources to advance general chinese embedding. Preprint, arXiv:2309.07597.",
+            "Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. 2023. Fingpt: Open-source financial large language models. arXiv preprint arXiv:2306.06031."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "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, 54(11s):1-38.",
+            "Chunting Zhou, Pengfei Liu, Puxin Xu, Smini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. 2023. Lima: Less is more for alignment. arXiv preprint arXiv:2305.11206."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            203
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A The Importance of Repetition Figures",
+        "text_level": 1,
+        "bbox": [
+            510,
+            218,
+            873,
+            234
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/12792e2864fcfa37c5f44540720b0d87f810beb9b91141a21e66232ea13a280a.jpg",
+        "image_caption": [
+            "Figure 3: Training loss over time for Mistral-7B."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            256,
+            878,
+            426
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/ed1e6a4dfe8dc9cb8902ca2bb37b34f5a3d44c9f70742ff3c29834d14f6e494f.jpg",
+        "image_caption": [
+            "Figure 4: Model accuracy on the current events task as a function of the number of paraphrases."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            514,
+            483,
+            878,
+            663
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B RAG Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            509,
+            734,
+            726,
+            750
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Section 5, we compared various values of  $K \\in \\{0, \\dots, 5\\}$ , shown in Table 3. We were unable to find an optimal value of  $K$  per model, per  $0/5$ -shot, or per task. In fact, other than Anatomy that worked well with  $K = 2$  consistently, there seems to be no patterns that aid in predicting the performance per  $K$ , unlike the results presented in (Lewis et al., 2020) for other setups. Moreover, the gap between the best and worst performing  $K$ 's can be large.",
+        "bbox": [
+            507,
+            760,
+            882,
+            921
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "247",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Unfortunately, we must conclude that this additional hyperparameter is unstable. This is a downside of using RAG in practice, and the choice of  $K$  cannot be ignored.",
+        "bbox": [
+            112,
+            84,
+            489,
+            149
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Paraphrase Examples",
+        "text_level": 1,
+        "bbox": [
+            114,
+            162,
+            342,
+            180
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Below is the prompt we used to generate paraphrases with GPT-4:",
+        "bbox": [
+            112,
+            190,
+            489,
+            221
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Your task is to paraphrase a text paragraph. The paragraph is given below.",
+        "bbox": [
+            149,
+            237,
+            453,
+            269
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Make sure to keep the same meaning but change the wording. Do not change any factual information.",
+        "bbox": [
+            149,
+            275,
+            450,
+            322
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Try to keep roughly the same length of the original text.",
+        "bbox": [
+            149,
+            331,
+            450,
+            362
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Give NUM _ PARAPHRASES different paraphrases for each text.",
+        "bbox": [
+            149,
+            370,
+            448,
+            401
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "These paraphrases should be as different from each other as possible.",
+        "bbox": [
+            149,
+            409,
+            448,
+            439
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Return a JSON formatted string with one key, called 'paraphrases', and a list of paraphrases.",
+        "bbox": [
+            149,
+            448,
+            448,
+            495
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Input paragraph:",
+        "bbox": [
+            149,
+            504,
+            278,
+            519
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Below are some examples. First, an example taken from the following Wikipedia page: \"51st International Emmy Awards\".",
+        "bbox": [
+            112,
+            535,
+            487,
+            582
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Paraphrase I",
+        "text_level": 1,
+        "bbox": [
+            151,
+            598,
+            258,
+            614
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The fifty-first edition of the International Emmy Awards Competition accepted submissions across all categories from December 7, 2022, until the closing date of February 16, 2023.",
+        "bbox": [
+            149,
+            621,
+            450,
+            700
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Paraphrase II",
+        "text_level": 1,
+        "bbox": [
+            151,
+            708,
+            265,
+            724
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The International Emmy Awards Competition, now in its 51st year, accepted submissions across all categories starting from December 7, 2022, until the deadline on February 16, 2023.",
+        "bbox": [
+            149,
+            730,
+            453,
+            810
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Paraphrase III",
+        "text_level": 1,
+        "bbox": [
+            151,
+            818,
+            273,
+            834
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The fifty-first edition of the International Emmy Awards Competition accepted submissions across all categories from December 7, 2022, until the closing date of February 16, 2023.",
+        "bbox": [
+            149,
+            841,
+            450,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The next example was taken from the following Wikipedia page: \"2023 Indianapolis mayoral election\".",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Paraphrase I The mayoral election in Indianapolis took place on November 7, 2023, with preliminary elections occurring on May 2. The sitting mayor, Democrat Joe Hogsett, successfully ran for a third term. Both Hogsett and his Republican opponent, Jefferson Shreve, moved on to the main election.",
+        "bbox": [
+            544,
+            143,
+            848,
+            272
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Paraphrase II On November 7, 2023, citizens of Indianapolis cast their votes to elect their Mayor, following primary elections on May 2. Joe Hogsett, the Democrat already in office, won his bid for a third term. Hogsett and the Republican candidate, Jefferson Shreve, were the two contenders in the final electoral round.",
+        "bbox": [
+            544,
+            279,
+            848,
+            423
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Paraphrase III",
+        "text_level": 1,
+        "bbox": [
+            544,
+            430,
+            668,
+            445
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The mayoral election in Indianapolis took place on the 7th of November, 2023, following primary elections that occurred on the 2nd of May. Joe Hogsett, the incumbent Democrat, successfully ran for a third term. Both Hogsett and his Republican challenger, Jefferson Shreve, made it through to the final round of the election.",
+        "bbox": [
+            544,
+            451,
+            847,
+            595
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Current Events Existing Knowledge Examples",
+        "text_level": 1,
+        "bbox": [
+            509,
+            609,
+            858,
+            642
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To give a better understanding of how a model might be able to answer questions about new information, with better than random success, we present three possible scenarios as examples. These scenarios show how models with stronger reasoning skills can infer the correct answer even for unseen information.",
+        "bbox": [
+            507,
+            651,
+            884,
+            763
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The first scenario involves questions about previously unseen information, where basic reasoning abilities allow a model to make an educated guess.",
+        "bbox": [
+            507,
+            780,
+            882,
+            844
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: What was a key issue that led to the 2023 United Auto Workers strike?",
+        "bbox": [
+            544,
+            857,
+            847,
+            888
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Answers:",
+        "text_level": 1,
+        "bbox": [
+            546,
+            906,
+            623,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "248",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/3a1ca520cf30c08b3b74bd64a5fc69776b5d4f7ab3de0e38a095ab1e934de9a0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Task</td><td rowspan=\"2\">Model</td><td colspan=\"5\">#Retrieved documents (k)</td></tr><tr><td>1</td><td>2</td><td>3</td><td>4</td><td>5</td></tr><tr><td rowspan=\"3\">Anatomy (0-shot)</td><td>Mistral 7B</td><td>0.615</td><td>0.681</td><td>0.630</td><td>0.644</td><td>0.622</td></tr><tr><td>Llama2 7B</td><td>0.444</td><td>0.489</td><td>0.467</td><td>0.474</td><td>0.481</td></tr><tr><td>Orca2 7B</td><td>0.607</td><td>0.637</td><td>0.600</td><td>0.585</td><td>0.637</td></tr><tr><td rowspan=\"3\">Anatomy (5-shot)</td><td>Mistral 7B</td><td>0.659</td><td>0.667</td><td>0.659</td><td>0.681</td><td>0.674</td></tr><tr><td>Llama2 7B</td><td>0.496</td><td>0.563</td><td>0.541</td><td>0.526</td><td>0.526</td></tr><tr><td>Orca2 7B</td><td>0.630</td><td>0.659</td><td>0.600</td><td>0.600</td><td>0.600</td></tr><tr><td rowspan=\"3\">Astronomy (0-shot)</td><td>Mistral 7B</td><td>0.651</td><td>0.678</td><td>0.678</td><td>0.664</td><td>0.664</td></tr><tr><td>Llama2 7B</td><td>0.447</td><td>0.434</td><td>0.447</td><td>0.434</td><td>0.467</td></tr><tr><td>Orca2 7B</td><td>0.711</td><td>0.730</td><td>0.730</td><td>0.750</td><td>0.730</td></tr><tr><td rowspan=\"3\">Astronomy (5-shot)</td><td>Mistral 7B</td><td>0.704</td><td>0.684</td><td>0.658</td><td>0.684</td><td>0.724</td></tr><tr><td>Llama2 7B</td><td>0.461</td><td>0.447</td><td>0.474</td><td>0.428</td><td>0.454</td></tr><tr><td>Orca2 7B</td><td>0.730</td><td>0.737</td><td>0.750</td><td>0.743</td><td>0.763</td></tr><tr><td rowspan=\"3\">Biology (0-shot)</td><td>Mistral 7B</td><td>0.736</td><td>0.722</td><td>0.757</td><td>0.743</td><td>0.736</td></tr><tr><td>Llama2 7B</td><td>0.438</td><td>0.472</td><td>0.493</td><td>0.479</td><td>0.472</td></tr><tr><td>Orca2 7B</td><td>0.639</td><td>0.618</td><td>0.639</td><td>0.625</td><td>0.639</td></tr><tr><td rowspan=\"3\">Biology (5-shot)</td><td>Mistral 7B</td><td>0.722</td><td>0.778</td><td>0.778</td><td>0.771</td><td>0.743</td></tr><tr><td>Llama2 7B</td><td>0.500</td><td>0.521</td><td>0.507</td><td>0.465</td><td>0.472</td></tr><tr><td>Orca2 7B</td><td>0.625</td><td>0.639</td><td>0.625</td><td>0.660</td><td>0.660</td></tr><tr><td rowspan=\"3\">Chemistry (0-shot)</td><td>Mistral 7B</td><td>0.450</td><td>0.470</td><td>0.470</td><td>0.500</td><td>0.470</td></tr><tr><td>Llama2 7B</td><td>0.320</td><td>0.320</td><td>0.300</td><td>0.380</td><td>0.360</td></tr><tr><td>Orca2 7B</td><td>0.370</td><td>0.420</td><td>0.400</td><td>0.410</td><td>0.440</td></tr><tr><td rowspan=\"3\">Chemistry (5-shot)</td><td>Mistral 7B</td><td>0.540</td><td>0.490</td><td>0.500</td><td>0.510</td><td>0.470</td></tr><tr><td>Llama2 7B</td><td>0.280</td><td>0.320</td><td>0.340</td><td>0.340</td><td>0.380</td></tr><tr><td>Orca2 7B</td><td>0.390</td><td>0.430</td><td>0.400</td><td>0.430</td><td>0.470</td></tr><tr><td rowspan=\"3\">Prehistory (0-shot)</td><td>Mistral 7B</td><td>0.728</td><td>0.725</td><td>0.750</td><td>0.735</td><td>0.728</td></tr><tr><td>Llama2 7B</td><td>0.481</td><td>0.460</td><td>0.457</td><td>0.457</td><td>0.429</td></tr><tr><td>Orca2 7B</td><td>0.648</td><td>0.645</td><td>0.660</td><td>0.670</td><td>0.679</td></tr><tr><td rowspan=\"3\">Prehistory (5-shot)</td><td>Mistral 7B</td><td>0.710</td><td>0.750</td><td>0.759</td><td>0.756</td><td>0.762</td></tr><tr><td>Llama2 7B</td><td>0.512</td><td>0.485</td><td>0.525</td><td>0.519</td><td>0.531</td></tr><tr><td>Orca2 7B</td><td>0.660</td><td>0.688</td><td>0.685</td><td>0.698</td><td>0.688</td></tr></table>",
+        "bbox": [
+            208,
+            83,
+            789,
+            718
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 3: RAG ablation study.",
+        "bbox": [
+            396,
+            726,
+            596,
+            741
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Dissatisfaction with the quality of cafeteria food.",
+            "2. Disagreements over employee dress codes.",
+            "3. Discontent with stagnant wages and tiered employment systems.",
+            "4. Debates over the color scheme of the factories."
+        ],
+        "bbox": [
+            164,
+            766,
+            450,
+            897
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "In this case it is easy to guess that the third",
+        "bbox": [
+            132,
+            904,
+            485,
+            920
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "option is the most likely, even without knowledge of this specific strike.",
+        "bbox": [
+            507,
+            766,
+            880,
+            797
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A second scenario involves questions where prior knowledge about a topic may aid a model in answering.",
+        "bbox": [
+            507,
+            815,
+            880,
+            863
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: What environmental concern was raised by some scientists as a result of the 2023 Hawaii wildfires?",
+        "bbox": [
+            544,
+            873,
+            845,
+            920
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "249",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Answers:",
+        "text_level": 1,
+        "bbox": [
+            151,
+            101,
+            231,
+            116
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Rising temperatures.",
+            "2. Melting ice caps.",
+            "3. Charred soils running off into the shoreline.",
+            "4. Increased air pollution."
+        ],
+        "bbox": [
+            166,
+            124,
+            448,
+            212
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "In this case, knowing the geography of Hawaii, as well as immediate effects of wildfires, enables a model to give the first two options a lower likelihood. This process of elimination increases the probability of choosing one of the remaining options (the third option is the correct answer).",
+        "bbox": [
+            112,
+            225,
+            487,
+            322
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A third scenario arises due to the automatic question generation process, some questions strongly rely on pre-existing knowledge.",
+        "bbox": [
+            112,
+            338,
+            485,
+            385
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Question: What event in 2021 was compared to the September 2023 New York floods?",
+        "bbox": [
+            149,
+            398,
+            450,
+            445
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answers:",
+        "text_level": 1,
+        "bbox": [
+            151,
+            463,
+            228,
+            478
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Hurricane Katrina.",
+            "2. Hurricane Ida.",
+            "3. Hurricane Sandy.",
+            "4. Hurricane Harvey."
+        ],
+        "bbox": [
+            166,
+            487,
+            327,
+            558
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Since only one of these events occurred in 2021 (Hurricane Ida), and all the models tested have been exposed to events from 2021 during pretraining, this question can potentially be answered without using additional current information.",
+        "bbox": [
+            112,
+            571,
+            487,
+            650
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Finally, to demonstrate why it is reasonable to assume that models cannot generally answer questions about new information, with better than random success, look at the following example:",
+        "bbox": [
+            112,
+            668,
+            487,
+            733
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Question: How did Matthew Belk, a National Weather Service meteorologist, describe the September 2023 northeastern U.S. floods?",
+        "bbox": [
+            149,
+            745,
+            453,
+            808
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answers:",
+        "text_level": 1,
+        "bbox": [
+            151,
+            826,
+            228,
+            840
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. 50-year event.",
+            "2. 100-year event.",
+            "3. 200-year event.",
+            "4. 500-year event."
+        ],
+        "bbox": [
+            166,
+            850,
+            304,
+            920
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Even with some knowledge about floods and their statistical properties, it would be very difficult to guess that this specific meteorologist would call the flood a '200-year event'. This is especially true if the model was not exposed to information about the details of the flood.",
+        "bbox": [
+            507,
+            84,
+            882,
+            180
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "250",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 13
+    }
+]