Add Batch d7dba58c-cb42-4187-a755-a201760f00dd data

.gitattributes

@@ -4527,3 +4527,67 @@ saved_model/**/* 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/HEART-felt[[:space:]]Narratives_[[:space:]]Tracing[[:space:]]Empathy[[:space:]]and[[:space:]]Narrative[[:space:]]Style[[:space:]]in[[:space:]]Personal[[:space:]]Stories[[:space:]]with[[:space:]]LLMs/52ff86dc-0c94-4304-a369-06a14517252d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/HELPD_[[:space:]]Mitigating[[:space:]]Hallucination[[:space:]]of[[:space:]]LVLMs[[:space:]]by[[:space:]]Hierarchical[[:space:]]Feedback[[:space:]]Learning[[:space:]]with[[:space:]]Vision-enhanced[[:space:]]Penalty[[:space:]]Decoding/e64b5f27-e4c8-415c-9fbb-28824e138b14_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/HalluMeasure_[[:space:]]Fine-grained[[:space:]]Hallucination[[:space:]]Measurement[[:space:]]Using[[:space:]]Chain-of-Thought[[:space:]]Reasoning/4dd725bd-76ce-42b7-bd26-6a35e1b8ce5b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Hate[[:space:]]Personified_[[:space:]]Investigating[[:space:]]the[[:space:]]role[[:space:]]of[[:space:]]LLMs[[:space:]]in[[:space:]]content[[:space:]]moderation/98ff3b1a-68ac-4b9d-91bf-9c764205c8dc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Hateful[[:space:]]Word[[:space:]]in[[:space:]]Context[[:space:]]Classification/95a8e7f6-2b9f-4389-9551-280f7c5f4ca9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Heterogeneous[[:space:]]LoRA[[:space:]]for[[:space:]]Federated[[:space:]]Fine-tuning[[:space:]]of[[:space:]]On-Device[[:space:]]Foundation[[:space:]]Models/d6c62a2e-d2cd-45b7-a876-95885365f893_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/HiFT_[[:space:]]A[[:space:]]Hierarchical[[:space:]]Full[[:space:]]Parameter[[:space:]]Fine-Tuning[[:space:]]Strategy/8be09224-3564-4845-996a-57391f9ec216_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Hidden[[:space:]]Persuaders_[[:space:]]LLMs’[[:space:]]Political[[:space:]]Leaning[[:space:]]and[[:space:]]Their[[:space:]]Influence[[:space:]]on[[:space:]]Voters/9092ce3a-456e-4ac2-bb0b-5a622f61e8b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Hierarchical[[:space:]]Deconstruction[[:space:]]of[[:space:]]LLM[[:space:]]Reasoning_[[:space:]]A[[:space:]]Graph-Based[[:space:]]Framework[[:space:]]for[[:space:]]Analyzing[[:space:]]Knowledge[[:space:]]Utilization/08993fbe-2ae5-4a5c-8118-cc62910ca405_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Holistic[[:space:]]Automated[[:space:]]Red[[:space:]]Teaming[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Top-Down[[:space:]]Test[[:space:]]Case[[:space:]]Generation[[:space:]]and[[:space:]]Multi-turn[[:space:]]Interaction/86bbe5b8-cb8f-4b74-87d2-9621ab82e944_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Holistic[[:space:]]Evaluation[[:space:]]for[[:space:]]Interleaved[[:space:]]Text-and-Image[[:space:]]Generation/f8eb0448-29f0-4f5b-aa8f-9cf0305bf576_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Hopping[[:space:]]Too[[:space:]]Late_[[:space:]]Exploring[[:space:]]the[[:space:]]Limitations[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]on[[:space:]]Multi-Hop[[:space:]]Queries/65ec8836-1550-46f2-ba65-02952a60667d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Householder[[:space:]]Pseudo-Rotation_[[:space:]]A[[:space:]]Novel[[:space:]]Approach[[:space:]]to[[:space:]]Activation[[:space:]]Editing[[:space:]]in[[:space:]]LLMs[[:space:]]with[[:space:]]Direction-Magnitude[[:space:]]Perspective/161ff080-646f-4f4f-b321-dac0cd3737f3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]Do[[:space:]]Humans[[:space:]]Write[[:space:]]Code_[[:space:]]Large[[:space:]]Models[[:space:]]Do[[:space:]]It[[:space:]]the[[:space:]]Same[[:space:]]Way[[:space:]]Too/bcab7780-51ce-4e6d-a96e-544341cfdd57_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]Do[[:space:]]Your[[:space:]]Code[[:space:]]LLMs[[:space:]]perform_[[:space:]]Empowering[[:space:]]Code[[:space:]]Instruction[[:space:]]Tuning[[:space:]]with[[:space:]]Really[[:space:]]Good[[:space:]]Data/c3130af5-1c68-49da-8e9f-5fb738437989_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]Does[[:space:]]the[[:space:]]Disclosure[[:space:]]of[[:space:]]AI[[:space:]]Assistance[[:space:]]Affect[[:space:]]the[[:space:]]Perceptions[[:space:]]of[[:space:]]Writing_/21822ca9-b1b9-47f0-af22-502f4a53c24e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]Does[[:space:]]the[[:space:]]Textual[[:space:]]Information[[:space:]]Affect[[:space:]]the[[:space:]]Retrieval[[:space:]]of[[:space:]]Multimodal[[:space:]]In-Context[[:space:]]Learning_/5298ee0c-4669-4d87-bfbb-96b5f5ae3d6d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]Far[[:space:]]Can[[:space:]]We[[:space:]]Extract[[:space:]]Diverse[[:space:]]Perspectives[[:space:]]from[[:space:]]Large[[:space:]]Language[[:space:]]Models_/547cf537-ae60-4fc6-b05e-1218f8d92378_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]Hard[[:space:]]is[[:space:]]this[[:space:]]Test[[:space:]]Set_[[:space:]]NLI[[:space:]]Characterization[[:space:]]by[[:space:]]Exploiting[[:space:]]Training[[:space:]]Dynamics/a07b0b23-cf0f-4ee2-9e6f-c71926ac7277_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]Susceptible[[:space:]]are[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Ideological[[:space:]]Manipulation_/36e1bead-fc00-4832-bcd4-444ff4aacaa7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]do[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Learn[[:space:]]In-Context_[[:space:]]Query[[:space:]]and[[:space:]]Key[[:space:]]Matrices[[:space:]]of[[:space:]]In-Context[[:space:]]Heads[[:space:]]are[[:space:]]Two[[:space:]]Towers[[:space:]]for[[:space:]]Metric[[:space:]]Learning/05fbb172-9b13-42ec-8593-20c146571e57_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]to[[:space:]]Compute[[:space:]]the[[:space:]]Probability[[:space:]]of[[:space:]]a[[:space:]]Word/f9b4d5e0-7e26-4508-aa8a-946a515b8d5e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/How[[:space:]]to[[:space:]]Leverage[[:space:]]Demonstration[[:space:]]Data[[:space:]]in[[:space:]]Alignment[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Model_[[:space:]]A[[:space:]]Self-Imitation[[:space:]]Learning[[:space:]]Perspective/12bec327-2766-4965-bd73-0bd6cbac6a62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Human-LLM[[:space:]]Hybrid[[:space:]]Text[[:space:]]Answer[[:space:]]Aggregation[[:space:]]for[[:space:]]Crowd[[:space:]]Annotations/01d86178-7a61-4c4b-9676-259275c9ae85_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Humans[[:space:]]or[[:space:]]LLMs[[:space:]]as[[:space:]]the[[:space:]]Judge_[[:space:]]A[[:space:]]Study[[:space:]]on[[:space:]]Judgement[[:space:]]Bias/12ad6463-4ce1-4995-bbb3-a6da9008fbb8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/I[[:space:]]Could’ve[[:space:]]Asked[[:space:]]That_[[:space:]]Reformulating[[:space:]]Unanswerable[[:space:]]Questions/a303d495-b113-49e4-8111-9ff60ae869a6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/I[[:space:]]Learn[[:space:]]Better[[:space:]]If[[:space:]]You[[:space:]]Speak[[:space:]]My[[:space:]]Language_[[:space:]]Understanding[[:space:]]the[[:space:]]Superior[[:space:]]Performance[[:space:]]of[[:space:]]Fine-Tuning[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]LLM-Generated[[:space:]]Responses/e246e7ae-9e68-487a-b3cc-2a5c4cd9de15_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/I[[:space:]]Need[[:space:]]Help![[:space:]]Evaluating[[:space:]]LLM’s[[:space:]]Ability[[:space:]]to[[:space:]]Ask[[:space:]]for[[:space:]]Users’[[:space:]]Support_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]on[[:space:]]Text-to-SQL[[:space:]]Generation/07b89491-3d76-4df6-a824-3975a6a93aaa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/I[[:space:]]love[[:space:]]pineapple[[:space:]]on[[:space:]]pizza[[:space:]]!=[[:space:]]I[[:space:]]hate[[:space:]]pineapple[[:space:]]on[[:space:]]pizza_[[:space:]]Stance-Aware[[:space:]]Sentence[[:space:]]Transformers[[:space:]]for[[:space:]]Opinion[[:space:]]Mining/4b50926f-5527-4e0b-abe2-4ff156ae017e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/I-AM-G_[[:space:]]Interest[[:space:]]Augmented[[:space:]]Multimodal[[:space:]]Generator[[:space:]]for[[:space:]]Item[[:space:]]Personalization/a78e3ff0-6061-4889-b5a4-bc4f851e025c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/IDEAW_[[:space:]]Robust[[:space:]]Neural[[:space:]]Audio[[:space:]]Watermarking[[:space:]]with[[:space:]]Invertible[[:space:]]Dual-Embedding/826f4cd0-4588-481c-9c2e-0deae0fbac80_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/IFCap_[[:space:]]Image-like[[:space:]]Retrieval[[:space:]]and[[:space:]]Frequency-based[[:space:]]Entity[[:space:]]Filtering[[:space:]]for[[:space:]]Zero-shot[[:space:]]Captioning/e151b0b9-1282-4d23-9d2d-8eb8283fb748_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/IM-BERT_[[:space:]]Enhancing[[:space:]]Robustness[[:space:]]of[[:space:]]BERT[[:space:]]through[[:space:]]the[[:space:]]Implicit[[:space:]]Euler[[:space:]]Method/b5821ab0-473c-4b75-9dec-904b1a89a306_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/If[[:space:]]CLIP[[:space:]]Could[[:space:]]Talk_[[:space:]]Understanding[[:space:]]Vision-Language[[:space:]]Model[[:space:]]Representations[[:space:]]Through[[:space:]]Their[[:space:]]Preferred[[:space:]]Concept[[:space:]]Descriptions/3a3042bf-2835-4eb3-babe-19e243ff93a6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ImageInWords_[[:space:]]Unlocking[[:space:]]Hyper-Detailed[[:space:]]Image[[:space:]]Descriptions/ce921883-5420-451f-8747-1aab3cbe3e97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Images[[:space:]]Speak[[:space:]]Louder[[:space:]]than[[:space:]]Words_[[:space:]]Understanding[[:space:]]and[[:space:]]Mitigating[[:space:]]Bias[[:space:]]in[[:space:]]Vision-Language[[:space:]]Model[[:space:]]from[[:space:]]a[[:space:]]Causal[[:space:]]Mediation[[:space:]]Perspective/30576b10-3cee-466f-b396-6f59ea324c72_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Impeding[[:space:]]LLM-assisted[[:space:]]Cheating[[:space:]]in[[:space:]]Introductory[[:space:]]Programming[[:space:]]Assignments[[:space:]]via[[:space:]]Adversarial[[:space:]]Perturbation/9882655a-8a2b-4118-a108-c5f556b81faa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improve[[:space:]]Dense[[:space:]]Passage[[:space:]]Retrieval[[:space:]]with[[:space:]]Entailment[[:space:]]Tuning/9a4b3a2e-f4b8-481c-842e-de711003feeb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improve[[:space:]]Student’s[[:space:]]Reasoning[[:space:]]Generalizability[[:space:]]through[[:space:]]Cascading[[:space:]]Decomposed[[:space:]]CoTs[[:space:]]Distillation/3b9d8aa2-522e-4f20-b502-0010d73fea81_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improving[[:space:]]Discriminative[[:space:]]Capability[[:space:]]of[[:space:]]Reward[[:space:]]Models[[:space:]]in[[:space:]]RLHF[[:space:]]Using[[:space:]]Contrastive[[:space:]]Learning/5360ce99-0abd-45ef-859a-11b95d294c37_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improving[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Completion[[:space:]]with[[:space:]]Structure-Aware[[:space:]]Supervised[[:space:]]Contrastive[[:space:]]Learning/f097f273-a8e8-4b67-87a1-f0b8706c00a5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improving[[:space:]]Minimum[[:space:]]Bayes[[:space:]]Risk[[:space:]]Decoding[[:space:]]with[[:space:]]Multi-Prompt/0a5e62fe-fa9e-4a3f-ab75-1c10ad86313f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improving[[:space:]]Multi-party[[:space:]]Dialogue[[:space:]]Generation[[:space:]]via[[:space:]]Topic[[:space:]]and[[:space:]]Rhetorical[[:space:]]Coherence/678a0b50-dac8-4912-afbf-857ca696c03b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improving[[:space:]]Retrieval-augmented[[:space:]]Text-to-SQL[[:space:]]with[[:space:]]AST-based[[:space:]]Ranking[[:space:]]and[[:space:]]Schema[[:space:]]Pruning/a4acee9a-3913-4803-912b-f9cebee5a106_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improving[[:space:]]Spoken[[:space:]]Language[[:space:]]Modeling[[:space:]]with[[:space:]]Phoneme[[:space:]]Classification_[[:space:]]A[[:space:]]Simple[[:space:]]Fine-tuning[[:space:]]Approach/080ced85-d6fc-46a6-99c4-4475adcb000a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Improving[[:space:]]Zero-shot[[:space:]]LLM[[:space:]]Re-Ranker[[:space:]]with[[:space:]]Risk[[:space:]]Minimization/29bf6d49-6fef-427d-b88e-4e1295d1a0ee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/In[[:space:]]Search[[:space:]]of[[:space:]]the[[:space:]]Long-Tail_[[:space:]]Systematic[[:space:]]Generation[[:space:]]of[[:space:]]Long-Tail[[:space:]]Inferential[[:space:]]Knowledge[[:space:]]via[[:space:]]Logical[[:space:]]Rule[[:space:]]Guided[[:space:]]Search/e7a95ac7-3b45-4d73-9d06-08113fb3ac7d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/In-Context[[:space:]]Compositional[[:space:]]Generalization[[:space:]]for[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/25eec163-bb61-4646-abdb-ac90d7428122_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/In-context[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Event[[:space:]]Causality[[:space:]]Identification/83c3dfce-4bfe-4c7c-a433-cd3ac18b01f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Incomplete[[:space:]]Utterance[[:space:]]Rewriting[[:space:]]with[[:space:]]Editing[[:space:]]Operation[[:space:]]Guidance[[:space:]]and[[:space:]]Utterance[[:space:]]Augmentation/dfc15a96-bc31-48bd-871f-ed7212ef5784_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Incubating[[:space:]]Text[[:space:]]Classifiers[[:space:]]Following[[:space:]]User[[:space:]]Instruction[[:space:]]with[[:space:]]Nothing[[:space:]]but[[:space:]]LLM/1aedd604-8727-4206-8147-2ad281800265_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Induct-Learn_[[:space:]]Short[[:space:]]Phrase[[:space:]]Prompting[[:space:]]with[[:space:]]Instruction[[:space:]]Induction/cae76993-19b2-4e16-849c-fac166c52471_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Inductive-Deductive[[:space:]]Strategy[[:space:]]Reuse[[:space:]]for[[:space:]]Multi-Turn[[:space:]]Instructional[[:space:]]Dialogues/69acf8de-2087-47f9-be9b-a5ef2eceec84_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/InferAligner_[[:space:]]Inference-Time[[:space:]]Alignment[[:space:]]for[[:space:]]Harmlessness[[:space:]]through[[:space:]]Cross-Model[[:space:]]Guidance/f7d3f2ea-861b-4fc8-b7f5-b8959378d729_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Inference[[:space:]]Helps[[:space:]]PLMs’[[:space:]]Conceptual[[:space:]]Understanding_[[:space:]]Improving[[:space:]]the[[:space:]]Abstract[[:space:]]Inference[[:space:]]Ability[[:space:]]with[[:space:]]Hierarchical[[:space:]]Conceptual[[:space:]]Entailment[[:space:]]Graphs/06e09a16-132e-4777-a308-49e31a8a629c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/InfiniPot_[[:space:]]Infinite[[:space:]]Context[[:space:]]Processing[[:space:]]on[[:space:]]Memory-Constrained[[:space:]]LLMs/cd383772-7cc5-46ad-b613-2e96016ba688_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Information[[:space:]]Flow[[:space:]]Routes_[[:space:]]Automatically[[:space:]]Interpreting[[:space:]]Language[[:space:]]Models[[:space:]]at[[:space:]]Scale/6c15aeef-2fdd-42e8-bda1-336f7b8b4558_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Initialization[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Reparameterization[[:space:]]to[[:space:]]Mitigate[[:space:]]Loss[[:space:]]Spikes/940c6998-5838-4b93-8c76-6cb51ebb7a9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Instruction[[:space:]]Fine-Tuning_[[:space:]]Does[[:space:]]Prompt[[:space:]]Loss[[:space:]]Matter_/640ee643-f003-4473-bcbc-75e5a394bcd0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Instruction[[:space:]]Matters_[[:space:]]A[[:space:]]Simple[[:space:]]yet[[:space:]]Effective[[:space:]]Task[[:space:]]Selection[[:space:]]for[[:space:]]Optimized[[:space:]]Instruction[[:space:]]Tuning[[:space:]]of[[:space:]]Specific[[:space:]]Tasks/d7a15f98-1057-434b-9c95-266d4833dcbb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Instruction[[:space:]]Pre-Training_[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Supervised[[:space:]]Multitask[[:space:]]Learners/3a290c64-0860-4f96-8d35-7be94aa18747_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/IntCoOp_[[:space:]]Interpretability-Aware[[:space:]]Vision-Language[[:space:]]Prompt[[:space:]]Tuning/4cef5d07-4740-411d-8511-819d85e49de2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Integrating[[:space:]]Argumentation[[:space:]]and[[:space:]]Hate-Speech-based[[:space:]]Techniques[[:space:]]for[[:space:]]Countering[[:space:]]Misinformation/f63b5785-c0b2-437d-bc07-6f24b9b4ff83_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Integrating[[:space:]]Plutchik’s[[:space:]]Theory[[:space:]]with[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]for[[:space:]]Enhancing[[:space:]]Emotion[[:space:]]Classification/99242a6d-163a-446c-8308-d6024dc67f00_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/HEART-felt Narratives_ Tracing Empathy and Narrative Style in Personal Stories with LLMs/52ff86dc-0c94-4304-a369-06a14517252d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:728519b7c565b7ff94acdc223717e8cf0c495083c98dd84c3f415a0a6d56ac4d
+size 1064601

2024/HEART-felt Narratives_ Tracing Empathy and Narrative Style in Personal Stories with LLMs/full.md

@@ -0,0 +1,804 @@
+# HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs
+
+Jocelyn Shen<sup>1</sup> Joel Mire<sup>2</sup> Hae Won Park<sup>1</sup> Cynthia Breazeal<sup>1</sup> Maarten Sap<sup>2</sup>
+
+<sup>1</sup>Massachusetts Institute of Technology, Cambridge, MA, USA
+
+2Carnegie Mellon University, Pittsburgh, PA, USA
+
+joceshen@mit.edu, jmire@andrew.cmu.edu, haewon@mit.edu,
+
+breazeal@mit.edu, msap2@andrew.cmu.edu
+
+# Abstract
+
+Empathy serves as a cornerstone in enabling prosocial behaviors, and can be evoked through sharing of personal experiences in stories. While empathy is influenced by narrative content, intuitively, people respond to the way a story is told as well, through narrative style. Yet the relationship between empathy and narrative style is not fully understood. In this work, we empirically examine and quantify this relationship between style and empathy using LLMs and large-scale crowdsourcing studies. We introduce a novel, theory-based taxonomy, HEART (Human Empathy and Narrative Taxonomy) that delineates elements of narrative style that can lead to empathy with the narrator of a story. We establish the performance of LLMs in extracting narrative elements from HEART, showing that prompting with our taxonomy leads to reasonable, human-level annotations beyond what prior lexicon-based methods can do. To show empirical use of our taxonomy, we collect a dataset of empathy judgments of stories via a large-scale crowdsourcing study with  $N = 2$ , 624 participants.<sup>1</sup> We show that narrative elements extracted via LLMs, in particular, vividness of emotions and plot volume, can elucidate the pathways by which narrative style cultivates empathy towards personal stories. Our work suggests that such models can be used for narrative analyses that lead to human-centered social and behavioral insights.
+
+# 1 Introduction
+
+Empathy, which is a foundational psychological process that drives many prosocial functions, (Zaki, 2019; Morelli et al., 2015), is often delivered through storytelling and sharing of personal experiences (Coplan, 2004; Keen, 2014). Empathetic responses evoked by stories are affected by factors
+
+Story: It was a long and difficult pregnancy. I felt like my insides were being ripped apart. But at 4:15 pm, I gave birth to a beautiful baby. I was totally exhausted, with cold tears streaming down my face. But looking into my baby's eyes, all the pain disappeared, and I just felt warmth in my heart.
+
+![](images/58621cc4e9104aa7eb491831e5c57d1d2ce692f4f4b70d890d1b3ee88a84231a.jpg)  
+Figure 1: Narrative empathy can be evoked through the way a story is told (narrative style). This work introduces HEART, a theory-driven taxonomy of narrative elements that contribute to empathy.
+
+beyond the content of the story alone – delivery, context, and reader characteristics all contribute to the emotional resonance of a narrative. Most studies of narrative empathy and its related constructs focus on reader characteristics, and content of a story (Sharma et al., 2020; Shen et al., 2023). However, intuitively, people also respond to the way a story is told, or the stylistic devices used within a narrative (Figure 1).<sup>2</sup>
+
+A key challenge in narrative analysis within the NLP community is that extracting stylistic features relevant to empathy is not trivial. Prior works use word-count-based (e.g., lexica; Roshanaei et al., 2019; Zhou et al., 2021) or hand-crafted features on extremely limited story sets (Kuzmičová et al., 2017; Fernandez-Quintanilla, 2020; Fernandez-Quintanilla and Stradling, 2023; Eekhof et al., 2023; Mangen et al., 2018; Hartung et al., 2016) to quantify narrative elements. However, more complex stylistic narrative devices, such as plot shifts (Nabi and Green, 2015) or vividness of emotions
+
+(Pillemer, 1992) are harder to summarize with lex-ica alone. While a few works have explored using LLMs for more complex narrative analysis tasks (Zhu et al., 2023; Michelmann et al., 2023; Sap et al., 2022), to what extent LLMs can effectively model stylistic devices, and how LLM-extracted features might be leveraged for downstream social insights, remains underexplored.
+
+In this work, we fill this gap by presenting the following contributions. (1) We introduce HEART (Human Empathy and Narrative Taxonomy), a theory-driven taxonomy of narrative style elements that relate to empathy. (2) We use LLMs to quantify aspects of narrativity in our taxonomy and evaluate how well LLMs represent these elements in line with human judgments. For a subset of narrative elements with available lexica, we compare lexical measures with LLM measures, finding that in most cases, GPT-4 and Llama 3 outperform lexica. (3) Through a human study of  $N = 2$ , 624 participants, we introduce a new crowdsourced dataset (HEART-felt Stories Dataset) of empathetic reactions to personal narratives, including annotated narrative style elements, reader characteristics and narrative reactions. (4) With our dataset, we conduct an analysis of pathways through narrative style and reader characteristics leading to empathy, demonstrating the value of HEART in exploring empirical behavioral insights around narrative empathy. In particular, we find that narrative styles with heightened vividness of emotions, character development and action, and plot volume, are tied to narrative empathy. We additionally show that empathy is personalized, with high variability even for the same story, and that beyond narrative style, factors like a reader's trait empathy and similarity of experiences to the narrator also significantly impact empathy.
+
+# 2 Related Work
+
+Computational linguistic methods can be used to analyze many aspects of narrativity across a large corpus of stories (Sap et al., 2022). Prior works have used lexicon-based approaches to extract psychologically-grounded word categories and relate these to empathy (Roshanaei et al., 2019; Xiao et al., 2016). Zhou et al. (2021) use linguistic style features such as degree of interdependent thinking and integrative complexity (the ability of a person to recognize multiple perspectives and connect them) to predict a viewer's empathy towards a specific situation. Antoniak et al. (2019) ap
+
+ply narrative analysis techniques to birth stories online, and show patterns of affective and event-based sequences over time. More recently, Yaden et al. (2024) used linguistic features, such as word phrases and topics, and leveraged LDA to analyze language that separates more empathetic people from more compassionate people, showing that compassionate people use more other-focused language than empathetic people. Other works leverage recent natural language processing (NLP) methods to predict empathy and prosociality from text (Shen et al., 2023; Buechel et al., 2018; Sharma et al., 2020; Bao et al., 2021), but do not explore pathways via which readers feel empathy.
+
+A few works have explored the power of LLMs in characterizing aspects of narrative. In particular, Michelmann et al. (2023) show that LLMs serve as good approximations of human annotators in narrative event segmentation. Other works show that LLMs achieve reasonable performance on character profiling tasks for fictional narratives, particularly in factual consistency and motivation understanding. However, Subbiah et al. (2024) indicate that LLMs fail to perform authentic summarization of stories in line with feedback from writers, apart from successfully drawing on thematic components of the stories. Ultimately, LLMs demonstrate growing potential for narrative understanding tasks (Zhu et al., 2023), but how well they perform, what types of tasks they succeed in, and how they can reveal human behavioral insights, is an active area of research (Agnew et al., 2024).
+
+Our work leverages LLMs to extract narrative style elements that may play a role in narrative empathy through our grounded taxonomy. We evaluate the performance of prompting LLMs to extract such elements against expert human raters. Our empirical study using LLM-extracted narrative elements focuses more on the scientific and behavioral question of how to untangle aspects of narrative style and reader characteristics to understand their contribution towards empathy, rather than improving performance on empathy prediction alone.
+
+# 3 Background
+
+Empathy in the context of narratives has been the subject of many studies in psychology and literary studies. We briefly summarize those below.
+
+Narrative Style and its Role in Empathy. Prior works have theorized how shifts in narrative style impact empathic effect of a story. Keen (2006) pro
+
+posed a theory of narrative empathy that draws on narrative techniques to enhance empathy, such as flatness or roundness of a character, the character's mode of consciousness, and vivid use of settings. van Krieken et al. (2017) presented a framework of linguistic cues to measure identification with narrative characters, including character dimensions such as the emotional or perceptual subject of the story. This framework covers both background elements of a story, which can facilitate immersive experiences, and foregrounded elements (such as figurative language), which facilitate aesthetic experiences with the text (Jacobs, 2015).
+
+However, many of these narrative techniques, particularly those that are more abstract in nature, such as plot structure or emotional shifts (Nabi and Green, 2015), have yet to be tested empirically. Researchers in narratology have explored the impact of literary quality on reader empathy, varying aspects such as foregrounding, point of view/viewpoint words, emotion and discourse presentation, and characterisation techniques, but have found mixed results in small-scale studies (Kuzmičová et al., 2017; Fernandez-Quintanilla, 2020; Fernandez-Quintanilla and Stradling, 2023; Eekhof et al., 2023; Mangen et al., 2018; Hartung et al., 2016). Other studies have looked at how aspects of literary reading contribute to transportation, or the ability to absorb in a narrative, which further predicts empathy towards a story (Walkington et al., 2020; van Laer et al., 2014, 2019). Koopman (2015) conducted a larger-scale study to investigate the role of genre, personal factors, and affective responses on both empathic understanding and pro-social behavior, finding that genre affected prosocial behaviors. However, narrative style encompasses many aspects beyond genre alone, and each of these elements couples with one another to enhance or diminish narrative empathy.
+
+# Reader Characteristics and Narrative Empathy.
+
+While narrative style can have an effect on empathy, other factors such as the reader's characteristics or experiences during reading can affect empathy as well. For example, psychology, economics, and neuroscience have suggested that gender has a significant influence on people's cognitive empathy, with women exhibiting higher cognitive empathy than men across a variety of age groups (Christov-Moore et al., 2014; Michalska et al., 2013; O'brien et al., 2013). Levels of narrative empathy can also be modulated by one's trait empathy level (Kon
+
+rath et al., 2018), emotional state during reading (Roshanaei et al., 2019), or general exposure to literature (Mar et al., 2006). Untangling the effects of these fixed can be challenging, and has been attempted by a few prior works, but with varied results (Koopman and Hakemulder, 2015; Fernandez-Quintanilla, 2020; Roshanaei et al., 2019).
+
+In our work, we propose a taxonomy of narrative empathy based on theories and empirical results presented in the aforementioned works, then scientifically explore what pathways through both narrative style and reader characteristics and life experiences and to overall empathy towards a story. In contrast to prior works, which often vary a single element of narrative style, we construct a thorough taxonomy of narrative elements related to empathy.
+
+# 4 HEART Taxonomy for Empathy and Narrative Style
+
+Based on the aforementioned theoretical and empirical research, we propose HEaRt, a taxonomy of narrative style elements that can lead to empathy. In A Theory of Narrative Empathy, Keen posits that aspects of characterization, narrative situation, internal perspective, and techniques to represent character consciousness can contribute to narrative empathy. We use these concepts as precursors for developing HEaRt. Our theoretical model serves as a starting point for understanding what aspects of narrative characteristics might lead to empathy and how we can measure these factors using computational approaches.
+
+Figure 2 shows our full taxonomy, which delineates narrative style as it relates to narrative empathy via four main categories: (1) Character identification (2) Plot (3) Point of view and (4) Setting. In the remainder of this section, we outline each element of our taxonomy and the theoretical and empirical roots of how each element may contribute to narrative empathy.
+
+Character Identification We refer to character identification elements as story aspects that draw readers into the narrator's perspective, whether this be across internal dimensions (emotion/cognition) or external dimensions (perception/time). We define 6 high-level elements of our taxonomy that can contribute to identification with a character in a story, primarily rooted in (van Krieken et al., 2017)'s work on character identification:
+
+1. Flatness/roundness (Keen, 2006) of the charac
+
+![](images/082891f850d646619c8049c21f500fdc2a0dd3fcde122ce9b1787b735da79c34.jpg)  
+Figure 2: Narrative Empathy and Style Taxonomy delineating aspects of narrative style that theoretically relate to empathy towards a narrative.
+
+ter, including depth of the character expressed through character development over the course of the story or character vulnerability.
+
+2. Emotional subject (van Krieken et al., 2017; Roshanaei et al., 2019; Pillemer, 1992), refers to the way emotions are expressed both in tone and vividness of emotions.  
+3. Cognitive subject (Schweitzer and Waytz, 2021; van Krieken et al., 2017), captures expressions of cognition such as thinking, planning, and decision making.  
+4. Moral subject (van Krieken et al., 2017; Saldias and Roy, 2020) primarily refers to how evaluations or expressions of the narrator's opinion are conveyed through the story.  
+5. Action subject (van Krieken et al., 2017), refers to expressions of character action.  
+6. Subject perception (van Krieken et al., 2017) captures the vividness of perception and bodily sensations experienced by the character.  
+7. Temporal references (Pillemer, 1992) contain expressed nostalgia (looking to the past) or forecasting and anticipation (looking to the future).
+
+Plot Defining plot has been a key task in narrative analysis (Toubia et al., 2021; Reagan et al., 2016), and can foster empathy through enhancing the narrator's story via shifts at critical junctures. We delineate 3 aspects of plot that relate to narrative empathy:
+
+1. Plot volume (Keen, 2014; van Laer et al., 2014, 2019) captures the frequency and significance of events in a story.  
+2. Emotion shifts (Nabi and Green, 2015) indicate fluctuations in the overall emotional trajectory
+
+of the story (such as from low to high valence and vice versa).
+
+3. Resolution (Mcadams, 2006) captures the release of tension after the main conflict that a character experiences.
+
+Point of view Prior works suggest that point of view can affect empathy towards a narrator (Eekhof et al., 2023; Fernandez-Quintanilla, 2020; Spitale et al., 2022). For example, first-person perspective can emphasize the personal nature of the story and draw readers into the shoes of the narrator.
+
+Setting Finally, the environment and context of the narrator can facilitate narrative empathy (Pillemer, 1992; van Krieken et al., 2017), for example through world-building to enhance narrative transportation. We capture this element via the vividness of the setting description in a narrative.
+
+# 5 HEART-felt Stories Dataset Annotation
+
+With our theory-grounded taxonomy, we next evaluate how well LLMs can approximate narrative style elements. In order to do so, we annotate the HEART-felt Stories Dataset, a corpus of personal narratives with expert ratings on a subset of stories.
+
+# 5.1 Story Dataset
+
+To empirically observe the narrative elements of HEART, we started with a seed dataset of personal narratives from the EMPATHICSTORIES (Shen et al., 2023) and the EMPATHICSTORIES++ (Shen et al., 2024) dataset, which were specifically designed to include meaningful and vulnerable personal stories with diverse narrators, shared across diverse topics (e.g. relationships, mental health, career and school, etc.). The EMPATHICSTORIES
+
+<table><tr><td>Feature</td><td>KA</td><td>PPA</td><td>ρ</td></tr><tr><td>Optimistic tone</td><td>49.27</td><td>81.50</td><td>72.21***</td></tr><tr><td>Vivid setting</td><td>48.48</td><td>76.00</td><td>64.23***</td></tr><tr><td>Plot volume</td><td>45.97</td><td>83.50</td><td>60.32***</td></tr><tr><td>Resolution</td><td>44.29</td><td>79.00</td><td>58.97***</td></tr><tr><td>Character vulnerability</td><td>38.17</td><td>75.00</td><td>50.06***</td></tr><tr><td>Character development</td><td>28.55</td><td>72.50</td><td>45.24**</td></tr><tr><td>Cognition</td><td>27.56</td><td>70.00</td><td>39.18**</td></tr><tr><td>Evaluations</td><td>26.29</td><td>74.00</td><td>31.3*</td></tr><tr><td>Emotion shifts</td><td>23.49</td><td>74.50</td><td>46.34**</td></tr><tr><td>Vivid emotions</td><td>21.17</td><td>66.00</td><td>31.8*</td></tr><tr><td>Temporal references</td><td>18.29</td><td>77.00</td><td>27.96*</td></tr><tr><td>Bodily sensations</td><td>3.79</td><td>60.33</td><td>34.25*</td></tr></table>
+
+Table 1: Agreement between 2 expert human annotators on the narrative elements of our taxonomy. Scores are multiplied by 100 and rounded for readability and sorted by KA. Spearman's correlation  $\rho$  indicates significance.
+
+dataset consists of  $\sim 1,500$  personal narratives collected from social media sites (Facebook, Reddit), crowdsourced personal narratives, and transcribed podcasts. The EMPATHICSTORIES++ dataset contains  $\sim 500$  conversational personal stories that were automatically transcribed from storytelling interactions with an AI. We filtered stories to remove potentially harmful topics (e.g. mentions of sexual assault, excessive swearing), and filtered stories that were under 200 words (which might not contain rich narrative style elements), resulting in a final dataset of 874 personal stories.
+
+# 5.2 Expert Narrative Style Annotation
+
+We randomly sampled 50 stories from our final dataset of 874 stories to obtain expert annotations of the narrative elements and validate LLM performance on the task. We selected a subset of 12 narrative elements from our taxonomy that are nontrivial to extract from existing NLP toolkits, and which required human judgments given the subjectivity of the task. Three independent members of our research team with expertise in text analysis and annotation iteratively designed a codebook (Appendix C) with instructions and examples for gauging the presence of each element.
+
+Subsequently, two independent expert annotators rated the presence of each of the 12 narrative elements in the 50 sampled stories. Table 1 shows the agreement between the 2 raters using Krippendorf's alpha (KA), percent pairwise agreement (PPA), and Spearman's correlation  $(\rho)$ . All ratings are positively correlated to each other, but different narrative elements have varying degrees of agreement. We observe the lowest agreement between human annotators for TEMPORAL REFER-
+
+ENCES and BODILY SENSATIONS, where unrealis events and mentions of body sensations across multiple characters caused confusion. Moreover, while some human agreements may appear low using the KA metric, these scores are consistent with prior NLP tasks with more subjectivity (Shen et al., 2023; Rashkin et al., 2018; Sap et al., 2017). In our subsequent empirical analysis, we do not use features with low agreement (below 0.2 KA).
+
+# 6 LLMs for Narrative Style Extraction
+
+Our work explores how LLM-extracted narrative features can be used to yield empirical social insights around empathy and storytelling. As such, we validate whether LLMs are capable of narrative style annotations in line with expert human judgments. To this end, we prompt GPT-4<sup>3</sup> and the instruction-tuned variant of Llama 3 8B<sup>4</sup> with the same instructions and codebook given to human annotators (Appendix C). In Table 2, we report agreement between averaged human ratings and the LLM-based ratings on the same 50 sampled stories.
+
+We observe similar patterns in agreement between GPT-4 and human raters as we do in agreement between our two expert annotators. GPT-4 provides ratings with substantial agreement for narrative features such as CHARACTER VULNERABILITY, OPTIMISTIC TONE, and RESOLUTION. For most features, the GPT-4 ratings are more positively correlated with human annotations than are the Llama 3 ratings. As such, we use GPT-4 to extract the narrative elements for all the remaining stories in our corpus and exclude features that have low agreement with human gold labels in our subsequent empirical study.
+
+# 6.1 Performance of LLMs vs. Lexica
+
+As prior works use lexica (Roshanaei et al., 2019; Zhou et al., 2021) to quantify narrative elements, we compare whether GPT-4 and Llama 3 can outperform psychologically validated lexica in capturing features of HEART. We select 4 dimensions in our taxonomy that readily map to lexicon-based dimensions in LIWC-22 (Boyd, 2022; Pennebaker et al., 1999) and compare correlation to human expert ratings in Table 3. We find that GPT-4-extracted features for OPTIMISTIC TONE, VIVID EMOTIONS, and CHARACTER VULNERABILITY
+
+<table><tr><td rowspan="2">Feature</td><td colspan="3">GPT-4</td><td colspan="3">Llama 3 8B Instruct</td></tr><tr><td>KA</td><td>PPA</td><td>ρ</td><td>KA</td><td>PPA</td><td>ρ</td></tr><tr><td>Character vulnerability</td><td>62.89</td><td>86.50</td><td>80.15***</td><td>27.08</td><td>79.00</td><td>70.55***</td></tr><tr><td>Optimistic tone</td><td>50.97</td><td>82.25</td><td>68.06***</td><td>48.41</td><td>82.75</td><td>67.14***</td></tr><tr><td>Resolution</td><td>44.55</td><td>80.00</td><td>61.59***</td><td>7.26</td><td>71.83</td><td>34.93*</td></tr><tr><td>Character development</td><td>44.09</td><td>79.25</td><td>61.64***</td><td>20.99</td><td>77.25</td><td>46.51**</td></tr><tr><td>Vivid setting</td><td>42.12</td><td>78.00</td><td>67.31***</td><td>-31.07</td><td>57.03</td><td>41.84**</td></tr><tr><td>Plot volume</td><td>33.00</td><td>79.25</td><td>44.88**</td><td>-4.00</td><td>76.08</td><td>27.51</td></tr><tr><td>Emotion shifts</td><td>32.25</td><td>82.25</td><td>45.5**</td><td>25.13</td><td>80.58</td><td>52.13***</td></tr><tr><td>Vivid emotions</td><td>27.25</td><td>75.00</td><td>59.21***</td><td>25.80</td><td>76.00</td><td>42.13**</td></tr><tr><td>Cognition</td><td>19.83</td><td>73.00</td><td>34.91*</td><td>24.98</td><td>76.00</td><td>52.89***</td></tr><tr><td>Evaluations</td><td>-9.76</td><td>75.00</td><td>22.69</td><td>-27.16</td><td>73.00</td><td>NaN</td></tr></table>
+
+Table 2: Agreement between aggregated human annotators (gold ratings) and GPT-4 and Llama 3 8B Instruct ratings of narrative elements in our taxonomy. Rows are sorted by GPT-4 KA.  
+
+<table><tr><td>Feature</td><td>ρLlWC</td><td>ρGPT-4</td><td>ρLlama3</td></tr><tr><td>Optimistic tone</td><td>47.35***</td><td>68.06***</td><td>67.14***</td></tr><tr><td>Cognition</td><td>41.29**</td><td>34.91*</td><td>52.89***</td></tr><tr><td>Vivid emotions</td><td>37.63**</td><td>59.21***</td><td>42.13**</td></tr><tr><td>Character vulnerability</td><td>-6.95</td><td>80.15***</td><td>70.55***</td></tr></table>
+
+Table 3: Comparison of correlations with human annotations for LIWC, GPT-4, and Llama 3 8B Instruct.
+
+are better aligned with human ratings than LIWC correspondents, although only CHARACTER VULNERABILITY is statistically significantly higher  $(p < 0.001$  as measured by Fisher's exact test). However, LIWC outperforms GPT-4 in the COGNITION category, although not statistically significantly so. We discuss the source of potential errors in using GPT-4 to extract COGNITION level of narratives in our error analyses below. Notably, although Llama 3 annotations are generally relatively less correlated with human annotations, the Llama 3 extracted features consistently outperform the LIWC correspondents.
+
+# 6.2 Error Analysis
+
+We observe that GPT-4 consistently over-rates the level of EVALUATIONS and COGNITION expressed in a story as compared to human annotators. Through qualitative examples of stories where GPT-4 and human disagreements are large (Appendix D), GPT-4 typically conflates emotional reactions with evaluations, attributions, or desires (e.g. "...it really got me thinking about when I first went to College...How excited my parents were for me and scared. And I was both excited and scared..."). For COGNITION errors, we see that these systematic errors are typically due to GPT-4 conflating recollection with demonstrations of cognition when overall, the story did not contain more internal thinking processes.
+
+Regarding Llama 3, we observe that when human annotators and GPT-4 agree, but Llama 3 disagrees, it tends to assign higher scores to a minority
+
+of features (e.g., CHARACTER VULNERABILITY) while giving lower scores to a majority of features (e.g., VIVID EMOTIONS, VIVID SETTING). The lower ratings for imagery-related features suggest a lesser adeptness with figurative language.
+
+Ultimately, our validation study demonstrates that LLMs – in particular, GPT-4 – can approximate extracting narrative elements relevant to empathy as corroborated by prior work (Shen et al., 2023; Ziems et al., 2024), but some features are more challenging for the model to identify. We show in the following section that GPT-4 narrative ratings still reveal interesting behavioral insights around narrative empathy, even without perfect agreement.
+
+# 7 Human Study for Measuring Empathy
+
+To demonstrate the empirical use of our taxonomy and how extracted narrative elements can be used to explore behavioral insights around narrative empathy, we conduct a large-scale user study presenting stories to different participants and asking them to rate their empathy towards the story. In this section, we discuss our study participants, the task procedure, and our data collection and measures used.
+
+# 7.1 Participants
+
+We recruited  $N = 2$ , 624 participants on Prolific<sup>5</sup> to read and rate empathy towards personal stories. An overview of participant demographics is shown in Appendix A. Participants were balanced by sex, predominantly white, and had high trait empathy on average.
+
+# 7.2 Study Procedure
+
+Our study procedure was determined exempt by our institution's ethics review board. At the beginning of the study, participants rated their current
+
+emotional state (arousal/valence), before reading a personal story. After reading the story, they were asked to rate their empathy towards the story, and to check which of the narrative elements within our taxonomy based on which elements contributed most to their emotional reaction towards the story. We asked a qualitative, open-ended question asking what aspects of the narrative's style made them relate to the story.
+
+After this, we asked participants to answer questions related to (1) narrative-reader interaction effects, which encompass reader factors that are tied to the process of reading the narrative (narrative transportation, prior experience with something that happened in the story, and perceived similarity to the narrator, and (2) reader characteristics (age, gender, ethnicity, trait empathy, how often they read for pleasure, fluent languages, and education level). Survey measurements and reasoning for selecting such measurements are detailed in the following section. All participants were paid $1 for answering the survey, and participants spent on average 7 minutes completing the entire task. Each of the 874 stories was rated at least 3 times by independent readers, resulting in 2,624 empathetic reactions to stories in total.
+
+# 7.3 Data Collection and Measures
+
+Our user study aims to capture empathy towards a diverse set of narratives with a diverse set of participants with varying reader characteristics in addition to variables that might moderate the effect of narrative style on empathy. Based on related empirical work exploring factors related to empathy (Figure 3), we designed the following surveys (all surveys are included in Appendix E for reproducibility). We make our dataset publicly available to open up deeper research in narrative empathy analysis.
+
+Empathy and Narrative Style Preferences We measure empathy towards the story through the State Empathy Scale (Shen, 2010). To gauge narrative style preferences, participants check off relevant elements from our taxonomy that they felt contributed to empathy towards the story. In addition, we ask for qualitative free-response feedback on what narrative style elements contributed to empathy towards the story.
+
+Narrative-Reader Interaction Effects We define effects at the intersection of reader characteristics and the experience of reading the narrative as narrative-reader interaction effects. These include
+
+![](images/7fb964c112823d77d713f32aa1d09e3c069c620a5ceed268da9e5fda49348a6c.jpg)  
+Figure 3: Visualization of how narrative style elements and reader characteristics influence the experience a reader has with a narrative (narrative-reader interaction effects). All of these components combined in turn influence downstream narrative empathy.
+
+(1) narrative transportation, measured by the Transportation Scale Short-Form / TS-SF (Appel et al., 2015; Walkington et al., 2020), (2) prior experience, measured by a Likert scale of how much the reader believes they have been in a similar situation as the narrator, and (3) perceived similarity to the narrator, measured by the Perceived Relational Diversity Scale (Clark, 2002). These features allow us to better understand the pathways via how narrative style elements interplay with narrative-reader interactions to lead to downstream empathy.
+
+Reader Characteristics We collect reader characteristics based on comprehensive literature review of properties that are related to empathy. These features include (1) the emotional state of the reader before reading the story, measured by the arousal/valence scale (Roshanaei et al., 2019), (2) basic demographic information including age, gender, ethnicity (Christov-Moore et al., 2014; Michal-ska et al., 2013; O'brien et al., 2013), (3) how often participants read for pleasure (Koopman, 2015; Mar et al., 2006), and (4) trait empathy, measured by the Single Item Trait Empathy Scale / SITES (Konrath et al., 2018) and the Toronto Empathy Questionnaire / TEQ (Spreng et al., 2009). Prolific automatically provides additional demographic information on participants such as fluent languages, nationality, and employment and student status.
+
+# 8 Empirical Insights on Narrative Empathy
+
+Next, we demonstrate the efficacy of our taxonomy in exploring empirical questions around empathy with a relevant subset of features from our dataset.
+
+Narrative Style Affects Empathy First, we aggregate empathy ratings for each story by taking
+
+![](images/b476eab3de4bc206f9db32cc412ef3071a2826337791a846acd8984faa30960a.jpg)  
+Figure 4: Structural equation modeling of how narrative style elements lead to narrative transportation, combined with effects of the reader sharing a similar experience with the narrator and the reader's baseline trait empathy.
+
+![](images/5ea7b8e495d4adf819dbc58fa06155c525afc898ad6d19c63855fc1aedc40ee5.jpg)  
+Figure 5: Comparing average empathy across high vs low presence of each narrative feature, we show that there are significant increases in empathy for stories with more character development and plot volume.
+
+the mean across the 3 raters. Then, we split stories into high vs. low presence of each narrative feature and apply Mann-Whitney u-tests to the averaged state empathy for the stories. Figure 5 shows that high aggregated empathy stories have more character development and plot volume. These results are statistically significant, after applying Benjamini-Hochberg correction to account for nine comparisons ( $p = 0.03$  for character development,  $p = 0.03$  for plot trajectory).
+
+Our work is, to the best of our knowledge, the first to empirically test the effect of character development and plot volume on narrative empathy. While some prior works (van Krieken et al., 2017) propose narrative features that relate to character identification, these are lower level than character development, such as the flatness/roundness or vulnerability of a character. Our findings regarding plot volume are in line with prior works that discuss how salient plot events can mark important moments in narratives that influence the emotional impact of the story (Sap et al., 2022). Prior works primarily from narrative studies use handcrafted features on smaller story sets (Fernandez-Quintanilla, 2020; Eekhof et al., 2023), but do not
+
+find significant effects of narrative features such as viewpoint and foregrounding. These studies focus primarily on literary texts rather than narratives that are more common online, and do not take into account other aspects of narrative style and narrative traits that are a part of our theorized taxonomy. These findings suggest future focused works, for example looking at how narrative style relates to empathy across narrative forms (literary vs. personal stories, spoken vs. textual, etc.)
+
+Narrative Empathy is not "One Size Fits All" While our previous analysis captures aggregated empathy, different people can have diverse emotional reactions to the same story. In Figure 6 (Appendix B), we show the standard deviations in state empathy scores for the same story, finding that on average this std. dev. is significantly greater than zero  $(p < 0.001)$ , indicating that the same narrative can evoke different levels of empathy. To address within-subject variance, we fit mixed effects models of empathy ratings using demographic groups, grouping individuals of similar Age, Sex, Trait Empathy, and Ethnicity and conditioning on multiple ratings for a single story. We find through a likelihood ratio test that empathy predicted by demographic group results in significantly better model fit  $(p = 0.002)$ . These two results indicate that there is high variance in empathy for the same story and that incorporating information regarding diverse demographic profiles can improve empathy model fit, aligning with prior works (August et al., 2020). Our findings have implications in broader empathy prediction tasks within NLP (Buechel et al., 2018; Sharma et al., 2020), which often optimize for a single objective empathy score assigned to a piece of text, aggregating empathy which can overlook individual factors.
+
+Vivid Emotional Expression of Narratives Leads to Narrative Empathy Given our finding that narrative empathy is not "one size fits all," we conduct analyses taking into account random effects for each story ID with structural equation modeling using the semopy library. Structural equation modeling (SEM) is a standard social science method for structured hypothesis testing and uses a formulation of generalized linear models to account for fixed and random effects when a theoretical model with relationships between elements is proposed.
+
+From our SEM results (Figure 4), we find that vividness of emotions significantly impacts narrative transportation, which in turn influences downstream empathy towards the story. The importance of vividness of emotions in personal stories is supported by other work in psychology. In particular, Pillemer (1992) elaborates that vivid descriptions of emotion in personal stories can convey believability in the experience, more readily evoking empathetic responses. While some computational works explore impact of narrative features on empathy (Roshanaei et al., 2019), they typically focus on positive/negative emotion words, rather than the narrative style or way in which emotions are conveyed through text, and may be better captured by current large-language models.
+
+Figure 4 shows how narrative features contribute to narrative transportation, leading to downstream empathy and taking into account non-stylistic factors like the reader sharing a similar experience as the narrator and the reader's trait empathy level. We find that both the narrator's previous experience with something happening in the story as well as their baseline trait empathy are significant predictors of empathy towards the story, but not as much as narrative transportation. In particular, our findings are in line with appraisal theory that suggests that feeling similar emotions is predicated on the target sharing similar experiences (Wondra and Ellsworth, 2015; Yang and Jurgens, 2024). While it is not particularly surprising that similar experience correlates with empathy, very few works have looked at narrative style interactions in tandem with fixed (trait empathy) and more dynamic traits (experiencing something similar), suggesting more holistic consideration of contextual factors related to narrative empathy.
+
+Narrative Style Preferences in Relation to Empathy are Personalized Finally, we show different demographic profiles might prefer different ways of telling a story, where preference is gauged by narrative empathy. Adding the interaction term TRAIT EMPATHY  $\times$  VIVIDNESS OF EMOTIONS to our structural model, we find a significant interaction effect of vivid emotions on the state empathy (est  $= 0.252$ ,  $p < 0.001$ ). This indicates that the relationship between vividness of emotions and state empathy increases as trait empathy increases, suggesting that narrative style preferences are personalized across demographic profiles.
+
+While certainly not exhaustive, our empirical analyses show how HEART can be used to yield interesting behavioral insights around how narrative style contributes to empathy. In particular, we note that looking at personalization in narrative empathy, as well as contextualizing reader factors such as their trait empathy level are important for empathy prediction, and are often overlooked in existing empathy tasks.
+
+# 9 Conclusion
+
+In this work, we quantify narrative style as it relates to narrative empathy. We introduce HEART, the first theory-driven taxonomy delineating elements of narrative style that can evoke empathy towards a story. We evaluate the performance of LLMs in extracting narrative elements from HEART, showing that prompting GPT-4 with our taxonomy leads to reasonable, human-level annotations beyond what prior lexicon-based methods can do, but that LLMs struggle in specific tasks, such as GPT-4's limited ability to extract expressions of cognition and evaluations. Through a crowdsourced study with over 2,000 participants, we demonstrate how HEART can be used to empirically understand the empathic role of narrative style factors. We find that vividness of emotions expressed, character development and plot volume are related to narrative empathy, and contextual factors such as a person's baseline trait empathy or sharing an experience with the narrator contribute to these effects. Additionally, we show that empathy responses are highly variable even in the same story, and that narrative style preferences are personalized to people with different demographic profiles (such as varying levels of trait empathy). Our findings show the promise of using LLMs for annotating complex story features that can yield interesting social and behavioral insights.
+
+# Limitations
+
+Narrative Style Annotation While most of the features in our taxonomy yielded reasonable consistency across human and LLM annotators, a few elements such as bodily perception and evaluations were less consistent. We excluded these features from our empirical analysis, but future work could make improvements to the annotation process for these specific elements. For example, our codebook makes use of Likert scale ratings for each of the narrative features within an entire story, but more granular annotations such as frequency of occurrences may have more consistency.
+
+Empirical Study Size and Reproducibility Findings in human behavior should be reproducible across different populations and contexts. While we conducted a large scale study with many participants, we did not ask participants to rate multiple stories. Additionally, the demographic distribution of Prolific crowdworkers is predominantly white. Future work should aim to reproduce our empirical insights with diverse populations and different types of stories.
+
+Statistical Modeling Our analysis methods involve interpretable statistical models commonly used in social science research. We chose to use structural equation modeling to gauge behavioral insights around how narrative style contributes to empathy, rather than achieving the best performance on narrative empathy prediction. Future work could improve upon narrative empathy prediction by incorporating narrative features in more complex transformer-based models and ablating different features.
+
+# Ethical Considerations
+
+Personal stories can contain intimate and vulnerable information, in addition to inducing emotions in readers. Our study protocol for showing sensitive stories to crowdworkers was approved by our institution's ethics review board as an exempt study. Participants gave informed consent that their survey ratings would be collected via Prolific. We ensured that all datasets we used were also collected via IRB-approved protocols, and will only distribute our dataset to IRB-approved protocols.
+
+More broadly, our work aims to advance research in narrative analysis as it relates to real-world human outcomes, such as empathy. Our
+
+findings corroborate that empathy is a highly personalized and contextualized experience. As such, in future work, we find that, rather than modeling the average person, it is important to value the rich diversity of human experiences.
+
+We recognize the ethical implications of modeling empathy in stories is double-edged. Empathy can be used in persuasion, marketing, or emotional manipulation. We encourage the findings from our work, and future work on narrative empathy analysis, to focus on improving human empathy for social good. For example, one could develop interactive tools to help a user convey a story more empathetically through understanding the role of narrative devices in reader empathy. Or one could use these insights to understand, at scale, social patterns behind storytelling, and how these might drive empathetic shifts online.
+
+# Acknowledgments
+
+We would like to thank all of our participants and teammates for their invaluable contributions to this project. Special thanks to Sue Holm for narrative annotation and Laura Vianna for study analysis guidance. This work was supported by an NSF GRFP under Grant No. 2141064 and partially funded by NSF grant No. 2230466.
+
+# References
+
+William Agnew, A. Stevie Bergman, Jennifer Chien, Mark Diaz, Seliem El-Sayed, Jaylen Pittman, Shakir Mohamed, and Kevin R. McKee. 2024. The illusion of artificial inclusion. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-12. ArXiv:2401.08572 [cs].  
+Maria Antoniak, David Mimno, and Karen Levy. 2019. Narrative Paths and Negotiation of Power in Birth Stories. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW):88:1-88:27.  
+Markus Appel, Timo Gnambs, Tobias Richter, and Melanie C. Green. 2015. The Transportation Scale-Short Form (TS-SF). Media Psychology, 18(2):243-266.  
+Tal August, Maarten Sap, Elizabeth Clark, Katharina Reinecke, and Noah A. Smith. 2020. Exploring the Effect of Author and Reader Identity in Online Story Writing: the STORIESINTHEWILD Corpus. In Proceedings of the First Joint Workshop on Narrative Understanding, Storylines, and Events, pages 46-54, Online. Association for Computational Linguistics.  
+Jiajun Bao, Junjie Wu, Yiming Zhang, Eshwar Chandrasekharan, and David Jurgens. 2021. Conversations Gone Alright: Quantifying and Predicting
+
+Prosocial Outcomes in Online Conversations. In Proceedings of the Web Conference 2021, pages 1134-1145, Ljubljana Slovenia. ACM.  
+Ryan L Boyd. 2022. The Development and Psychometric Properties of LIWC-22.  
+Sven Buechel, Anneke Buffone, Barry Slaff, Lyle Ungar, and João Sedoc. 2018. Modeling Empathy and Distress in Reaction to News Stories. arXiv preprint. ArXiv:1808.10399 [cs].  
+Leonardo Christov-Moore, Elizabeth A Simpson, Gino Coude, Kristina Grigaityte, Marco Iacoboni, and Pier Francesco Ferrari. 2014. Empathy: Gender effects in brain and behavior. *Neuroscience & biobehavioral reviews*, 46:604-627.  
+Mark Andrew Clark. 2002. Perceived relational diversity: A fit conceptualization. Ph.D., Arizona State University, United States - Arizona. ISBN: 9780493434216.  
+Amy Coplan. 2004. Empathic Engagement with Narrative Fictions. The Journal of Aesthetics and Art Criticism, 62(2):141-152. Publisher: [Wiley, American Society for Aesthetics].  
+Lynn S. Eekhof, Kobie van Krieken, José Sanders, and Roel M. Willems. 2023. Engagement with narrative characters: the role of social-cognitive abilities and linguistic viewpoint. Discourse Processes, 60(6):411-439. Publisher: Routledge_eprint: https://doi.org/10.1080/0163853X.2023.2206773.  
+Carolina Fernandez-Quintanilla. 2020. Textual and reader factors in narrative empathy: An empirical reader response study using focus groups. Language and Literature, 29(2):124-146. Publisher: SAGE Publications Ltd.  
+Carolina Fernandez-Quintanilla and Fransina Stradling. 2023. Introduction: stylistic approaches to narrative empathy. Journal of Literary Semantics, 52(2):103-121. Publisher: De Gruyter Mouton.  
+Franziska Hartung, Michael Burke, Peter Hagoort, and Roel M. Willems. 2016. Taking Perspective: Personal Pronouns Affect Experiential Aspects of Literary Reading. PLOS ONE, 11(5):e0154732. Publisher: Public Library of Science.  
+Arthur M. Jacobs. 2015. Neurocognitive poetics: methods and models for investigating the neuronal and cognitive-affective bases of literature reception. Frontiers in Human Neuroscience, 9. Publisher: Frontiers.  
+Suzanne Keen. 2006. A Theory of Narrative Empathy. Narrative, 14(3):207-236. Publisher: Ohio State University Press.  
+Suzanne Keen. 2014. Narrative Empathy. In *Narrative Empathy*, pages 521-530. De Gruyter.
+
+Sara Konrath, Brian P. Meier, and Brad J. Bushman. 2018. Development and validation of the single item trait empathy scale (SITES). Journal of Research in Personality, 73:111-122.  
+Eva Maria (Emy) Koopman. 2015. Empathic reactions after reading: The role of genre, personal factors and affective responses. Poetics, 50:62-79.  
+Eva Maria (Emy) Koopman and Frank Hakemulder. 2015. Effects of Literature on Empathy and Self-Reflection: A Theoretical-Empirical Framework. Journal of Literary Theory, 9(1):79-111. Publisher: De Gruyter.  
+Anežka Kuzmičová, Anne Mangen, Hildegunn Støle, and Anne Charlotte Begnum. 2017. Literature and readers' empathy: A qualitative text manipulation study. Language and Literature, 26(2):137-152. Publisher: SAGE Publications Ltd.  
+Anne Mangen, Anne Charlotte Begnum, Anežka Kuzmičová, Kersti Nilsson, Mette Steenberg, and Hildegunn Støle. 2018. Empathy and literary style. Orbis Litterarum, 73(6):471-486.  
+Raymond A. Mar, Keith Oatley, Jacob Hirsh, Jennifer dela Paz, and Jordan B. Peterson. 2006. Bookworms versus nerds: Exposure to fiction versus non-fiction, divergent associations with social ability, and the simulation of fictional social worlds. Journal of Research in Personality, 40(5):694-712.  
+Dan P. Mcadams. 2006. The Problem of Narrative Coherence. Journal of Constructivist Psychology. Publisher: Taylor & Francis Group.  
+Kalina J Michalska, Katherine D Kinzler, and Jean Decety. 2013. Age-related sex differences in explicit measures of empathy do not predict brain responses across childhood and adolescence. Developmental cognitive neuroscience, 3:22-32.  
+Sebastian Michelmann, Manoj Kumar, Kenneth A. Norman, and Mariya Toneva. 2023. Large language models can segment narrative events similarly to humans. arXiv preprint. ArXiv:2301.10297 [cs, q-bio].  
+Sylvia A. Morelli, Matthew D. Lieberman, and Jamil Zaki. 2015. The Emerging Study of Positive Empathy. Social and Personality Psychology Compass, 9(2):57-68.  
+Robin L. Nabi and Melanie C. Green. 2015. The Role of a Narrative's Emotional Flow in Promoting Persuasive Outcomes. Media Psychology, 18(2):137-162. Publisher: Taylor & Francis Ltd.  
+Ed O'brien, Sara H Konrath, Daniel Grühn, and Anna Linda Hagen. 2013. Empathic concern and perspective taking: Linear and quadratic effects of age across the adult life span. *Journals of Gerontology Series B: Psychological Sciences and Social Sciences*, 68(2):168-175.  
+James Pennebaker, Martha Francis, and Roger Booth. 1999. Linguistic inquiry and word count (LIWC).
+
+David B. Pillemer. 1992. Remembering personal circumstances: A functional analysis. In *Affect and accuracy in recall: Studies of "flashbulb" memories*, Emory symposia in cognition, 4., pages 236-264. Cambridge University Press, New York, NY, US.  
+Hannah Rashkin, Antoine Bosselut, Maarten Sap, Kevin Knight, and Yejin Choi. 2018. Modeling Naive Psychology of Characters in Simple Commonsense Stories. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2289-2299, Melbourne, Australia. Association for Computational Linguistics.  
+Andrew J. Reagan, Lewis Mitchell, Dilan Kiley, Christopher M. Danforth, and Peter Sheridan Dodds. 2016. The emotional arcs of stories are dominated by six basic shapes. *EPJ Data Science*, 5(1):1-12. Number: 1 Publisher: SpringerOpen.  
+Mahnaz Roshanaei, Christopher Tran, Sylvia Morelli, Cornelia Caragea, and Elena Zheleva. 2019. Paths to Empathy: Heterogeneous Effects of Reading Personal Stories Online. In 2019 IEEE International Conference on Data Science and Advanced Analytics (DSAA), pages 570-579, Washington, DC, USA. IEEE.  
+Belen Saldias and Deb Roy. 2020. Exploring aspects of similarity between spoken personal narratives by disentangling them into narrative clause types. arXiv preprint. Number: arXiv:2005.12762 arXiv:2005.12762 [cs].  
+Maarten Sap, Anna Jafarpour, Yejin Choi, Noah A. Smith, James W. Pennebaker, and Eric Horvitz. 2022. Quantifying the narrative flow of imagined versus autobiographical stories. Proceedings of the National Academy of Sciences, 119(45):e2211715119.  
+Maarten Sap, Marcella Cindy Prasettio, Ari Holtzman, Hannah Rashkin, and Yejin Choi. 2017. Connotation Frames of Power and Agency in Modern Films. page 6.  
+Shane Schweitzer and Adam Waytz. 2021. Language as a window into mind perception: How mental state language differentiates body and mind, human and nonhuman, and the self from others. Journal of Experimental Psychology: General, 150(8):1642-1672.  
+Ashish Sharma, Adam S. Miner, David C. Atkins, and Tim Althoff. 2020. A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support. arXiv preprint. ArXiv:2009.08441 [cs].  
+Jocelyn Shen, Yubin Kim, Mohit Hulse, Wazeer Zulfikar, Sharifa Alghowinem, Cynthia Breazeal, and Hae Won Park. 2024. Empathicstories++: A multimodal dataset for empathy towards personal experiences. In Findings of the 62nd Annual Meeting of the Association for Computational Linguistics.  
+Jocelyn Shen, Maarten Sap, Pedro Colon-Hernandez, Hae Park, and Cynthia Breazeal. 2023. Modeling
+
+Empathic Similarity in Personal Narratives. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6237-6252, Singapore. Association for Computational Linguistics.  
+Lijiang Shen. 2010. On a Scale of State Empathy During Message Processing. Western Journal of Communication, 74(5):504-524. Publisher: Routledge _eprint: https://doi.org/10.1080/10570314.2010.512278.  
+Micol Spitale, Sarah Okamoto, Mahima Gupta, Hao Xi, and Maja J Matarić. 2022. Socially Assistive Robots as Storytellers That Elicit Empathy. ACM Transactions on Human-Robot Interaction, page 3538409.  
+R. Nathan Spreng, Margaret C. McKinnon, Raymond A. Mar, and Brian Levine. 2009. The Toronto Empathy Questionnaire. Journal of personality assessment, 91(1):62-71.  
+Melanie Subbiah, Sean Zhang, Lydia B. Chilton, and Kathleen McKeown. 2024. Reading Subtext: Evaluating Large Language Models on Short Story Summarization with Writers. arXiv preprint. ArXiv:2403.01061 [cs].  
+Olivier Toubia, Jonah Berger, and Jehoshua Eliashberg. 2021. How quantifying the shape of stories predicts their success. Proceedings of the National Academy of Sciences, 118(26):e2011695118. Publisher: Proceedings of the National Academy of Sciences.  
+Kobie van Krieken, Hans Hoeken, and José Sanders. 2017. Evoking and Measuring Identification with Narrative Characters - A Linguistic Cues Framework. Frontiers in Psychology, 8. Publisher: Frontiers.  
+Tom van Laer, Ko de Ruyter, Luca M. Visconti, and Martin Wetzels. 2014. The Extended Transportation-Imagery Model: A Meta-Analysis of the Antecedents and Consequences of Consumers' Narrative Transportation. Journal of Consumer Research, 40(5):797-817.  
+Tom van Laer, Jennifer Edson Escalas, Stephan Ludwig, and Ellis A van den Hende. 2019. What Happens in Vegas Stays on TripAdvisor? A Theory and Technique to Understand Narrativity in Consumer Reviews. Journal of Consumer Research, 46(2):267-285.  
+Zoë Walkington, Stefanie Ashton Wigman, and David Bowles. 2020. The impact of narratives and transportation on empathic responding. Poetics, 80:101425.  
+Joshua D. Wondra and Phoebe C. Ellsworth. 2015. An appraisal theory of empathy and other vicarious emotional experiences. 122(3):411-428. 123 citations (Crossref) [2024-09-24].  
+Bo Xiao, Zac E. Imel, Panayiotis Georgiou, David C. Atkins, and Shrikanth S. Narayanan. 2016. Computational Analysis and Simulation of Empathic Behaviors: A Survey of Empathy Modeling with Behavioral
+
+Signal Processing Framework. Current psychiatry reports, 18(5):49.
+
+David B. Yaden, Salvatore Giorgi, Matthew Jordan, Anneke Buffone, Johannes C. Eichstaedt, H. Andrew Schwartz, Lyle Ungar, and Paul Bloom. 2024. Characterizing empathy and compassion using computational linguistic analysis. Emotion, 24(1):106-115.
+
+Jiamin Yang and David Jurgens. 2024. Modeling empathetic alignment in conversation. Preprint, arxiv:2405.00948 [cs]. 0 citations (Semantic Scholar/arXiv) [2024-09-24].
+
+Jamil Zaki. 2019. The war for kindness: Building empathy in a fractured world. Crown.
+
+Ke Zhou, Luca Maria Aiello, Sanja Scepanovic, Daniele Quercia, and Sara Konrath. 2021. The Language of Situational Empathy. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1):1-19.
+
+Lixing Zhu, Runcong Zhao, Lin Gui, and Yulan He. 2023. Are NLP Models Good at Tracing Thoughts: An Overview of Narrative Understanding. arXiv preprint. ArXiv:2310.18783 [cs].
+
+Caleb Ziems, William Held, Omar Shaikh, Jiao Chen, Zhehao Zhang, and Diyi Yang. 2024. Can Large Language Models Transform Computational Social Science? Computational Linguistics, 50(1):237-291.
+
+# A Participant Demographics
+
+Table 4: Participants' demographic breakdown.  
+
+<table><tr><td>Gender</td><td>Age</td><td>Ethnicity</td><td>Trait Empathy</td><td>Reading for pleasure</td></tr><tr><td>Female: 1329</td><td>43 ± 14</td><td>White : 2234</td><td>4.14 ± 0.88</td><td>3.45 ± 1.29</td></tr><tr><td>Male: 1295</td><td>min : 18</td><td>Asian : 150</td><td>min : 1</td><td>min : 1</td></tr><tr><td></td><td>max : 80</td><td>Black : 109</td><td>max : 5</td><td>max : 5</td></tr><tr><td></td><td></td><td>Mixed : 86</td><td></td><td></td></tr><tr><td></td><td></td><td>Other : 38</td><td></td><td></td></tr><tr><td></td><td></td><td>NA : 8</td><td></td><td></td></tr></table>
+
+# B Distribution of Empathy Standard Deviation
+
+# C Codebook and LLM Prompts
+
+# C.1 Character development
+
+We define character development in terms of changes that a character undergoes through the course of narrative events.
+
+We define changes broadly to include cognitive, emotional, behavioral, spiritual, moral, bodily, and social changes.
+
+Notably, we do not consider environmental changes for characters sufficient for character development, but acknowledge that other types of change (e.g. emotional, social) often accompany or are caused by environmental changes.
+
+![](images/72f98ce35c9f881696f14de80e370b15034ee4a1377d9dfa494537f4c94075b9.jpg)  
+Histogram of Standard Deviation in Empathy Ratings for the Same Story  
+Figure 6: Distribution of standard deviation in empathy scores for the same story indicate that empathy can differ drastically for the same story.
+
+Rate the narrator's character development based on the following scale:
+
+- 1 - no change  
+2 - limited change  
+3 - moderate change  
+4 - significant change  
+- 5 - life-altering, dramatic change
+
+# Examples:
+
+- I watched the birds splashing in the puddle from a bench at the park. They were so playful and content, even as it started to drizzle. (1 - character does not change)  
+- It wasn't until my brother told me what he's been going through that I realized how distant I had been. I broke down in front of him at the time. From that day forward, I decided to be there for my family, no matter what—even if that meant quitting my job and moving home. (5 - multiple dramatic changes)
+
+Respond with a single integer.
+
+Story: [STORY]
+
+# C.2 Character vulnerability
+
+Rate how emotionally vulnerable the narrator is in telling their story. We define vulnerability as how personal or intimate the information shared by the narrator is.
+
+Use the following scale:
+
+- 1 - not vulnerable at all  
+- 2 - somewhat vulnerable  
+3 - very vulnerable
+
+# Examples:
+
+- But I just doubt myself a lot. It's inevitable. (3 - the author reveals their self doubt)  
+- I went on a very memorable trip to Crater Lake Oregon on July 8th. (1 - does not share any sensitive information)
+
+Respond with a single integer.
+
+Story: [STORY]
+
+# C.3 Optimistic tone
+
+Rate the level of optimistic/pessimistic tone in the narrator's story. This should be the tone from the narrator's perspective, not of other characters in the story.
+
+-2 - very pessimistic  
+- 1 - somewhat pessimistic  
+0-neutral  
+1 - somewhat optimistic  
+2 - very optimistic
+
+# Examples:
+
+- I feel alone. It is so frustrating. I used to be fine with it, but then for some reason I actually started wanting to have a friend. I have nobody. And nobody around me seems interesting enough to me. Life gets boring. And frustrating. (rating = -2, very pessimistic)  
+- He is grown up and I have done my job to get him out into the world. I will miss his teenage years (somewhat), but I am proud of him. (rating = 2, very optimistic)
+
+Respond with a single integer.
+
+Story: [STORY]
+
+# C.4 Vivid emotions
+
+Rate the vividness of emotions described in the story. For example, vividness can be characterized by metaphor, simile, imagery, or strong language. Use the following scale:
+
+- 1 - not vivid at all
+
+- 2 - somewhat vivid  
+3 - very vivid
+
+# Examples:
+
+- I didn't feel great about the situation. (1)  
+- He was a hard-hitter in business, but outside of work he was completely different. (2)  
+- The pain of losing someone is like being stabbed in the chest. I was devastated when I lost her. (3)  
+- I was totally exhausted, tears running down my face (3).
+
+Respond with a single integer.
+
+Story: [STORY]
+
+# C.5 Expressions of cognition
+
+Rate how prominent descriptions of cognitive processes are in the story. We define descriptions of cognitive processes as statements that reveal the mental state or thinking pattern of the narrator.
+
+Use the following scale:
+
+- 1 - minimal or no cognitive processes  
+- 2 - moderate prominence of cognitive processes  
+3 - high prominence of cognitive processes
+
+# Examples:
+
+- I was born in the United States. (rating = 1)  
+- I wondered if I had seen him before. (rating = 2)  
+- I was thinking about how I could do it but I couldn't focus because I kept remembering what Sam said to me yesterday. (rating = 3)
+
+Respond with a single integer.
+
+Story: [STORY]
+
+# C.6 Temporal references
+
+Rate the extent to which the character focuses on the past (such as expressing nostalgia or reflections on memories) vs on the future (anticipation, looking forward) in the context of the story.
+
+Note that we are not asking whether the story is a past-tense, present-tense, or future-tense story. We are concerned with the orientation the narrator
+
+has toward the past, present, or future. We define 'extent' as the amount of narration time oriented toward the relative past, present, or future.
+
+Use the following scale:
+
+-2 - heavy focus on the past  
+- 1 - light focus on the past  
+- 0 - focus on the present  
+1 - light focus on the future  
+2 - heavy focus on the future
+
+# Examples:
+
+"I was stuck in bureaucratic processes for a year, and the whole time I was dreaming of the day my application processing was complete." (2)  
+"I went to the mall and saw a parade on my way." (0)  
+- "When I started taking my test, I regretted how little I had studied." (-1)
+
+Respond with a single integer
+
+Story: [STORY]
+
+# C.7 Plot volume
+
+Stories are structured by a sequence of events.
+
+We define plot trajectory as the amount and significance of events in the story.
+
+If the events are banal or insignificant and do not have a big impact on characters, then the plot trajectory is relatively small. If the events significantly impact characters or setting, then the story has a large plot trajectory.
+
+Rate the degree to which characters and setting are transformed through the course of the story based on the following scale:
+
+1 - no change  
+2 -trivial change  
+3 - moderate change  
+4 - significant change  
+- 5 - life-altering, dramatic change
+
+# Examples:
+
+- I stared out the window absent-mindedly for three hours. It was a lovely day. (1)
+
+- I heard a crash outside. I ran outside to see what had happened. It turned out the wind had blown over a box of garden tools. (3)  
+After a long and difficult pregnancy, I gave birth to a beautiful baby at 4:15pm. It was a crazy day at the hospital, but thanks to my family and the medical staff, we got through it! (5)
+
+Respond with a single integer.
+
+Story: [STORY]
+
+# C.8 Emotion shifts
+
+Most (but not all) emotions have either a positive (high) or negative (low) valence.
+
+For example, "anger" and "disgust" are low valence, whereas "happy" and "content" are high valence.
+
+Other emotions like "ambivalent" or "surprised" could be neutral, low, high, or ambiguous depending on the context.
+
+We consider 5 different types of emotional shifts that can occur in a story:
+
+- low-to-high valence (e.g. sad to happy)  
+- high-to-low valence (e.g. happy to sad)  
+- high-to-high valence (e.g. happy to hopeful)  
+- low-to-low valence (e.g. sad to angry)  
+- ambiguous-to-any valence (e.g. bittersweet to excited)
+
+We are interested in relatively straightforward emotion shifts that are either explicitly asserted in the text or easily inferrable based on information in the text. We are less interested in extremely subtle emotional shifts (e.g. joyful to content).
+
+Rate the degree of emotional shifts in the story below.
+
+Use the following scale:
+
+- 1 - no emotional shifts  
+- 2 - limited or trivial emotional shifts  
+3 - moderate emotional shifts  
+- 4 - significant emotional shifts  
+- 5 - life-altering, dramatic emotional shifts
+
+Example Story:
+
+"I went to college for 1 year before dropping out." (1)  
+"I was surprised to see my friend show up at the cafe where I was working" (2)  
+"I was frustrated with Ben for not inviting me, but when I ran into him a few weeks later, our conversation went fine." (3)  
+"I worked hard all semester and was mentally and physically exhausted by the end. It was such a relief to see my grades come in and see that all of my hard work paid off." (4)  
+"I was so excited to get out of class but before the bell rang the principal called me to his office. I was in trouble. I was stressed out of my mind walking to his office, but when I got there, he gave me the good news: I won the school-wide design contest!" (5)
+
+Respond with a single integer.
+
+Story: [STORY]
+
+# C.9 Resolution
+
+In the course of events and interactions between characters, stories introduce conflict. Stories also raise questions about the motives of characters, the meaning of events, and more. Conflict can be explicitly or implicitly referenced by narrators. Alternatively, the reader may subjectively perceive conflict in the situations described by narrators.
+
+Resolution refers to the extent to which conflict is addressed and questions are answered by the end of the story. There are many ways a story may be resolved, partially or completely. Resolution can occur for the narrator, characters within the story, or the reader.
+
+A story with low resolution may not have much conflict or leave conflict unaddressed by the end of the story. A story with high resolution will involve conflict that is addressed or raise questions that are ultimately answered.
+
+Rate the degree of resolution by the end of the story based on the following scale:
+
+1 - no resolution  
+2 - limited resolution  
+3 - moderate resolution  
+4 - significant resolution  
+- 5 - complete resolution
+
+# Examples:
+
+- I couldn't believe that he didn't apologize. How can someone just pretend that nothing happened? (1)  
+- I was homeless and finally found a new job, but I hate it and want to find a new one. (3)  
+- I looked for love my entire life, and had almost given up, when I met them. Now I couldn't be more in love. (5)
+
+Respond with a single integer. Do not include any words or punctuation marks in your answer.
+
+Story: [STORY]
+
+# C.10 Vividness of setting
+
+Rate the vividness of the setting described in the story. For example, vividness can be characterized by metaphor, simile, imagery, or strong language. Examples:
+
+I went to the restaurant to grab a bite to eat. (1)  
+- The sun cast warm rays onto the concrete in the park. (3)  
+- The waves in Palos Verdes crashed against the shore, making beautiful ribbons (3)  
+There was a house, with music playing in it (2)  
+- 1 - not vivid at all  
+- 2 - somewhat vivid  
+3 - very vivid
+
+Story: [STORY]
+
+# D GPT-4 Error Analysis Story Examples
+
+Scores range from 0 to 1, where 0 indicates low presence of narrative feature and 1 indicates high presence.
+
+# D.1 Stories with EVALUATIONS Disagreement
+
+Human: 0.0
+
+GPT-4: 1.0
+
+Yeah, so this is the beginning of the school year, and I've seen a lot of people moving into their dorms and apartments, and it really got me thinking about when I first went to College, when
+
+I was moving into the dorm. How excited my parents were for me and scared. And I was both excited and scared, moving away from home and my parents, and knowing that I'd probably get really homesick. Watching all those kids moving in really made me think about how that felt for me. It was really important for me, there was a lot of pressure for me to do well in College because my dad came here from a developing country and wasn't able to get an education past second grade. I was the first person in my family to go to College, and to make it that far. So I had a lot of emotions, definitely some anxiety, some stress about the pressure of performing and doing well. But also the excitement, and kind of the normal fear that you get doing something you've never done before and not having parents who had never experienced College like that before. I didn't really have anyone to go to, to understand what that meant, like what to expect. So watching those kids just brought me back to that moment.
+
+Human: 0.25
+
+GPT-4: 1.0
+
+When covid first hit I had to move from my city to my home town. Wasn't a huge deal as it was for others. It was a year and a half and a lot happened and eventually I came back to my original town, new fiancee and cat in tow. I couldn't find a job once I came back and when I did that place got shut down. My fiancee did have one, but any paycheck he had didn't go to our shared place, mostly his phone bill or groceries once in a while, while he stayed out till 2 am drinking. So I started. Excessively.
+
+I don't know now how I did, it was only a few months ago, how I managed to. I borrowed from family or friends, I took out loans to make sure rent was paid. It got so bad I was hospitalized for two weeks because for over a day I was throwing up every hour. I had torn my esophagus. No food. My fiance didn't visit, saying he was scared, but his aunt visited. I cried every night.
+
+When I got home he was working, he came back and was clearly drinking. The next day my mother texted me telling me she was disappointed in me drinking so much I got myself in that situation. I saw that and realized when I was in the hospital, alone, afraid, and not wanting to live like this. Thus me and my mother didn't speak for months.
+
+Family stuff happened, got back in touch. We both never apologized but there's a whole lot of
+
+cans there to be opened. I ruined relationships I can't take back. Lost contact with friends.
+
+After this stint my anxiety has been on high alert, making it hard for me to eat or even drink water. Public transit has been scary as I do have those impulse "what if I ran on the track" thoughts which I know are ridiculous.
+
+Thankfully now I do have a job I enjoy, I have my cat, my own place I pay rent.
+
+Human: 0.25
+
+GPT-4: 1.0
+
+About four months ago, my wife and I sold our first family home. We have a large family. It is my wife and I plus five children. Our oldest daughter started asking us about having her own room. Although we loved that house, we knew it was time to get something bigger.
+
+Luckily, we sold it after being on the market for only three days. We found a house with more bedrooms quickly, and the whole process was as smooth as it could have been.
+
+However, it is bittersweet looking back on everything. That house was very special to me. I did a lot of work on it. I saw my children grow and learn and love there. We made so many good memories. We charted how our children grew on a closet door (which is probably still there). It was a wonderful house while it lasted, but things happen and we had to let it go.
+
+As of today, I can still remember every little nook and cranny of that place. After all, it has only been a few months. However, it is sad to think these memories will eventually fade.
+
+I love our new house, but that first one will always hold a special place in my heart.
+
+Human: 0.0
+
+GPT-4: 1.0
+
+A couple months ago my younger brother got married. I traveled back to my home town of St. Louis, Missouri for the event. I took my girlfriend along with me. It was her first trip to my hometown ever. The trip started out great, we picked up our rental car and went to grab a pizza.
+
+The following day was my brother's wedding ceremony. We thought it had all been planned out thoroughly, but it turns out that nobody had checked the weather report. The day of the wedding came, it started out sunny, a hot day in late
+
+July. Clear blue skies and not a cloud to be seen. We were all so optimistic about the big day. The ceremony was scheduled for 6 PM at sunset, so it would start to cool down and allow for some relieve from the heat of the day. In theory, that was a great idea.
+
+However, when my girlfriend and I pulled out of the driveway we noticed something new that we hadn't seen yet on the trip. Storm clouds moving in fast, and lots of them. Dark gray giants rose onto the horizon at a frightening pace. Lightning was visible in the distance as we began our drive to the wedding venue. We hoped and prayed that the storm would blow the other way, and that the outdoor wedding venue would be spared from this particular storm. Would we be able to get away with having the ceremony in decent weather?
+
+It became a race with time. As we drove to the wedding ceremony, it felt as though the clouds were following us and growing larger. As we arrived, I greeted my brother in the parking lot and asked if he thought it would rain. He said maybe, it depends how fast we can get this done. Everyone was present except for the minister, one of the few people who was completely essential to the process.
+
+As the minister arrived, it finally began to rain. It was raining on my brother's wedding day, I couldn't believe it, but luckily the ceremony was completed and we had a wonderful sunny reception the next day.
+
+Human: 0.25
+
+GPT-4: 1.0
+
+I was hiking near Lake Ontario with my partner and our two grandkids. It was a beautiful sunny day. The lake sparkled brilliantly. I have a bad knee, so I was struggling with some of the physical activity. My partner suggested that I rest a bit on a fallen log. I was nervous, because I would not be able to get up by myself, but I agreed.
+
+My partner and my granddaughter wanted to hike further to see the bluffs. I said I would be OK for a bit, but my sweet grandson insisted on staying with me. He said, "I won't let my granny sit in the forest all alone."
+
+Well it was a good thing. My partner and granddaughter didn't return in a reasonable amount of time. We got very nervous! My grandson is 10 years old, but not strong enough to help me up. He searched for a stout walking stick and found one
+
+nearby. I used it to prop myself up, and managed to get my feet underneath me.
+
+Together we went down the shore to find the rest of our party. Luckily all was well, but I would truly have been distraught if I had been all alone waiting for so long!
+
+# D.2 Stories with COGNITION Disagreement
+
+Human: 0.25
+
+GPT-4: 1.0
+
+I've been hearing a lot of people saying that MIT students aren't successful as Stanford or Harvard students because there aren't as many well-known MIT CEOs. It seems rather unfair of them to say that because MIT students have contributed a lot to this world from Nobel-prize winning theorems to groundbreaking algorithms.
+
+Also there are lot of MIT grads like David Siegel who went on to found great companies that don't necessarily have a face to the brand like Jobs' Apple or Zuckerberg's Facebook. And on top of that, there many of MIT grads who go on to be CTOs or other types of product managers (sorry for the emphasis on course 6), and without them, the companies would not be the same.
+
+Above all, out of the "top" institutions, MIT does the most to help lower-income students attain social and economic mobility (I remember reading an article, but can't find the link).
+
+This is not to say that MIT doesn't have problems, but at the end of the day, I wish people didn't equate fame/status with success. You don't have to be a famous CEO or a CEO in general to be successful. And I'm sure a lot of the people I mentioned didn't end up becoming crazy famous because they value privacy, which is fine!
+
+And a lot of alumns end up doing what they're interested in regardless of status, which is amazing and also indicative of success! Success can look different for people.
+
+Human: 0.25
+
+GPT-4: 1.0
+
+When covid first hit I had to move from my city to my home town. Wasn't a huge deal as it was for others. It was a year and a half and a lot happened and eventually I came back to my original town, new fiancee and cat in tow. I couldn't find a job once I came back and when I did that place got shut down. My fiancee did have one, but any paycheck he
+
+had didn't go to our shared place, mostly his phone bill or groceries once in a while, while he stayed out till 2 am drinking. So I started. Excessively.
+
+I don't know now how I did, it was only a few months ago, how I managed to. I borrowed from family or friends, I took out loans to make sure rent was paid. It got so bad I was hospitalized for two weeks because for over a day I was throwing up every hour. I had torn my esophagus. No food. My fiance didn't visit, saying he was scared, but his aunt visited. I cried every night.
+
+When I got home he was working, he came back and was clearly drinking. The next day my mother texted me telling me she was disappointed in me drinking so much I got myself in that situation. I saw that and realized when I was in the hospital, alone, afraid, and not wanting to live like this. Thus me and my mother didn't speak for months.
+
+Family stuff happened, got back in touch. We both never apologized but there's a whole lot of cans there to be opened. I ruined relationships I can't take back. Lost contact with friends.
+
+After this stint my anxiety has been on high alert, making it hard for me to eat or even drink water. Public transit has been scary as I do have those impulse "what if I ran on the track" thoughts which I know are ridiculous.
+
+Thankfully now I do have a job I enjoy, I have my cat, my own place I pay rent.
+
+Human: 0.25
+
+GPT-4: 1.0
+
+Today was one of the saddest days of my life. It started early in the day, and my parents came by and picked me up at my house. Everyone was in a very somber mood, but it was sunny and quite warm. We drove out to a church about thirty minutes away near where my mom grew up, and while driving I couldn't help but think back to all the good memories I had with my cousin. She was always so happy and nice and just fun to be around. But now, that was all gone, and all I had were the memories that were going over in my mind.
+
+Arriving at the church and seeing all of my family, it was hard. It was just so sad, all of it. Seeing my aunt was the hardest part I think, but I knew then that she was strong and was going to be able to get past this. My uncle is an ordained pastor, so he was able to help with the service and I think that helped ease some of the pain.
+
+After the service we all went to the cemetery
+
+and gathered up on the hill in the shade. Seeing the final resting place really hit me hard, I started to cry much harder than I had been all day at that point. All of the memories and the final shock to my brain that she was never coming back, made me very sad, and made me miss her dearly.
+
+We then all met at a local place where they served a late lunch and we had some drinks. It was good to see so many of my family, but at the same time, so sad, because I thought that we shouldn't be seeing each other, at least not for this reason.
+
+I didn't really know how to feel when we left and I made it back home. I was deeply saddened, and just thought of how my aunt, uncle, and cousins felt. I know that life had changed for them forever, and now life was starting again without their dear one, and that hurt me again. But my family is strong, and stronger together, and I know we will get through this like we will any other tragedy that comes our way.
+
+# E Surveys
+
+# E.1 Empathy and Narrative Style Preferences
+
+State Empathy Scale (Shen, 2010)
+
+Please indicate the level to which you agree with each of the following statements - Strongly disagree (1) to Strongly agree (5)
+
+1. The narrator's emotions are genuine.  
+2. I experienced the same emotions as the narrator while reading this story.  
+3. I was in a similar emotional state as the narrator when reading this story.  
+4. I can feel the narrator's emotions.  
+5. I can see the narrator's point of view.  
+6. I recognize the narrator's situation.  
+7. I can understand what the narrator was going through in the story.  
+8. The narrator's reactions to the situation are understandable.  
+9. When reading the story, I was fully absorbed.  
+10. I can relate to what the narrator was going through in the story.  
+11. I can identify with the situation described in the story.
+
+12. I can identify with the narrator in the story.
+
+# Narrative Style Preferences
+
+Check the aspects of narrative style (the way the story was told) that made you resonate with the story.
+
+1. Flatness/roundness of the character (the character shows development/is vulnerable/subverts expectations)  
+2. References to the past (nostalgia) or the future  
+3. The vividness of emotions described in the story  
+4. The way thoughts/cognition of the character are expressed  
+5. The way moral judgments of the character are expressed  
+6. The way the character's perception and physical sensations are expressed  
+7. The way the character's actions are expressed  
+8. The way the setting of the story is expressed  
+9. The way the character's point of view is expressed  
+10. The overall plot trajectory of the story  
+11. The presence of a resolution in the story  
+12. The flow and readability of the story  
+13. The overall shifts in the emotional tone of the story
+
+[FREE RESPONSE] What about the narrative style (the way the story was told) made you resonate with it (if any)?
+
+# E.2 Narrative-Reader Interaction
+
+Transportation Scale Short-Form / TS-SF (Appel et al., 2015)
+
+Rate the extent to which you agree with the following statements - Not at all (1) to Very much (7)
+
+1. I could picture myself in the scene of the events described in the narrative.  
+2. I was mentally involved in the narrative while reading it.
+
+3. I wanted to learn how the narrative ended.  
+4. The narrative affected me emotionally.  
+5. While reading the narrative I had a vivid image of the narrator.
+
+# Similar Experience
+
+I have experienced a similar situation as the narrator in my life before – Strongly disagree (1) to Strongly agree (5)
+
+Similar to Narrator (Clark, 2002) I am similar to the narrator – Strongly disagree (1) to Strongly agree (5)
+
+Rate how similar you believe you are to the narrator in terms of the following characteristics - Not similar at all (1) to Highly similar (5)
+
+1. Age  
+2. Race/ethnicity  
+3. Sex  
+4. Religion  
+5. Sexual orientation  
+6. Socio-economic status  
+7. Geographic origin
+
+# E.3 Reader Characteristics
+
+Education Level
+
+What is the highest level of education you have completed?
+
+1. Some high school or less  
+2. High school diploma or GED  
+3. Some college, but no degree  
+4. Associates or technical degree  
+5. Bachelor's degree  
+6. Graduate or professional degree (MA, MS, PhD, JD, MD, DDS etc.)  
+7. Prefer not to say
+
+# Reading for pleasure
+
+How often do you read for pleasure?
+
+1. Almost never  
+2. A couple of times a year
+
+3. A couple of times a month  
+4. At least once a week  
+5. Once or more a day
+
+Trait Empathy (Konrath et al., 2018; Spreng et al., 2009)
+
+To what extent does the following statement describe you: "I am an empathetic person" - Strong disagree (1) to Strongly agree (5)
+
+Below is a list of statements. Please read each statement carefully and rate how frequently you feel or act in the manner described – Never (1) to Always (5)
+
+1. When someone else is feeling excited, I tend to get excited too  
+2. Other people's misfortunes do not disturb me a great deal  
+3. It upsets me to see someone being treated disrespectfully  
+4. I remain unaffected when someone close to me is happy  
+5. I enjoy making other people feel better  
+6. I have tender, concerned feelings for people less fortunate than me  
+7. When a friend starts to talk about his/her problems, I try to steer the conversation towards something else  
+8. I can tell when others are sad even when they do not say anything  
+9. I find that I am "in tune" with other people's moods  
+10. I do not feel sympathy for people who cause their own serious illnesses  
+11. I become irritated when someone cries  
+12. I am not really interested in how other people feel  
+13. I get a strong urge to help when I see someone who is upset  
+14. When I see someone being treated unfairly, I do not feel very much pity for them  
+15. I find it silly for people to cry out of happiness  
+16. When I see someone being taken advantage of, I feel kind of protective towards him

2024/HEART-felt Narratives_ Tracing Empathy and Narrative Style in Personal Stories with LLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:510a516bfc3c483925c365c16804802b23eb7e6a54d2a86f1027e8dd8bac6ee0
+size 283150

2024/HELPD_ Mitigating Hallucination of LVLMs by Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding/e64b5f27-e4c8-415c-9fbb-28824e138b14_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7edc5fd6c1d3d2ffe05e533a6aca60d68b615b930d9952fbbdfbff372df0d725
+size 3001919

2024/HELPD_ Mitigating Hallucination of LVLMs by Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding/full.md

@@ -0,0 +1,523 @@
+# HELPD: Mitigating Hallucination of LVLMs by Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding
+
+Fan Yuan $^{1,2}$ , Chi Qin $^{1,2}$ , Xiaogang Xu $^{3}$ , Piji Li $^{1,2*}$
+
+<sup>1</sup> College of Artificial Intelligence,
+
+Nanjing University of Aeronautics and Astronautics, Nanjing, China
+
+$^{2}$  MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, Nanjing, China
+
+<sup>3</sup> The Chinese University of Hong Kong, Hong Kong, China
+
+{fanyuan, qinchi, pjli}@nuaa.edu.cn, xiaogangxu00@gmail.com
+
+# Abstract
+
+Large Vision-Language Models (LVLMs) have shown remarkable performance on many visual-language tasks. However, these models still suffer from multimodal hallucination, which means the generation of objects or content that violates the images. Many existing work detects hallucination by directly judging whether an object exists in an image, overlooking the association between the object and semantics. To address this issue, we propose Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding (HELPD). This framework incorporates hallucination feedback at both object and sentence semantic levels. Remarkably, even with a marginal degree of training, this approach can alleviate over  $15\%$  of hallucination. Simultaneously, HELPD penalizes the output logits according to the image attention window to avoid being overly affected by generated text. HELPD can be seamlessly integrated with any LVLMs. Our experiments demonstrate that the proposed framework yields favorable results across multiple hallucination benchmarks. It effectively mitigates hallucination for different LVLMs and concurrently improves their text generation quality.
+
+# 1 Introduction
+
+Large Language Models (LLMs) (Brown et al., 2020; OpenAI, 2023; Touvron et al., 2023a,b), guided by human instruction, have demonstrated impressive performance in numerous Natural Language Processing (NLP) tasks (Qin et al., 2023). In light of the success of LLMs, researchers aspire to integrate the powerful capabilities of LLMs into multimodal domains, consequently introducing Large Vision-Language Models (LVLMs) (Li et al., 2023a; Dai et al., 2023; Zhu et al., 2023; Ye et al., 2023; Liu et al., 2023b). Particularly,
+
+![](images/6c3397f13c72615201d0a7c2f23da52bf89203f9b10a8c40e93aadb11d14d929.jpg)  
+Describe the content of the image in detail.
+
+![](images/0a40553261d8d961c8ad688c070f519989e6448cc342bcac7dbe4e549b10defd.jpg)
+
+![](images/443c12999db062f39179a8a25aee50041091faff02116b0859681997146f6e3c.jpg)  
+LVLM  
+Figure 1: A case of LVLM hallucination. The parts marked in red are, in fact, hallucinations. The parts marked in blue would be mistaken for hallucinations by detection methods that focus only on objects.
+
+The image features a squirrel jumping in the air, leaping over snow-covered ground. It is likely that this action was taken to avoid dangerous predators or catch food hidden beneath it. This action takes place during winter, with new snowflakes covering everything around, except for a few spots of grass and trees that can be seen on both sides. This scene happens after the squirrel's jump across an open area, where everything, including roofs, walls, and fences, is covered by thick layers of snow.
+
+GPT-4 (OpenAI, 2023) has been endowed with the capability to engage in complex, image-based dialogues with humans, while also being proficient in resolving a series of visual-language tasks.
+
+Despite the fact that LVLMs have achieved quite considerable results on various tasks, problems with these models have gradually emerged. Within these problems, the hallucination (Rohrbach et al., 2018; Li et al., 2023c) has attracted significant attention. This is a phenomenon that LVLMs tend to generate content contradictory to the image, such as non-existent objects. In order to alleviate this phenomenon, many explorations have been carried out in recent work (Li et al., 2023c; Wang et al., 2023; Liu et al., 2023a; Zhou et al., 2023; Zhai et al., 2023; Lee et al., 2023; Huang et al., 2023; Wang et al., 2024). CoVe (Dhuliawala et al.,
+
+![](images/84b3a0d67c9d41e352d59e45f1752252596d6736dea2e7912928baedcc1e6450.jpg)  
+(a) InstructBLIP
+
+![](images/55064d03fa0f86c68b8e50b2b466d365c663decd97385e6b0f3b7b5f3076cbb6.jpg)
+
+![](images/e2a778eedf7ea8432e3159904e1fe447b07ed80ce218a7d67e6a98c582d7c2df.jpg)  
+(b) mPLUG-Owl  
+(c) MiniGPT-4  
+Figure 2: Attention visualization of LVLMs. For the same input, each image represents the attention matrix of a specific LVLM generation instance. Red indicates the attention of the image, while green represents the phenomenon of "Over-trust" in the generated text.
+
+2023) proposes a Chain-of-Verification method, it first generates verification questions, then executes them to check for hallucination, and finally gets a revised response. Additionally, some approaches aim to alleviate the hallucination from the decoding strategy (O'Brien and Lewis, 2023; Chuang et al., 2023; Huang et al., 2023). Dola decoding (Chuang et al., 2023) is a strategy of contrasting the mature layer and the immature layer of the model, followed by the determination of the next token based on the differences in logits. Opera (Huang et al., 2023) is a recently proposed decoding method that employs an Over-trust Penalty to determine the occurrence of hallucination, and utilizes a Retrospection-Allocation rollback mechanism for decoding.
+
+However, most of the existing work focuses on alleviating object-level hallucinations. During this process, some methods excessively concentrate on whether the generated objects exist in the image, neglecting the association between these objects and the semantics of the whole sentence. As illustrated in Figure 1, words marked in red, such as "trees", are real object hallucinations. Solely considering the presence of objects, the parts marked in blue, such as "predators" and "food", would also be defined as hallucinations. Nevertheless, combined with context and semantics, such a definition is deemed inappropriate. Meanwhile, as indicated by the green box in Figure 2, "Over-trust" (Huang et al., 2023) does exist in LVLMs, which means certain generated tokens receive excessive attention, leading to a subsequent generation that deviates from the image. We initially assumed that insuffi
+
+cient focus on the visual part of the input might be one cause of this phenomenon. However, further observation of the attention matrix of these models reveals strong focus on the visual input (see red boxes in Figure 2). This indicates that considering the over-trust penalty only accounts for the impact of the text, additional focus on the image is therefore required to balance it.
+
+Based on the observations above, we propose HELPD, a novel LVLM framework that utilizes Hierarchical FeEdback Learning with Vision-enhanced Penalty Decoding. We note the necessity of integrating both the inherent properties of an object and semantic meaning to determine the presence of hallucination. Thus, we propose the hierarchical feedback learning, which only requires a small amount of training, and we add this feedback mechanism at the end of the training period. On the one hand, the collection of objects is extracted from the sampled sentences and label sentences, and the object-level feedback is obtained through the comparison of the object sets. On the other hand, leveraging the powerful few-shot inference capabilities of GPT-4, we conduct a semantic comparison to obtain sentence-level feedback.
+
+The manner of sampling constitutes a crucial component in the hierarchical feedback learning process. Opera decoding (Huang et al., 2023) predicts the next word by subtracting the over-trust penalty score from the logits, where the penalty score is computed based on the attention window of the generated text, disregarding the potent influence of visual attention. Consequently, we propose the Vision-Enhanced Penalty Decoding, which incorporates visual attention into the penalty score computation and makes the final logits place more emphasis on the image input. This approach effectively mitigates an over-reliance on the textual modality during the decoding process, and enhances the influence of visual modality, thereby alleviating the hallucination.
+
+Our contributions can be summarized as follows:
+
+- We propose a hierarchical feedback learning method, incorporating object-level and sentence-level hallucination feedback. It can mitigate the occurrence of hallucinations with only a minimal amount of training.  
+- With the analysis of the attention matrix during decoding, we introduce the vision-enhanced penalty decoding to enhance the influence of images on the generation process.
+
+- Extensive experimental results indicate that our proposed framework shows better performance on multiple hallucination metrics and can effectively alleviate the hallucination of LVLMs.
+
+# 2 Related Work
+
+# 2.1 Large Vision-Language Models (LVLMs)
+
+Owing to the success of Large Language Models (LLMs) (Brown et al., 2020; OpenAI, 2023; Touvron et al., 2023a,b; Chung et al., 2022; Zeng et al., 2023; Sun et al., 2021; Yang et al., 2023) in many Natural Language Processing (NLP) tasks (Qin et al., 2023), many researchers have endowed it with multimodal perception capabilities. Among these, Large Vision-Language Models (LVLMs) (Li et al., 2023a; Dai et al., 2023; Zhu et al., 2023; Ye et al., 2023; Liu et al., 2023b; OpenAI, 2023; Peng et al., 2023; Anil et al., 2023) have shown particularly notable performance.
+
+LVLMs primarily consist of three components: a visual encoder, a modality alignment module, and a Large Language Model (LLM) (Yin et al., 2023; Zhang et al., 2024). Visual encoders include Vision Transformers (ViT) (Dosovitskiy et al., 2021), CLIP ViT (Radford et al., 2021), and others (Brock et al., 2021; Fang et al., 2023). Specifically, ViT splits images into patches, which are then input into Transformer blocks through linear mapping for feature learning. Since there exists a modality gap between the visual encoders and the LLMs, the modal alignment module is required as a bridge. Models such as Flamingo (Alayrac et al., 2022), BLIP-2 (Li et al., 2023a), and InstructBLIP (Dai et al., 2023) apply the Q-former, a method that extracts visual features in a query-based manner by employing a set of learnable vectors. Another more direct method involves using a linear interface for modality alignment. For instance, LLaVA (Liu et al., 2023b) employs a linear layer to map images to the textual embedding space.
+
+# 2.2 Hallucination in LVLMs
+
+Multimodal hallucination is a significant challenge faced by LVLMs, severely impairing the reliability and robustness of these models. It typically manifests as generating content that is inconsistent with the image or contradicts common sense. Generally, hallucinations can be divided into Intrinsic Hallucination and Extrinsic Hallucination. Intrinsic hallucination refers to the generation of content
+
+that conflicts with the input. On the other hand, extrinsic hallucination represents the generation of additional content that does not actually exist, such as objects not present in the image.
+
+Recently, numerous efforts have been dedicated to the elimination of multimodal hallucination (Li et al., 2023c; Wang et al., 2023; Liu et al., 2023a; Zhou et al., 2023; Zhai et al., 2023; Lee et al., 2023; Huang et al., 2023; Wang et al., 2024; Gunjal et al., 2024; Zhao et al., 2023). CHAIR (Rohrbach et al., 2018) is an early proposed metric for evaluating object hallucinations in image captioning tasks. It assesses the degree of hallucination by calculating the proportion of objects that appear in the generated descriptions but not in the image itself. POPE (Li et al., 2023c) introduces a polling-based object probing evaluation method, which assesses the degree of hallucination based on the responses to questions like "Is there a <object> in the image?" that are posed based on the objects. CoVe (Dhuliawala et al., 2023) introduces a chain-of-verification that considers its own responses and self-corrects hallucination. Liu et al. (2023a) conducts visual instruction tuning on LVLMs with the newly proposed LRV-Instruction dataset to mitigate hallucinations.
+
+# 3 Method
+
+# 3.1 Hierarchical Feedback Learning
+
+As we have illustrated in Section 1, to determine the occurrence of hallucination, it is necessary to consider not only whether the mentioned object appears in the image, but also to judge whether it is a reasonable association in combination with semantics. To address the aforementioned issue, we propose Hierarchical Feedback Learning, a learning method that enhances the model's intrinsic ability to avoid hallucination through different granularity hallucination detection feedback (see Figure 3).
+
+In practice, we conduct minimal further training on the LVLMs and incorporate this feedback-learning mechanism towards the end of the training process. More specifically, after every fixed number of training steps, we sample the logits of the model's output to obtain the actions (which means the sampled tokens)  $A = \{a_{ij}\}_{j=1}^{t}, \quad i = 1, \ldots, b$ , where  $t$  is the length of the sampled sentence and  $b$  is batch size. With the help of NLTK $^2$  and GPT-4, we extract objects from the sampled sentence and label sentence, respectively, obtaining the sampled
+
+![](images/cd99efeb65df510d134445643e416c75d57d779c442fe0d14c01da278c240dc3.jpg)  
+Figure 3: This diagram illustrates the framework of HELPD. The Hierarchical Feedback Learning detects hallucination by obtaining object-level feedback from comparing object sets extracted from sampled and label sentences, and sentence-level feedback through semantic comparison using GPT-4's few-shot inference capabilities. To improve the effectiveness of sampling, the Vision Penalty Decoding augments the over-trust penalty score with a vision-enhanced penalty score, making the final logits closer to the image.
+
+object set  $S_{sam} = \{obj_1, obj_2, \ldots, obj_m\}$  and the label object set  $S_{lab} = \{obj_1, obj_2, \ldots, obj_n\}$ , where  $m$  and  $n$  represent the number of objects. Subsequently, we calculate the F1 score of these two sets as the object-level feedback scores  $R_{obj}$ :
+
+$$
+\text {P r e c i s i o n} = \frac {\left| S _ {s a m} \cap S _ {l a b} \right|}{\left| S _ {s a m} \cap S _ {l a b} \right| + \left| S _ {s a m} \setminus S _ {l a b} \right|} \tag {1}
+$$
+
+$$
+\text {R e c a l l} = \frac {\left| S _ {s a m} \cap S _ {l a b} \right|}{\left| S _ {s a m} \cap S _ {l a b} \right| + \left| S _ {l a b} \setminus S _ {s a m} \right|} \tag {2}
+$$
+
+$$
+R _ {o b j} = 2 * \frac {\text {P r e c i s i o n} * \text {R e c a l l}}{\text {P r e c i s i o n} + \text {R e c a l l}} \tag {3}
+$$
+
+Sentence-level feedback is obtained through the few-shot inference capability of GPT-4. We provide a detailed evaluation method and pre-announce several sentence pairs as context (see Appendix A for detail) to instruct it on discerning hallucination from semantics. The score ranging from 0 to 1, returned by GPT-4, is defined as  $R_{sen}$ .
+
+Given that  $R_{sen}$  and  $R_{obj}$  are non-differentiable, they cannot be directly incorporated into training using the gradient method. Inspired by (Sutton et al., 1999), we introduce the reinforce algorithm to handle this problem. Specifically, based on the tokens sampled, we first retrieve their corresponding log probabilities from the original logits:
+
+$$
+P _ {i, j} = \log \left(\frac {e ^ {\text {l o g i t s} _ {i , j , A _ {i , j}}}}{\sum_ {k = 1} ^ {V} e ^ {\text {l o g i t s} _ {i , j , k}}}\right), \tag {4}
+$$
+
+where  $i$  is the index within the batch,  $j$  is the index within the sequence length,  $A_{i,j}$  is the corresponding sampled action, and  $k$  is the index within the vocabulary of size  $V$ . A hyperparameter  $\sigma$  is set to determine the relative importance of the two types of feedback mentioned above (see Equation (5)). By summing the product of the feedback and the corresponding log probabilities of the actions, we obtain a negative weighted log-likelihood loss. To prevent the loss from infinitely increasing with the number of actions, the total loss is divided by the number of sampled actions to yield the loss function for reinforce algorithm, denoted as  $\mathcal{L}_{\mathrm{RL}}$ :
+
+$$
+R _ {i} = \sigma R _ {s e n, i} + (1 - \sigma) R _ {o b j, i}, \tag {5}
+$$
+
+$$
+\mathcal {L} _ {\mathrm {R L}} = - \frac {1}{N} \sum_ {i = 1} ^ {b} \sum_ {j = 1} ^ {t} P _ {i j} \cdot R _ {i}. \tag {6}
+$$
+
+In the early stage of training, we employ cross-entropy loss  $\mathcal{L}_{\mathrm{CE}}$ . When the training step reaches  $c *$  total steps,  $\mathcal{L}_{\mathrm{RL}}$  is added to the total loss. The total loss is defined as:
+
+$$
+\mathcal {L} _ {\mathrm {C E}} = - \sum_ {h} ^ {H} \log P \left(x _ {h} \mid x _ {<   h}\right), \tag {7}
+$$
+
+$$
+\mathcal {L} = \left\{ \begin{array}{l l} \mathcal {L} _ {\mathrm {C E}}, & \text {i f s t e p s <   c * t o t a l s t e p s ,} \\ \frac {\mathcal {L} _ {\mathrm {C E}}}{\| \mathcal {L} _ {\mathrm {C E}} \|} + \frac {\mathcal {L} _ {\mathrm {R L}}}{\| \mathcal {L} _ {\mathrm {R L}} \|}, & \text {o t h e r w i s e ,} \end{array} \right. \tag {8}
+$$
+
+where  $x$  is the generated token.
+
+![](images/6b6359e50738073747df30e2e1b4277fca22effe6deafe6d9656b98edef352d7.jpg)  
+Figure 4: The illustration of Vision-enhanced Penalty Decoding. The total penalty is composed of the vision penalty and the over-trust penalty. The over-trust penalty is computed based on the generated text (the upper region), while the vision penalty is computed from the vision attention window (the lower area).
+
+# 3.2 Vision-enhanced Penalty Decoding
+
+With the hierarchical feedback learning, we can effectively detect object-level and sentence-level hallucinations and correct them through back propagation. To obtain the sampled sentences, we need to sample each token from the model's logits. Based on our analysis of the attention matrix, we propose the Vision-enhanced Penalty Decoding based on Opera (Huang et al., 2023).
+
+Over-Trust Logit Penalty. First, we provide a brief description of the original process. It sets a local window of length  $h$  for the attention matrix, where  $h$  represents the length of the current generated sequence. Upon obtaining such a lower triangular matrix, it pads the upper triangular part with zeros and scales up the values to avoid excessively small values. Subsequently, it conducts columnwise multiplication on this matrix and select the maximum value of the vector as the over-trust penalty  $\phi(\omega_{\leq h})$ . Finally, it subtracts this penalty from the original logits to predict the next token.
+
+Vision-enhanced Penalty Decoding. As illustrated in Section 1, we note that the over-trust penalty establishes a local window whose size is limited to the length of the generated text. This approach can effectively extract over-trust patterns from past tokens, but it inadvertently amplifies the model's reliance on the text modality, thereby passively diminishing its focus on images.
+
+To foster a greater focus on images during the sampling process, we set an additional local window  $\mathbf{W}_l^h$  beyond the local window of the over-trust penalty, as shown in Figure 4, purported for storing the image components within the attention matrix:
+
+$$
+\mathbf {W} _ {l} ^ {h} = \left\{\mathbf {w} ^ {i} \right\} _ {i = 1} ^ {h}, \quad \text {s . t .} \mathbf {w} ^ {i} = \left\{\omega_ {i, j} \right\} _ {j = 1} ^ {l}, \tag {9}
+$$
+
+where  $h$  is the length of the over-trust penalty window,  $l$  means the length of the visual input within the attention matrix, and  $\omega_{i,j}$  represents the attention weight from  $i_{th}$  token to  $j_{th}$  token. Subsequently, we conduct the column-wise multiplication on the  $\mathbf{W}_l^h$  to obtain a vector of column-wise scores, which represents the accumulated attention values of image:
+
+$$
+\psi \left(\omega_ {\leq h}\right) = \sum_ {i = 1} ^ {h} \omega_ {i}, \quad \text {s . t .} \omega_ {i} = \prod_ {j = 1} ^ {l} \omega_ {i, j}, \tag {10}
+$$
+
+where  $\psi (\omega_{< h})$  means the vision-enhanced penalty.
+
+Given the difference in numerical magnitudes, the initial step involves scaling  $\psi (\omega_{\leq h})$  to match the order of magnitude of  $\phi (\omega_{\leq h})$ , then calculating the overall penalty weight  $\rho (\omega_{\leq h})$ , as:
+
+$$
+\begin{array}{l} \rho (\omega_ {\leq h}) = \phi (\omega_ {\leq h}) - \beta \psi (\omega_ {\leq h}), \\ s. t. \beta = \frac {\overline {{\sum_ {j \leq h} \phi \left(\omega_ {j}\right)}}}{\overline {{\sum_ {j \leq h} \psi \left(\omega_ {j}\right)}}}. \tag {11} \\ \end{array}
+$$
+
+Then, this penalty weight is added to the original logits for the prediction of the next token  $\hat{x}_h$ , as:
+
+$$
+\hat {x} _ {h} = \arg \max  _ {x \in V} [ p (x | x _ {<   h}) - \rho (\omega_ {\leq h}) ], \tag {12}
+$$
+
+where  $V$  is the size of vocabulary and  $x$  represents the predicted token.
+
+# 4 Experimental Setups
+
+# 4.1 Hallucination Benchmarks
+
+CHAIR. Caption Hallucination Assessment with Image Relevance (CHAIR) (Rohrbach et al., 2018) is an evaluation metric employed for assessing hallucination in image captioning, and it is often used to evaluate LVLMs. CHAIR obtains scores for the degree of hallucination by calculating what proportion of objects generated are actually in the image according to the ground truth sentences and object segmentations. Specifically, it computes the hallucination at both instance level (defined as  $\mathrm{CHAIR}_i$ ) and sentence level (defined as  $\mathrm{CHAIR}_s$ ):
+
+$$
+\mathrm {C H A I R} _ {s} = \frac {\left| \{\text {h a l l u c i n a t e d o b j e c t s} \} \right|}{\left| \{\text {a l l m e n t i o n e d o b j e c t s} \} \right|}, \tag {13}
+$$
+
+$$
+\mathrm {C H A I R} _ {i} = \frac {\left| \left\{\text {c a p t i o n s w / h a l l u c i n a t e d o b j e c t s} \right\} \right|}{\left| \left\{\text {a l l c a p t i o n s} \right\} \right|}, \tag {14}
+$$
+
+where  $\mathrm{CHAIR}_s$  represents the proportion of hallucinated objects among all mentioned objects, and  $\mathrm{CHAIR}_i$  denotes the proportion of captions with hallucinated objects among all captions.
+
+<table><tr><td>POPE</td><td>Model</td><td>Accuracy</td><td>Precision</td><td>Recall</td><td>F1 Score</td><td>Yes (%)</td></tr><tr><td rowspan="6">Random</td><td>MiniGPT-4</td><td>53.06</td><td>51.58</td><td>99.60</td><td>67.97</td><td>96.53</td></tr><tr><td>InstructBLIP</td><td>86.28</td><td>84.11</td><td>95.02</td><td>89.23</td><td>55.63</td></tr><tr><td>mPLUG-Owl2</td><td>87.71</td><td>89.13</td><td>85.92</td><td>87.50</td><td>48.21</td></tr><tr><td>mPLUG-Owl2 (w/ours)</td><td>88.02</td><td>89.90</td><td>85.67</td><td>87.73</td><td>47.62</td></tr><tr><td>LLaVA-1.5</td><td>89.44</td><td>87.21</td><td>90.11</td><td>88.63</td><td>53.09</td></tr><tr><td>LLaVA-1.5 (w/ours)</td><td>89.65</td><td>87.84</td><td>91.98</td><td>89.86</td><td>52.41</td></tr><tr><td rowspan="6">Popular</td><td>MiniGPT-4</td><td>50.53</td><td>50.26</td><td>99.60</td><td>66.81</td><td>89.06</td></tr><tr><td>InstructBLIP</td><td>81.67</td><td>74.12</td><td>93.31</td><td>82.61</td><td>65.43</td></tr><tr><td>mPLUG-Owl2</td><td>84.10</td><td>82.81</td><td>85.23</td><td>84.00</td><td>51.90</td></tr><tr><td>mPLUG-Owl2 (w/ours)</td><td>85.67</td><td>85.81</td><td>85.27</td><td>85.53</td><td>49.66</td></tr><tr><td>LLaVA-1.5</td><td>84.91</td><td>81.02</td><td>90.71</td><td>85.59</td><td>55.51</td></tr><tr><td>LLaVA-1.5 (w/ours)</td><td>85.79</td><td>81.85</td><td>91.97</td><td>86.62</td><td>56.26</td></tr><tr><td rowspan="6">Adversarial</td><td>MiniGPT-4</td><td>50.46</td><td>50.23</td><td>99.61</td><td>66.78</td><td>99.13</td></tr><tr><td>InstructBLIP</td><td>72.12</td><td>65.63</td><td>95.27</td><td>77.32</td><td>73.26</td></tr><tr><td>mPLUG-Owl2</td><td>81.70</td><td>79.21</td><td>85.93</td><td>82.43</td><td>54.23</td></tr><tr><td>mPLUG-Owl2 (w/ours)</td><td>81.64</td><td>80.28</td><td>85.65</td><td>82.87</td><td>53.31</td></tr><tr><td>LLaVA-1.5</td><td>77.61</td><td>71.72</td><td>92.55</td><td>80.81</td><td>63.83</td></tr><tr><td>LLaVA-1.5 (w/ours)</td><td>78.15</td><td>72.04</td><td>92.91</td><td>81.15</td><td>63.88</td></tr></table>
+
+Table 1: Results of LVLMs under three evaluation settings of POPE on the validation set of MSCOCO. “Yes” denotes the proportion of answering “Yes” to the given question. “w/ours” means the application of HELPD.
+
+POPE. POPE (Li et al., 2023c) converts hallucination assessment into asking the model to answer a series of true or false questions about whether an object is present in the image. Specifically, given an image set and the object annotations contained in each image, POPE will construct a series of triples consisting of images, questions, and answers. It considers three polling strategies by sampling the objects randomly, from popular objects, and among those frequently co-occurring objects, respectively. Finally, POPE involves 3K questions for the captions of 500 images and uses the Accuracy, Precision, Recall, and F1 scores for evaluation.
+
+GAVIE. GPT4-Assisted Visual Instruction Evaluation (GAVIE) (Liu et al., 2023a) is an approach to measure the hallucination without the need for human-annotated ground-truth answers. GPT-4 takes the generated captions with bounding box coordinates as the image content and compares human instructions and model response. Then, ask GPT-4 to score the answers based on two criteria: (1) Accuracy: whether the response hallucinates with the image content. (2) Relevancy: whether the response directly follows the instruction. It is composed of 1k questions and uses accuracy and relevancy for evaluation.
+
+MMHal-Bench. MMHal-Bench (Sun et al., 2023) has a focus on penalizing hallucinations
+
+with 96 image-question pairs, ranging in 8 question categories and 12 object topics from OpenImages (Kuznetsova et al., 2018). It uses GPT-4 to compare the model's response to the correct answer based on the given object information. If the score is below 3, it is considered to have hallucinations.
+
+# 4.2 Baselines
+
+We use 4 recently released LVLMs as baselines: (1) MiniGPT4 (Zhu et al., 2023); (2) InstructBLIP (Dai et al., 2023); (3) LLaVA-1.5 (Liu et al., 2023b); (4) mPLUG-Owl2 (Ye et al., 2023). All models above have been tuned on their visual instruction data.
+
+# 4.3 Implementation Details
+
+We randomly select 5,000 images from the training sets of MSCOCO 2014 (Lin et al., 2014) and Flickr30k (Plummer et al., 2017). Given that each image corresponds to multiple short captions, we prompt GPT-4 to synthesize a longer caption for each image based on these short captions (see Appendix A). Then, we employ LoRA-tuning (Hu et al., 2022) and deepspeed zero stage 3 to conduct minimal training on LLaVA-1.5-7b and mPLUG-Owl2-7b for 1 epoch. We use the AdamW (Loshchilov and Hutter, 2019) optimizer for optimization purposes. The learning rate and weight decay are set to 0.0001 and 0.1, respectively. Dur
+
+<table><tr><td>Method</td><td>Model</td><td>Cs↓</td><td>Ci↓</td><td>Len</td></tr><tr><td rowspan="4">Beam5</td><td>mPLUG-Owl2</td><td>46.6</td><td>14.5</td><td>68.4</td></tr><tr><td>mPLUG-Owl2 (w/ ours)</td><td>22.4</td><td>8.4</td><td>67.1</td></tr><tr><td>LLaVA-1.5</td><td>15.4</td><td>8.2</td><td>63.2</td></tr><tr><td>LLaVA-1.5 (w/ ours)</td><td>14.6</td><td>6.1</td><td>57.8</td></tr><tr><td rowspan="4">Opera</td><td>mPLUG-Owl2</td><td>39.8</td><td>13.1</td><td>64.5</td></tr><tr><td>mPLUG-Owl2 (w/ ours)</td><td>21.4</td><td>8.2</td><td>53.3</td></tr><tr><td>LLaVA-1.5</td><td>14.1</td><td>6.1</td><td>58.9</td></tr><tr><td>LLaVA-1.5 (w/ ours)</td><td>13.7</td><td>5.0</td><td>59.9</td></tr><tr><td rowspan="4">Vep</td><td>mPLUG-Owl2</td><td>36.2</td><td>13.0</td><td>65.1</td></tr><tr><td>mPLUG-Owl2 (w/ ours)</td><td>20.6</td><td>7.8</td><td>54.0</td></tr><tr><td>LLaVA-1.5</td><td>11.0</td><td>6.2</td><td>59.7</td></tr><tr><td>LLaVA-1.5 (w/ ours)</td><td>9.6</td><td>4.9</td><td>60.8</td></tr></table>
+
+Table 2: CHAIR hallucination evaluation results. "w/ours" means the use of hierarchical feedback learning, and "Vep" is the vision-enhanced penalty decoding.  
+
+<table><tr><td>Model</td><td>Relevancy</td><td>Accuracy</td></tr><tr><td>MiniGPT-4</td><td>3.84</td><td>5.35</td></tr><tr><td>InstructBLIP</td><td>6.27</td><td>5.83</td></tr><tr><td>mPLUG-Owl2</td><td>8.29</td><td>5.68</td></tr><tr><td>mPLUG-Owl2 (w/ ours)</td><td>8.88</td><td>6.12</td></tr><tr><td>LLaVA-1.5</td><td>7.56</td><td>5.49</td></tr><tr><td>LLaVA-1.5 (w/ ours)</td><td>7.98</td><td>6.01</td></tr></table>
+
+Table 3: Evaluation results on GAVIE. The metric scores of Relevancy and Accuracy are from 0 to 10. "w/ours" means the application of HELPD.
+
+ing the training process, we initiate a warm-up ratio of 0.03, after which we apply the cosine schedule to decay the learning rate. We set  $\sigma$  to 0.6. The values of  $c$  for LLaVA-1.5 and mPLUG-Owl2 are set to 0.7 and 0.8. Each model requires approximately 4 hours to train with 2 NVIDIA 3090 24Gb GPUs.
+
+# 5 Results
+
+# 5.1 Main Results
+
+In general, it is noticeable that the application of HELPD with various LVLMs is able to enhance their performance across different evaluation metrics, compared to the original LVLMs.
+
+Upon examining the results from the POPE benchmark, as detailed in Table 1, it is evident that the hierarchical feedback learning has led to enhancements in the accuracy, precision, and F1 score. This suggests that our proposed framework can provide effective hallucination detection and feedback during training, combining object entities and semantic information to guide the model in enhancing its ability to discern hallucinated objects.
+
+As shown in Table 2, from the score of  $\mathrm{CHAIR}_s$  and  $\mathrm{CHAIR}_i$ , it is evident that with the help of
+
+![](images/c56490ccb58401722e9b3077573c9ec18fb4d424503ace3a98d896b4a708565a.jpg)  
+Figure 5: Detailed performance of LVLMs on the eight categories in MMHAL-Bench, where “Overall” indicates the averaged performance across all categories. “w/ ours” means the application of HELPD.
+
+HELPD, both mPLUG-Owl2 and LLaVA-1.5 have demonstrated varying degrees of hallucination reduction. Specifically, the trained mPLUG-Owl2, under various decoding methods, is able to reduce the  $\mathrm{CHAIR}_s$  by an average of 19.4 and the  $\mathrm{CHAIR}_i$  by an average of 5.4. This indicates that our proposed framework can effectively mitigate the generation of hallucinations, whether at the instance level or the sentence level. Moreover, it can be observed that the trained LVLMs do not exhibit significant fluctuations in the length of generated text. In most cases, LLaVA-1.5 can even increase the average generated length by 1.1 to 4.0. This illustrates that HELPD, while enhancing the anti-hallucination capabilities of LVLMs, does not excessively interfere with the generated length.
+
+Table 3 shows the performance of different LVLMs on the GAVIE benchmark, which asks GPT-4 to pretend to be a smart teacher and scores (0-10) the answers according to the image content and instructions. The trained models achieve improvements in both accuracy and relevancy. Specifically, the trained mPLUG-Owl2 attains scores of 8.88 and 6.12 in relevancy and accuracy respectively, surpassing the provided baseline models. This demonstrates that our proposed framework can aid LVLMs in more directly following instructions, and the generated responses are more accurate concerning the image content.
+
+Detailed performance of LVLMs on the eight categories in MMHAL-Bench is shown in Figure 5. It is evident that the trained models surpass their corresponding baseline models in performance across all eight question categories, and achieve a score of over 3 in five categories, including Object attribute
+
+<table><tr><td>Model</td><td>\( R_{obj} \)</td><td>\( R_{sen} \)</td><td>\( C_s \downarrow \)</td><td>\( C_i \downarrow \)</td><td>Rel</td><td>Acc</td></tr><tr><td rowspan="4">mPLUG-Owl2</td><td>X</td><td>X</td><td>46.6</td><td>14.5</td><td>8.2</td><td>5.6</td></tr><tr><td>✓</td><td>X</td><td>31.1</td><td>11.2</td><td>8.4</td><td>5.9</td></tr><tr><td>X</td><td>✓</td><td>25.9</td><td>9.9</td><td>8.7</td><td>5.9</td></tr><tr><td>✓</td><td>✓</td><td>22.4</td><td>8.4</td><td>8.8</td><td>6.1</td></tr><tr><td rowspan="4">LLaVA-1.5</td><td>X</td><td>X</td><td>15.4</td><td>8.2</td><td>7.5</td><td>5.4</td></tr><tr><td>✓</td><td>X</td><td>14.9</td><td>7.1</td><td>7.6</td><td>5.7</td></tr><tr><td>X</td><td>✓</td><td>15.8</td><td>6.8</td><td>7.7</td><td>5.8</td></tr><tr><td>✓</td><td>✓</td><td>14.6</td><td>6.1</td><td>7.9</td><td>6.0</td></tr></table>
+
+and Spatial relation. This implies that their generated texts are somewhat informative and exhibit almost no hallucination.
+
+# 5.2 Further Analysis
+
+Break-down Study of Hierarchical Feedback Learning. As illustrated in Section 3.1. In order to detect hallucinations of different granularities during training and provide feedback for parameter updates, we introduce hallucination feedback at both the object and sentence levels. To verify whether these two types of feedback contribute to the mitigation of hallucination, we conduct ablation experiments, and the results are presented in Table 4. Both object-level and sentence-level feedback can aid in alleviating hallucination, making the generated text adhere more closely to instructions and rendering it more accurate concerning the image content. It can also be observed that, compared to object-level feedback, sentence-level feedback can more effectively enhance the model's ability to resist hallucination. We hypothesize that this is because object-level feedback is more uncontrollable, such as possible omissions in the process of object extraction, or score reductions due to the presence of synonyms. However, the sentence-level feedback generated by prompting GPT-4 can effectively compensate for the deficiencies of the object-level feedback, thereby enhancing the overall performance of hierarchical feedback learning.
+
+The Timing of Incorporating Hierarchical Feedback Learning. To investigate at which stage of training the integration of hierarchical feedback learning can better enhance the model's antihallucination capabilities, we also conduct an ablation study on the hyperparameter  $c$ . The experimental results of LLaVA-V1.5 on the random set of POPE are shown in Table 5, with more details available in the Appendix B. It indicates that LLaVA-
+
+Table 4: Ablation results on different levels of feedback on CHAIR and GAVIE.  $R_{obj}$  and  $R_{sen}$  represent object-level and sentence-level feedback, respectively.  
+
+<table><tr><td>Model</td><td>c</td><td>Precision</td><td>Recall</td><td>F1 Score</td></tr><tr><td rowspan="5">LLaVA-1.5</td><td>0.6</td><td>87.01</td><td>92.05</td><td>89.45</td></tr><tr><td>0.7</td><td>87.84</td><td>91.98</td><td>89.86</td></tr><tr><td>0.8</td><td>86.21</td><td>93.07</td><td>89.50</td></tr><tr><td>0.9</td><td>86.09</td><td>93.33</td><td>89.56</td></tr><tr><td>1.0</td><td>86.11</td><td>91.33</td><td>88.74</td></tr></table>
+
+Table 5: Ablation results on the timing of incorporating HELPD. We only show the random set results on LLaVA-v1.5, more details can be seen in Appendix B.  
+
+<table><tr><td>Method</td><td>MiniGPT-4</td><td>InstructBLIP</td><td>mPLUG-Owl2</td><td>LLaVA-v1.5</td></tr><tr><td>Nucleus</td><td>58.6</td><td>78.9</td><td>82.9</td><td>82.3</td></tr><tr><td>Beam5</td><td>69.2</td><td>82.1</td><td>84.7</td><td>84.7</td></tr><tr><td>Opera</td><td>73.3</td><td>84.7</td><td>85.1</td><td>85.4</td></tr><tr><td>Vep</td><td>74.1</td><td>85.0</td><td>85.3</td><td>85.6</td></tr></table>
+
+Table 6: Ablation results on the decoding strategy. We exhibit the average F1-score computed on random, popular, and adversarial splits of POPE.
+
+V1.5 exhibits fewer hallucinations when  $c = 0.7$ , while mPLUG-Owl2 performs better when  $c = 0.8$ . Therefore, we default to assigning  $c = 0.7$  for LLaVA-V1.5 and  $c = 0.8$  for mPLUG-Owl2.
+
+Different Decoding Strategy. Based on the observations of the attention matrix, we propose the vision-enhanced penalty decoding based on opera. To validate its effectiveness, we conduct an ablation study on LVLMs. The experimental results are shown in Table 2 and Table 6. As can be observed, compared to the baseline decoding strategy, the vision-enhanced penalty decoding demonstrates superior performance on benchmarks such as POPE and CHAIR, and has a smaller impact on the length of the generated text. It should be noted that this decoding strategy pays more attention to the hallucination performance of long texts.
+
+# 6 Conclusion
+
+In this paper, we aim to alleviate hallucinations in Large Vision-Language Models (LVLMs), and propose the HELPD framework, which employs the hierarchical feedback learning for small amounts of training on the model. To enhance attention to the visual modality, we also propose a vision-enhanced penalty decoding strategy. To evaluate the effectiveness of our approach, we conduct evaluations on numerous benchmarks. Experimental results demonstrate that our proposed framework effectively mitigates hallucination for different LVLMs without impacting sentence length and concurrently improves their text generation quality. Future work could focus on a more comprehensive evaluation of hallucination at different granularities.
+
+# Limitations
+
+Although HELPD effectively mitigates the hallucination in VLVMs, it remains subject to certain limitations. Firstly, to further train VLVMs, even for minimal training, a rich corpus of modality-aligned data is required. Secondly, compared to traditional decoding strategies, our proposed vision-enhanced penalty decoding may slightly increase decoding time, thereby potentially limiting inference speed.
+
+# Ethics Statement
+
+The data (Lin et al., 2014; Plummer et al., 2017) used in our work is all drawn from open-source datasets. The data and text involved in our research do not involve private information and social issues.
+
+# Acknowledgments
+
+This research is supported by the National Natural Science Foundation of China (No.62476127, No.62106105), the Natural Science Foundation of Jiangsu Province (No.BK20242039), the CCF-Baidu Open Fund (No.CCF-Baidu202307), the CCF-Zhipu AI Large Model Fund (No.CCF-Zhipu202315), the Fundamental Research Funds for the Central Universities (No.NJ2023032), the Scientific Research Starting Foundation of Nanjing University of Aeronautics and Astronautics (No.YQR21022), and the High Performance Computing Platform of Nanjing University of Aeronautics and Astronautics.
+
+# References
+
+Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. 2022. Flamingo: a visual language model for few-shot learning. 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.  
+Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. 2016. SPICE: semantic propositional image caption evaluation. In Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings,
+
+Part V, volume 9909 of Lecture Notes in Computer Science, pages 382-398. Springer.  
+Rohan Anil, Sebastian Borgeaud, Yonghui Wu, JeanBaptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Haut, Katie Millican, David Silver, Slav Petrov, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy P. Lillicrap, Angeliki Lazaridou, Orhan Firat, James Molloy, Michael Isard, Paul Ronald Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anaïs White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, and et al. 2023. Gemini: A family of highly capable multimodal models. CoRR, abs/2312.11805.  
+Satanjeev Banerjee and Alon Lavie. 2005. METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization@ACL 2005, Ann Arbor, Michigan, USA, June 29, 2005, pages 65-72. Association for Computational Linguistics.  
+Andy Brock, Soham De, Samuel L. Smith, and Karen Simonyan. 2021. High-performance large-scale image recognition without normalization. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 1059-1071. PMLR.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James R. Glass, and Pengcheng He. 2023. Dola: Decoding by contrasting layers improves factuality in large language models. CoRR, abs/2309.03883.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Y. Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling instruction-finetuned language models. CoRR, abs/2210.11416.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang,
+
+Boyang Li, Pascale Fung, and Steven C. H. Hoi. 2023. Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR, abs/2305.06500.  
+Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-verification reduces hallucination in large language models. CoRR, abs/2309.11495.  
+Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.  
+Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. 2023. EVA: exploring the limits of masked visual representation learning at scale. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19358-19369. IEEE.  
+Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. 2024. Multi-modal hallucination control by visual information grounding. CoRR, abs/2403.14003.  
+Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. 2023. MME: A comprehensive evaluation benchmark for multimodal large language models. CoRR, abs/2306.13394.  
+Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6325-6334. IEEE Computer Society.  
+Anisha Gunjal, Jihan Yin, and Erhan Bas. 2024. Detecting and preventing hallucinations in large vision language models. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pages 18135-18143. AAAI Press.  
+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.  
+Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. 2023. OPERA: alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. CoRR, abs/2311.17911.  
+Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Tom Duerig, and Vittorio Ferrari. 2018. The open images dataset V4: unified image classification, object detection, and visual relationship detection at scale. CoRR, abs/1811.00982.  
+Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. 2023. Volcano: Mitigating multimodal hallucination through self-feedback guided revision. CoRR, abs/2311.07362.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. 2023a. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 19730-19742. PMLR.  
+Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2023b. Contrastive decoding: Open-ended text generation as optimization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 12286-12312. Association for Computational Linguistics.  
+Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023c. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 292-305. Association for Computational Linguistics.  
+Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.  
+Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. 2014. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, volume 8693 of Lecture Notes in Computer Science, pages 740-755. Springer.
+
+Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. 2023a. Mitigating hallucination in large multi-modal models via robust instruction tuning. arXiv preprint arXiv:2306.14565, 1(2):9.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023b. Improved baselines with visual instruction tuning. CoRR, abs/2310.03744.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.  
+Sean O'Brien and Mike Lewis. 2023. Contrastive decoding improves reasoning in large language models. CoRR, abs/2309.09117.  
+OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.  
+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, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318. ACL.  
+Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. 2023. Kosmos-2: Grounding multimodal large language models to the world. CoRR, abs/2306.14824.  
+Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. 2017. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. Int. J. Comput. Vis., 123(1):74-93.  
+Chengwei Qin, Aston Zhang, Zhuosheng Zhang, Jiao Chen, Michihiro Yasunaga, and Diyi Yang. 2023. Is chatgpt a general-purpose natural language processing task solver? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 1339-1384. Association for Computational Linguistics.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.  
+Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural
+
+Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 4035-4045. Association for Computational Linguistics.  
+Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, Weixin Liu, Zhihua Wu, Weibao Gong, Jianzhong Liang, Zhizhou Shang, Peng Sun, Wei Liu, Xuan Ouyang, Dianhai Yu, Hao Tian, Hua Wu, and Haifeng Wang. 2021. ERNIE 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. CoRR, abs/2107.02137.  
+Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darrell. 2023. Aligning large multimodal models with factually augmented RLHF. CoRR, abs/2309.14525.  
+Richard S. Sutton, David A. McAllester, Satinder Singh, and Yishay Mansour. 1999. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems 12, [NIPS Conference, Denver, Colorado, USA, November 29 - December 4, 1999], pages 1057-1063. The MIT Press.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023a. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal 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, 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, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open foundation and fine-tuned chat models. CoRR, abs/2307.09288.  
+Liam van der Poel, Ryan Cotterell, and Clara Meister. 2022. Mutual information alleviates hallucinations in abstractive summarization. In Proceedings of
+
+the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 5956-5965. Association for Computational Linguistics.  
+Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, Jitao Sang, and Haoyu Tang. 2023. Evaluation and analysis of hallucination in large vision-language models. CoRR, abs/2308.15126.  
+Lei Wang, Jiabang He, Shenshen Li, Ning Liu, and Ee-Peng Lim. 2024. Mitigating fine-grained hallucination by fine-tuning large vision-language models with caption rewrites. In MultiMedia Modeling - 30th International Conference, MMM 2024, Amsterdam, The Netherlands, January 29 - February 2, 2024, Proceedings, Part IV, volume 14557 of Lecture Notes in Computer Science, pages 32-45. Springer.  
+Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, Fan Yang, Fei Deng, Feng Wang, Feng Liu, Guangwei Ai, Guosheng Dong, Haizhou Zhao, Hang Xu, Haoze Sun, Hongda Zhang, Hui Liu, Jiaming Ji, Jian Xie, Juntao Dai, Kun Fang, Lei Su, Liang Song, Lifeng Liu, Liyun Ru, Luyao Ma, Mang Wang, Mickel Liu, MingAn Lin, Nuolan Nie, Peidong Guo, Ruiyang Sun, Tao Zhang, Tianpeng Li, Tianyu Li, Wei Cheng, Weipeng Chen, Xiangrong Zeng, Xiaochuan Wang, Xiaoxi Chen, Xin Men, Xin Yu, Xuehai Pan, Yanjun Shen, Yiding Wang, Yiyu Li, Youxin Jiang, Yuchen Gao, Yupeng Zhang, Zenan Zhou, and Zhiying Wu. 2023. Baichuan 2: Open large-scale language models. CoRR, abs/2309.10305.  
+Dingchen Yang, Bowen Cao, Guang Chen, and Changjun Jiang. 2024. Pensieve: Retrospect-then compare mitigates visual hallucination. CoRR, abs/2403.14401.  
+Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. 2023. MPLug-owl2: Revolutionizing multi-modal large language model with modality collaboration. CoRR, abs/2311.04257.  
+Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. 2023. A survey on multimodal large language models. CoRR, abs/2306.13549.  
+Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, Zhiyuan Liu, Peng Zhang, Yuxiao Dong, and Jie Tang. 2023. GLM-130B: an open bilingual pre-trained model. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.
+
+Bohan Zhai, Shijia Yang, Xiangchen Zhao, Chenfeng Xu, Sheng Shen, Dongdi Zhao, Kurt Keutzer, Manling Li, Tan Yan, and Xiangjun Fan. 2023. Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. CoRR, abs/2310.01779.  
+Duzhen Zhang, Yahan Yu, Chenxing Li, Jiahua Dong, Dan Su, Chenhui Chu, and Dong Yu. 2024. Mmllms: Recent advances in multimodal large language models. CoRR, abs/2401.13601.  
+Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Jiaqi Wang, and Conghui He. 2023. Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimization. CoRR, abs/2311.16839.  
+Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. 2023. Analyzing and mitigating object hallucination in large vision-language models. CoRR, abs/2310.00754.  
+Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592.
+
+# A Prompts
+
+We show the prompt for generating sentence-level feedback score in Figure 6 and synthesizing longer captions for each image based on corresponding five short captions in Figure 7 with GPT-4.
+
+# B More Experimental Results
+
+# B.1 More Experimental Results Compared to Existing Methods.
+
+For POPE benchmark in Table 9, compared to existing methods (Liu et al., 2023a; Favero et al., 2024; Yang et al., 2024), our approach performs well on both the random and popular test sets, but falls short in terms of accuracy on the adversarial test set. This is because the method by Zhou et al. (2023) uses multiple similar images for comparison, enabling multi-angle judgment of the test problem. In contrast, our proposed method requires less computational resources and is able to surpass the performance in terms of F1 score. While for CHAIR in Table 10, compared to existing work, we are able to demonstrate better performance.
+
+# B.2 Ablation Results About the Timing of Incorporating Hierarchical Feedback Learning.
+
+In Section 5.2, to investigate at which stage of training the integration of Hierarchical Feedback
+
+Learning can better enhance the model's antihallucination capabilities, we conduct an ablation study on the hyperparameter  $c$ . We show the additional results in Table 7.
+
+# B.3 Further Analysis About the Quality of the Generated Text.
+
+Considering that interventions in the decoding process can impact the quality of the generated text, we also conduct experiments to measure the impact caused by the vision-enhanced penalty decoding. Evaluation was carried out on the generated text from CHAIR, using the BLEU (Papineni et al., 2002), ROUGE-L (Lin, 2004), METEOR (Banerjee and Lavie, 2005), and SPICE (Anderson et al., 2016) metrics, with the experimental results presented in Table 8. It can be observed that our proposed vision-enhanced penalty decoding performs comparably to beam search in terms of text generation metrics, without demonstrating excessive decline. It even surpasses beam search on BLEU1, BLEU4, and ROUGE-L. As can also be seen from Table 2, compared to beam search and opera decoding, our proposed method is able to maintain the length of the generated sentences as well. This elucidates that our method can maintain the quality of the generated text while mitigating hallucinations. Additionally, to verify whether HELPD can mitigate hallucinations while preserving general capabilities, we conduct evaluations on VQA-v2 (Goyal et al., 2017) and MME (Fu et al., 2023). As shown in Table 11 and Table 12, LVLMs with HELPD can maintain relatively stable performance across various metrics, demonstrating that the framework does not significantly impair the model's foundational abilities.
+
+# C Cases
+
+We show some generation cases in Figure 8, 9, and 10.
+
+You will be presented with two pieces of text that describe the same image. First text comes from a dataset and can be considered as the label, and the other is generated by a model. Your task is to compare these two pieces of text and evaluate them from the perspective of hallucination. If the contents described by the two pieces of text are completely consistent and there is no hallucination, please give a score of 1. If hallucination is present in the model-generated text, i.e., it describes content not present in the label text, please give a score closer to 0. Note that the lower the score, the less obvious the hallucination. The score range is between 0 and 1. Please only provide a score and do not provide reason.
+
+Output example 1:
+
+Dataset text (Label): Label 1; Model generated text: Text 1; score: score 1
+
+Output example 2:
+
+Dataset text (Label): Label 2; Model generated text: Text 2; score: score 2
+
+Output example 3:
+
+Dataset text (Label): Label 3; Model generated text: Text 3; score: score 3
+
+Output format:
+
+Dataset text (Label): Model generated tex score:
+
+Figure 6: Prompt for generating sentence-level feedback score.
+
+Given the following five captions of the same image, please combine them into a single, comprehensive caption that includes all the information, especially objects, without any repetition:
+
+Output example 1:
+
+Caption 1: caption_1;
+
+Caption 2: caption_2;
+
+Caption 3: caption_3;
+
+Caption 4: caption_4;
+
+Caption 5: caption_5;
+
+Combined caption 1: response_1
+
+Output example 2:
+
+Caption 1: caption_1;
+
+Caption 2: caption_2;
+
+Caption 3: caption_3;
+
+Caption 4: caption_4;
+
+Caption 5: caption_5;
+
+Combined caption 2: response_2
+
+Output example 3:
+
+Caption 1: caption_1;
+
+Caption 2: caption_2;
+
+Caption 3: caption_3;
+
+Caption 4: caption_4;
+
+Caption 5: caption_5;
+
+Combined caption 3: response_3
+
+Output format:
+
+Caption 1:
+
+Caption 2:
+
+Caption 3:
+
+Caption 4:
+
+Caption 5:
+
+Combined caption:
+
+Figure 7: Prompt for synthesizing longer captions for each image based on corresponding five short captions.
+
+<table><tr><td>POPE</td><td>Model</td><td>c</td><td>Precision</td><td>Recall</td><td>F1 Score</td><td>Yes (%)</td></tr><tr><td rowspan="8">Random</td><td rowspan="4">mPLUG-Owl2</td><td>0.6</td><td>88.7</td><td>86.1</td><td>87.3</td><td>46.5</td></tr><tr><td>0.7</td><td>88.9</td><td>85.9</td><td>87.3</td><td>46.4</td></tr><tr><td>0.8</td><td>89.9</td><td>85.6</td><td>87.7</td><td>47.6</td></tr><tr><td>0.9</td><td>90.1</td><td>85.2</td><td>87.6</td><td>47.3</td></tr><tr><td rowspan="4">LLaVA-1.5</td><td>0.6</td><td>87.0</td><td>92.1</td><td>89.5</td><td>52.9</td></tr><tr><td>0.7</td><td>87.8</td><td>91.9</td><td>89.9</td><td>52.4</td></tr><tr><td>0.8</td><td>86.2</td><td>93.1</td><td>89.5</td><td>54.6</td></tr><tr><td>0.9</td><td>86.1</td><td>93.3</td><td>89.7</td><td>54.3</td></tr><tr><td rowspan="8">Popular</td><td rowspan="4">mPLUG-Owl2</td><td>0.6</td><td>82.8</td><td>85.9</td><td>84.3</td><td>52.0</td></tr><tr><td>0.7</td><td>84.2</td><td>86.1</td><td>85.1</td><td>52.0</td></tr><tr><td>0.8</td><td>85.6</td><td>85.6</td><td>85.6</td><td>50.0</td></tr><tr><td>0.9</td><td>85.8</td><td>85.2</td><td>85.5</td><td>49.6</td></tr><tr><td rowspan="4">LLaVA-1.5</td><td>0.6</td><td>80.1</td><td>93.3</td><td>86.2</td><td>58.2</td></tr><tr><td>0.7</td><td>81.4</td><td>92.0</td><td>86.4</td><td>56.5</td></tr><tr><td>0.8</td><td>79.7</td><td>93.0</td><td>85.8</td><td>58.3</td></tr><tr><td>0.9</td><td>79.5</td><td>93.5</td><td>85.9</td><td>58.8</td></tr><tr><td rowspan="8">Adversarial</td><td rowspan="4">mPLUG-Owl2</td><td>0.6</td><td>79.2</td><td>85.9</td><td>82.5</td><td>54.1</td></tr><tr><td>0.7</td><td>79.8</td><td>85.9</td><td>82.7</td><td>54.2</td></tr><tr><td>0.8</td><td>80.2</td><td>85.6</td><td>82.8</td><td>53.3</td></tr><tr><td>0.9</td><td>80.7</td><td>85.2</td><td>82.9</td><td>52.8</td></tr><tr><td rowspan="4">LLaVA-1.5</td><td>0.6</td><td>71.2</td><td>93.3</td><td>80.8</td><td>65.6</td></tr><tr><td>0.7</td><td>72.1</td><td>92.0</td><td>80.9</td><td>63.8</td></tr><tr><td>0.8</td><td>72.1</td><td>91.9</td><td>80.8</td><td>63.8</td></tr><tr><td>0.9</td><td>70.8</td><td>93.5</td><td>80.6</td><td>66.0</td></tr></table>
+
+Table 7: Additional ablation results on the timing of incorporating HELPD.  
+
+<table><tr><td>Model</td><td>Method</td><td>B1</td><td>B4</td><td>M</td><td>R-L</td><td>S</td></tr><tr><td rowspan="3">LLaVA-1.5</td><td>greedy</td><td>20.8</td><td>5.1</td><td>19.9</td><td>22.4</td><td>22.3</td></tr><tr><td>beam5</td><td>21.6</td><td>5.3</td><td>21.2</td><td>23.2</td><td>23.4</td></tr><tr><td>vep</td><td>22.3</td><td>5.5</td><td>20.8</td><td>23.3</td><td>22.9</td></tr><tr><td rowspan="3">mPLUG-Owl2</td><td>greedy</td><td>18.9</td><td>4.8</td><td>16.7</td><td>17.1</td><td>16.9</td></tr><tr><td>beam5</td><td>19.8</td><td>5.1</td><td>17.2</td><td>19.7</td><td>18.3</td></tr><tr><td>vep</td><td>20.1</td><td>5.2</td><td>17.2</td><td>18.9</td><td>18.1</td></tr></table>
+
+Table 8: Evaluation of the quality of the generated text from CHAIR. B1, B4, M, R-L, and S are abbreviations for BLEU1, BLEU4, METEOR, ROUGE-L, and SPICE. "vep" represents the vision-enhanced penalty decoding.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Random</td><td colspan="2">Popular</td><td colspan="2">Adversarial</td></tr><tr><td>Accuracy</td><td>F1 Score</td><td>Accuracy</td><td>F1 Score</td><td>Accuracy</td><td>F1 Score</td></tr><tr><td>Liu et al. (2023a)</td><td>87.3</td><td>87.3</td><td>73.4</td><td>80.1</td><td>65.0</td><td>73.9</td></tr><tr><td>Favero et al. (2024)</td><td>81.2</td><td>65.6</td><td>73.9</td><td>67.3</td><td>68.2</td><td>75.4</td></tr><tr><td>Yang et al. (2024)</td><td>87.5</td><td>86.1</td><td>85.1</td><td>84.8</td><td>81.7</td><td>80.7</td></tr><tr><td>HELPD (ours)</td><td>89.6</td><td>89.8</td><td>85.7</td><td>86.6</td><td>78.1</td><td>81.1</td></tr></table>
+
+Table 9: More experimental results on POPE benchmark.
+
+<table><tr><td>Method</td><td>CHAIRs ↓</td><td>CHAIRi ↓</td><td>Len</td></tr><tr><td>Zhou et al. (2023)</td><td>27.1</td><td>6.4</td><td>58.8</td></tr><tr><td>van der Poel et al. (2022)</td><td>16.2</td><td>6.7</td><td>59.6</td></tr><tr><td>Li et al. (2023b)</td><td>14.8</td><td>6.3</td><td>60.7</td></tr><tr><td>Liu et al. (2023a)</td><td>13.8</td><td>5.9</td><td>60.1</td></tr><tr><td>HELPD (ours)</td><td>9.6</td><td>4.9</td><td>60.8</td></tr></table>
+
+Table 10: More experimental results on CHAIR benchmark.  
+
+<table><tr><td>Method</td><td>Yes/No</td><td>Number</td><td>Other</td><td>Overall</td></tr><tr><td>LLaVA-1.5</td><td>92.23</td><td>60.01</td><td>71.07</td><td>78.53</td></tr><tr><td>LLaVA-1.5 (w/ours)</td><td>92.88</td><td>60.78</td><td>68.86</td><td>77.50</td></tr><tr><td>mPLUG-Owl2</td><td>91.96</td><td>63.24</td><td>70.51</td><td>79.05</td></tr><tr><td>mPLUG-Owl2 (w/ours)</td><td>92.56</td><td>60.21</td><td>69.82</td><td>78.20</td></tr></table>
+
+Table 11: Experimental results on VQA-v2 benchmark.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Category</td><td rowspan="2">Total</td></tr><tr><td>Perception</td><td>Cognition</td></tr><tr><td>LLaVA-1.5</td><td>1508.5</td><td>355.7</td><td>1864.2</td></tr><tr><td>LLaVA-1.5 (w/ ours)</td><td>1505.6</td><td>342.8</td><td>1848.4</td></tr><tr><td>mPLUG-Owl2</td><td>1446.5</td><td>313.3</td><td>1759.8</td></tr><tr><td>mPLUG-Owl2 (w/ ours)</td><td>1434.5</td><td>312.7</td><td>1747.2</td></tr></table>
+
+Table 12: Experimental results on MME benchmark.
+
+mPLUG-Owl2  
+mPLUG-OwI2 w/ HELPD  
+![](images/bee3d023a37fae4c2842099df62d323768e7173c0fecbff1b16380b9e2556943.jpg)  
+Describe the content of the image in detail.
+
+![](images/dd53963cc2b1ea3f67f501fad904bfab22b174af60364eba351eebb420126c72.jpg)
+
+The image features a lion and a cow running in a meadow. They are all in motion, one chasing the other, running together in their natural habitat. In this wild scene of natural beauty, the two animals seem to rush at full speed. Additionally, there are some other animals in the distance, such as horses, creating a natural atmosphere.
+
+![](images/4b42a92201ccf89bcd871b03b40b830a3852ab274f88a08a869a35987ab6510e.jpg)  
+Figure 8: An illustrative case is presented to compare the output of mPLUG-Owl2 and mPLUG-Owl2 with HELPD.
+
+This photograph depicts a lion and a cow running in a meadow. They are engaged in a chase, each animal exploring their natural surroundings in pursuit or flight. Amidst this wild display of natural splendor, the two creatures appear to interact, charging full-tilt towards an object of interest.
+
+![](images/f590e64e6c9c4b86e78050b3395288fda7763d51ba32a91744ff0f3d6f609fa7.jpg)  
+Figure 9: An illustrative case is presented to compare the output of LLaVA-v1.5 and LLaVA-v1.5 with HELPD.
+
+Describe the content of the image in detail.
+
+![](images/4ef787ca7f45cf6e1cef58ff862e0697c5248831eebd58d7b62452735bd22093.jpg)
+
+LLaVA-1.5
+
+The image features a young woman standing on the bank of an outdoor body water, such as river or lake. She is holding onto her fishing rod and appears to be enjoying herself while casting for something in the waterside area near trees that surround it. There are several other people visible around this scene who may also have joined together with others nearby.
+
+![](images/173f88ea2ee7c2666cec7cb545057d0f8575e1a3030c12856a324a94b1c28f6a.jpg)
+
+LLAVA-1.5 w/ HELPD
+
+The image showcases a young woman standing alone on the bank of an outdoor body of water, possibly a river or a lake. She holds a fishing rod, seemingly enjoying her time as she casts her line into the water. The area is surrounded by trees, adding to the serenity of the scene. With no other individuals are present in this area.
+
+![](images/e0d4b42dd77b93a14cb04db21a96fb096c5d66e6c2614603c85e5910958c1f42.jpg)  
+Figure 10: An illustrative case is presented to compare the output of mPLUG-Owl2 and mPLUG-Owl2 with HELPD.
+
+Describe the content of the image in detail.
+
+![](images/7eeaa0cf904007cdd13f09b96c567b47a53cda06643272430aefd6ca2f3083da.jpg)
+
+mPLUG-Owl2
+
+The image features a man with other players in white and blue clothing, paddling an oar-powered boat through rough water. They are wearing protective gear such as helmets on their head to ensure safety, while navigating the rapids of this challenging course with precision skillfulness required for competitive kayaking. Additionally, there are a lot of spectators on both sides cheering for them.
+
+![](images/3dca22e4368ccc820f05142fccc0ad34a61defc80c6cebfbfcbcaebcb59c8ccb.jpg)
+
+nPLUG-Owl2 w/ HELPD
+
+The image features a man in white and blue clothing, paddling an oar-powered boat through rough water. The man is wearing protective gear such as helmets to ensure safety, while navigating the rapids of this challenging course with precision skillfulness required for competitive kayaking.

2024/HELPD_ Mitigating Hallucination of LVLMs by Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6a12ccff69624de3b2d5c84d896d5687575af221b22808127d8da33d51a1f59
+size 999953

2024/HalluMeasure_ Fine-grained Hallucination Measurement Using Chain-of-Thought Reasoning/4dd725bd-76ce-42b7-bd26-6a35e1b8ce5b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0b37f5d20a4a53ceed04aae699d762fa91d517452c54df36f6714e208a4d783
+size 8873841

2024/HalluMeasure_ Fine-grained Hallucination Measurement Using Chain-of-Thought Reasoning/full.md

@@ -0,0 +1,565 @@
+# HalluMeasure: Fine-grained Hallucination Measurement Using Chain-of-Thought Reasoning
+
+Shayan A. Akbar*, Md Mosharaf Hossain*, Tess Wood*, Si-Chi Chin, Erica Salinas, Victor Alvarez, Erwin Cornejo
+
+Customer Experience and Business Trends, Amazon.com {shayaakb, hosmdmos, tesswoo, sichi, erislin, miranvi, eccornej}@amazon.com
+
+# Abstract
+
+Automating the measurement of hallucinations in LLM-generated responses is a challenging task as it requires careful investigation of each factual claim in a response. In this paper, we introduce HalluMeasure, a new LLM-based hallucination detection mechanism that decomposes an LLM response into atomic claims, and evaluates each atomic claim against the provided reference context. The model uses a step-by-step Chain-of-Thought reasoning process and can identify 3 major categories of hallucinations (e.g., contradiction) as well as 10 more specific subtypes (e.g., overgeneralization) which help to identify reasons behind the hallucination errors. Specifically, we explore four different configurations for HalluMeasure's classifier: with and without CoT prompting, and using a single classifier call to classify all claims versus separate calls for each claim. The best-performing configuration (with CoT and separate calls for each claim) demonstrates significant improvements in detecting hallucinations, achieving a 10-point increase in F1 score on our TechNewsSumm dataset, and a 3-point increase in AUC ROC on the SummEval dataset, compared to three baseline models (RefChecker, AlignScore, and Vectara HHEM). We further show reasonable accuracy on detecting 10 novel error subtypes of hallucinations (where even humans struggle in classification) derived from linguistic analysis of the errors made by the LLMs.
+
+# 1 Introduction
+
+Hallucinations in Large Language Models (LLMs) are inevitable (Xu et al., 2024) and can cause significant harm. For instance, non-existent legal cases generated by an LLM in court papers submitted by a law firm resulted in sanctions by a judge (CNBC, 2023). Wrongly claiming the Webb Space Telescope was the first to photograph an exoplanet during an LLM-based product demo led to a  $7.7\%$  drop
+
+in a company's stock value, wiping out $100 billion in market capitalization (Reuters, 2023). Moreover, misleading information about bereavement travel policy generated by an airline chatbot led to a court order for partial refund to a passenger (Technica, 2024). Consequently, detecting and measuring hallucinations in LLMs has become a crucial research area in recent years to prevent their potential harmful effects.
+
+We measure hallucinations in LLM responses by comparing the claims made in a response against a reference context document. Since manually annotating LLM responses against such context documents is time-consuming and expensive, there is a need to develop automatic approaches to measure hallucination at scale. In recent years, several studies (Zha et al., 2023; Hu et al., 2023; Vectara, 2023; Min et al., 2023; Huang et al., 2023) have proposed solutions to automate measurement of hallucinations in LLM responses. These solutions involve classifying pairs of context and LLM response texts into hallucination vs. non-hallucination classes using machine learning models.
+
+We highlight key differences between prior solutions and our HalluMeasure method below:
+
+- LLM-based classification approach: Unlike some of the prior works that use traditional machine learning models and BERT-based classifiers (Zha et al., 2023; Vectara, 2023), we use LLMs with prompt engineering for hallucination detection and measurement, achieving better performance.
+
+- Claim-level classification: Many studies (Zha et al., 2023; Vectara, 2023) propose solutions measuring hallucinations at the response level, some (Kryscinski et al., 2020; Laban et al., 2021) segment responses into sentences, and a few (Min et al., 2023; Chern et al., 2023) operate at the claim level. We employ claim-level classification, enabling fine-grained mea
+
+![](images/96e62ca64dbbbae3b3f11e68b7b4be7dca8cae62bb27efe29006d22abe70bf72.jpg)  
+Figure 1: HalluMeasure Overview. The LLM response goes through claim extraction model to produce a list of claims which are classified using a claim classification model. Finally scores are aggregated and returned to the user.
+
+![](images/f6bd206532eaab9e1a49a726845d16785d5e23e5e38625892ad8fd195af3522c.jpg)  
+Figure 2: HalluMeasure assigns each claim a thought for hallucination labeling, then a label from 5 labels (supported, partially supported, absent, contradicted, unsupported, unevaluable). Each claim also gets a thought for error type classification, then an error type label from 10 labels (see Table 2).
+
+surement. When hallucinated information exists within lengthy LLM responses, evaluating individual extracted claims improves hallucination detection accuracy.
+
+- Chain-of-Thought Prompting: We leverage few-shot Chain-of-Thought (CoT) prompting (Wei et al., 2022) to teach reasoning abilities to our claim classifier model, enabling it to accurately classify claims through exemplar demonstrations. This enhances claim classification accuracy over prior works that use simple few-shot prompting (Hu et al., 2023).  
+- Single classification call for list of claims: Previous studies processed claims individually (Hu et al., 2023). However, we lever
+
+age batch prompting (Cheng et al., 2023) and demonstrate an LLM-based classifier that classifies multiple claims extracted from the same response simultaneously with reasonable accuracy. This approach reduces LLM calls, associated latency, and costs, enabling more scalable hallucination detection while maintaining reasonable performance.
+
+- Fine-grained error types: Previous works often classify results into binary or NLI (ternary) classes. Our study demonstrates the value of granular hallucination error types (Section 3.2). By providing deeper insights into the type of hallucinations produced, HalluMeasure enables more targeted solutions to enhance LLM reliability.
+
+<table><tr><td>LLM-in-test response</td><td>Atomic claims</td></tr><tr><td>Samsung&#x27;s Gear Blink could have a projected keyboard that allows you to type in the air. Ralph Lauren&#x27;s Polo Tech Shirt uses bio-sensing fabrics to monitor physical activity. Hush earplugs filter out unwelcome sounds while allowing phone calls and alarms to intrude.</td><td>1. Samsung has a product called Gear Blink.
+2. Gear Blink could have a projected keyboard.
+3. Gear Blink&#x27;s projected keyboard would allow typing in air.
+4. Ralph Lauren has a product called Polo Tech Shirt.
+5. Polo Tech Shirt uses bio-sensing fabrics.
+6. Polo Tech Shirt bio-sensing fabrics monitor physical activity.
+7. There is a product called Hush earplugs.
+8. Hush earplugs filter out unwelcome sounds.
+9. Hush earplugs allow phone calls to be heard.
+10. Hush earplugs allow alarms to be heard.</td></tr></table>
+
+Table 1: An example summary text of a news article & extracted atomic claims by our claim extractor.
+
+Our HalluMeasure method works by first decomposing the LLM response into a set of claims using a claim extraction model. Then, we classify the claims into 5 key classes (e.g., supported, absent, contradiction, partially supported, and unevaluable) by comparing them against the contexts using our claim classification model. Additionally, we classify the claims into 10 novel distinct error types (e.g., entity, temporal, over-generalization, etc.) that provide a fine-grained analysis of hallucination errors. Finally, we produce an aggregated hallucination score by measuring the rate of unsupported claims (i.e., those assigned classes other than supported), and calculate the distribution of fine-grained error types. This distribution provides LLM builders with valuable insights into the nature of errors their LLM is making, facilitating targeted improvements. Figure 1 illustrates the main components and process behind HalluMeasure.
+
+Our results demonstrate that HalluMeasure outperforms existing solutions in terms of F1 score and AUC ROC metric on two datasets: TechNewsSumm (our own curated dataset) and a popular public benchmark dataset SummEval (Fabbri et al., 2021). We attribute HalluMeasure's superior performance to (1) our improved prompting strategy that utilizes Chain-of-Thought (CoT) reasoning, and (2) our claim-level classification approach that measures hallucination based on fine-grained information present in the response text.
+
+We attempt to answer the following 6 key research questions as part of this study:
+
+- RQ1: How effectively can HalluMeasure extract claims from LLM responses?  
+- RQ2: How does HalluMeasure method compare against state-of-the-art methods?  
+- RQ3: Is a single call to classify all claims effective for hallucination classification?
+
+- RQ4: Can HalluMeasure effectively detect fine-grained hallucination error types?  
+- RQ5: Does the use of CoT prompting improve hallucination measurement performance?  
+- RQ6: How effectively can HalluMeasure's method generalize to different underlying LLMs for hallucination classification?
+
+Our key contributions are (1) a novel HalluMeasure method that automatically measures hallucinations using fine-grained analysis of LLM responses using Chain-of-Thought reasoning, (2) experimental results of our HalluMeasure method that outperforms existing solutions (RefChecker, AlignScore, and Vectara HHEM), and (3) a novel TechNewsSumm dataset containing fine-grained claim level labels for news summarization task with tech news articles taken from CNN/DailyMail dataset.
+
+# 2 Related Work
+
+Hallucination is a topic of growing research interest, and a range of prior studies have addressed its identification and measurement. Several survey papers provide a useful overview and analysis (Huang et al., 2023; Wang et al., 2024; Rawte et al., 2023). A number of works provide either a hallucination measurement dataset (Li et al., 2023; Lin et al., 2022; Tam et al., 2023), an automatic evaluation metric (Zha et al., 2023; Chern et al., 2023; Min et al., 2023; Manakul et al., 2023; Mündler et al., 2024; Gekhman et al., 2023; Krysciński et al., 2019) or a meta-evaluation (i.e., evaluation of different hallucination metric performances) (Honovich et al., 2022; Gabriel et al., 2021). Many of the use cases addressed focus on summarization (e.g., news summarization or headline generation) and use popular news datasets for testing (CNN/DailyMail news articles corpus (See
+
+et al., 2017), or XSUM news headline dataset (Narayan et al., 2018)). Another popular use case is Wikipedia-style biography generation (e.g., as in WikiBio dataset (Lebret et al., 2016)). Previous approaches to measurement include using pretrained or finetuned models or NLI- and Question-Answer-Generation (QAG)-based metrics. More recent studies employ LLMs to classify responses. Most of these (Zha et al., 2023; Vectara, 2023) classify at the response level while a smaller number classify at the sentence (Kryscinski et al., 2020; Laban et al., 2021) or fine-grained claim level (Min et al., 2023; Chern et al., 2023). Further, they differ in whether they use a binary hallucination/non-hallucination classification, ternary NLI classes (Chern et al., 2023; Min et al., 2023), or perform fine-grained multi-class classification to divide hallucination into different error types (e.g., negation error, number swap, or entity swap, etc.) (Rawte et al., 2023). A number of different taxonomies of hallucination error types have been proposed (Tang et al., 2023; Zhu et al., 2023; Tang et al., 2024). We build on these earlier approaches by combining claim-level analysis with fine-grained error types (distinguished based on their impact and potential causes or mitigations), and by using current LLMs with few-shot learning and CoT prompting to produce strong results in measuring hallucinations at both the claim and response levels.
+
+# 3 Method
+
+In this section we present methodology for HalluMeasure. We use a claim-level hallucination measurement approach inspired by Chern et al. (2023), Hu et al. (2023), and Min et al. (2023). We decompose LLM responses into smaller units ('claims') for more precise measurement using our claim extraction model based on Claude 2.1. Then, we classify claims and assign high-level labels, further identifying 10 more specific types of hallucination errors for unsupported claims using CoT reasoning with Claude 3 Sonnet (See Table 9).
+
+HalluMeasure calculates the percentage of hallucinations at both the claim and response level, and provides the distribution of hallucination types identified. Below, we discuss specific components of our hallucination measurement method that takes in context and response pairs as input and produces hallucination scores as output. Figure 1 shows the HalluMeasure process.
+
+# 3.1 Extracting Claims from LLM Responses
+
+As noted above, our approach decomposes an LLM response into a set of claims. An intuitive definition of claim is 'the smallest unit of information that can be evaluated against a context'; in the prototypical case, this consists of a single predicate with a subject and (optionally) an object. Several recent works on hallucination and factuality decompose sentences into claims for evaluation; however, as noted by Wanner et al. (2024), the method of decomposition affects the number of claims extracted from a given model response, and therefore impacts hallucination metrics. In general, higher atomicity of claims allows for more precise measurement and localization of hallucinations. Table 1 shows an example response with the claims extracted by our claim extractor.
+
+We develop a claim extraction model which, given an LLM response, decomposes that response into claims to be evaluated. We prompt the claim extraction model using a small set of demo example responses with manually extracted claims, which have been judged to be both atomic and comprehensive (i.e., the claims list covers all significant information from the response text). Note that unlike some existing approaches (Min et al., 2023), we don't use sentences in responses to decompose into claims. Rather, we directly extract the list of claims from the full response text since a single claim may incorporate information from more than one sentence (e.g. entity resolution, reasoning).
+
+We utilize the Amazon Bedrock hosted Claude 2.1 model (with temperature  $= 0.8$  and top  $p = 0.9$ ) (Amazon Web Services) to develop our claim extractor. The process involves developing a prompt that enables the model to learn how to extract claims from an LLM response. The prompt structure begins with an initial instruction, followed by a set of rules outlining the task requirements. It also includes a selection of example texts accompanied by their manually extracted claims. Finally, the prompt ends with the target response (i.e., LLM response under evaluation) from which the model needs to extract the relevant claims. By providing this comprehensive prompt, we aim to effectively teach (without updating weights) Claude to accurately extract claims from any given response. Once Claude returns the claims as a text string, we convert it into a Python list to store as a list of claims associated with the response. Figure 3 in the appendix provides the prompt for claim extraction.
+
+<table><tr><td>Number</td><td>A claim has a different number than the original context (e.g. 20% vs. 0.7%). Any number, including year, dimensions, ages, etc.</td></tr><tr><td>Entity</td><td>A claim includes swapped, incorrectly specified, or inserted noun phrases (e.g. one named entity used in a context where another word is expected).</td></tr><tr><td>False Concatenation</td><td>A claim incorrectly combines information about multiple entities or events.</td></tr><tr><td>Attribution Failure</td><td>A claim lacks proper attribution, either crediting the wrong source or presenting information as fact without citation.</td></tr><tr><td>Overgeneralization</td><td>A claim is based on accurate contextual information but is too broad or too general to be supported by the context.</td></tr><tr><td>Reasoning Error</td><td>A claim is based on accurate contextual information but contains a reasoning error or makes an unsupported conclusion.</td></tr><tr><td>Hyperbole</td><td>A claim is based on accurate information but exaggerated or overstated.</td></tr><tr><td>Temporal</td><td>A claim does not accurately incorporate tense, modality (e.g. might vs. will), or time reference in relation to the context.</td></tr><tr><td>Context-based meaning error</td><td>A claim includes incorrect interpretation of idiomatic language, homonyms, or words with multiple meanings, therefore failing to capture the intended meaning.</td></tr><tr><td>Other</td><td>All other types of errors are captured in this category. This includes too-far inferences, circumstantial errors, or incoherent sentences or paragraphs.</td></tr></table>
+
+Table 2: Specific subtypes of hallucinations.
+
+# 3.2 Classifying Claims into Hallucination Labels
+
+When comparing claims against a reference context, our primary distinction is between claims that are supported vs. unsupported by the context. A claim is supported if, under normal circumstances, a reader would believe the claim to be true given the context. We divide unsupported claims into three main types: 1. Contradicted claims: the context contains information which is explicitly inconsistent with the claim ('intrinsic hallucination'). 2. Absent claims: there is no evidence in the context to support or refute the claim ('extrinsic hallucination'), and 3. Partially Supported claims: the claim is almost fully supported by the context but has a minor error. We note that the categories Supported, Contradicted, and Absent are similar to Entailed, Contradicted, and Neutral in NLI terminology. However, the NLI labels may imply a continuum from entailment to contradiction, in which neutral appears to be a lesser error - yet many of the most stereotypical and most problematic examples of hallucination are those in which a model adds completely new, unsupported content in its output. At the same time, we recognize that there are in fact degrees of severity of errors, and we therefore distinguish a class of Partially Supported claims. Distinguishing these claims which are unsupported in subtle ways allows us to both quantify the presence of hallucination errors, and, at least at a basic level, distinguish their severity.
+
+Identifying Partially Supported claims is unique to our approach. Examples include: missing/incorrect attribution, number/conjunction misinterpretation, and mild hyperbole. For example, if the context states, "According to the company, their revenue increased" then the claim "The company's revenue increased" would be Partially Supported. In addition to the above claim types, we have an Unevaluable class label for claims that do not fall into any of the high-level types (e.g., questions).
+
+Within unsupported claims, we identify specific subtypes of hallucinations in order to understand and compare hallucinations from specific models in more detail, and to potentially apply such information to strategies for reducing or mitigating LLM hallucinations. We currently distinguish 10 subtypes (see Table 2), though we will continue to refine our categorization based on emerging data as part of our future work. Subtypes of hallucination are exemplified in Table 9 in the appendix.
+
+Unlike the traditional BERT-based approaches in some recent studies (Zha et al., 2023; Vectara, 2023), we leverage in-context prompting and develop the claim classifier with the Amazon Bedrock-hosted Claude 3 Sonnet model (with temperature  $\approx 0.1$  for reproducibility). Notably, our classifier not only detects the main hallucination labels but also identifies specific subtypes (as elaborated in Table 9), and provides an explanation for the claim label. To analyze the effectiveness of our approach, we have developed four prompting
+
+<table><tr><td>Prompting LLM with CoT</td><td>Prompting LLM without CoT</td></tr><tr><td>Step 1: Read and fully understand the claim. It is a short, standalone sentence containing a single piece of information related to the source text.</td><td>Step 1: Read and fully understand the claim. It is a short, standalone sentence containing a single piece of information related to the source text.</td></tr><tr><td>Step 2: Thoroughly analyze how the claim relates to the information in the source text. Then, write your reasoning in 1-3 sentences to determine the most appropriate label to describe the claim&#x27;s truthfulness based on the source text.</td><td>Step 2: Write the most appropriate label for the claim based on the source text.</td></tr><tr><td>Step 3: Write the label for the claim based on your reasoning in Step 2.</td><td>Step 3: If the label in Step 2 is ‘contradicted’, ‘absent’, or ‘partially supported’, then write the specific error type (i.e., sublabel) for the claim.</td></tr><tr><td>Step 4: If the label in Step 3 is ‘contradicted’, ‘absent’, or ‘partially supported’, then thoroughly analyze the specific errors (i.e., sublabels) present in the claim based on provided source text. Then, provide your reasoning in 1-3 sentences to determine the error. However, if the label in Step 3 is ‘supported’, simply write ‘None - claim is supported’, and set the sublabel to ‘None’.</td><td></td></tr><tr><td>Step 5: Write sublabel based on your reasoning in Step 4.</td><td></td></tr></table>
+
+Table 3: Instruction steps for prompting LLM with and without CoT.
+
+strategies to investigate two key aspects: 1) the potential benefits of incorporating Chain-of-Thought (CoT) reasoning (Wei et al., 2022), and 2) whether evaluating claims individually (requiring multiple Claude API calls) offers advantages over evaluating all claims together (requiring a single API call). We provide the four prompt setups below.
+
+(1) With and (2) Without CoT Reasoning. We check if asking the model to think and analyze before deciding the labels and specific error subtypes (i.e., sublabels) is beneficial. By following Wei et al. (2022), we develop a 5 step CoT prompt, including steps to thoroughly examine each claim's faithfulness to the reference context, and writing down the reasoning behind the thoughts. Similar to other reasoning tasks where CoT is useful (Wei et al., 2022) like mathematical and commonsense reasoning, we hypothesize that these written thoughts provide insights into the model's reasoning process prior to selecting the final hallucination label for each claim. Table 3 shows the steps for with and without CoT prompting strategies. See Figures 5 and 4 for prompt templates. Figure 2 shows CoT model output with Thoughts generated for hallucination label and error type sublabels.
+
+(3) One-claim-eval and (4) All-claims-eval. We check if evaluating each claim separately is better than evaluating all claims together. The first approach evaluates each claim independently against the context by making multiple API calls to Claude
+
+Depending on the number of claims in a response). The second approach includes all the claims in the same prompt and makes a single API call to Claude. We refer to the former as one-claim-eval and the latter as all-claims-eval. While all-claims-eval is better for reducing latency, one-claim-eval performs better as it allows the model to focus on only one claim at the time of evaluation. See Figures 6 and 7 for prompt templates.
+
+# 3.3 Aggregating Scores for Hallucination Measurement
+
+After classifying claims, each claim has a label (out of 5 labels) and an error type (out of 10 types) assigned to it. Now, we assign a score for hallucination and for each error type by aggregating across all claims in the responses in the dataset. We produce two scores: 1) Response hallucination rate, calculated by dividing the number of unsupported claims (combining different error types) by the total number of claims for the response. 2) Error type distribution, calculated by scoring subtype or class errors separately.
+
+# 4 Experiments and Results
+
+# 4.1 Dataset
+
+We present the datasets used to evaluate our HalluMeasure approach and compare with existing models. We create the first dataset, TechNewsSumm, by sampling 30 tech news articles from the CNN/Dailymail dataset, collecting summaries (20
+
+<table><tr><td rowspan="2"></td><td colspan="3">Response-level</td><td colspan="3">Claim-level</td></tr><tr><td>Precision</td><td>Recall</td><td>F1</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>Vectara HHEM</td><td>0.75</td><td>0.15</td><td>0.25</td><td>-</td><td>-</td><td>-</td></tr><tr><td>AlignScore</td><td>0.57</td><td>0.20</td><td>0.30</td><td>-</td><td>-</td><td>-</td></tr><tr><td>RefChecker (AlignScore Checker)</td><td>0.68</td><td>0.75</td><td>0.71</td><td>-</td><td>-</td><td>-</td></tr><tr><td>RefChecker (NLI Checker)</td><td>0.75</td><td>0.75</td><td>0.75</td><td>-</td><td>-</td><td>-</td></tr><tr><td>RefChecker (Claude Checker)</td><td>0.79</td><td>0.75</td><td>0.77</td><td>-</td><td>-</td><td>-</td></tr><tr><td>W/o CoT + all-claims-eval (ours)</td><td>0.73</td><td>0.80</td><td>0.76</td><td>0.72</td><td>0.59</td><td>0.65</td></tr><tr><td>W/ CoT + all-claims-eval (ours)</td><td>0.85</td><td>0.85</td><td>0.85</td><td>0.73</td><td>0.67</td><td>0.70</td></tr><tr><td>W/o CoT + one-claim-eval (ours)</td><td>0.83</td><td>0.75</td><td>0.79</td><td>0.86</td><td>0.56</td><td>0.68</td></tr><tr><td>W/ CoT + one-claim-eval (ours)</td><td>0.89</td><td>0.85</td><td>0.87</td><td>0.87</td><td>0.66</td><td>0.75</td></tr></table>
+
+Table 4: Results on TechNewsSumm Dataset: Response- and claim-level evaluation metrics for the unsupported label obtained with the existing methods and the four settings of our approach. For a fair comparison, we convert our four main labels into binary labels (i.e., supported and unsupported).  
+
+<table><tr><td></td><td>API Call Time</td><td>#Prompt Tokens (K)</td><td>#Output Tokens (K)</td></tr><tr><td>W/o CoT + all-claims-eval</td><td>9.45</td><td>6.27</td><td>0.51</td></tr><tr><td>W/o CoT + one-claim-eval</td><td>42.24</td><td>81.51</td><td>0.52</td></tr><tr><td>W/ CoT + all-claims-eval</td><td>40.21</td><td>10.14</td><td>1.27</td></tr><tr><td>W/ CoT + one-claim-eval</td><td>77.17</td><td>133.10</td><td>1.58</td></tr></table>
+
+Table 5: Latency and Token Stats on TechNewsSumm Experiments: Claude API call duration and Input/Output tokens for HalluMeasure's Claim Classifier. Duration in seconds, token counts in thousands (K). For one-claim-eval setups, prompt and output tokens from all API calls for a single response are summed.
+
+from the Cohere Command model and 10 human-written), extracting atomic claims (400 claims) from the summaries using our claim extractor (Section 2), and manually evaluating the claims against the reference context to identify three main types and 10 specific types (i.e., subtypes) of hallucinations (Table 9). We observe moderate agreement between the annotators in main labels (Kappa: 0.44) and subtypes annotations (Kappa: 0.45). Our second dataset is the SummEval dataset (Fabbri et al., 2021) with 1600 annotated samples (hallucination vs. non-hallucination) at the response level from the TRUE benchmark (Honovich et al., 2022).
+
+# 4.2 Baselines
+
+We compare HalluMeasure with three state-of-the-art hallucination measurement approaches: Vectara HHEM (Vectara, 2023) outputs a factual consistency score (0-1) using a finetuned cross-encoder model. We use 1 - factual consistency score as hallucination score and threshold at 0.5 for label assignment. AlignScore (Zha et al., 2023) classifies claims against contexts into aligned/not-aligned classes using a RoBERTa model trained on 7 NLP tasks. We use 1 - aligned class score as hallucination score and threshold at 0.5. RefChecker (Hu et al., 2023) splits responses into claim-triplets and checks them against references using a claim
+
+checker (GPT4, Claude2, NLI, or AlignScore). It produces hallucination scores based on strict (any contradicted claim means the response is labeled as hallucination), soft (ratio of contradicted and neutral claims), or majority voting criteria. We use the soft criteria for hallucination score and strict for label assignment.
+
+# 4.3 Hallucination Measurement Experiments
+
+We provide experimental results for these four hallucination measurement approaches (HalluMeasure, RefChecker, AlignScore, and Vectara HHEM). Tables 4 and 6 present evaluation results on binary classification and error subtype classification on our own curated TechNewsSumm dataset. Table 7 shows results on the SummEval dataset.
+
+Through our experiments, we attempt to answer our research questions (RQs) below:
+
+# RQ1: How effectively can HalluMeasure extract claims from LLM responses?
+
+We validate our claim extractor's accuracy by evaluating its performance on 25 tech news articles from the CNN/DailyMail dataset. We generate summaries using Cohere's Command model and extract claims from these summaries. Four researchers manually annotated the claims; each claim was annotated by two annotators, with disagreements adjudicated. The rate of claims with
+
+<table><tr><td></td><td>#Adj.</td><td>#Pred.</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>Number</td><td>1</td><td>0</td><td>0.00</td><td>0.00</td><td>0.00</td></tr><tr><td>Entity</td><td>3</td><td>9</td><td>0.11</td><td>0.50</td><td>0.18</td></tr><tr><td>False Concatenation</td><td>6</td><td>1</td><td>0.00</td><td>0.00</td><td>0.00</td></tr><tr><td>Attribution Failure</td><td>1</td><td>0</td><td>0.00</td><td>0.00</td><td>0.00</td></tr><tr><td>Overgeneralization</td><td>6</td><td>13</td><td>0.15</td><td>0.50</td><td>0.24</td></tr><tr><td>Reasoning Error</td><td>10</td><td>7</td><td>0.14</td><td>0.17</td><td>0.15</td></tr><tr><td>Hyperbole</td><td>3</td><td>2</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td>Temporal</td><td>3</td><td>4</td><td>0.50</td><td>1.00</td><td>0.67</td></tr><tr><td>Context-based meaning</td><td>6</td><td>0</td><td>0.00</td><td>0.00</td><td>0.00</td></tr><tr><td>Other</td><td>48</td><td>28</td><td>0.93</td><td>0.58</td><td>0.71</td></tr><tr><td>Macro-average</td><td></td><td></td><td>0.32</td><td>0.47</td><td>0.33</td></tr><tr><td>Weighted-average</td><td></td><td></td><td>0.71</td><td>0.52</td><td>0.58</td></tr></table>
+
+Table 6: Error Subtypes Classification Performance on TechNewsSumm Dataset: Evaluation results on the specific subtypes of hallucinations obtained with our best setup of HalluMeasure model (i.e., W/ CoT + one-claim-eval). #Adj. denotes the count of adjudicated subtypes, and #Pred. denotes the count of predicted subtypes.
+
+disagreements is  $6.5\%$  (12/185). The adjusted Cohen's kappa (PABAK) score of 0.87 indicates strong annotator agreement. Using the adjudicated claims as ground truth, the claim extractor's precision is 0.96. Since there is no definitive gold set of claims, we compute a revised recall metric as correctly extracted claims / (correctly extracted claims + missing correct claims). The revised recall and F1-score of 0.97 indicate the claim extractor accurately extracts claims from responses.
+
+# RQ2: How does HalluMeasure method compare against state-of-the-art methods?
+
+We answer this question by comparing our method with RefChecker, AlignScore, and Vectara HHEM. Out of these three methods, RefChecker's approach is similar to HalluMeasure with a key distinction that we use CoT few-shot prompting instead of simple few-shot prompting employed by RefChecker. The remaining two methods (AlignScore and Vectara HHEM) are specifically trained to measure hallucinations using BERT-based models. While RefChecker performs similarly to some of our prompting setups (e.g. without CoT), HHEM and AlignScore achieve lower performance on our TechNewsSumm dataset. Our best setup (HalluMeasure W/ CoT + one-claim-eval) outperforms existing models by at least  $13\%$  F1 score on TechNewsSumm dataset (See Table 4).
+
+To demonstrate performance on a public benchmark dataset, we show experimental results on the popular SummEval dataset in Table 7. Note that we report AUC ROC scores on the SummEval dataset as computed using the TRUE benchmark software package (Honovich et al., 2022). Our HalluMeasure model achieves an AUC ROC value of 0.80, outperforming the baseline models by 3 to 9
+
+points on 1600 samples of SummEval. The baseline models' AUC ROC values are: Vectara HHEM: 0.77, AlignScore: 0.71, RefChecker (Alignscore checker): 0.75, RefChecker (NLI checker): 0.75, and RefChecker (Claude 2): 0.74. Our results on SummEval show that HalluMeasure significantly outperforms several existing baseline models. So, results on two datasets show that HalluMeasure outperforms existing state-of-the-art models.
+
+# RQ3: Is a single call to classify all claims effective for hallucination classification?
+
+We answer this question by comparing the results for one-claim-eval vs. all-claims-eval from Table 4. Evaluating one claim at a time is beneficial over evaluating all claims together based on results on both response and claim level evaluations (e.g., F1: 0.75 vs. 0.70). However, all-claims-eval comes with benefit of improved latency and cost compared to one-claim-eval (9.45 secs vs. 42.24 secs for W/o CoT prompt; 40.21 secs vs. 77.71 secs for W/ CoT prompt). See Table 5 for details.
+
+# RQ4: Can HalluMeasure effectively detect fine-grained hallucination error types?
+
+While HalluMeasure's best setup shows excellent overall results, it struggles to accurately classify the specific error types, with macro-F1 and weighted-F1 scores of only 0.33 and 0.58, respectively (Table 6). This is likely due to having 10 different error types, with similar classification issues demonstrated by low human agreement (Kappa: 0.45) as mentioned in Section 4.1. Moreover, some error types are hard to distinguish and may not in fact be mutually exclusive, such as reasoning error vs. context-based meaning, and false concatenation vs. overgeneralization. Improving the accurate identification of these specific error types remains
+
+<table><tr><td></td><td>AUC ROC</td></tr><tr><td>Vectara HHEM</td><td>0.77</td></tr><tr><td>AlignScore</td><td>0.71</td></tr><tr><td>RefChecker (AlignScore Checker)</td><td>0.75</td></tr><tr><td>RefChecker (NLI Checker)</td><td>0.75</td></tr><tr><td>RefChecker (Claude Checker)</td><td>0.74</td></tr><tr><td>HalluMeasure (ours; W/o CoT)</td><td>0.78</td></tr><tr><td>HalluMeasure (ours; W/ CoT)</td><td>0.80</td></tr></table>
+
+Table 7: Performance Comparison on SummEval Benchmark: AUC ROC scores obtained with existing models and HalluMeasure (W/ CoT and W/o CoT + one-claim-eval) on the SummEval dataset (Fabbri et al., 2021).
+
+a priority for future work.
+
+# RQ5: Does few-shot CoT prompting improve the hallucination measurement performance?
+
+We answer this question by comparing the performance of our claim classifier model with and without CoT prompting. Table 4 shows that in both one-claim-eval and all-claims-eval settings, CoT prompting improves model performance on our TechNewsSumm dataset at both response-level (F1: 0.85 vs. 0.76) and claim-level (0.7 vs. 0.65). Table 7 also shows that CoT prompting improves model performance on SummEval public benchmark dataset (AUC ROC: 0.78 vs. 0.80).
+
+# RQ6: How effectively can HalluMeasure's method generalize to different LLMs for hallucination classification?
+
+We present additional results with Cohere's Command R+ and Mistral Large models on the TechNewsSumm dataset in Table 8. We only experiment with the one-claim-eval setup due to read timeout issues with the all-claims-eval setup; and we obtain only response-level results due to the additional annotation effort required for claim-level analysis. The results show that, unlike Claude Sonnet 3.0, Command R+ and Mistral Large exhibit similar results for both with and without CoT approaches, with F1 scores ranging from 0.77 to 0.79. However, these scores are significantly lower than the best results we achieved with Claude Sonnet (0.87). We also check HalluMeasure's performance with these LLMs on the SummEval dataset. Surprisingly, Mistral Large performs slightly better than Claude, achieving an AUC ROC score of 0.81, while Command R+ scores 0.73. These results show that performance may be sensitive to characteristics of the dataset, and proprietary LLMs do not always out-perform open source models.
+
+<table><tr><td></td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td colspan="4">Cohere&#x27;s Command R+</td></tr><tr><td>W/o CoT + one-claim-eval</td><td>0.79</td><td>0.75</td><td>0.77</td></tr><tr><td>W/ CoT + one-claim-eval</td><td>0.76</td><td>0.80</td><td>0.78</td></tr><tr><td colspan="4">Mistral large</td></tr><tr><td>W/o CoT + one-claim-eval</td><td>0.68</td><td>0.95</td><td>0.79</td></tr><tr><td>W/ CoT + one-claim-eval</td><td>0.69</td><td>0.90</td><td>0.78</td></tr></table>
+
+Table 8: Response-level results from Cohere's Command R+ and Mistral Large LLMs on TechNewsSumm.
+
+# 5 Conclusion
+
+We introduce HalluMeasure, a novel approach to automatically measure hallucinations in LLM responses. HalluMeasure decomposes an LLM response into set of claims using a claim extraction model based on Claude with few-shot prompting. It compares the extracted claims against a context document using a claim classification model leveraging few-shot Chain-of-Thought prompting with Claude to enhance classification performance. An aggregated response-level score is produced by measuring the rate of unsupported claims and distribution of specific error types. We demonstrate the effectiveness of HalluMeasure on: TechNewsSumm (our own curated dataset with detailed claim-level error labels) and SummEval (a popular benchmark dataset). Our results demonstrate HalluMeasure's superior performance over baseline models, with at least a 10-point F1 score improvement on TechNewsSumm and a 3-point AUC ROC increase on SummEval. For future work, we plan to employ dynamic few-shot prompting and use optimized prompts with fast-inference LLMs.
+
+# Limitations
+
+Our study detects 10 hallucination error types, including an Other class. We plan to further explore and refine error categorization for better detection and to support mitigation. Additionally, we have focused mainly on hallucination detection in plain text responses, and have yet to explore detecting and measuring hallucinations in other formats like tables and code. While our focus to date has been on news article benchmarks, we aim to include specialized domains like medicine, law, and finance.
+
+# Acknowledgement
+
+The authors would like to thank Lisa Baytler, UX designer, for her assistance in creating figures for this paper.
+
+# References
+
+Amazon Web Services. Amazon bedrock. https://aws.amazon.com/bedrock/?refid=36201f68-a9b0-45cc-849b-8ab260660e1c. Accessed: 05-30-2024.  
+Zhoujun Cheng, Jungo Kasai, and Tao Yu. 2023. Batch prompting: Efficient inference with large language model apis. arXiv preprint arXiv:2301.08721.  
+I-Chun Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, and Pengfei Liu. 2023. Factool: Factuality detection in generative ai - a tool augmented framework for multi-task and multi-domain scenarios. Preprint, arXiv:2307.13528.  
+CNBC. 2023. Judge sanctions lawyers for brief written by a.i. with fake citations.  
+Alexander R. Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. Summeval: Re-evaluating summarization evaluation. Preprint, arXiv:2007.12626.  
+Saadia Gabriel, Asli Celikyilmaz, Rahul Jha, Yejin Choi, and Jianfeng Gao. 2021. Go figure: A meta evaluation of factuality in summarization. Preprint, arXiv:2010.12834.  
+Zorik Gekhman, Jonathan Herzig, Roee Aharoni, Chen Elkind, and Idan Szpektor. 2023. Trueteacher: Learning factual consistency evaluation with large language models. Preprint, arXiv:2305.11171.  
+Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and Yossi Matias. 2022. True: Re-evaluating factual consistency evaluation. Preprint, arXiv:2204.04991.  
+Xiangkun Hu, Dongyu Ru, Qipeng Guo, Lin Qiu, and Zheng Zhang. 2023. Refchecker for fine-grained hallucination detection.  
+Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. Preprint, arXiv:2311.05232.  
+Wojciech Kryscinski, Bryan McCann, Caiming Xiong, and Richard Socher. 2020. Evaluating the factual consistency of abstractive text summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9332-9346, Online. Association for Computational Linguistics.  
+Wojciech Krysciński, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Evaluating the factual consistency of abstractive text summarization. Preprint, arXiv:1910.12840.
+
+Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2021. Summac: Re-visiting nlibased models for inconsistency detection in summarization. Preprint, arXiv:2111.09525.  
+R. Lebret, D. Grangier, and M. Auli. 2016. Neural Text Generation from Structured Data with Application to the Biography Domain. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023. Halueval: A large-scale hallucination evaluation benchmark for large language models. Preprint, arXiv:2305.11747.  
+Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Truthfulqa: Measuring how models mimic human falsehoods. Preprint, arXiv:2109.07958.  
+Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. 2023. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. Preprint, arXiv:2303.08896.  
+Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettle-moyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076-12100, Singapore. Association for Computational Linguistics.  
+Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin Vechev. 2024. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. Preprint, arXiv:2305.15852.  
+Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. ArXiv, abs/1808.08745.  
+Vipula Rawte, Swagata Chakraborty, Agnibh Pathak, Anubhav Sarkar, S. M Towhidul Islam Tonmoy, Aman Chadha, Amit P. Sheth, and Amitava Das. 2023. The troubling emergence of hallucination in large language models - an extensive definition, quantification, and prescriptive remediations. Preprint, arXiv:2310.04988.  
+Reuters. 2023. Alphabet shares dive after google air chatbot bard flubs answer in ad.  
+Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073-1083, Vancouver, Canada. Association for Computational Linguistics.  
+Derek Tam, Anisha Mascarenhas, Shiyue Zhang, Sarah Kwan, Mohit Bansal, and Colin Raffel. 2023. Evaluating the factual consistency of large language
+
+models through news summarization. Preprint, arXiv:2211.08412.  
+Liyan Tang, Tanya Goyal, Alexander R. Fabbri, Philippe Laban, Jiacheng Xu, Semih Yavuz, Wojciech Kryscinski, Justin F. Rousseau, and Greg Durrett. 2023. Understanding factual errors in summarization: Errors, summarizers, datasets, error detectors. Preprint, arXiv:2205.12854.  
+Liyan Tang, Igor Shalyminov, Amy Wing mei Wong, Jon Burnsky, Jake W. Vincent, Yu'an Yang, Siffi Singh, Song Feng, Hwanjun Song, Hang Su, Lijia Sun, Yi Zhang, Saab Mansour, and Kathleen McKeown. 2024. Tofueval: Evaluating hallucinations of llms on topic-focused dialogue summarization. Preprint, arXiv:2402.13249.  
+Ars Technica. 2024. Air canada must honor refund policy invented by airline's chatbot.  
+Vectara. 2023. Vectara hughes hallucination evaluation model.  
+Yuxia Wang, Minghan Wang, Muhammad Arslan Manzoor, Fei Liu, Georgi Georgiev, Rocktim Jyoti Das, and Preslav Nakov. 2024. Factuality of large language models in the year 2024. Preprint, arXiv:2402.02420.  
+Miriam Wanner, Seth Ebner, Zhengping Jiang, Mark Dredze, and Benjamin Van Durme. 2024. A closer look at claim decomposition. arXiv preprint arXiv:2403.11903.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. 2024. Hallucination is inevitable: An innate limitation of large language models. Preprint, arXiv:2401.11817.  
+Yuheng Zha, Yichi Yang, Ruichen Li, and Zhiting Hu. 2023. AlignScore: Evaluating factual consistency with a unified alignment function. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11328-11348, Toronto, Canada. Association for Computational Linguistics.  
+Rongxin Zhu, Jianzhong Qi, and Jey Han Lau. 2023. Annotating and detecting fine-grained factual errors for dialogue summarization. Preprint, arXiv:2305.16548.
+
+# A Prompts
+
+We show the following prompts: claim extraction in Figure 3; claim classification without CoT + all-claims-eval in Figure 4; with CoT + all-claims-eval prompt in Figure 5; without CoT + one-claim-eval prompt in Figure 6; with CoT + one-claim-eval
+
+prompt in Figure 7. Note that we have truncated the examples from the prompts to save space.
+
+# B Error Type Classification Examples
+
+See Table 9 for error type examples.
+
+# C Qualitative Example
+
+We show example #1 in Table 10. Note that news article is the context input document, and Cohere Command output summary is the response under evaluation. We show one claim extracted from the response for analysis. Note that the human annotated label for the claim is "Absent" since the news article does not mention the status of "Chris Hadfield". However, HalluMeasure W/o CoT model labels the claim as "Supported". This could be because the model does not reason properly when generating the class label for a claim. When HalluMeasure is executed W/ CoT prompting, the claim is correctly labeled as "Absent". In addition, we show the explanation automatically generated by our model about why the claim is labeled "Absent".
+
+<table><tr><td>Error type</td><td>Example context</td><td>Example claim with error</td></tr><tr><td>Number</td><td>The company&#x27;s drones will begin delivering products in Sydney in early 2014.</td><td>The company&#x27;s drone deliveries will begin in 2013.</td></tr><tr><td>Entity</td><td>New Zealand company Martin Jetpack...</td><td>Martin Jetpack is based in Sydney.</td></tr><tr><td>False Concatenation</td><td>The company will be listed on the stock exchange within the next few months.... Their product will be released early next year.</td><td>The product will be released within the next few months.</td></tr><tr><td>Attribution Failure</td><td>According to Cooper, drone technology is currently under-regulated...</td><td>Drone technology is under-regulated.</td></tr><tr><td>Overgeneralization</td><td>This product is designed with teachers and students of STEM in mind...</td><td>This product develops math skills.</td></tr><tr><td>Reasoning Error</td><td>We depend on the Atlas V rocket, which carries many of our most important satellites and is powered by the Russian-made RD-180 rocket engine.</td><td>We rely on Russia&#x27;s Atlas V rocket</td></tr><tr><td>Hyperbole</td><td>The technology will significantly improve driver safety.</td><td>The technology will revolutionize driver safety.</td></tr><tr><td>Temporal</td><td>The company will use drones to deliver packages.</td><td>The company uses drones.</td></tr><tr><td>Context-based meaning error</td><td>The package includes a Bluetooth system that lets users turn their Roomba into a DJ.</td><td>DJs can instruct the robot.</td></tr><tr><td>Other</td><td>[Context contains no information about the army unveiling a robot called Atlas.]</td><td>The army previously unveiled the Atlas robot.</td></tr></table>
+
+Table 9: Examples of specific subtypes of hallucinations.
+
+# Task
+
+```txt
+A claim is a short sentence containing a single piece of information.  
+You will extract claims from a given text inside <text></text> XML tags.
+```
+
+# Task-rules
+
+```txt
+Here are the "Task-rules" you must follow when generating the claims. <task-rules> <rule>The claim should be entirely self-contained. For instance, the claim should be comprehended without relying on other claims.</rule> <rule>The claim should not contain pronouns. If there are pronouns in the input text, replace them with their corresponding nouns when generating the claims.</rule> <rule>The claim should not exceed 15 words.</rule> <rule>You will always output a list of the extracted claims.</rule> <rule>You will always change double quotes to single quotes in the claims. For example, write 'glass' instead of "glass".</rule> </task-rules>
+```
+
+# Example
+
+```txt
+An example is given below:  
+<example>  
+...  
+</example>
+```
+
+# Input
+
+```yaml
+<text>:
+    #TARGET_TEXT# </text>
+\n\nAssistant:
+```
+
+Figure 3: Claim extractor prompt.
+
+# Task
+
+# <task>
+
+You will act as an expert annotator to evaluate a claim against a provided source text.
+
+The source text will be given within <source>...</source> XML tags
+
+The claims to evaluate will be provided within <claims>...</claims> XML tags.
+
+For each claim, follow these steps:
+
+Step 1: Read and fully understand the claim. It is a short, standalone sentence containing a single piece of information related to the source text.
+
+Step 2: Write the most appropriate label for the claim based on the source text. The definitions of the labels are provided below.
+
+1. supported: The claim is definitely true according to the source text.  
+2. contradicted: The claim is definitely false according to the source text.  
+3. absent: The source text provides no information to confirm or deny the claim.  
+4. partially supported: The claim is mostly consistent with the source text, but some words are added, omitted, or inaccurate. This is a 'near miss' label. If either '2-contradicted' or '3-absent' seems like an appropriate label, use these instead. Note that claims that should be attributed to a particular source (e.g. quotes) are '4-partially supported' if the attribution is missing.  
+5. unevaluatable: The claim cannot be interpreted as a statement to evaluate against the source text (e.g. it is a question or instruction).
+
+Step 3: If the label in Step 2 is '2-contradicted', '3-absent', or '4-partially supported', then write the specific error type (i.e., sublabel) for the claim. Given below are the definitions of those specific error types.
+
+1. number: the claim has a different number than the original context (e.g., 50m vs. 60m).  
+2. entity: the claim includes swapped or incorrectly selected entities (e.g., one named entity vs. another).  
+3. false-concat: the claim inappropriately combines information about two or more different entities or events.  
+4. attribution-failure: the claim does not correctly attribute the information it contains, e.g. failing to attribute quoted material to the correct person or entity named in the article.  
+5. overgeneralization: the claim is based on accurate contextual information but is too broad or too general to be supported by the context.  
+6. reasoning-error: the claim is based on accurate contextual information but includes a reasoning error which makes the claim inaccurate.  
+7. hyperbole: the claim is based on accurate contextual information but is inappropriately strengthened or overstated.  
+8. temporal: the claim does not accurately incorporate tense, modality (e.g. might vs. will) or time reference in relation to the context.  
+9. context-based-meaning: the claim includes incorrect interpretation of idiomatic language, homonyms, or words with multiple meanings, therefore failing to capture the intended meaning  
+10. other: all other types of errors are captured in this category.
+
+Repeat these steps for each claim provided.
+
+```txt
+You will output a list of dictionaries. Each dictionary will have below format:  
+{  
+"claim": "<CLAIM TEXT>",  
+"label": "<CLAIM LABEL>",  
+"sublabel": "CLAIM SUBLABEL"  
+}  
+</task>
+```
+
+# Example
+
+```txt
+Given below are the examples for you to comprehend the task:  
+<example>  
+...  
+</example>
+```
+
+# Source
+
+```txt
+<source>:
+{[SOURCE_TEXT]}</source>
+\n
+<claims>
+{[TARGET Claims]}</claims>
+```
+
+Figure 4: Claim classifier prompt for without  $CoT +$  all-claim-eval setup.
+
+# Task
+
+# <task>
+
+You will act as an expert annotator to evaluate claims against a provided source text. The source text will be given within <source>...</source> XML tags  
+The claims to evaluate will be provided within <claims>...</claims> XML tags.
+
+For each claim, follow these steps:
+
+Step 1: Read and fully understand the claim. It is a short, standalone sentence containing a single piece of information related to the source text.  
+Step 2: Thoroughly analyze how the claim relates to the information in the source text. Then, write your reasoning in 1-3 sentences to determine the most appropriate label to describe the claim's truthfulness based on the source text. The definitions of the labels are provided below.
+
+1. supported: The claim is definitely true according to the source text.  
+2. contradicted: The claim is definitely false according to the source text.  
+3. absent: The source text provides no information to confirm or deny the claim.  
+4. partially supported: The claim is mostly consistent with the source text, but some words are added, omitted, or inaccurate. This is a 'near miss' label. If either '2-contradicted' or '3-absent' seems like an appropriate label, use these instead. Note that claims that should be attributed to a particular source (e.g. quotes) are '4-partially supported' if the attribution is missing.  
+5. unevaluatable: The claim cannot be interpreted as a statement to evaluate against the source text (e.g. it is a question or instruction).
+
+Step 3: Write the label for the claim based on your reasoning in Step 2.
+
+Step 4: If the label in Step 3 is '2-contradicted', '3-absent', or '4-partially supported', then thoroughly analyze the specific error or inaccuracy present in the claim based on provided source text. Then, provide your reasoning in 1-3 sentences to determine the error. Given below are the definitions of those specific error types (i.e., sublabels). However, if the label in Step 3 is '1-supported', simply write 'None - claim is supported', and set the sublabel to 'None'.
+
+1. number: the claim has a different number than the original context (e.g., 50m vs. 60m).  
+2. entity: the claim includes swapped or incorrectly selected entities (e.g., one named entity vs. another).  
+3. false-concat: the claim inappropriately combines information about two or more different entities or events.  
+4. attribution-failure: the claim does not correctly attribute the information it contains, e.g. failing to attribute quoted material to the correct person or entity named in the article.  
+5. overgeneralization: the claim is based on accurate contextual information but is too broad or too general to be supported by the context.  
+6. reasoning-error: the claim is based on accurate contextual information but includes a reasoning error which makes the claim inaccurate.  
+7. hyperbole: the claim is based on accurate contextual information but is inappropriately strengthened or overstated.  
+8. temporal: the claim does not accurately incorporate tense, modality (e.g. might vs. will) or time reference in relation to the context.  
+9. context-based-meaning: the claim includes incorrect interpretation of idiomatic language, homonyms, or words with multiple meanings, therefore failing to capture the intended meaning
+
+10. other: all other types of errors are captured in this category.
+
+Step 5: Write the sublabel based on your reasoning in Step 4.
+
+Repeat these steps for each claim provided.
+
+You will output a list of dictionaries. Each dictionary will have below format:
+
+```javascript
+"claim": "<CLAIM TEXT>", "thought_for_label": "<THOUGHT FOR LABEL FOR THE CLAIM>", "label": "<CLAIM LABEL>", "thought_for_sublabel": "<THOUGHT FOR SUBLABEL FOR THE CLAIM>", "sublabel": "CLAIM SUBLABEL" } </task>
+```
+
+# Example
+
+```txt
+Given below are the examples for you to comprehend the task:  
+<example>  
+...  
+</example>
+```
+
+# Source
+
+```txt
+<source>:
+{[SOURCE_TEXT]}</source>
+\n
+<claims>
+{TARGET Claims} </claims>
+```
+
+Figure 5: Claim classifier prompt for with  $CoT +$  all-claims-eval setup.
+
+# Task
+
+<task>
+
+You will act as an expert annotator to evaluate a claim against a provided source text.
+
+The source text will be given within <source>...</source> XML tags
+
+The claim to evaluate will be provided within <claim>...</claim> XML tags.
+
+For each claim, follow these steps:
+
+Step 1: Read and fully understand the claim. It is a short, standalone sentence containing a single piece of information related to the source text.
+
+Step 2: Write the most appropriate label for the claim based on the source text. The definitions of the labels are provided below.
+
+1. supported: The claim is definitely true according to the source text.  
+2. contradicted: The claim is definitely false according to the source text.  
+3. absent: The source text provides no information to confirm or deny the claim.  
+4. partially supported: The claim is mostly consistent with the source text, but some words are added, omitted, or inaccurate. This is a 'near miss' label. If either '2-contradicted' or '3-absent' seems like an appropriate label, use these instead. Note that claims that should be attributed to a particular source (e.g. quotes) are '4-partially supported' if the attribution is missing.  
+5. unequalatable: The claim cannot be interpreted as a statement to evaluate against the source text (e.g. it is a question or instruction).
+
+Step 3: If the label in Step 2 is '2-contradicted', '3-absent', or '4-partially supported', then write the specific error type (i.e., sublabel) for the claim. Given below are the definitions of those specific error types.
+
+1. number: the claim has a different number than the original context (e.g., 50m vs. 60m).  
+2. entity: the claim includes swapped or incorrectly selected entities (e.g., one named entity vs. another).  
+3. false-concat: the claim inappropriately combines information about two or more different entities or events.  
+4. attribution-failure: the claim does not correctly attribute the information it contains, e.g. failing to attribute quoted material to the correct person or entity named in the article.  
+5. overgeneralization: the claim is based on accurate contextual information but is too broad or too general to be supported by the context.  
+6. reasoning-error: the claim is based on accurate contextual information but includes a reasoning error which makes the claim inaccurate.  
+7. hyperbole: the claim is based on accurate contextual information but is inappropriately strengthened or overstated.  
+8. temporal: the claim does not accurately incorporate tense, modality (e.g. might vs. will) or time reference in relation to the context.  
+9. context-based-meaning: the claim includes incorrect interpretation of idiomatic language, homonyms, or words with multiple meanings, therefore failing to capture the intended meaning
+
+10. other: all other types of errors are captured in this category.
+
+```txt
+For the claim, you will output a dictionary with the below format: { "claim": "<CLAIM TEXT>", "label": "<CLAIM LABEL>", "sublabel": "CLAIM SUBLABEL" } </task>
+```
+
+# Example
+
+Given below are the examples for you to comprehend the task. Each example includes a list of claims and a corresponding list of dictionaries as responses to aid your understanding. However, the input you will receive consists of a single claim, and you will need to provide a dictionary response for that particular claim.
+
+```twig
+<example>  
+...  
+</example>
+```
+
+# Source
+
+```txt
+<source>:
+{<SOURCE_TEXT>} </source>
+\n
+<claims>
+{<TARGET Claims>} </claims>
+```
+
+Figure 6: Claim classifier prompt for without CoT + one-claim-eval setup.
+
+# Task
+
+# <task>
+
+You will act as an expert annotator to evaluate claims against a provided source text. The source text will be given within <source>...</source> XML tags The claim to evaluate will be provided within <claim>...</claim> XML tags.
+
+For each claim, follow these steps:
+
+Step 1: Read and fully understand the claim. It is a short, standalone sentence containing a single piece of information related to the source text.  
+Step 2: Thoroughly analyze how the claim relates to the information in the source text. Then, write your reasoning in 1-3 sentences to determine the most appropriate label to describe the claim's truthfulness based on the source text. The definitions of the labels are provided below.
+
+1. supported: The claim is definitely true according to the source text.  
+2. contradicted: The claim is definitely false according to the source text.  
+3. absent: The source text provides no information to confirm or deny the claim.  
+4. partially supported: The claim is mostly consistent with the source text, but some words are added, omitted, or inaccurate. This is a 'near miss' label. If either '2-contradicted' or '3-absent' seems like an appropriate label, use these instead. Note that claims that should be attributed to a particular source (e.g. quotes) are '4-partially supported' if the attribution is missing.  
+5. unevaluable: The claim cannot be interpreted as a statement to evaluate against the source text (e.g. it is a question or instruction).
+
+Step 3: Write the label for the claim based on your reasoning in Step 2.
+
+Step 4: If the label in Step 3 is '2-contradicted', '3-absent', or '4-partially supported', then thoroughly analyze the specific error or inaccuracy present in the claim based on provided source text. Then, provide your reasoning in 1-3 sentences to determine the error. Given below are the definitions of those specific error types (i.e., sublabels). However, if the label in Step 3 is '1-supported', simply write 'None - claim is supported', and set the sublabel to 'None'.
+
+1. number: the claim has a different number than the original context (e.g., 50m vs. 60m).  
+2. entity: the claim includes swapped or incorrectly selected entities (e.g., one named entity vs. another).  
+3. false-concat: the claim inappropriately combines information about two or more different entities or events.  
+4. attribution-failure: the claim does not correctly attribute the information it contains, e.g. failing to attribute quoted material to the correct person or entity named in the article.  
+5. overgeneralization: the claim is based on accurate contextual information but is too broad or too general to be supported by the context.  
+6. reasoning-error: the claim is based on accurate contextual information but includes a reasoning error which makes the claim inaccurate.  
+7. hyperbole: the claim is based on accurate contextual information but is inappropriately strengthened or overstated.  
+8. temporal: the claim does not accurately incorporate tense, modality (e.g. might vs. will) or time reference in relation to the context.  
+9. context-based-meaning: the claim includes incorrect interpretation of idiomatic language, homonyms, or words with multiple meanings, therefore failing to capture the intended meaning. 10. other: all other types of errors are captured in this category.
+
+Step 5: Write the sublabel based on your reasoning in Step 4.
+
+```txt
+You will output a list of dictionaries. Each dictionary will have below format:  
+{  
+    "claim": "<CLAIM TEXT>",  
+    "thought_for_label": "<THOUGHT FOR LABEL FOR THE CLAIM>",  
+    "label": "<CLAIM LABEL>",  
+    "thought_for_sublabel": "<THOUGHT FOR SUBLABEL FOR THE CLAIM>",  
+    "sublabel": "CLAIM SUBLABEL"  
+}  
+</task>
+```
+
+# Example
+
+```txt
+Given below are the examples for you to comprehend the task. Each example includes a list of claims and a corresponding list of dictionaries as responses to aid your understanding. However, the input you will receive consists of a single claim, and you will need to provide a dictionary response for that particular claim. <example> ... </example>
+```
+
+# Source
+
+```txt
+<source>:
+{[SOURCE_TEXT]}</source>
+\n
+<claims>
+{TARGET Claims} </claims>
+```
+
+Figure 7: Claim classifier prompt for with  $CoT +$  one-claim-eval setup.
+
+Context (News Article): (CNN)Elon Musk has built a $12 billion company in an endeavor to pave the way to Mars for humanity. He insists that Mars is a long-term insurance policy for the light of consciousness in the face of climate change, extinction events, and our recklessness with technology. On the other hand, astronaut Chris Hadfield is skeptical: Humanity is not going extinct, he told me. He added: There's no great compelling reason to go, apart from curiosity, and that's not going to be enough to sustain the immense cost necessary with the technology that exists right now. But I question our future, stuck here on Earth. Our environment is a highly balanced system and we are the destabilizing element. Pursuing greeninitiatives is no long-term solution to the wall we're hurting towards, they're speeding bumps. If this is where humankind is destined to remain, then we shall find ourselves fighting over whatever is left of it. Politically speaking, sending humans into space brings nations together – the International Space Station stood as the physical manifestation of the reunification of the USA and Russia and is now a platform for broader international cooperation. Space exploration is also inspiring: during NASA's Apollo program to the Moon, the number of graduates in mathematics, engineering and the sciences in the US doubled. Igniting the imagination of that generation helped propel the US into the dominant position it's held since the 1960s. What could a Mars program do? Wouldn't the Moon, so much nearer than Mars, be a better first step? Actually, no – it's just too different. It's better to test hardware and train people in analogs on Earth, such as the geologically similar high-altitude desert in Utah or the cold and dry Canadian Arctic desert. Why the European Space Agency has declared the Moon a stepping-stone to Mars is beyond me, as doing so increases the cost of a Mars program hugely. It takes about  $50\%$  more energy to put something on the surface of the Moon than it does on Mars. The Martian atmosphere can be used to slow down approaching spacecraft, instead of the need for extra fuel to slow the descent. It would also mean developing two different sets of landing techniques and hardware. There are reasons to go to the Moon, just not if your ultimate destination is Mars. Even colonizing the Moon is questionable: it simply hasn't the resources to sustain an advanced colony. Mars has fertile soil, an abundance of water (as ice), a carbon-dioxide rich atmosphere and a 24-and-a-half hour day. The Moon's soil is not fertile, water is as rare, it has no effective atmosphere, and a 708-hour day. It's feasible to introduce biological life to Mars, but not the Moon. With only a relatively small push, Mars could be returned to its former warm, wet, hospitable state. Raising the temperature at the south pole by a few degrees would see frozen CO2 in the soil begin to gasify. As a greenhouse gas, it would further raise the temperature, gasifying more CO2 in a self-sustained global-warming process. Eventually, water frozen into the soil would liquefy, covering half of the planet. After about a century, Mars would settle down with an atmosphere about as dense as the lowland Himalayas and a climate suitable for T-shirts. Hadfield warns that we need to invent a lot of things before going to Mars, and that there's no great advantage to being the early explorers who die. Few would disagree with that, but what are the challenges a crewed mission to Mars faces? Radiation: An astronaut would receive a lifetime allowable dose of radiation in a single 30-month round-trip, including 18 months on the surface. But this is only equivalent to increasing the lifetime cancer risk from about  $20\%$  to  $23\%$ . As the majority of this is received in transit between planets, with proper radiological protection on the ship, it would actually be (radiologically speaking) healthier for an astronaut to live on Mars with a radiation dose of 0.10 sieverts per year than to smoke on Earth at 0.16 sieverts per year. There is no single practical solution to the radiation problem. One strategy I helped develop was to optimise the internal layout of the equipment and structures in the Mars habitat module to minimise exposure – placing existing bulk in all the right places. This reduced exposure by about  $20\%$ , without adding any mass. Even taking empty sandbags, packing them with Martian soil and putting them on the roof would be a simple and effective measure on Mars. Radiation is an issue to tackle, but it's not a deal-breaker. Power: We need a compact energy source, says Hadfield. We cannot be relying on the tiny bit of solar power that happens to arrive at that location. While the solar energy reaching the surface of Mars is about half that on Earth, this isn't a show-stopper. A quick back-of-the-envelope calculation shows that to power the equivalent of an average U.S. household on Mars, even through dust storms, one would need an array of solar panels totalling six metres square – very achievable. Reduced gravity: The effects of microgravity on astronauts' health have been studied for decades, and a range of techniques have been developed to mitigate the wasting effects on muscle and bone. With Martian gravity around a third of that on Earth, it would take astronauts a couple of days to acclimatize, and perhaps a few months to fully adapt. NASA and ESA have been developing an under-suit that compresses the body to overcome the negative effects of a reduction in pressure and gravity. However, biological adaption could be made easier if microgravity were avoided altogether. The spacecraft could be spun in-transit to generate an artificial gravity that slowly decayed, simulating a transition from Earth to Mars gravity (and vice versa) over the six-month journey. Ultimately, until humans are actually living on other planets it's unlikely we'll solve or even recognise all the subtle long-term health problems associated with reduced gravity. And who's to say what the advances in bio-engineering and technology will make the human body capable of when that time comes? Life on Mars: If there's life on Mars, even if it's microbial, should we be allowed to spread to the planet, potentially risking its extinction? I find this question strange – as Chris McKay put it: We commit microbial genocide every time we wash our hands. We engineer and farm the complex life around us as systematically and as cheaply as possible. Billions of people eat the carcasses of organisms that were thinking and breathing only days before. Why, all of a sudden, should Martian microbes be given such sanctity? It should certainly be studied, but it shouldn't prevent our spreading. Back contamination: Conversely, the question of whether some Martian plague might accidentally be introduced to Earth should be taken seriously – but not blown out of proportion. There's only a remote chance that Martian life might be hazardous. The things that kill us do so because they've evolved in lock-step with us in a continual evolutionary arms race. Any Martian life will have evolved independently and is unlikely to be capable even of interacting with Earth life on a molecular level. As Robert Zubrin put it: Trees don't get colds and humans don't get Dutch Elm Disease. Psychology: Depending on relative orbits, sending a message between Earth and Mars can take between three and 22 minutes. This loss of real-time communication will leave astronauts feeling cut-off and alone. Hadfield says that it's vital to keep up crew morale and motivation: Once you get any distance away on any sort of voyage, the epic-ness disappears, the reality becomes the foreground, and the applause is long gone. Cost: A crewed Mars program would cost the equivalent of a few weeks of the U.S. defense budget. The US plans on spending about 10 times more on nuclear weapons than on space exploration over the coming decade. The UK government spends about as much on gastric band surgery through the NHS as it does on its space activities. So while a Mars program certainly has challenges to overcome, the technological gap between us and Mars is far smaller than it was for the Moon program in the 1960s. And the prospects the Red Planet holds for humanity are far greater. Copyright 2015 The Conversation. Some rights reserved."
+
+Response (Cohere Summary): The Conversation article presents an argument in favor of pursuing space exploration, specifically a manned mission to Mars. It explains that while Chris Hadfield, a retired astronaut, is skeptical about the necessity of such an endeavor, the author believes that it is imperative for humanity to explore space. The author points out that Earth's environment is delicately balanced and that human activity has the potential to disrupt it significantly. Therefore, humans should look to expand their presence beyond Earth in order to ensure their long-term survival. In addition, the article addresses common concerns and challenges of a Mars mission, such as the high cost, potential dangers, and lack of practical benefits. However, it argues that these challenges are not insurmountable, and that the benefits of exploring Mars are immense. Ultimately, the author concludes that while Mars may seem like a distant and impractical goal, it is a worthwhile endeavor that has the potential to propel humanity into a new era of innovation and progress.
+
+HalluMeasure Extracted Claim: Chris Hadfield is a retired astronaut.
+
+Human Annotated Label: Absent
+
+HalluMeasure W/o CoT Label: Supported
+
+HalluMeasure W/ CoT Label: Absent
+
+HalluMeasure W/ CoT Label Explanation: The source text refers to Chris Hadfield as an astronaut and quotes him directly, but does not explicitly state whether he is currently active or retired.
+
+Table 10: Sample Result: HalluMeasure W/ and W/o CoT comparison on a claim from news article context document in TechNewsSumm dataset.

2024/HalluMeasure_ Fine-grained Hallucination Measurement Using Chain-of-Thought Reasoning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66dacc9be9aaf23b3f7d9ac3a4365701e150a6e5715ea0387f5f90d666871005
+size 771095

2024/Hate Personified_ Investigating the role of LLMs in content moderation/98ff3b1a-68ac-4b9d-91bf-9c764205c8dc_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf9cea8b7fb5322f4a0453c3c5311c118e72b9c65645e3b951791cb3812f1d7d
+size 954717

2024/Hate Personified_ Investigating the role of LLMs in content moderation/full.md

@@ -0,0 +1,451 @@
+# Hate Personified: Investigating the role of LLMs in content moderation
+
+Sarah Masud\*, Sahajpreet Singh\*, Viktor Hangya\*, Alexander Fraser\*, Tanmoy Chakraborty\*
+
+$^{1}$ IIT Delhi,  $^{2}$ IIT Delhi,  $^{3}$ LMU,  $^{4}$ TUM
+
+sarahm@iitd.ac.in, sahaj.phy@gmail.com, hangyav@cis.lmu.de, alexander.fraser@tum.de, tanchak@iitd.ac.in
+
+# Abstract
+
+For subjective tasks such as hate detection, where people perceive hate differently, the Large Language Model's (LLM) ability to represent diverse groups is unclear. By including additional context in prompts, we comprehensively analyze LLM's sensitivity to geographical priming, persona attributes, and numerical information to assess how well the needs of various groups are reflected. Our findings on two LLMs, five languages, and six datasets reveal that mimicking persona-based attributes leads to annotation variability. Meanwhile, incorporating geographical signals leads to better regional alignment. We also find that the LLMs are sensitive to numerical anchors, indicating the ability to leverage community-based flagging efforts and exposure to adversaries. Our work provides preliminary guidelines and highlights the nuances of applying LLMs in culturally sensitive cases. $^{1}$
+
+# 1 Introduction
+
+Human evaluators from diverse backgrounds are necessary to provide coverage against hate speech, which the UN defines as "any kind of communication that attacks or uses pejorative or discriminatory language with reference to a person or a group on the basis of who they are." Variation in annotations with respect to demographics matters as it is reflective of their lived experiences. However, the background of the evaluators also contributes to annotation biases (Rottger et al., 2022; Aroyo et al., 2019a; Munn, 2020). On the other hand, the role of Large Language models (LLMs) (Zhang et al., 2023) to help content moderation is now being explored (Huang et al., 2023; Roy et al., 2023).
+
+Motivation. In a recently proposed hate speech dataset CREHate (Lee et al., 2023), annotators
+
+![](images/d9c86d9b10a235ffdcca3665322bc3ab4c7628c95d7c6fbb7ce7e32ebdaf7b7a.jpg)
+
+![](images/60c2cf2a15b132fca95b8aad28b762e9f4b9df45c6f8fed0851f151fdfa8c822.jpg)
+
+![](images/e6ad690e834541811e38c9ffecf2c80efa9abeb6287da649bae8650116ce1c54.jpg)
+
+![](images/3dd45d4aa2e7664882b9d55b0e5233f6a84fc94bb9a1dd2a9729adcf545534d3.jpg)
+
+![](images/0f9fa16fc0766e12879880d62ca93e98f1d632fb1e93ae26a4e2187e637d0b9d.jpg)
+
+I like my girlfriends like I like my dogs Rescued from a young age and stays in their cage.
+
+![](images/63f6a4fe7c2883f2a2614c3de36ca3e9f5e6266b794e9f71b882c6ccd17b4018.jpg)
+
+![](images/bd439463eae8ff1ea9f80e9b77cd1f099c81f47dbdb92b6e3e68e3afd5a11529.jpg)
+
+![](images/cd6ff7db3432498cd1c0d411c05ff2372680c55fd8bc2ffecd8d839c74ae281f.jpg)
+
+![](images/e330633a43f40fd10faf9b41fad0ea9b41a08c499c79a85e630297b8a032348d.jpg)
+
+![](images/4fcc9f0c0be6da8fc8ba5b0c7fc88f7b38d2e13a5642bc7e5042bee928ef13eb.jpg)
+
+"Red Pill" cuck gets used for money on a date, writes a field report on it Imfao.
+
+![](images/eeed26f674368fd318d422ace9cafce6c0e6cb052ba701a444118650d3282942.jpg)
+
+![](images/40ae943f0fea21b631853612007bafc3a78df0fd952e72aeb26ee0cb606ace3b.jpg)
+
+![](images/53d30fe960a279d2b5ac1014c459bd327ee8c58083cd75f4a5e2b5cfe22f8583.jpg)
+
+![](images/a556f2dec8ce6cd59d7af0e671c1658a749af53638dea08c23e42ffc5f3d2a04.jpg)
+
+![](images/087d91d83b5c17001a15900f139b60df3dc220a073eae5cf5f59f272304ba63a.jpg)
+
+This is how they work. They are domestic terrorists. They are taking over corporations world wide and nothing good will come of it.
+
+![](images/08bb3c73fe93d347616f6090502dd9a91ad687788c503ff84069d4ce5860a4ff.jpg)
+
+![](images/e35cc3923d8d8bc1f50bfb9915aa9d231eef75ab24425ae92ded86a432af3493.jpg)
+
+![](images/577b880f59131e197e4618df47e581273b7864bcb05c52442ad86243866b34f4.jpg)
+
+![](images/380c40a910fdc754a59f8e114c6875bc3921b8d20d624edfe97b5a963e77afa6.jpg)
+
+![](images/bd82514e83a9de3df8c169f252cb37b83ad0dd77cc52c24a9835fe5adfa7befd.jpg)
+
+Figure 1: Annotations of hate/non-hate (red/green) for USA, Australia, UK, South Africa, and Singapore, by a) annotators from respective countries (circle) and b) prompting GPT-3.5 with 'The following statement was made in <country>: <POST>. Is the given statement hateful?' (square). The posts and human labels are verbatim from the CREHate dataset (Lee et al., 2023).
+
+of five countries perceived the same post (in English) differently. The authors also observed variation upon introducing the country when prompting LLM for similar annotations. Reproducing results from CREHate, we observe (Figure 1) a gap between the human-LLM alignment of hatefulness. We, thus, investigate variations in LLM's output when primed with context, which, under a similar setting for humans, causes variability. In a two-party setup consisting of a human and a zero-shot LLM as annotators, we prompt the LLM with contextual information and observe the inter-annotator variation (w.r.t human label) in the LLM's output.
+
+Borrowing from the demographic identities of CREHate, we assess if LLMs qualify as a proxy for annotators of a given identity. The direct implication is obtaining cheaper and faster soft labels via LLMs for a demographic that is underrepresented among human annotators. We also investigate the role of geographical cues, similar to the use of geolocation metadata by online platforms, in annotating hate. Apart from the post's content, numerical
+
+![](images/601e0569c9e9c3351374c17f5611ba65b2899e25712cf8857746c945d7ba9b1f.jpg)  
+Figure 2: Overview of the research pipeline: an incoming post is prefixed with context to form the prompt for the LLM. The predicted label is then evaluated against ground truth to examine variability arising from context.
+
+metadata (Founta et al., 2019), such as the number of likes, number of times a post has been flagged as offensive, etc, is often available internally to content moderators. We examine if such statistics can better guide LLMs to reflect a community's needs.
+
+Prompting Format. Decision-making in the human (Slovic et al., 1977) is influenced by prior and current information. However, for general-purpose LLMs to replicate the same effect, they must be externally cajoled/cued. The prompts are formatted as 'cue + post + query', where  $p_{\text{base}} =$  'Statement: <POST>. Is the given statement 'hateful?' forms the base prompt. Further, we translate the cue and query into the respective language via Google Translate + manual verification for the multilingual datasets. RQ1 (Section 4) delves into the geographical nuances like the country or language of the post (e.g., 'The following statement was made in country x...') as a cue for regional context. RQ2 (Section 5) explores diverse demographic facets (e.g., 'Suppose you are a person of x ethnicity...') as a proxy for different human evaluator groups. Finally, RQ3 (Section 6) examines the variability introduced by a numerical context, expressed as 'x% of individuals labeled this post as 'hateful', as a proxy for anchoring bias in LLMs. The RQs allow us to investigate the difference between implicit (base prompt) and explicit (contextual prompt) nudging (Figure 2).
+
+Research Scope. Firstly, the aim is not to establish SOTA for a given dataset. We do not engineer the highest-performing prompt for a specific dataset. Instead, our study helps provide a general assessment across the datasets. Secondly, finetuning on a single hate speech dataset does not necessarily transfer to out-of-distribution samples (Yin and Zubiaga, 2021). Lastly, the variability and nuances probed in this study are not feasible
+
+to replicate under in-context and finetuned setups where multiple hyperparameters already impact the output. Therefore, in this study, we examine only the zero-shot prompt setting.
+
+Observations and Implications. We perform an exhaustive analysis spanning two LLMs, five languages, and six hate speech datasets. We investigate 86 prompts (Table 12) in English and 40 multilingual prompts (Table 13). We caution against the blindsided use of LLMs for crowdsourcing in subjective tasks. In summary, this work can help practitioners gauge the LLM vs. manual effort needed in their content moderation pipeline.
+
+- Geographical Cues. We observe that geographical cues lead to visible and significant increases in human-LLM agreement (Section 4). As these metadata are readily available within a platform, including them in the prompting does not require additional effort.  
+- Persona Cues. As no persona cue + LLM combination encompasses sensitivity toward all groups, practitioners employing LLMs as proxies must do so cautiously. Further, we establish that 'the manner/format' of imbibing the persona is equally crucial (Section 5).  
+- Numerical Cues. We observe that pseudovoting values influence predicted labels, highlighting both positive (community flagging) and negative (adversarial attacks) effects. The results question the use of numerical features in zero-shot prompting (Section 6).  
+- Multilinguality: Under multilingual prompting, we observe the above patterns to persist, albeit with an expected loss in annotation agreements (Sections 4 and 5). As native speakers would prefer to engage with LLMs in their native language, this calls for improving the low-resource specificity in LLMs.
+
+<table><tr><td rowspan="2">Dataset (Language) (Reference)</td><td colspan="3"># Samples in original dataset</td><td colspan="3"># Samples used in RQs</td></tr><tr><td># Hate</td><td># Non-hate</td><td>Total</td><td># Hate</td><td># Non-hate</td><td>Total</td></tr><tr><td>HateXplain (En) (Mathew et al., 2021)</td><td>4748</td><td>6251</td><td>10999</td><td>4748</td><td>6251</td><td>10999</td></tr><tr><td>CREHate (En) (Lee et al., 2023)</td><td>709</td><td>871</td><td>1580</td><td>709</td><td>871</td><td>1580</td></tr><tr><td>MLMA (Ar) (Ousidhoum et al., 2019)</td><td>460</td><td>915</td><td>1375</td><td>250</td><td>250</td><td>500</td></tr><tr><td>MLMA (Fr) (Ousidhoum et al., 2019)</td><td>207</td><td>821</td><td>1028</td><td>207</td><td>293</td><td>500</td></tr><tr><td>HASOC-2020 (De) (Mandl et al., 2020)</td><td>146</td><td>1700</td><td>1846</td><td>146</td><td>354</td><td>500</td></tr><tr><td>HASOC-2020 (Hi) (Mandl et al., 2020)</td><td>234</td><td>2116</td><td>2350</td><td>234</td><td>266</td><td>500</td></tr></table>
+
+Table 1: Dataset statistics employed in this study. Here, Hate  $\sim$  Hateful & Non-Hate  $\sim$  Normal  $\sim$  None.
+
+At the intersection of annotation priming and zero-shot prompting, the current research acts as a guideline for configuring the LLM-assisted content moderation pipeline. Our study establishes the role of context in making LLM-based hate speech annotations reflect the preferences of a given vulnerable community or different cultural groups more closely. Our study illustrates that explicit cues align better with human annotations.
+
+# 2 Related Work
+
+Annotation Biases. Labeling for hate speech datasets is primarily led by human annotators (Waseem, 2016; Founta et al., 2018). Human annotations for subjective tasks (Rottger et al., 2022) are rife with ambiguity (Kanclerz et al., 2022; Aroyo et al., 2019b) and biases (Garg et al., 2023). Of particular interest is the annotation bias (Wich et al., 2021). In hate speech, annotation bias manifests due to differences among the annotator's belief (Sap et al., 2022), experience, world knowledge (Yin and Zubiaga, 2021; Abercrombie et al., 2023), and social-demographic conditioning (Orlikowski et al., 2023). Disparity in access to additional context (Ljubesic et al., 2022; Ihtiyar et al., 2023) and annotation guidelines (Ross et al., 2016), in some instances, reduce the bias and, in some cases, confirm the annotator's biases. Analysis and mitigation of biases in hate speech is an active area of research (Biester et al., 2022; Wojatzki, 2018; Sap et al., 2019). Parallel research seeks to model diverse opinions (Braylan and Lease, 2020; Li et al., 2021; Weerasooriya et al., 2023) as a way to reduce the annotation bias. Variability in labeling is unavoidable, even if considered on a continuous scale (Sachdeva et al., 2022) or augmented with signals (Koufakou et al., 2020) of sentiment or emotion, both of which are hard to annotate.
+
+LLMs for Annotations. The rise of instruction-tuned LLMs has further facilitated prompt-based labeling of hate speech (AlKhamissi et al., 2022;
+
+Yang et al., 2023; Huang et al., 2023; Roy et al., 2023). Consequently, chain-of-thought (AlKhamissi et al., 2022) and few-shot (Khullar et al., 2023) prompting are also being investigated for hate speech detection. However, the use of LLMs for annotations in NLP tasks is still in its early stages (He et al., 2023; Ostyakova et al., 2023) owing to ethical, legal, and interpretability concerns (Zini and Awad, 2022). The initial research does hint at the efficiency of zero-shot annotations via GPT-3.5 (Chung et al., 2022) for myriad standard NLP tasks (Gilardi et al., 2023; Kocón et al., 2023). However, parallel research has shed light on adversarial attacks (Le et al., 2022; Zou et al., 2023; Nookala et al., 2023) and generative biases while prompting (Griffin et al., 2023; Lin and Ng, 2023; Wang et al., 2023; Zhao et al., 2021). It should be noted that while in-context learning is a promising area of research, the  $n$ -shot samples may exhibit repetition and sampling biases (Zhao et al., 2021; Zamfirescu-Pereira et al., 2023). Hence, our analysis concentrates on the zero-shot setting as a first step to control for additional variability.
+
+# 3 Experimental Setup
+
+This section outlines the datasets, models, and evaluation metrics employed in the study<sup>2</sup>.
+
+Datasets. An overview of the datasets employed in this study is provided in Table 1. All these datasets are publicly available. These datasets contain multi-class labels, including hate speech, offensive, and normal for HateXplain, and various mixed labels, such as abusive_hateful for multilingual datasets. To remove the subjectivity of these umbrella terms, we classify instances as hate when the label is either 'hate' or 'hateful,' and as non-hate when the label is 'normal,' 'none,' or 'non-hate.' We utilize the exact texts from the original datasets. We have not attempted to identify or remove any previously mentioned entities, such as
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2"># of parameters</td><td colspan="4">HateXplain</td><td colspan="4">CREHate</td></tr><tr><td># Samples</td><td># Hal</td><td>F1</td><td>IAA</td><td># Samples</td><td># Hal</td><td>F1</td><td>IAA</td></tr><tr><td>FlanT5-Small</td><td>60M</td><td>≈11k</td><td>2</td><td>0.412</td><td>0.000</td><td>≈1.5k</td><td>2</td><td>0.391</td><td>0.000</td></tr><tr><td>FlanT5-Base</td><td>250M</td><td>≈11k</td><td>85</td><td>0.649</td><td>0.341</td><td>≈1.5k</td><td>156</td><td>0.536</td><td>0.166</td></tr><tr><td>FlanT5-Large</td><td>780M</td><td>≈11k</td><td>4545</td><td>0.339</td><td>0.136</td><td>≈1.5k</td><td>572</td><td>0.411</td><td>0.187</td></tr><tr><td>FlanT5-XL</td><td>3B</td><td>≈11k</td><td>0</td><td>0.588</td><td>0.293</td><td>≈1.5k</td><td>4</td><td>0.638</td><td>0.292</td></tr><tr><td>Mistral</td><td>7B</td><td>≈11k</td><td>135</td><td>0.531</td><td>0.228</td><td>≈1.5k</td><td>198</td><td>0.568</td><td>0.303</td></tr><tr><td>Zephyr</td><td>7B</td><td>≈11k</td><td>3948</td><td>0.343</td><td>0.123</td><td>≈1.5k</td><td>560</td><td>0.323</td><td>0.102</td></tr><tr><td>Llama 3</td><td>8B</td><td>≈11k</td><td>1971</td><td>0.439</td><td>0.180</td><td>≈1.5k</td><td>679</td><td>0.357</td><td>0.150</td></tr><tr><td>FlanT5-XXL</td><td>11B</td><td>≈11k</td><td>0</td><td>0.731</td><td>0.476</td><td>≈1.5k</td><td>0</td><td>0.649</td><td>0.297</td></tr><tr><td>FlanT5-XXL</td><td>11B</td><td>500</td><td>0</td><td>0.738</td><td>0.487</td><td>500</td><td>0</td><td>0.649</td><td>0.297</td></tr><tr><td>GPT-3.5-Turbo*</td><td>&gt;150B</td><td>500</td><td>0</td><td>0.780</td><td>0.576</td><td>500</td><td>2</td><td>0.758</td><td>0.517</td></tr></table>
+
+Table 2: Performance of LLMs when prompted with  $p_{base} =$ '<POST>. Is the given statement hateful?''. We report the number of samples in the data set used for prompting (# samples), the number of hallucinated outputs (# hal), the rectified weighted-F1, and the rectified inter-annotator agreement (IAA). *close-sourced model.
+
+the name of the target person in the given hate speech, etc. This decision ensures that the LLM receives the exact text for annotation as presented to the human annotator, maintaining consistency with the ground truth. For RQ1 and RQ2, we use all 1580 samples from CREHate (Lee et al., 2023). Each sample in English is labeled as hateful or not by annotators from different nations (the United States, Australia, the United Kingdom, South Africa, and Singapore). For RQ3, we employ HateXplain (Mathew et al., 2021), which contains English instances, encompassing three labels – toxic, hateful, and normal. We take samples with a majority label as hateful or normal, leading to  $\approx 11k$  instances. We also investigate four datasets containing multilingual and code-mixed (with English) posts. We use Arabic (Ar) and French (Fr) datasets by Ousidhoum et al. (2019) and Hindi (Hi) and German (De) by Mandl et al. (2020). Here again, we binarize the labels wherever applicable.
+
+Evaluation Metric. For  $p_{base}$ , we have the ground labels (majority voted gold labels). So, we employ a weighted F1 score to compare the  $p_{base}$  outputs. Among the prompt variants in RQs, where we may have a direct ground truth, we analyze the performance disparities via the Cohen-  $\kappa$  inter-annotator agreement (IAA) (Cohen, 1960). It has been observed that IAA and F1 are positively correlated when the dataset is not skewed (Richie et al., 2022). In our case, the skewness is controlled by having samples that are almost equal in both classes (as noted in Table 1). When choosing the IAA metric over the F1 score, our primary goal is not to emphasize high precision and recall. Instead, we aim to demonstrate how closely gold human annotations align with those generated by the LLM. This is why we also use the predicted hate-label
+
+ratio (PHLR), which represents the proportion of all samples labeled as hate. It is calculated as the ratio of the total number of predicted hate labels to the total generated labels, excluding hallucinations and empty outputs.
+
+Rectified Scores for Hallucination. As the aim is to know if the post is hateful, for our use-case, any output not in the form of 'yes/no', 'hate/non-hate', or 'hateful/non-hateful' can be considered as a 'hallucinated' label falling outside the range of the expected answers. We specify the prompt suffix 'answer in one word only.' We also perform a manual evaluation to access the unique outputs per setup, and where the output could be salvaged, they were updated. In line with the existing literature (Lee et al., 2023), after all filters, the outputs that still did not qualify as acceptable are discarded. Recording the number of hallucinations (discarded outputs), we introduce 'rectified F1/IAA scores',  $score_{rectified} = \left(1 - \frac{h}{t}\right) \times score$ , where  $h$  and  $t$  are the hallucinated and total samples, respectively. Any mention of F1 and IAA in our study means 'rectified weighted F1' and 'rectified IAA.'
+
+Models. We begin with FlanT5 (Chung et al., 2022), Mistral (Jiang et al., 2023), Zephyr (Tunstall et al., 2023), Llama 3 (Touvron et al., 2023) and InstructGPT (Ouyang et al., 2022) variant GPT-3.5-Turbo (hereby referred to as GPT-3.5). Performance metrics on all shortlisted LLMs (with the total number of parameters in that model) are captured in Table 2 for the  $p_{base}$ . From Table 2, we observe that even with  $p_{base}$ , most models generate noisy and ill-formatted outputs. To reduce the influence of noisy results, we select only FlanT5-XXL and GPT-3.5 to analyze more intricate setups in our RQs. Both these models portray negligible hallucinations and produce relatively higher F1 and
+
+IAA. To test F1anT5-XXL, we employ all the samples. For GPT-3.5, due to API rate limits, we test random 500 stratified samples per dataset.
+
+We also examine language-specific LLMs to corroborate the role of zero-shot prompting setup of general-purpose LLM (GPT-3.5) with cultural context vs language-specific models. The results (Table 3) highlight that not just demographic-based training corpus but the manner of training and size of models impact multilingualism.
+
+Prompts. Tables 12 and 13 provide the list of verbalized prompts under English and multilingual settings, respectively. For the latter, after employing Google Translate, the sample prompts were verified by two linguistic experts.
+
+Note on Hate Speech Definition. Existing literature has observed that providing explicit definitions to LLMs may not lead to performance gain (Guo et al., 2023; Roy et al., 2023). We, thus, conduct our experiments building on the latent notions of hate speech in LLMs. For a sanity check, we prompt F1anT5-XXL and GPT-3.5 with 'Define: Hate Speech' and obtain following responses:
+
+- FlanT5-XXL: “Hate Speech is a form of speech intended to incite hatred or violence against a group of people.”  
+- GPT-3.5: "Hate speech refers to any form of communication, whether spoken, written, or otherwise expressed, that promotes or encourages hatred, hostility, discrimination, or violence against individuals or groups based on attributes such as race, ethnicity, religion, nationality, gender, sexual orientation, disability, or other characteristics. It often involves derogatory language, stereotypes, or dehumanizing rhetoric intended to intimidate, offend, or incite harm against the targeted individuals or communities. Hate speech is typically viewed as a violation of ethical principles and may also be prohibited by law in many jurisdictions due to its potential to inflame tensions, provoke violence, and undermine social cohesion."
+
+Hardware and Software(s). All experiments are conducted on an NVIDIA-A100 (80GB) GPU. Model weights are loaded from the HuggingFace pretrained models using float16 precision. For GPT-3.5, OpenAI's API was utilized. All models are employed in a zero-shot setting. The temperature variation experiment takes a total of 100 hours. The rest of the experiments collectively take  $\approx 4 - 5$  hours of GPU time.
+
+<table><tr><td>Dataset (Lang)</td><td>Model (# Params)</td><td># Samples</td><td># Hal</td><td>F1</td><td>IAA</td></tr><tr><td rowspan="2">MLMA (Ar)</td><td>Mistral-Ar (7B)</td><td>500</td><td>378</td><td>0.027</td><td>0.0</td></tr><tr><td>GPT-3.5 (&gt;150B)</td><td>500</td><td>222</td><td>0.359</td><td>0.140</td></tr><tr><td rowspan="2">HASOC-2020 (Hi)</td><td>Airavata (7B)</td><td>500</td><td>477</td><td>0.046</td><td>0.0</td></tr><tr><td>GPT-3.5 (&gt;150B)</td><td>500</td><td>131</td><td>0.257</td><td>0.018</td></tr></table>
+
+Table 3: Performance of LLMs when prompted with  $p_{base} =$ '<POST>. Is the given statement hateful?' in the respective language. We report the number of samples in the dataset used for prompting (#Samples), the number of hallucinated outputs (#Hal), and the rectified weighted-F1/inter-annotator agreement (IAA).
+
+# 4 Do LLMs Pick on Geographical Cues?
+
+Background. Humans from different countries are prone to variability when flagging the same post as 'hateful' or 'non-hateful' (Lee et al., 2023). Given that AI-assisted content moderation systems often have access to geolocation and language markers along with a post, we investigate whether 'akin to humans, do geographical cues influence LLM's predictions for an incoming hateful post.'
+
+Setup. Inspired by Lee et al. (2023), we reproduce and extend their analysis on the language tag and multilingual datasets. As described in Section 3, we modify their prompt to suit our format. We compare the change in IAA when considering the human annotator of the respective country vs.  $p_{\text{base}}$  against the human annotator vs.  $p_{\text{con}} + p_{\text{base}}$  or  $p_{\text{lang}} + p_{\text{base}}$ . Here,  $p_{\text{con}} =$  'The following statement was made in <country>.' where country = {United States (USA), Australia (AUS), United Kingdom (UK), South Africa (SA), Singapore (SG)}. Meanwhile,  $p_{\text{lang}} =$  'The following statement was made in <lang> language.' where lang = {Arabic (Ar), French (Fr), German (De), Hindi (Hi)}. For analyzing  $p_{\text{con}}$ , we employ CREHate for both F1anT5-XXL and GPT-3.5. For  $p_{\text{lang}}$ , we employ the HASOC-2020 (German & Hindi) and MLMA (Arabic & French) on GPT-3.5's English and language-specific prompting.
+
+Insights. We discuss the results in two broad settings: for country and language tags.
+
+Country. From Figure 3 (b) with  $p_{con} + p_{base}$ , GPT-3.5 exhibits disparity in lack of context about social constructs of the Global South (SA, SG) compared to higher alignment with so-called Western nations (USA, UK, AUS) (Zhou et al., 2022; Li et al., 2022; Lee et al., 2023). Surprisingly, for F1anT5-XXL (Figure 3 (a)), we observe an improvement in all countries. It is further corroborated by significance testing on F1anT5-XXL, where we observe (Appendix D) that adding the country cue leads to a notable change in output. One hypothesis
+
+![](images/48770eac7d1d703be5708a5b36a98552f28869ac0c7b1e62c6066449a0ccf9f8.jpg)  
+(a) F1anT5-XXL
+
+![](images/ba02b059b84648bc5ffb2607bac93e63395bc18ff275c8343e943f0f25940747.jpg)  
+(b) GPT-3.5
+
+![](images/ce5830b6eda10a061289070f07c50d892730ccd89b2496e80c1f07bd9bc710e2.jpg)  
+(c) GPT-3.5  
+Figure 3: [RQ1] (a-b) The IAA w.r.t human annotation for each country F1anT5-XXL and GPT-3.5, respectively, for English posts. (c) Captures each language's IAA w.r.t human labels via GPT-3.5 with posts in the language and prompts in English. Here, without (w/o) is  $p_{base}$  and with (w/) is  $p_{con} / p_{lang} + p_{base}$ .
+
+esis for the difference in  $p_{con}$  results can be the nature of training the LLM. Due to only instruction tuning of F1anT5-XXL, it develops country-specific bias implicitly from the training data; however, for GPT-3.5 the implicit bias is augmented with explicit human feedback. It surely calls into question how LLM pretraining mechanisms impact the subjective (non-GLEU) downstream tasks such as hate speech detection (Roy et al., 2023).
+
+Language. From Figure 3 (c), we conjecture that  $p_{lang}$  nudges GPT-3.5 to language-specific subspaces leading to a visible higher IAA with ground truth labels when we add  $p_{lang}$ . As we infer in Appendix A, the delta increase in performance in  $p_{lang} + p_{base}$  vs.  $p_{base}$  follows the same order of magnitude irrespective of whether the prompts are in English or the respective language.
+
+Takeaways. Firstly, our observations emphasize the fact that both corpus and paradigm in training play a role in the geographical sensitivity of the LLMs. Judging why some cues and LLMs garner more improvement than others is challenging to access from prompting alone. This calls for more transparency in the LLM training to decode the geographical bias during training epochs. Secondly, as an actional insight, the findings encourage incorporating geographical cues in the zero-shot prompt to ensure higher human-LLM alignment. Interestingly, even if the prompt is in the respective language, explicitly nudging helps.
+
+# 5 Can LLMs Mimic Annotator Persona?
+
+Background. Socio-cultural experiences of humans color their outlook about hate and cause variation in annotations (Sap et al., 2019; Orlikowski et al., 2023). However, without direct signals about mental state, we rely on demographic attributes as markers of human conditioning (Aroyo et al.,
+
+2019b; Sap et al., 2022) a.k.a personas. Meanwhile, LLMs only possess statistical socio-cultural experience. When employing LLMs for crowdsourcing, we need to assess 'Do LLM's emulation of demographics at a fine-grained level leads to variations in annotating hate?'. We further hypothesize that differences in a vulnerable group's projection) can lead to variability in the hate perceptiveness of LLMs. We consider a vulnerable group as those who have historically been mistreated based on identity (Kulkarni et al., 2023).
+
+Setup. Borrowing the annotator demographic list from CREHate, we define our persona attributes in terms of  $D = \{Gender(D_g),\text{Ethnicity}(D_e),\text{Political Orientation}(D_p),\text{Religion}(D_r),\text{Education Level}(D_q)\}$ . Each demographic in  $D_{*} \in D$  has further sub-classes as outlined in Table 4. For each  $D_{*}$ , we augment the prefix  $p_{trait} = 'A person who is < D^{*}>$ , annotated the following statement as  $< H / N >$ .' Operated via the  $< H / N >$  tag, we run two persona variants, calling the annotated statement either hateful  $p_{trait}^{H}$  or non-hateful  $p_{trait}^{N}$ . We also examine a third variant  $p_{trait}^{A}$  with the prompt  $p_{trait}^{A} = 'Suppose you are a person who is < D >$  and given the following statement to annotate.' The 3 personas prompts  $(p_{trait}^{*} + p_{base})$  are examined via the CREHate dataset on F1anT5-XXL and GPT-3.5.
+
+Intrigued by the success of the language tag in RQ1 (Section 4), we deep dive into persona traits that closely represent each language's demographic. Here,  $\forall l\in lang = \{Arabic, French, German, Hindi\}$ , we introduce prompts  $p_{trait}^{L*} =$  'The following statement is in  $<l>$  language, and  $a < I>$  annotated this as  $<H/N>$ . In the prompt,  $I$  enlists a base/majoritarian persona vs. a vulnerable/minority persona of the respective geography examined on GPT-3.5. Steps for obtaining the following personas are provided in Appendix B.
+
+<table><tr><td rowspan="3">Annotator demographics</td><td rowspan="3">Sub-classes</td><td colspan="6">Flan-T5-XXL</td><td colspan="6">GPT-3.5</td></tr><tr><td colspan="2">\( p_{trait}^{H} \)</td><td colspan="2">\( p_{trait}^{N} \)</td><td colspan="2">\( p_{trait}^{A} \)</td><td colspan="2">\( p_{trait}^{H} \)</td><td colspan="2">\( p_{trait}^{N} \)</td><td colspan="2">\( p_{trait}^{A} \)</td></tr><tr><td>IAA</td><td>PHLR</td><td>IAA</td><td>PHLR</td><td>IAA</td><td>PHLR</td><td>IAA</td><td>PHLR</td><td>IAA</td><td>PHLR</td><td>IAA</td><td>PHLR</td></tr><tr><td rowspan="3">Gender</td><td>Male</td><td>0.42</td><td>0.53</td><td>0.00</td><td>0.00</td><td>0.31</td><td>0.29</td><td>0.40</td><td>0.70</td><td>0.55</td><td>0.44</td><td>0.57</td><td>0.46</td></tr><tr><td>Female</td><td>0.42</td><td>0.53</td><td>0.00</td><td>0.00</td><td>0.33</td><td>0.31</td><td>0.39</td><td>0.72</td><td>0.46</td><td>0.35</td><td>0.52</td><td>0.54</td></tr><tr><td>Non-binary</td><td>0.42</td><td>0.42</td><td>0.01</td><td>0.01</td><td>0.32</td><td>0.29</td><td>0.31</td><td>0.77</td><td>0.45</td><td>0.38</td><td>0.53</td><td>0.58</td></tr><tr><td rowspan="5">Ethnicity</td><td>Asian</td><td>0.46</td><td>0.56</td><td>0.03</td><td>0.02</td><td>0.33</td><td>0.23</td><td>0.37</td><td>0.75</td><td>0.55</td><td>0.51</td><td>0.51</td><td>0.59</td></tr><tr><td>Black</td><td>0.43</td><td>0.61</td><td>0.03</td><td>0.01</td><td>0.33</td><td>0.23</td><td>0.37</td><td>0.74</td><td>0.54</td><td>0.51</td><td>0.50</td><td>0.64</td></tr><tr><td>Hispanic</td><td>0.45</td><td>0.56</td><td>0.03</td><td>0.01</td><td>0.36</td><td>0.24</td><td>0.39</td><td>0.71</td><td>0.56</td><td>0.49</td><td>0.51</td><td>0.62</td></tr><tr><td>Middle Eastern</td><td>0.46</td><td>0.52</td><td>0.03</td><td>0.01</td><td>0.29</td><td>0.19</td><td>0.40</td><td>0.70</td><td>0.54</td><td>0.54</td><td>0.49</td><td>0.64</td></tr><tr><td>White</td><td>0.46</td><td>0.54</td><td>0.03</td><td>0.01</td><td>0.36</td><td>0.24</td><td>0.40</td><td>0.69</td><td>0.51</td><td>0.57</td><td>0.52</td><td>0.56</td></tr><tr><td rowspan="3">Political orientation</td><td>Liberal</td><td>0.42</td><td>0.52</td><td>0.01</td><td>0.01</td><td>0.32</td><td>0.29</td><td>0.49</td><td>0.63</td><td>0.54</td><td>0.53</td><td>0.61</td><td>0.53</td></tr><tr><td>Independent</td><td>0.43</td><td>0.58</td><td>0.00</td><td>0.00</td><td>0.32</td><td>0.29</td><td>0.46</td><td>0.65</td><td>0.58</td><td>0.48</td><td>0.54</td><td>0.50</td></tr><tr><td>Conservative</td><td>0.44</td><td>0.53</td><td>0.00</td><td>0.00</td><td>0.31</td><td>0.26</td><td>0.49</td><td>0.63</td><td>0.53</td><td>0.56</td><td>0.48</td><td>0.47</td></tr><tr><td rowspan="6">Religion</td><td>Christian</td><td>0.39</td><td>0.53</td><td>0.03</td><td>0.03</td><td>0.33</td><td>0.31</td><td>0.44</td><td>0.68</td><td>0.53</td><td>0.58</td><td>0.53</td><td>0.48</td></tr><tr><td>Buddhism</td><td>0.41</td><td>0.54</td><td>0.03</td><td>0.03</td><td>0.32</td><td>0.30</td><td>0.45</td><td>0.66</td><td>0.52</td><td>0.59</td><td>0.51</td><td>0.45</td></tr><tr><td>Islam</td><td>0.43</td><td>0.49</td><td>0.02</td><td>0.01</td><td>0.37</td><td>0.26</td><td>0.47</td><td>0.65</td><td>0.52</td><td>0.54</td><td>0.52</td><td>0.53</td></tr><tr><td>Judaism</td><td>0.45</td><td>0.48</td><td>0.02</td><td>0.01</td><td>0.30</td><td>0.22</td><td>0.46</td><td>0.68</td><td>0.50</td><td>0.60</td><td>0.52</td><td>0.55</td></tr><tr><td>Hinduism</td><td>0.42</td><td>0.49</td><td>0.04</td><td>0.03</td><td>0.32</td><td>0.24</td><td>0.48</td><td>0.65</td><td>0.52</td><td>0.53</td><td>0.51</td><td>0.49</td></tr><tr><td>Irreligious</td><td>0.40</td><td>0.39</td><td>0.02</td><td>0.02</td><td>0.30</td><td>0.25</td><td>0.44</td><td>0.68</td><td>0.54</td><td>0.57</td><td>0.54</td><td>0.43</td></tr><tr><td rowspan="6">Education level</td><td>&lt;High school</td><td>0.42</td><td>0.46</td><td>0.01</td><td>0.01</td><td>0.31</td><td>0.30</td><td>0.52</td><td>0.60</td><td>0.54</td><td>0.51</td><td>0.50</td><td>0.49</td></tr><tr><td>High school</td><td>0.41</td><td>0.52</td><td>0.00</td><td>0.01</td><td>0.29</td><td>0.29</td><td>0.45</td><td>0.65</td><td>0.56</td><td>0.52</td><td>0.53</td><td>0.50</td></tr><tr><td>College</td><td>0.42</td><td>0.52</td><td>0.00</td><td>0.01</td><td>0.29</td><td>0.29</td><td>0.44</td><td>0.67</td><td>0.53</td><td>0.44</td><td>0.52</td><td>0.46</td></tr><tr><td>Bachelor</td><td>0.42</td><td>0.52</td><td>0.00</td><td>0.00</td><td>0.28</td><td>0.29</td><td>0.43</td><td>0.68</td><td>0.56</td><td>0.47</td><td>0.50</td><td>0.44</td></tr><tr><td>Master&#x27;s</td><td>0.42</td><td>0.52</td><td>0.00</td><td>0.00</td><td>0.28</td><td>0.29</td><td>0.41</td><td>0.69</td><td>0.54</td><td>0.45</td><td>0.48</td><td>0.46</td></tr><tr><td>PhD</td><td>0.42</td><td>0.51</td><td>0.00</td><td>0.01</td><td>0.27</td><td>0.28</td><td>0.47</td><td>0.65</td><td>0.53</td><td>0.42</td><td>0.51</td><td>0.49</td></tr></table>
+
+Table 4: [RQ2] The inter-annotator agreement (IAA) and predicted hate label ratio (PHLR) w.r.t majority voted gold label in CREHate, for F1anT5-XXL and GPT-3.5. The demographic attributes  $(D)$  are compared under the hateful  $p_{\text{trait}}^{H}$ , non-hateful  $p_{\text{trait}}^{N}$ , and assumed persona  $p_{\text{trait}}^{A}$  settings.  $\forall D_{*} \in D$ , we combine the  $p_{\text{trait}}^{*} + p_{\text{base}}$ .
+
+![](images/bc4cf2d79be6c3d18c972e7a5579e19be48fbc291117a4112fc00862850b4c4d.jpg)  
+(a) GPT-3.5 (H)
+
+![](images/63c2a939b66e8026557583e65f94d492c99e578d4eb5cab59c2256a07814af92.jpg)  
+(b) GPT-3.5 (N)  
+Figure 4: [RQ2] Predicted hate label ratio (PHLR) from GPT-3.5 comparing  $p_{\text{trait}}^{L_*} + p_{\text{base}}$  for Arabic, French, German, and Hindi. (a) and (b) capture the base vs. vulnerable persona for hate/non-hate queries, respectively. (c) and (d) capture the native vs. non-native speaker persona for hate/non-hate queries, respectively.
+
+![](images/65eccdba755c339be0321ed1dab32d89d701b7eea8589143b9e044779882f460.jpg)  
+(c) GPT-3.5 (H)
+
+![](images/a2813dbb5be5ffd50a518b4b425534d22e9f1f335bb2205aee732f1987de404d.jpg)  
+(d) GPT-3.5 (N)
+
+1. For Arabic,  $p_{\text{trait}}^{L_{Ar}} \in \{ \text{Muslim/Non-mulsim} \}$  
+2. For French,  $p_{\text{trait}}^{L_{Fr}} \in \{\text{French/Mediterranean descent}\}$  
+3. For German,  $p_{\text{trait}}^{L_{De}} \in \{\text{Native/Non-native German speaker}\}$  
+4. For Hindi,  $p_{\text{trait}}^{L_{Hi}} \in \{\text{Upper/Lower caste}\}$
+
+We examine the hateful  $(H)$  and non-hateful  $(N)$  queries here as well, both with English and multilingual prompts. As none of the datasets provide demographic-specific labels, we rely on IAA among predicted and the majority-voted gold labels for our assessment. For  $p_{\text{trait}}^{H/N}$ , we provide in the prompt if the persona identifies a statement as hate/non-hate speech. Our objective here is to assess the role of these traits in persuading LLMs to increase/decrease the number of hate labels in their responses. We thus utilize the Predicted Hate Label Ratio (PHLR) metric in addition to IAA.
+
+Insights. We discuss the results broadly for English and multilingual datasets.
+
+English: Table 4 corroborates that nudging the model to assume a persona  $(p_{\text{trait}}^{A})$  is different from presenting the LLMs with a persona  $(p_{\text{trait}}^{H/N})$ . Further, it is evident from the predicted hate-label ratio (PHLR) that LLMs are sensitive towards some demographic subclasses more than others. Significance testing corroborates the same (Appendix D).
+
+Under Gender demographic for GPT-3.5, the percentage of hate labels is higher for 'Non-binary' than 'Males' (for  $p_{\text{trait}}^H$ ). It aligns with the former being a more susceptible subclass of gender in the real world. Meanwhile, for F1anT5-XXL, the presence of the non-hate tag  $p_{\text{trait}}^N$  dominates the demographic information in the context. However, the opposite is not valid for  $p_{\text{trait}}^H$ . Despite the ground labels associated with samples being balanced across classes,  $p_{\text{trait}}^N$  for F1anT5-XXL still predicts the majority of the labels as 'non-hate.'
+
+(a) FlanT5-XXL (H)  
+![](images/8c2a240d33b2ab3a78e340b4664628f8255c74212bbef4fef8813726d7d4bcdf.jpg)  
+Note: diagonals in heatmaps represent IAA between  $p_{base}$  and  $p_{vote} + p_{base}$ . Meanwhile, when x and y are different, it represents IAA b/w ( $P_{vote} = x + P_{base}$ ) and ( $P_{vote} = y + P_{base}$ ).
+
+![](images/a9a5b490d942d96fb338caf48aaf93f16fa903b352327e06200fd317d687cbac.jpg)  
+(b) GPT-3.5 (H)
+
+![](images/8206e02101ea010f1fffd9c9f6337b1d566b04e0e4e9e06cf7895088b87a23f9.jpg)  
+(c) F1anT5-XXL (N)  
+Figure 5: [RQ3] For the  $p_{vote}^{H}$ : (a) and (b) capture the IAA among various hateful voting percentages, i.e.,  $z\%$  for F1anT5-XXL and GPT-3.5, respectively. For  $p_{vote}^{N}$ : (c), (d) function analogously to (a), (b), respectively.
+
+![](images/4ccd1b319faa973a555e1ce7899563166eb70731f1a4b47f3cb4d37e34da4943.jpg)  
+(d) GPT-3.5 (N)
+
+Multilingual. From Figure 4 (a-b), we hypothesize that explicitly known vulnerabilities like Islamophobia (Arabic + Muslim) and Casteism (Hindi + Lower caste) are better captured by the LLM than ethnicity, leading to higher sensitivity of these pairs towards hate when prompted in English. While the patterns persist under multilingual prompting for the rest of the languages, its lack of variation in  $p_{\text{trait}}^H$  Hindi is puzzling (Table 6 in Appendix B).
+
+As the highest gap in  $p_{\text{trait}}^*$  English is observed for German, we deep dive into this prompt (native speaker vs non-native speaker) and repeat this for all languages under English and multilingual prompting. From Figure 4 (c-d) and Table 7 in Appendix B, one can conjecture that when a native-speaking persona considers a post as hateful, the model may be contextualizing the higher acuity of the speaker to understand the geographical context.
+
+Takeaways. Given the sensitivity of LLMs toward the combination of persona + label, we urge the practitioners not mindlessly to rely on LLM as a substitute for demographic attributes. The persistence of results in English and multilingualism is an advantage for the researchers working on multilingual datasets. It can allow faster assessment of the LLMs without requiring separate prompt engineering for both setups.
+
+# 6 Are LLMs Sensitive to Anchoring Bias?
+
+Background. Anchoring bias occurs when humans rely too heavily on the anchor information (relevant or not) to influence their decision-making. While some anchors like language and geolocation, as established in RQ1 (Section 4), are helpful for LLM-based content moderation, the influence of numerical cues under the zero-shot setting is unknown. Numerical features/cues can be defined as aggre
+
+gated real-world or simulated values representing countable metadata associated with the posts. It can range from the total/average views, likes, and comments a post/user receives to the number of people who have in the past reported/flagged the post/user as hateful. The former set of metadata is publicly available and has been successfully employed in hate speech detection to improve finetuned accuracy (Founta et al., 2019; Kulkarni et al., 2023). Meanwhile, the influence of previously flagged counts is not known, as such metrics are not available to the public. Here, using voting statistics as a proxy for the crowd's opinion about the post, we are motivated to examine whether 'made-up voting percentages in the prompt lead to manipulation in LLM's output?'
+
+Setup. The base prompt  $p_{base}$ , is prefixed with  $p_{vote}^{H/N} =$  'The following statement was labeled  $< H/N>$  by  $< z>$  annotators,' where  $z \in \{0\%, 25\%, 50\%, 75\%, 100\% \}$ . The two variants  $p_{vote}^{H/N}$  capture the hateful (H) or non-hateful voting (N) label. We represent  $z$  in the percentage to give a relative sense of majority voting. At  $z = 50\%$  saying,  $50\%$ , annotators consider the post is hateful, and  $50\%$  say it is non-hateful is not the same as  $p_{base}$ . We conduct this experiment with HateXplain on FlanT5-XXL and GPT-3.5 with  $p_{vote} + p_{base}$ . The  $z\%$  alludes to annotators in general and not a specific persona. Further, it should be noted that these percentages are not available as a part of HateXplain, and added by us to introduce the "anchoring" information.
+
+Insights. In line with existing evidence (Zhao et al., 2021) of majority labeling bias in few-shot learning, we also establish that LLMs are prone to labeling bias even under zero-shot settings if the context mentions voting percentages. From Figures
+
+5 (a-b) regarding  $p_{vote}^{H}$ , it is evident that alignment in hate labeling is more consistent when the percentages lie closer to each other and decrease as one moves away. Succinctly,  $IAA_{ij} > IAA_{ik}$  if  $|z_{j} - z_{i}| < |z_{k} - z_{i}|$ . A similar pattern is observed for  $p_{vote}^{N}$  (Figures 5 (c-d)).
+
+Takeways. Corroborated by significance testing (Appendix D) and controlling for decoding temperature (Appendix C), we discover that while LLMs understand relative percentages, they are prone to emphasize this information over the post's content. As in our RQ, voting serves as a proxy of numerical metadata in the content moderation pipeline; it implies that corruption (accidental or intentional) can lead to instance misclassification. Despite numerical/statistical features being used earlier to enhance the finetuning in hate speech detection (Founta et al., 2019; Kulkarni et al., 2023), they cannot be directly extrapolated as features in prompting for hate speech annotations.
+
+# 7 Discussion
+
+This section summarizes the implications of the research. We also provide recommendations for LLM-assisted content moderation. Except for the F1anT5-XXL non-hate persona, the results across models, datasets, languages, and prompts are aligned. Cognizant of adversarial attacks on prompting, the study balances both feature importance (corroborated by significance testing) and cautions against adversary features. Despite how the perception of hatefulness manifests via  $p_{\text{base}}$ , we observe that the inclusion of cues nudges the LLMs to explore a contextual definition of hate.
+
+Geographical Sensitivity. How the latent spaces are triggered at the mention of geographical cues (Zhou et al., 2022) is intractable from prompting alone. This becomes especially tricky for overlapping signals like 'Arabic,' 'Muslims,' and 'Islamophobia' (Figure 3 (c)). It once again highlights the need for more transparency in the training and cultural alignment of LLMs. As a quick fix, we suggest augmenting the language/country tags for an incoming post being prompted to improve alignment with people representative of a given region/geography.
+
+Demographic Sensitivity. Our work establishes that the manner of personification of demographic attributes can lead to variation in hate labeling, with assumed persona  $(p_{trait}^{A})$  being closer to the inherent knowledge and biases an LLM possesses. Here,
+
+we caution practitioners looking to adopt LLMs for crowdsourcing to experiment with different framing of the personas and identity traits.
+
+Numerical Sensitivity. As evident from our experiments, LLMs do not have a clear way of discovering noise from the informativeness in the prompt. Our advice to practitioners is to refrain from adding any numerical statistics in the prompt unless they are quality-checked and to mask such adversarial expressions written by the authors of the input samples.
+
+Multilingual Sensitivity. Following the same prompt format, we observe similar performance patterns across both English and Multilingual settings (refer Table 5 in Appendix B). The results are encouraging, allowing content moderators to work in the native language of the post without extraneous prompt engineering. The degradation in performance from English to Multilingual (Jin et al., 2024), however, calls for more investment in non-English training and evaluation of LLMs.
+
+LLMs for Hate Speech Annotations. It is imperative to point out that hate speech is a human-centric phenomenon steeped in historical and cultural contexts. As such, any computational attempt to flag it can only be assistive. Our experiments over three contextual settings indicate that LLM cannot outright substitute a demographic group in the annotation process (Section 5). Further, if LLMs are prone to anchoring, that is not always beneficial. Then, our study opens up the questions around using few-shot/guideline examples for labeling hate (Sections 5 and 6).
+
+# 8 Conclusion
+
+In the current LLM space, both alignment and robustness do not have a singular non-overlapping definition, which makes our study sit at the intersection of robustness in hate speech detection as well as conceptualization of the human-LLM alignment. Over multiple RQs and prompt setups, we explore how well-suited LLMs are for assisting humans at different stages of the content moderation pipeline. Our results on demographic sensitivity, cultural priming, and anchoring bias are evident over multiple annotations on 6 datasets, 5 languages, and 2 LLMs. Our analysis reiterates using LLMs as an assistive system rather than substituting human moderators. We would like to work on more datasets and explore intersectional demographic attributes in the future.
+
+# 9 Limitations
+
+First and foremost, the list of research questions and prompts analyzed, while significant in number, is not exhaustive. Given that the adoption of LLMs in the hate speech moderation pipeline is nascent, the scope for research is more vast than what we can account for in current research. We hope our findings and primary analysis motivate future research. Besides, there is a shortage of open-source instruction-tuned multilingual LLMs, which restricts our multilingual and multicultural analysis to GPT-3.5. Owing to the lack of datasets with ground truth labels where multiple annotators are considered and demographic-specific annotators are released, performing one-to-one mapped inter-annotator agreement analysis between LLMs and demographic personas is challenging. The current study only utilizes textual content. In the future, we need to conduct assessments for content moderation on other modalities like memes, short videos, etc. It should be noted that geography or language only acts as a proxy for cultural and linguistic recalls and does not represent an absolute assessment for multilingual LLM evaluations. Consequently, when employing geographical or demographic attributes, we need to be cognizant of the fact that the sensitivity of the LLMs can also be a source of their implicit bias.
+
+# 10 Ethical Considerations
+
+This work does not produce any new scientific artifacts regarding datasets or proposed models, and no human evaluators were involved. Our work utilizes LLMs to annotate (Gilardi et al., 2023) hate speech, which comes with legal, technical, and ethical challenges. There is an inherent issue of magnifying human behaviors and biases by LLMs. Our findings corroborate the sensitivity of LLMs towards different demographics and identity traits. Hence, their generations/predictions should be taken cautiously. Further, this study is underpinned by gold labels obtained from human annotators, which themselves could be erroneous. The intent of releasing our prompt list is to broaden the examination of biases and fallacies an LLM is prone to. While F1anT5-XXL is an opensource LLM, GPT-3.5 is not, which raises concerns around reproducibility (Ollion et al., 2024), and is a broader challenge for the NLP community.
+
+# Acknowledgements
+
+Sarah Masud acknowledges the support of the Prime Minister Doctoral Fellowship in association with Wipro AI and Google PhD Fellowship. Viktor Hangya and Alexander Fraser acknowledge funding from the German Research Foundation (DFG; grant FR 2829/7-1). Tanmoy Chakraborty acknowledges Anusandhan National Research Foundation (CRG/2023/001351) for the financial support. The authors thank Ahmed Tarek Sobhi Abdelrahman and Nicolas Payen for validating the French, German, and Arabic prompts.
+
+# References
+
+Gavin Abercrombie, Dirk Hovy, and Vinodkumar Prabhakaran. 2023. Temporal and second language influence on intra-annotator agreement and stability in hate speech labelling. In Proc. of the 17th Linguistic Annotation Workshop (LAW-XVII), pages 96-103, Toronto, Canada.  
+Badr AlKhamissi, Faisal Ladhak, Srinivasan Iyer, Veselin Stoyanov, Zornitsa Kozareva, Xian Li, Pascale Fung, Lambert Mathias, Asli Celikyilmaz, and Mona Diab. 2022. ToKen: Task decomposition and knowledge infusion for few-shot hate speech detection. In Proc. of the 2022 Conference on EMNLP, pages 2109-2120, Abu Dhabi, United Arab Emirates. ACL.  
+Lora Aroyo, Lucas Dixon, Nithum Thain, Olivia Redfield, and Rachel Rosen. 2019a. Crowdsourcing subjective tasks: The case study of understanding toxicity in online discussions. In *Companion Proceedings of The 2019 World Wide Web Conference*, page 1100–1105, New York, NY, USA. Association for Computing Machinery.  
+Lora Aroyo, Anca Dumitrache, Oana Inel, Zoltán Szlák, Benjamin Timmermans, and Chris Welty. 2019b. Crowdsourcing inclusivity: Dealing with diversity of opinions, perspectives and ambiguity in annotated data. In *Companion Proceedings of The 2019 World Wide Web Conference*, page 1294–1295, New York, NY, USA. Association for Computing Machinery.  
+Laura Biester, Vanita Sharma, Ashkan Kazemi, Naihao Deng, Steven Wilson, and Rada Mihalcea. 2022. Analyzing the effects of annotator gender across NLP tasks. In Proc. of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 10-19, Marseille, France. European Language Resources Association.  
+Alexander Braylan and Matthew Lease. 2020. Modeling and aggregation of complex annotations via annotation distances. In Proc. of The Web Conference 2020, WWW '20, page 1807-1818, New York, NY, USA. Association for Computing Machinery.
+
+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.  
+Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37-46.  
+Antigoni Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large scale crowdsourcing and characterization of twitter abusive behavior. In Proceedings of the international AAAI conference on web and social media, volume 12.  
+Antigoni Maria Founta, Despoina Chatzakou, Nicolas Kourtellis, Jeremy Blackburn, Athena Vakali, and Ilias Leontiadis. 2019. A unified deep learning architecture for abuse detection. In Proceedings of the 10th ACM Conference on Web Science, WebSci '19, page 105-114, New York, NY, USA. Association for Computing Machinery.  
+Tanmay Garg, Sarah Masud, Tharun Suresh, and Tan-moy Chakraborty. 2023. Handling bias in toxic speech detection: A survey. ACM Comput. Surv., 55(13s).  
+Fabrizio Gilardi, Meysam Alizadeh, and Mael Kubli. 2023. Chatgpt outperforms crowd workers for text-annotation tasks. Proc. of the National Academy of Sciences, 120(30):e2305016120.  
+Lewis Griffin, Bennett Kleinberg, Maximilian Mozes, Kimberly Mai, Maria Do Mar Vau, Matthew Caldwell, and Augustine Mavor-Parker. 2023. Large language models respond to influence like humans. In Proc. of the First Workshop on Social Influence in Conversations (SICON 2023), pages 15-24, Toronto, Canada. ACL.  
+Keyan Guo, Alexander Hu, Jaden Mu, Ziheng Shi, Ziming Zhao, Nishant Vishwamitra, and Hongxin Hu. 2023. An investigation of large language models for real-world hate speech detection. In 2023 International Conference on Machine Learning and Applications (ICMLA), pages 1568-1573.  
+Xingwei He, Zhenghao Lin, Yeyun Gong, Alex Jin, Hang Zhang, Chen Lin, Jian Jiao, Siu Ming Yiu, Nan Duan, Weizhu Chen, et al. 2023. Annollm: Making large language models to be better crowdsourced annotators. arXiv preprint arXiv:2303.16854.  
+Fan Huang, Haewoon Kwak, and Jisun An. 2023. Is chatgpt better than human annotators? potential and limitations of chatgpt in explaining implicit hate speech. In Companion Proceedings of the ACM Web Conference 2023, WWW '23 Companion, page 294-297, New York, NY, USA. Association for Computing Machinery.
+
+Musa İhtiyar, Ömer Özdemir, Mustafa Erengül, and Arzucan Özgür. 2023. A dataset for investigating the impact of context for offensive language detection in tweets. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1543-1549, 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. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Yiqiao Jin, Mohit Chandra, Gaurav Verma, Yibo Hu, Munmun De Choudhury, and Srijan Kumar. 2024. Better to ask in english: Cross-lingual evaluation of large language models for healthcare queries. In Proceedings of the ACM on Web Conference 2024, WWW '24, page 2627-2638, New York, NY, USA. Association for Computing Machinery.  
+Kamil Kanclerz, Marcin Gruza, Konrad Karanowski, Julita Bielaniewicz, Piotr Milkowski, Jan Kocon, and Przemyslaw Kazienko. 2022. What if ground truth is subjective? personalized deep neural hate speech detection. In Proc. of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 37-45, Marseille, France. European Language Resources Association.  
+Aman Khullar, Daniel Nkemelu, Cuong V. Nguyen, and Michael L. Best. 2023. Hate speech detection in limited data contexts using synthetic data generation. ACM J. Comput. Sustain. Soc.  
+Jan Kocón, Igor Cichecki, Oliwier Kaszyca, Mateusz Kochanek, Dominika Szydlo, Joanna Baran, Julita Bieleniewicz, Marcin Gruza, Arkadiusz Janz, Kamil Kanclerz, Anna Kocón, Bartlomiej Koptyra, Wiktoria Mieleszczenko-Kowszewicz, Piotr Miłkowski, Marcin Oleksy, Maciej Piasecki, Łukasz Radlinski, Konrad Wojtasik, Stanisław Wozniak, and Przemysław Kazienko. 2023. Chatgpt: Jack of all trades, master of none. Information Fusion, 99:101861.  
+Anna Koufakou, Endang Wahyu Pamungkas, Valerio Basile, and Viviana Patti. 2020. HurtBERT: Incorporating lexical features with BERT for the detection of abusive language. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 34-43, Online. Association for Computational Linguistics.  
+Atharva Kulkarni, Sarah Masud, Vikram Goyal, and Tanmoy Chakraborty. 2023. Revisiting hate speech benchmarks: From data curation to system deployment. In Proc. of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '23, page 4333-4345, New York, NY, USA. Association for Computing Machinery.  
+Thai Le, Long Tran-Thanh, and Dongwon Lee. 2022. Socialbots on fire: Modeling adversarial behaviors of socialbots via multi-agent hierarchical reinforcement learning. In Proc. of the ACM Web Conference 2022,
+
+WWW '22, page 545-554, New York, NY, USA. Association for Computing Machinery.  
+Nayeon Lee, Chani Jung, Junho Myung, Jiho Jin, Juho Kim, and Alice Oh. 2023. CReHate: Cross-cultural Re-annotation of English Hate Speech Dataset. arXiv preprint arXiv:2308.16705.  
+Dan Li, Zhaochun Ren, and Evangelos Kanoulas. 2021. Crowdgp: A gaussian process model for inferring relevance from crowd annotations. In Proc. of the Web Conference 2021, WWW '21, page 1821-1832, New York, NY, USA. Association for Computing Machinery.  
+Yizhi Li, Ge Zhang, Bohao Yang, Chenghua Lin, Anton Ragni, Shi Wang, and Jie Fu. 2022. HERB: Measuring hierarchical regional bias in pre-trained language models. In *Findings of the ACL: AACL-IJCNLP* 2022, pages 334-346, Online only. ACL.  
+Ruixi Lin and Hwee Tou Ng. 2023. Mind the biases: Quantifying cognitive biases in language model prompting. In Findings of the ACL: ACL 2023, pages 5269-5281, Toronto, Canada. ACL.  
+Nikola Ljubesic, Igor Mozetic, and Petra Kralj Novak. 2022. Quantifying the impact of context on the quality of manual hate speech annotation. Natural Language Engineering, pages 1-14.  
+Thomas Mandl, Sandip Modha, Anand Kumar M, and Bharathi Raja Chakravarthi. 2020. Overview of the HASOC Track at FIRE 2020: Hate Speech and Offensive Language Identification in Tamil, Malayalam, Hindi, English and German. In Proc. of the 12th Annual Meeting of the Forum for Information Retrieval Evaluation, pages 29-32.  
+Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. HateXplain: A Benchmark Dataset for Explanable Hate Speech Detection. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 14867-14875.  
+Luke Munn. 2020. Angry by design: toxic communication and technical architectures. *Humanities and Social Sciences Communications*, 7(1):53.  
+Venkata Prabhakara Sarath Nookala, Gaurav Verma, Subhabrata Mukherjee, and Srijan Kumar. 2023. Adversarial robustness of prompt-based few-shot learning for natural language understanding. In *Findings of the ACL: ACL* 2023, pages 2196-2208, Toronto, Canada. ACL.  
+Étienne Ollion, Rubing Shen, Ana Macanovic, and Arnault Chatelain. 2024. The dangers of using proprietary LLMs for research. Nature Machine Intelligence, 6(1):4-5.  
+Matthias Orlikowski, Paul Röttger, Philipp Cimiano, and Dirk Hovy. 2023. The ecological fallacy in annotation: Modeling human label variation goes beyond sociodemographics. In Proc. of the 61st Annual
+
+Meeting of the ACL (Volume 2: Short Papers), pages 1017-1029, Toronto, Canada. ACL.  
+Lidiia Ostyakova, Veronika Smilga, Kseniia Petukhova, Maria Molchanova, and Daniel Kornev. 2023. ChatGPT vs. crowdsourcing vs. experts: Annotating open-domain conversations with speech functions. In Proc. of the 24th Meeting of the Special Interest Group on Discourse and Dialogue, pages 242-254, Prague, Czechia. ACL.  
+Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, and Dit-Yan Yeung. 2019. Multilingual and multi-aspect hate speech analysis. In Proc. of EMNLP. ACL.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, volume 35, pages 27730-27744. Curran Associates, Inc.  
+Russell Ritchie, Sachin Grover, and Fuchiang (Rich) Tsui. 2022. Inter-annotator agreement is not the ceiling of machine learning performance: Evidence from a comprehensive set of simulations. In Proceedings of the 21st Workshop on Biomedical Language Processing, pages 275–284, Dublin, Ireland. Association for Computational Linguistics.  
+Bjorn Ross, Michael Rist, Guillermo Carbonell, Benjamin Cabrera, Nils Kurowsky, and Michael Wojatzki. 2016. Measuring the reliability of hate speech annotations: The case of the european refugee crisis. In NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication, Bochumer Linguistische Arbeitsberichte, pages 6-9, Germany. Ruhr-Universitat Bochum.  
+Paul Rottger, Bertie Vidgen, Dirk Hovy, and Janet Pierrehumbert. 2022. Two contrasting data annotation paradigms for subjective NLP tasks. In Proc. of the 2022 Conference of the North American Chapter of the ACL: Human Language Technologies, pages 175-190, Seattle, United States. ACL.  
+Sarthak Roy, Ashish Harshvardhan, Animesh Mukherjee, and Punyajoy Saha. 2023. Probing LLMs for hate speech detection: strengths and vulnerabilities. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 6116-6128, Singapore. Association for Computational Linguistics.  
+Pratik Sachdeva, Renata Barreto, Geoff Bacon, Alexander Sahn, Claudia von Vacano, and Chris Kennedy. 2022. The measuring hate speech corpus: Leveraging rasch measurement theory for data perspectivism. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 83-94, Marseille, France. European Language Resources Association.
+
+Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. The risk of racial bias in hate speech detection. In Proc. of the 57th Annual Meeting of the ACL, pages 1668-1678, Florence, Italy. ACL.  
+Maarten Sap, Swabha Swayamdipta, Laura Vianna, Xuhui Zhou, Yejin Choi, and Noah A. Smith. 2022. Annotators with attitudes: How annotator beliefs and identities bias toxic language detection. In Proc. of the 2022 Conference of the North American Chapter of the ACL: Human Language Technologies, pages 5884-5906, Seattle, United States. ACL.  
+P Slovic, B Fischhoff, and S Lichtenstein. 1977. Behavioral decision theory. Annual Review of Psychology, 28(Volume 28, 1977):1-39.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models.  
+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.  
+Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. 2023. Large language models are not fair evaluators. ArXiv, abs/2305.17926.  
+Zeerak Waseem. 2016. Are you a racist or am I seeing things? annotator influence on hate speech detection on Twitter. In Proc. of the First Workshop on NLP and Computational Social Science, pages 138-142, Austin, Texas. ACL.  
+Tharindu Weerasooriya, Sujan Dutta, Tharindu Ranasinghe, Marcos Zampieri, Christopher Homan, and Ashiqur KhudaBukhsh. 2023. Vicarious offense and noise audit of offensive speech classifiers: Unifying human and machine disagreement on what is offensive. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 11648-11668, Singapore. Association for Computational Linguistics.  
+Maximilian Wich, Christian Widmer, Gerhard Hagerer, and Georg Groh. 2021. Investigating annotator bias in abusive language datasets. In Proc. of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1515-1525, Held Online. INCOMA Ltd.  
+Michael Maximilian Wojatzki. 2018. Do women perceive hate differently: Examining the relationship between hate speech, gender, and agreement judgments. In Proceedings of the 14th Conference on Natural Language Processing (KONVENS 2018). Deutsche Nationalbibliothek.
+
+Yongjin Yang, Joonkee Kim, Yujin Kim, Namgyu Ho, James Thorne, and Se-Young Yun. 2023. HARE: Explainable hate speech detection with step-by-step reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5490-5505, Singapore. Association for Computational Linguistics.  
+Wenjie Yin and Arkaitz Zubiaga. 2021. Towards generalisable hate speech detection: a review on obstacles and solutions. PeerJ Computer Science, 7:e598.  
+J.D. Zamfirescu-Pereira, Richmond Y. Wong, Bjoern Hartmann, and Qian Yang. 2023. Why johnny can't prompt: How non-ai experts try (and fail) to design llm prompts. In Proc. of the 2023 CHI Conference on Human Factors in Computing Systems, CHI '23, New York, NY, USA. Association for Computing Machinery.  
+Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. 2023. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792.  
+Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In Proc. of the 38th International Conference on Machine Learning, volume 139 of Proc. of Machine Learning Research, pages 12697-12706. PMLR.  
+Kaitlyn Zhou, Kawin Ethayarajh, and Dan Jurafsky. 2022. Richer countries and richer representations. In *Findings of the ACL: ACL* 2022, pages 2074–2085, Dublin, Ireland. ACL.  
+Julia El Zini and Mariette Awad. 2022. On the explainability of natural language processing deep models. ACM Comput. Surv., 55(5).  
+Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. 2023. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043.
+
+# A RQ1 Multilingual Prompting
+
+We run additional experiments where the cues and query are in the same language as the multilingual post. The comparative results are enlisted in Table 5. Despite the expected loss in performance (Jin et al., 2024), the inclusion of  $p_{lang}$  leads to a better human-AI alignment.
+
+# B RQ2 Multilingual Persona Prompting
+
+Personas. We start with Wikipedia to get a general sense of the demographics of each geography. From there, we narrow down the most prominent demography of the nation. We further use census and news articles and consult social experts about each geography before narrowing down the vulnerable minority based on the hypothesis that minority groups get more hate than the majority.
+
+Prompting. From Tables 6 and 7, we again observe the same pattern as in English, thereby showcasing that the role of context is significant irrespective of the language under consideration. Although we have seen a similar pattern in PHLR, there is a significant degradation in the observed metrics compared to English.
+
+# C RQ3 Temperature Probing
+
+We simulate multiple annotators under each  $z$  value to further corroborate the results. Due to resource constrain, we run this analysis on only F1anT5-XXL. We generate 100 output for each sample, by uniformly sampling 100 decoder temperature values  $(t \in (0,2))$ . For reference, in  $p_{base}$ , we obtain a mean (std. dev) percentage of hate label as  $0.560(\pm 0.0169)$ . While the spread is low for  $p_{base}$ , we observe from Figure 6 (a) and (b) that for some  $z$ , the spread varies. It is an indicator that decoding temperature can lead to variations in the LLM's predictions when prompted on numerical anchors. We observe more spread on average when using  $p_{vote}$  as recorded in Table 8. For  $p_{vote}^H$ , with varying temperatures, as the value of  $z$  increases, the percentage of hate predicted increases, as evident from the shift towards the right for  $z = 100\%$  in Figure 6 (a). The reverse trend is observed for  $p_{vote}^N$  in Figure 6 (b) where the curve for  $z = 100\%$  is left shifted, leading to a decrease in the percentage of hate predicted as the majority of non-hateful increases.
+
+# D Statistical Testing on FlanT5-XXL
+
+Process. For RQ1 (Sec. 4) and RQ3 (Sec. 6), we perform the paired t-test and report the effect size. Meanwhile, for RQ2 (Sec. 5), to capture the intrademography disparity among the subclasses, we use ANOVA. All experiments are run via the SciPy and NumPy libraries.
+
+Observations. For all the RQs, we observe significant disparity caused by context.
+
+- RQ1: From Table 9, comparing both adjusted F1 and IAA, we observe that adding the country cue leads to a significant change in the prompted output (as captured by the higher effect size (ES) as well as the p-values). Following the reference Figure 3 (a), South Africa registers the highest impact.  
+- RQ2: From Table 10 we observe that different modes of the persona  $(p^{H}, p^{N}, p^{A})$  are impacted by varying sub-classes. Interestingly, the subgroups within Religion show considerable variation for three persona prompts.  
+- RQ3: From Table 11, for both  $p^H$  and  $p^N$ , we observe significant differences (as captured by the higher effect size (ES) as well as the p-values) in performance among the various percentages in the prompt. This further corroborates the anchoring bias in LLMs employed for zero-shot labeling.
+
+<table><tr><td rowspan="2">Language</td><td colspan="2">Prompt in English</td><td colspan="2">Prompt in 
+same language</td></tr><tr><td>pbase</td><td>plang</td><td>pbase</td><td>plang</td></tr><tr><td>Arabic</td><td>0.580</td><td>0.660</td><td>0.140</td><td>0.305</td></tr><tr><td>French</td><td>0.344</td><td>0.425</td><td>0.272</td><td>0.356</td></tr><tr><td>German</td><td>0.502</td><td>0.537</td><td>0.412</td><td>0.423</td></tr><tr><td>Hindi</td><td>0.242</td><td>0.371</td><td>0.018</td><td>0.031</td></tr></table>
+
+Table 5: [RQ1] Extension of Figure 3 (c) for IAA comparison of the results without and with language cues for prompts in English and the respective language.  
+
+<table><tr><td rowspan="2">Language</td><td rowspan="2">Majority or vulnerable</td><td colspan="2">Prompt in English</td><td colspan="2">Prompt in same language</td></tr><tr><td>pH</td><td>pN</td><td>pH</td><td>pN</td></tr><tr><td rowspan="2">Arabic</td><td>Muslim</td><td>0.778</td><td>0.364</td><td>0.992</td><td>0.737</td></tr><tr><td>Non-muslim</td><td>0.586</td><td>0.492</td><td>0.525</td><td>0.483</td></tr><tr><td rowspan="2">French</td><td>French descent</td><td>0.558</td><td>0.262</td><td>0.666</td><td>0.170</td></tr><tr><td>Mediterranean descent</td><td>0.520</td><td>0.298</td><td>0.649</td><td>0.176</td></tr><tr><td rowspan="2">German</td><td>Native</td><td>0.724</td><td>0.076</td><td>0.566</td><td>0.152</td></tr><tr><td>Non-native</td><td>0.374</td><td>0.170</td><td>0.248</td><td>0.248</td></tr><tr><td rowspan="2">Hindi</td><td>Upper caste</td><td>0.528</td><td>0.192</td><td>0.998</td><td>0.282</td></tr><tr><td>Lower caste</td><td>0.768</td><td>0.014</td><td>0.998</td><td>0.094</td></tr></table>
+
+Table 6: [RQ2] PHLR for comparison of vulnerable persona cues. Extension of Figure 4 (a-b).
+
+![](images/e902dcaa56161f82e727ad223a3c91e17d99a737648337faaf2c3f9119ba0336.jpg)  
+(a) FlanT5-XXL temperatures (H)
+
+![](images/fe5deefa2fba6dfce402c9da0b30ad241997d2adf5a26dd7bf95d2fe70d35f67.jpg)  
+(b) FlanT5-XXL temperatures (N)  
+Figure 6: [RQ3] The impact of decoding temperature with varying voting percentages for  $p_{\text{context}}^H$  on F1anT5-XXL.
+
+<table><tr><td rowspan="2">Language</td><td rowspan="2">Speaker</td><td colspan="2">Prompt in English</td><td colspan="2">Prompt in same language</td></tr><tr><td>PH</td><td>PN</td><td>PH</td><td>PN</td></tr><tr><td rowspan="2">Arabic</td><td>Native</td><td>0.890</td><td>0.100</td><td>0.764</td><td>0.099</td></tr><tr><td>Non-native</td><td>0.646</td><td>0.380</td><td>0.567</td><td>0.901</td></tr><tr><td rowspan="2">French</td><td>Native</td><td>0.752</td><td>0.232</td><td>0.916</td><td>0.130</td></tr><tr><td>Non-native</td><td>0.462</td><td>0.286</td><td>0.702</td><td>0.106</td></tr><tr><td rowspan="2">German</td><td>Native</td><td>0.724</td><td>0.076</td><td>0.566</td><td>0.152</td></tr><tr><td>Non-native</td><td>0.374</td><td>0.170</td><td>0.248</td><td>0.248</td></tr><tr><td rowspan="2">Hindi</td><td>Native</td><td>0.852</td><td>0.004</td><td>1.000</td><td>0.753</td></tr><tr><td>Non-native</td><td>0.328</td><td>0.038</td><td>1.000</td><td>0.858</td></tr></table>
+
+Table 7: [RQ2] PHLR for comparison of speaker persona cues. Extension of Figure 4 (c-d).  
+
+<table><tr><td rowspan="2">&lt;z&gt;</td><td colspan="2">Distribution parameters (μ ± σ)</td></tr><tr><td>pHvote</td><td>PNvote</td></tr><tr><td>0%</td><td>0.445 ± 0.0078</td><td>0.77 ± 0.067</td></tr><tr><td>25%</td><td>0.67 ± 0.037</td><td>0.56 ± 0.024</td></tr><tr><td>50%</td><td>0.72 ± 0.037</td><td>0.57 ± 0.065</td></tr><tr><td>75%</td><td>0.79 ± 0.039</td><td>0.443 ± 0.0078</td></tr><tr><td>100%</td><td>0.81 ± 0.038</td><td>0.11 ± 0.079</td></tr></table>
+
+Table 8: [RQ3] Mean and standard deviation of temperature distribution in Figure 6.  
+
+<table><tr><td>Country</td><td>ES F1</td><td>ES IAA</td></tr><tr><td>United States</td><td>1.116*</td><td>1.165*</td></tr><tr><td>Australia</td><td>0.731*</td><td>0.729*</td></tr><tr><td>United Kingdom</td><td>0.723*</td><td>0.707*</td></tr><tr><td>South Africa</td><td>1.244*</td><td>1.495*</td></tr><tr><td>Singapore</td><td>0.782*</td><td>0.834*</td></tr></table>
+
+Table 9: [RQ1] Effect size to indicate the significance of including the country cue for F1anT5-XXL in English.  $(p\leq 0.05)$  and  $^{**}(p\leq 0.001)$  indicate whether the difference is significant.  
+
+<table><tr><td>Annotator demographics</td><td>pH</td><td>PN</td><td>PA</td></tr><tr><td>Gender</td><td>0.010**</td><td>0.007</td><td>0.013*</td></tr><tr><td>Ethnicity</td><td>0.031**</td><td>0.027*</td><td>0.070</td></tr><tr><td>Political orientation</td><td>0.021*</td><td>0.011</td><td>0.007**</td></tr><tr><td>Religion</td><td>0.054**</td><td>0.021**</td><td>0.075**</td></tr><tr><td>Education level</td><td>0.013*</td><td>0.003</td><td>0.041</td></tr></table>
+
+Table 10: [RQ2] The absolute difference between the minimum and maximum IAA obtained for the respective persona, demographic, and LLM combination.  $(p\leq 0.05)$  and  $^{**}(p\leq 0.001)$  indicate if ANNOVA is significant among the sub-classes within a demographic.  
+
+<table><tr><td rowspan="2">Key</td><td rowspan="2">Value</td><td colspan="2">pH</td><td colspan="2">pN</td></tr><tr><td>ES F1</td><td>ES IAA</td><td>ES F1</td><td>ES IAA</td></tr><tr><td rowspan="4">0</td><td>25</td><td>2.24*</td><td>1.878*</td><td>6.379</td><td>4.47*</td></tr><tr><td>50</td><td>2.732*</td><td>2.362*</td><td>1.25*</td><td>0.233</td></tr><tr><td>75</td><td>3.805*</td><td>3.475*</td><td>4.342*</td><td>2.661*</td></tr><tr><td>100</td><td>4.037*</td><td>3.583*</td><td>0.178</td><td>1.637**</td></tr><tr><td rowspan="3">25</td><td>50</td><td>1.53*</td><td>1.334*</td><td>4.416*</td><td>3.748*</td></tr><tr><td>75</td><td>2.933*</td><td>2.626*</td><td>1.065*</td><td>1.412*</td></tr><tr><td>100</td><td>3.23*</td><td>2.818*</td><td>3.064*</td><td>4.91*</td></tr><tr><td rowspan="2">50</td><td>75</td><td>2.516*</td><td>2.231*</td><td>3.701*</td><td>2.6*</td></tr><tr><td>100</td><td>2.738*</td><td>2.36*</td><td>0.451*</td><td>1.559*</td></tr><tr><td>75</td><td>100</td><td>0.875*</td><td>0.773*</td><td>2.922*</td><td>3.493*</td></tr></table>
+
+Table 11: [RQ3] Effect size for percentage pairs 'x' (key) and 'y' (value) to indicate the significance of including the percentage cues.  $*(p \leq 0.05)$  and  $**(p \leq 0.001)$  indicate if the difference is significant or not.
+
+<table><tr><td>RQ</td><td>Prompt type</td><td>Prompt(s)</td></tr><tr><td>Base</td><td>\(p_{base}\)</td><td>Statement:</td></tr><tr><td rowspan="2">\(RQ_1\)</td><td>\(p_{con}\)</td><td>The following statement was made in:c:</td></tr><tr><td>\(p_{lang}\)</td><td>The following statement was made inl language:</td></tr><tr><td rowspan="4">\(RQ_2\)</td><td rowspan="2">\(p_{trait}^{H/N}\)</td><td>A &lt;g&gt; annotated the following statement as &lt;H/N&gt;: &lt;POST&gt;.A person of &lt;e&gt; ethnicity annotated the following statement as &lt;H/N&gt;: &lt;POST&gt;.A person of &lt;p&gt; political orientation annotated the following statement as &lt;H/N&gt;: &lt;POST&gt;.A person who follows &lt;r&gt; annotated the following statement as &lt;H/N&gt;: &lt;POST&gt;.An irreligious person annotated the following statement as &lt;H/N&gt;: &lt;POST&gt;.A person with an education level less than high school annotated the following statement as &lt;H/N&gt;:</td></tr><tr><td>&lt;POST&gt;.A person with education level of &lt;q&gt; annotated the following statement as &lt;H/N&gt;: &lt;POST&gt;.</td></tr><tr><td>\(p_{trait}^A\)</td><td>Suppose you are a &lt;g&gt; and given the following statement to annotate: &lt;POST&gt;.Suppose you are a person of &lt;e&gt; ethnicity and given the following statement to annotate: &lt;POST&gt;.Suppose you are a person of &lt;p&gt; political orientation and given the following statement to annotate:&lt;POST&gt;.Suppose you are a person who follows &lt;r&gt; and given the following statement to annotate: &lt;POST&gt;.Suppose you are an irreligious person and given the following statement to annotate: &lt;POST&gt;.Suppose you are a person with an education level less than high school and given the following statement to annotate: &lt;POST&gt;.Suppose you are a person with education level of &lt;q&gt; and given the following statement to annotate:&lt;POST&gt;.</td></tr><tr><td>\(p_{trait}^{L*}\)</td><td>The following statement is in &lt;l&gt; language and a &lt;B/V&gt; annotated this as &lt;H/N&gt;: &lt;POST&gt;.The following statement is in &lt;l&gt; language and a &lt;N/NN&gt; speaker annotated this as &lt;H/N&gt;:</td></tr><tr><td>\(RQ_3\)</td><td>\(p_{vote}^{H/N}\)</td><td>The following statement was labeled &lt;H/N&gt; by &lt;z&gt; annotators: &lt;POST&gt;.</td></tr></table>
+
+Table 12: Prompts employed in this study for all settings under RQ1, RQ2, and RQ3. For F1anT5 (all), prompt  $\leftarrow$  prompt + "Is the given statement hateful?" For Llama3-8B, Mistral-7B, Zephyr-7B and GPT-3.5, prompt  $\leftarrow$  prompt + "Is the given statement hateful? Answer in one word only." Here,  $c\in \{United States,Australia,United Kingdom,South Africa,Singapore\}$ $l\in \{\text{Arabic,French,German,Hindi}\}$ $H / N\in \{\text{hateful,non-hateful}\}$ $g\in \{\text{male,}$  female, non-binary},  $e\in \{\text{asian,black, hispanic,middle eastern,white}\}$ $p\in \{\text{liberal,independent,conservative}\}$ $r\in \{\text{christian,buddhism,islam,judaism,hinduism}\}$ $q\in \{\text{high school,college,bachelor,master's,PhD}\}$ $B / V\in$  {{muslim, non-muslim}}, {french descent, mediterranean descent}, {native german, non-native}, {person belongs to upper caste, person belongs to lower caste}},  $N / NN\in \{\text{native,non-native}\}$ $z\in \{0\% ,25\% ,50\% ,75\% ,100\% \}$
+
+<table><tr><td>Post language</td><td>Prompt in English language</td><td>Prompt in same language</td><td></td></tr><tr><td>Arabic</td><td>The following statement is in Arabic language and a non-muslim annotated this as non-hateful: صحي كان كادب الحير واني بعس رورل مصطفى لابي علّة الحرفية بعس رورل مصطفى لتربيعة الحرفية بعس رورل مصطفى بعس رورل مصطفى لتربيعة الحرفية بعس رورل مصطفى بعس رورل مصطفergic.
+url@....الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله    الله   ​
+Is the given statement hateful? Answer in one word only.</td><td>الترجمة الحرفية ممنواني بعس رورل مصطفergic.
+الترجمة الحرفية ممنواني بعس رورل مصطفergic.
+الترجمة الحرفية ممنواني بعس رورل مصطفergic.
+الترجمة الحرفية ممنواني بعس رورل مصطفergic.
+الترجمة الحرفية ممنواني بعس رورل مصطفergic.
+الترجمة الحرفية ممنواني بعس رورl مصطفergic.
+الترجمة الحرفية ممنواني بعس رورl مصطفergic.
+الTRL@saitanBacha: @Kanatunga.
+الTRL@saitanBacha: @Kanatunga.
+الTRL@saitanBacha: @Kanatunga.
+الTRL@saitanBacha: @Kanatunga.
+الTRL@saitanBacha: @Kanatunga.
+الTRL@saitanBacha: @Kanatunga.
+الTRL@saitanBacha: @Kanatunga.
+الTRL@saitanBacha; @Kanatunga.
+الTRL@saitanBacha; @Kanatunga.
+الTRL@saitanBacha; @Kanatunga.
+الTRL@saitanBacha; @Kanatunga.
+الTRL@saitanBacha; @Kanatunga.
+الTRL@saitanBacha; @Kanatunga.
+الTRL@saitanBacha; @Kanatunga.
+الTRL@SaitanBacha; @Kanatunga.
+الTRL@SaitanBacha; @Kanatunga.
+الTRL@SaitanBacha; @Kanatunga.
+الTRL@SaitanBacha; @Kanatunga.
+الTRL@SaitanBacha; @Kanatunga.
+الTRL@SaitanBacha; @Kanatunga.
+الTRL@SaitanBacha; @Kanatappa.
+الTRL@SaitanBacha; @Kanatappa.
+الTRL@SaitanBacha; @Kanatappa.
+الTRL@SaitanBacha; @Kanatappa.
+الTRL@SaitanBacha; @Kanatappa.
+الTRL@SaitanBacha; @Kanatappa.
+الTRL@SaitanBacha; @Kanatappa.
+الTRL@SaitanBacha, @Kanatappa.
+Is the given statement hateful? Answer in one word only.</td><td>La déclaration suivante est en langue française et une descendance française l&#x27;a annotée comme haineuse: @user @user l&#x27;islamisation en marche! la pauvreté généralisée partout en france. La déclaration faite est elle âne à la justification de la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse à la vieillesse ânon la jérôme % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon% ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon %ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % ânon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon % âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% âNon% �</td></tr><tr><td>German</td><td>The following statement is in German language and a native german speaker annotated this as non-hateful: @neothemarius @niemehrgruen Nicht zu vergessen, einen moderaten ISlam... Is the given statement hateful? Answer in one word only.</td><td>Die folgende Aussage ist in deutscher Sprache und ein deutscher Muttersprachler hat dies als nicht hasserfüllt bezeichnet: @neothemarius @niemehrgruen Nicht zu vergessen, einen moderaten ISlam... Ist die gegebene Aussage hasserfüllt? Antwerten Sie nur mit einem Wort.</td><td></td></tr><tr><td>Hindi</td><td>The following statement is in Hindi language and a person belongs to upper caste annotated this as hateful: RT @saitanBacha: @Kanatunga.
+Is the given statement hateful? Answer in one word only.</td><td>लामिन्तलिकोर्वल कृधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति असχति शिकति असχति शिकति असχति शिकति असχति शिकति असχति शिकति असχति शिकति असχति शिकति असχति शिकति असχति शिजति असχति शिजति असχति शिजति असχति शिजति असχति शिजति असχति शिजति असχति शिजति असχति शिजति असχति शिजति लिकति वविकति वविकति वविकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजति व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia व�िजतdia शिकति
+Is the given statement hateful? Answer in one word only.</td><td>लामिन्तलिकोर्वल कृधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थली योधन शिकति अस्थलia व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति शिकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िकति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगति व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia व�िगतdia शिकति
+Is the given statement hateful? Answer in one word only.</td></tr></table>
+
+Table 13: Random verbatim examples of multilingual prompts in the same language as the post (red) for the corresponding cues and queries (black) in English.

2024/Hate Personified_ Investigating the role of LLMs in content moderation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7cfec1ba81ad12240260613ee07bed09ea1d109c91d01836e92c15169d5d3d0
+size 1172701

2024/Hateful Word in Context Classification/95a8e7f6-2b9f-4389-9551-280f7c5f4ca9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b116011a527f59a6bc6833aad104468c7b77fbc51c1888a3aa8883d0554c1014
+size 730066

2024/Hateful Word in Context Classification/full.md

@@ -0,0 +1,374 @@
+# Hateful Word in Context Classification
+
+Sanne Hoeken<sup>1</sup>, Sina Zarrieß<sup>1</sup> and Özge Alacam<sup>1,2</sup>
+
+<sup>1</sup>Computational Linguistics, Department of Linguistics, Bielefeld University, Germany
+
+$^{2}$ Center for Information and Language Processing, LMU Munich, Germany {sanne.hoeken, sina.zarriess, oezge.alacam}@uni-bielefeld.de
+
+# Abstract
+
+Hate speech detection is a prevalent research field, yet it remains underexplored at the level of word meaning. This is significant, as terms used to convey hate often involve non-standard or novel usages which might be overlooked by commonly leveraged LMs trained on general language use. In this paper, we introduce the Hateful Word in Context Classification (HateWiC) task and present a dataset of  $\sim 4000$  WiC-instances, each labeled by three annotators. Our analyses and computational exploration focus on the interplay between the subjective nature (context-dependent connotations) and the descriptive nature (as described in dictionary definitions) of hateful word senses. HateWiC annotations confirm that hatefulness of a word in context does not always derive from the sense definition alone. We explore the prediction of both majority and individual annotator labels, and we experiment with modeling context- and sense-based inputs. Our findings indicate that including definitions proves effective overall, yet not in cases where hateful connotations vary. Conversely, including annotator demographics becomes more important for mitigating performance drop in subjective hate prediction.
+
+# 1 Introduction
+
+This paper introduces the Hateful Word in Context Classification (HateWiC) task, which aims to determine the harmfulness of a word within a specific context, as illustrated in Figure 1. We argue that hateful word senses are not enough in focus within Hate Speech Detection (HSD) research, and not descriptive only, but highly subjective, asking for another approach than other lexical semantic tasks like Word Sense Disambiguation (WSD).
+
+Hateful senses are not enough in focus within HSD research. The current focus of HSD research predominantly revolves around the classification of entire utterances, such as social me
+
+![](images/7b71f79d90ee6e76a1b070c1484369b9c36c19127bf3483a2d43a8da66b24b81.jpg)  
+Figure 1: Illustration of the HateWiC Classification task and a conceptual semantic space that underlies the targeted phenomena of hate-heterogeneous word senses, highlighting the distinction between the descriptive aspects (e.g. cookie or person) and hateful connotation.
+
+dia posts (Waseem and Hovy, 2016; Davidson et al., 2017). Within these utterances, lexical cues frequently play a significant role in the decision-making process. Yet, the computational modeling of context-specific hateful word meanings remains largely unexplored, with a few exceptions in this direction (Dinu et al., 2021; Hoeken et al., 2023b).
+
+LMs commonly employed in HSD systems demonstrate effective word meaning modeling (Nair et al., 2020), but they tend to lack sensitivity to domain-specific, non-standard or novel word senses (Kumar et al., 2019; Blevins and Zettlemoyer, 2020). This insensitivity becomes particularly critical in detecting hateful word meanings, that are used in unconventional or emerging contexts as the evolution of societal events gives rise to the continuous invention of novel expressions of
+
+hate (Qian et al., 2021). Words within the established lexicon, like Oreo, whose primary meaning may not have any negative connotations (a cookie), are repurposed to convey hate towards particular groups or individuals (e.g. based on ethnicity).
+
+Hateful senses are not descriptive only. Following theoretic work by Frigerio and Tenchini (2019), hateful terms could be positioned along a meaning continuum from descriptive to expressive, closer to but not at the expressive outer end. The descriptive component comprises the truth-conditional attributes of a term, often recorded in dictionary definitions. The expressive component, i.e. the connotation of a term, concerns speakers' attitudes and emotions, making it highly context-specific and subjective. A word's sense definition could imply a hateful connotation, but this is not always the case, such as when used in a playful or self-identifying way (e.g. the third usage in Figure 1). Thus, a word's hateful connotation is not exclusively tied to its descriptive definition, a phenomena which we term as hate-heterogeneous senses, but depends on various contextual factors like conveyed content or the reader's identity. This aspect is often overlooked in HSD systems, typically developed using data reflecting a single (majority) perspective (Zampieri et al., 2019; Mathew et al., 2020).
+
+Our contributions. In this study we address the gap in HSD by focusing on subjective hateful word meanings within context. We introduce the HateWiC dataset, a dataset of  $\sim 4000$  WiC instances for which we collected three hatefulness ratings each. We design methods to classify sense representations and evaluate them both against the majority and the individual annotator's label. In doing so, we experiment with modeling descriptive and subjective aspects of hateful word senses by incorporating sense definitions (as also provided to annotators) and annotator information.
+
+# 2 Related Work
+
+In this section, we discuss previous work on the key aspects of this study: HSD at the word level (2.1), incorporating subjectivity in HSD (2.2), and methods for modeling word senses (2.3).
+
+# 2.1 Hate Speech Detection on Word Level
+
+Although the main body of research into HSD has focused on the level of utterances, some studies have delved into hate speech on the lexical level. Prior to LMs, feature-based HSD systems (e.g. Lee et al. (2018)) often incorporated hate speech lexicons. Wiegand et al. (2018) demonstrated the induction of an abusive word lexicon in a noncontextualized setting. A specific subset of hateful terms within context is addressed by Hoeken et al. (2023b), who modelled slur detection employing a dimension-based method similar to the identification of gender bias in word embeddings (Bolukbasi et al., 2016). This approach, that requires a pre-given set of minimal pairs, is much more complex when tackling the broader spectrum of hateful terms, including words with both hateful and non-hateful meanings.
+
+Qian et al. (2019) presented a framework aiming to predict the definition of hateful symbols, terms with a non-hateful surface form conveying hate, yet not covering the disambiguation between hate and non-hate. Mendelsohn et al. (2023) focused on a related phenomenon, dog whistles, examining whether GPT-3 can identify their covert meanings, surface them in text generation and detect them in real-world texts. Dinu et al. (2021) introduce the task of disambiguating pejorative word usage, presenting two small-scale datasets and evaluating several methods, with an MLP model classifying BERT embeddings (Devlin et al., 2019) as most effective approach. Muti et al. (2024) addressed pejorative word disambiguation as a preliminary step for misogyny detection in Italian texts.
+
+Our study focuses on the disambiguation of words with hateful meanings, which, although overlapping with dog whistles and pejorative words, belong to distinct categories. Unlike hateful words, dog whistles are always intentionally ambiguous, concealing one meaning from the out-group which is not exclusively hateful. Pejorative words, encompass any negatively connoted terms that may not be hateful when not targeted at an individual or group. More importantly, unlike the single-perspective annotations employed in the aforementioned studies, our focus is on subjective hate speech annotation and it is conducted on a much larger scale.
+
+# 2.2 Subjective Hate Speech Detection
+
+Most existing datasets and methods in HSD adopt a single, majority perspective, ignoring the inher
+
+ent subjectivity influenced by diverse social and cultural factors (Zampieri et al., 2019; Founta et al., 2018). This approach has been shown to result in problematic biases, concerning e.g. ethnicity, gender, and political beliefs and highlight the need for new methodologies that account for the varying interpretations of hateful connotations (Davidson et al., 2019; Kumar et al., 2021; Sap et al., 2022).
+
+Davani et al. (2022) took steps in this direction by training a model to predict individual annotations as subtasks, still ultimately aiming to predict the majority label. Kanclerz et al. (2022) addressed the task of predicting each individual annotator's label, by leveraging annotator's labeling statistics within the dataset. Alacam et al. (2024) study the incorporation of gaze features (on token- and sentence level) from human annotators for predicting their subjective hate ratings. Another more comprehensive approach is presented by Fleisig et al. (2023), who included annotators' demographics, preferences, and experiences as input, along with text. They utilized RoBERTa (Liu et al., 2019) to embed descriptions of these characteristics. Our research continues this line of work by predicting individual annotator labels and accounting for their demographics in the classification of hateful words.
+
+# 2.3 Modeling Word Senses
+
+Shifting the focus from modeling hateful utterances to the meaning of hateful words within utterances, touches upon various lexical semantic NLP tasks that involve the creation of word sense representations (Vulic et al., 2020a; Schlechtweg et al., 2020; Martelli et al., 2021). Approaches to these tasks often employ contextualized word embeddings extracted from pretrained (often BERT-based) LMs (Loureiro and Jorge, 2019; Martinc et al., 2020; Bommasani et al., 2020). Fine-tuning a model on particular data or tasks, such as WSD or sentiment classification, is performed to potentially inject relevant information into the resulting representations (Giulianelli et al., 2020; Hoeken et al., 2023a). Rachinskiy and Arefyev (2022) leveraged an effective WSD model developed by Blevins and Zettlemoyer (2020), which jointly optimizes two encoders for the context and gloss of a word sense, respectively. For the task of semantic change discovery, they extracted the representations of the context encoder of the WSD-finetuned model.
+
+Recently, Giulianelli et al. (2023) introduced an innovative approach to computational sense representations. Their method adopts the definition-as
+
+sense paradigm, utilizing definitions generated by a Flan-T5 model (Chung et al., 2022) fine-tuned on datasets of definitions with usage examples. Sentence embeddings of these generated context-specific definitions show promising results on lexical semantic similarity tasks. Despite these advancements focused on descriptive word senses, effective approaches for modeling highly connotative lexical phenomena remain unclear.
+
+# 3 The HateWiC Dataset
+
+We introduce the HateWiC dataset, which includes hate ratings for words within example usages along with their word sense definitions which may be hate-heterogeneous, as illustrated in Figure 1. We describe the dataset construction below.
+
+# 3.1 Wiktionary Data
+
+Data was scraped from the English Wiktionary in November 2023, an online dictionary where anyone can contribute to documenting and explaining words in use. Therefore, Wiktionary provides up-to-date insights from user perspectives and covers a broader range of terms from diverse domains than traditional dictionaries.
+
+Each entry (word or multi-word expression) includes information such as definitions, example uses, and category labels that provide additional context about a word's use (e.g., 'British slang' or 'Archaic'). Using the Wiktionary API, we extracted all entries that had at least one word sense tagged with the categories Offensive and Derogatory and were also members of the category People, to gather the most relevant terms for hate speech detection purposes. For each of the resulting 1087 terms, we scraped all its sense definitions along with all labeled categories and example sentences (using the WiktionaryParser library). This resulted in 3500 senses and 4671 examples.
+
+To suit the dataset for our HateWiC classification task, we manually excluded 642 examples due to historical spelling or other deficiencies, as detailed in Appendix A. After processing, the dataset comprised 4029 instances covering 1888 unique senses, averaging 4.88 examples per sense, and 826 unique terms, averaging 2.29 senses per term.
+
+# 3.2 Annotation
+
+The senses from the Wiktionary data include labels regarding their offensiveness or derogatory nature. However, these classifications do not represent the
+
+<table><tr><td>Example</td><td>Term</td><td>Definition</td><td>Annotations</td><td>Binary labels</td><td>Majority label</td><td>Hate-hetero- 
+genuine sense</td><td>Agreement 
+on binary</td></tr><tr><td>(1) “Me having an up to date style even though I’ve turned into a carrot cruncher.”</td><td>carrot cruncher</td><td>Someone from a rural background.</td><td>Nh, Nh, Nh</td><td>0, 0, 0</td><td>0</td><td>True</td><td>True</td></tr><tr><td>(2) “you’re a friggin’ carrot cruncher and you support the bloody scally’s.”</td><td>carrot cruncher</td><td>Someone from a rural background.</td><td>Sh, Sh, Sh</td><td>1, 1, 1</td><td>1</td><td>True</td><td>True</td></tr><tr><td>(3) “The bugger’s given me the wrong change.”</td><td>bugger</td><td>A foolish person or thing.</td><td>Wh, Sh, Sh</td><td>1, 1, 1</td><td>1</td><td>False</td><td>True</td></tr><tr><td>(4) “He’s a silly bugger for losing his keys.”</td><td>bugger</td><td>A foolish person or thing.</td><td>Nh, Wh, Sh</td><td>0, 1, 1</td><td>1</td><td>False</td><td>False</td></tr></table>
+
+Table 1: HateWiC examples with their annotations, illustrating the phenomena of annotator disagreement and hate-heterogeneous word senses (Nh = Not hateful, Wh = Weakly hateful, Sh = Strongly hateful)
+
+diverse interpretations of these labels due to their subjective nature. In this study, we aim to survey and model different beliefs, following a descriptive data annotation paradigm as proposed in the framework by Rottger et al. (2022). This paradigm highlights the value of using crowd-sourced annotators from diverse backgrounds to encourage annotator subjectivity and mitigate bias, without relaying on a predefined detailed definition of hate speech. Specifically, we collected crowd-sourced annotations using Prolific with a link integration to Argilla. Argilla, an open-source platform launched on HuggingFace Spaces, is used to set up the annotation task on HateWiC data.
+
+For each annotation instance, annotators are presented with an example sentence, the target term, and its sense definition. They are then prompted with the question: "How would you rate the hatefulness of the meaning of the target term within the specific example text?" Annotators respond by selecting from the labels: 'Not hateful', 'Weakly hateful', 'Strongly hateful' and 'Cannot decide'. An example of an annotation instance and the user interface are depicted in a screenshot provided in Appendix B. In the annotation guidelines (accessible on our repository), annotators are instructed to focus their evaluation on the specific usage of the term within the example sentence, rather than the overall connotation of the sentence, or the definition, which is only provided to aid in understanding the term's meaning. Additionally, we emphasize the subjective nature of their judgements.
+
+We aimed for three annotations per instance, with each annotator labeling 250 instances. Using Prolific's pre-screening filters, we selected annotators who indicated that their primary language is English. To improve the quality of the collected annotations, we excluded and replaced data from annotators who were too fast and/or failed control instances. Prolific provides demographic informa
+
+tion for each annotator, which can be connected to their annotations. The final pool of annotators, after exclusions, consisted of 48 individuals with diverse genders and ethnicities averaging 28 years old (more details in Appendix B).
+
+# 3.3 Dataset Results
+
+After excluding the 'Cannot decide' annotations<sup>4</sup>, the dataset yielded 11902 individual annotations, of which 5708 (48.0%) hateful and 6194 (52.0%) not hateful (after converting to binary by merging 'Weakly hateful' and 'Strongly hateful'). After applying majority voting, out of the 3845 example sentences with a clear majority binary label, 1815 (47.2%) were classified as hateful and 2030 (52.8%) as not hateful, yielding a balanced dataset with respect to hatefulness.
+
+Annotators agreed for  $60\%$  (i.e. 2414) of the binary classification with a Krippendorff's alpha of 0.45. For the three-class classification, agreement was  $51.3\%$  with a Krippendorff's alpha of 0.33. In comparison, Mathew et al. (2020) reported an agreement of 0.46 for a similar three-class task, and Vigna et al. (2017) 0.26 for their binary setting. The agreement scores underscore the inherent subjectivity of the task, motivating us to include individual demographics to our modeling.
+
+The high degree of context dependency regarding hate becomes even more apparent when we examine the relationship between word senses (the descriptive aspects outlined in their definitions) and the hatefulness ratings assigned to examples of those senses. We identified 319 hate-heterogeneous sense definitions, i.e. unique definitions for which example sentences exist in the dataset with both hateful and non-hateful majority annotations. Two examples from the annotated data given in Table 1 illustrate this phenomenon. Both examples mention the term carrot cruncher
+
+with the sense definition "Someone from a rural background." where (1) is unanimously annotated as not hateful and (2) is unanimously annotated as strongly hateful. This observation solidifies the idea, already implied by the inter-annotator agreement for individual labels (and exemplified by (4) in Table 1), that the hateful connotation of a word sense is not exclusively determined by its descriptive definition.
+
+# 4 HateWiC Classification
+
+Our HateWiC dataset enables the development and evaluation of computational methods for predicting whether the meaning of a target term is hateful within a specific context. Figure 2 provides an illustration of the primary methodological pipeline we present in this paper. We introduce various classification methods that differ with respect to the sense representations (outlined in 4.1) and incorporation of annotator information (4.2) as input to a classification model (4.3), or that leverage an instruction-tuned LLM (4.4).
+
+# 4.1 Sense Representations
+
+For representing the sense of a target term, we primarily follow a common procedure in lexical semantic NLP tasks and extract contextualized embeddings from pretrained LMs. To optimize effectiveness on the HateWiC task, we experiment with various encoder models and embedding types. Appendix C provides additional details on our employed methods.
+
+Encoder models. We experiment with three different encoder models, each trained on different data or tasks. We use the pretrained BERT (base) model (Devlin et al., 2019) and HateBERT (Caselli et al., 2021), a re-trained BERT model on hate speech<sup>5</sup>. As third, we utilize a trained bi-encoder model for Word Sense Disambiguation (Blevins and Zettlemoyer, 2020), which we refer to as WSD Biencoder. The model comprises a contextualized word encoder and a gloss encoder initialized with BERT-base encoders. We train it on WordNet data (Miller et al., 1994), following the same procedure as detailed in (Blevins and Zettlemoyer, 2020), for 7 epochs with a batch size of 8. Following Rachinskiy and Arefyev (2022), the WSD-optimized contextualized word encoder is then used for obtaining embeddings.
+
+Embeddings. The encoders are used to generate different word sense related representations. First, we compute word in context (WiC) embeddings. We feed the example sentence to the encoder model and extract the last hidden layer for the subword-tokenized position(s) that encode the target term (averaging over them in case of multi-subword target terms). Second, we test the incorporation of word sense definitions from Wiktionary. This definition (Def) embedding is obtained by averaging over all token embeddings, using the same procedure as for WiC embeddings but with the definition sentence as input. Third, considering that pre-given definitions may not be available in practical applications, we create T5-generated definition (T5Def) embeddings. We generate definitions using a FLAN-T5 Base (250M parameters) model developed by (Giulianelli et al., 2023)<sup>6</sup> which was fine-tuned on datasets of English definitions and usage examples. We prompt the model with the same template as it was trained on: "[SENTENCE] What is the definition of [TERM]?"'. Consequently, the generated generated are more context-specific than the Wiktionary definitions. These generated definitions are embedded the same way as the Def-embeddings.
+
+# 4.2 Annotator Information
+
+To address the subjective nature of the HateWiC classification task, highlighted by the inter-annotator agreement in our dataset, we incorporate this aspect into our modeling approaches. We experiment with a similar strategy as presented in Fleisig et al. (2023). For each individual annotation of a HateWiC instance, we concatenate an annotator (Ann) embedding to the corresponding sense embedding, that represent a description of annotator's demographics. This description is embedded through the same procedure as the definition embeddings and follows this template:
+
+"Reader is [AGE], [GENDER] and [ETHNICITY]."
+
+# 4.3 Classifying Embeddings
+
+We test the effectiveness of (the concatenation of combinations of) the embeddings proposed above on our HateWiC classification task by using them as input to a classification model. To this end, we train and test a four-layer multi-layer perceptron (MLP) model (a classification algorithm also used in Dinu et al. (2021)) on the HateWiC dataset.
+
+![](images/82a491d01918060ab10cd48825d0f5521a741faa7ee5df17378f7a8aee9b7c09.jpg)  
+Figure 2: Illustration of our main HateWiC classification pipeline.
+
+# 4.4 Classification with LLaMA 2
+
+In addition to the encoder-LM based approaches above, we also experiment with a LLaMA 2 model (Touvron et al., 2023). Due to their instruction-tuning training regime, and huge amount of training data, foundation models like LLaMA 2 are proven to be superior to LMs on many zero-shot settings, yet subjective HSD and WSD are by nature very challenging tasks. We aim to see the abilities of an instruction-tuned LLM on this task as a (strong) baseline. We test zero-shot classification with a 7B-sized LLaMA 2 model<sup>7</sup>. We run the inference of this model using the transformers library. In our prompt, we input the example sentence and the target term and instruct the model to classify the meaning of the term as hateful or not hateful (complete template and configuration parameters are provided in Appendix C).
+
+# 5 Evaluation Setup
+
+We evaluate our proposed methods using various test setups on the HateWiC dataset (5.1). Additionally, we compare our methods with the work of Dinu et al. (2021), as described in 5.2.
+
+# 5.1 HateWiC
+
+Our HateWiC dataset includes three hate ratings for each example sentence, allowing evaluation on two distinct tasks that vary in terms of subjectivity inclusion. For both tasks, we utilize binary labels.
+
+1. Majority label prediction: gold labels represent 4029 majority votes on each example.
+
+2. Subjective label prediction: gold labels consist of all 12442 individual annotations: a rating per example and annotator.
+
+We conduct evaluations for each task using a ten-fold cross-validation setup. For each fold, we divide the dataset into training, development, and test sets with an 80-10-10 ratio. We experiment with two variants:
+
+1. Random: The data is randomly split based on example sentences, testing performance on sentences not seen during training (similar to common practice in WSD-like tasks (Dinu et al., 2021)), which is particularly relevant for individual annotator prediction where multiple instances of the same sentence occur.  
+2. Out-of-Vocabulary (OoV): The data is split based on terms, testing performance on unseen terms, i.e. zero-shot capabilities.
+
+# 5.2 Comparison with Dinu et al. (2021)
+
+We also train and test on two small datasets of English tweets developed and used in Dinu et al. (2021). They collected these from existing hate speech datasets, focusing on tweets that mention one of the terms in a curated set of pejorative terms. Each tweet was labeled based on whether the term was used pejoratively. The first dataset, which we will refer to as DINU1 comprised 1004 tweets covering 31 terms. The second, which we name DINU2, consisted of 301 tweets covering 11 terms. Their reported best method involved MLP classification of BERTweet (Nguyen et al., 2020) and BERT (base) embeddings (extracted as the sum of all model layers for the target word position) on
+
+DINU1 and DINU2, respectively. We aimed to use the same evaluation set-up as described in their paper, using five-fold cross-validation and reporting the average over accuracies per term.
+
+# 6 Results
+
+This section presents the results of our proposed methods on the HateWiC classification, evaluated using the above outlined setups.
+
+# 6.1 Majority HateWiC Classification
+
+Table 2 presents the accuracy results on HateWiC classification compared to the majority label. Overall, the performance values demonstrate the effectiveness of all methods, with only minimal differences (max.  $2\%$  -points) between BERT, HateBERT and WSD biencoder models. Training BERT-based models on different types of information regarding hatefulness or word senses does not seem to have a substantial effect.
+
+<table><tr><td rowspan="2">Embeddings</td><td colspan="2">BERT</td><td colspan="2">HateBERT</td><td colspan="2">WSD bien.</td></tr><tr><td>Random</td><td>OoV</td><td>Random</td><td>OoV</td><td>Random</td><td>OoV</td></tr><tr><td>WiC</td><td>0.75</td><td>0.73</td><td>0.75</td><td>0.71</td><td>0.76</td><td>0.73</td></tr><tr><td>Def</td><td>0.77</td><td>0.75</td><td>0.78</td><td>0.73</td><td>0.78</td><td>0.73</td></tr><tr><td>T5Def</td><td>0.70</td><td>0.67</td><td>0.70</td><td>0.67</td><td>0.72</td><td>0.69</td></tr><tr><td>WiC+Def</td><td>0.78</td><td>0.77</td><td>0.80</td><td>0.77</td><td>0.79</td><td>0.78</td></tr><tr><td>WiC+T5Def</td><td>0.75</td><td>0.74</td><td>0.76</td><td>0.73</td><td>0.76</td><td>0.73</td></tr></table>
+
+Def-embeddings achieve slightly higher accuracies than WiC-embeddings, and a combination of the two yields the best results. For a test set with OoV terms only, all embedding types show only a slight drop in performance. WiC+Def-embeddings exhibit the smallest decline on the zero-shot setting and achieve  $2 - 5\%$  higher accuracy than WiC- and Def-embeddings. This indicates that definitions provide valuable information, performing better on their own than word information alone, and the combination of both is most effective, especially for OoV-terms. T5-generated definitions demonstrated the lowest accuracy on their own but perform equally or slightly better than WiC-embeddings when concatenated. An evaluation of T5-generated definitions compared to Wiktionary definitions showed a very low SacreBLEU score of 3.822 (in range 0 to 100), possibly explaining the differences in performance between them.
+
+The distinction between context-independent Def-embeddings and context-specific WiC
+
+and T5Def-embeddings becomes more clear upon examining their performance across hate-homogeneous and hate-heterogeneous instances (as defined in Section 3.3), presented in in Table 3. In the case of hate-heterogeneous instances, we observe an accuracy drop of up to  $47\%$  when using Def-including embeddings compared to the homogeneous instances. This drop is limited to  $24 - 29\%$  for the other embeddings, showcasing their superior ability in handling less descriptive scenarios. We define hate-homogeneous here as instances where definitions have example sentences in the dataset with either hateful or non-hateful (majority) annotations whereas hate-heterogeneous have both (as detailed in Section 3.3).
+
+Table 2: Accuracy on HateWiC classification compared to the majority label, with different input embeddings, tested on a random data split (best underlined) and a test split with OoV terms only (best in bold).  
+
+<table><tr><td rowspan="2">HateBERT embeddings</td><td colspan="2">Hate-homogeneous</td></tr><tr><td>True</td><td>False</td></tr><tr><td>WiC</td><td>0.82</td><td>0.55</td></tr><tr><td>Def</td><td>0.91</td><td>0.44</td></tr><tr><td>T5Def</td><td>0.76</td><td>0.52</td></tr><tr><td>WiC+Def</td><td>0.91</td><td>0.49</td></tr><tr><td>WiC+T5Def</td><td>0.84</td><td>0.55</td></tr></table>
+
+Table 3: Accuracy on HateWiC classification compared to the majority label w.r.t. hate homogeneity of the sense definition (best underlined).
+
+LLaMA 2 result. The accuracy score on the HateWiC classification using a LLaMa 2 model, following the zero-shot experimental setup detailed in Section 4.4, is 0.68. Unlike the superior performance on many downstream tasks, the LLaMA model falls short compared to the aforementioned models on our HateWiC task. This outcome highlights the subjective nature of the task, indicating that general-purpose models struggle to fully grasp its nuances and perform well on it.
+
+# 6.2 Subjective HateWiC Classification
+
+Performance of our designed methods on predicting individual annotation labels, which showed considerable variation in Section 3.3, are presented in Table 4. Overall, accuracy values are slightly lower (by  $2 - 5\%$  points) compared to predicting the majority label, but remain robust. The results exhibit the same patterns in terms of different models, test data setups, and tested embedding types. Adding the Annotator embedding has a minimal effect, generally resulting in equal or slightly improved performance compared to the same type of embedding without concatenated annotator information.
+
+To better understand the impact of subjectivity, we more closely examine instances where subjec
+
+<table><tr><td rowspan="2">Embeddings</td><td colspan="2">BERT</td><td colspan="2">HateBERT</td><td colspan="2">WSD bien.</td></tr><tr><td>Random</td><td>OoV</td><td>Random</td><td>OoV</td><td>Random</td><td>OoV</td></tr><tr><td>WiC</td><td>0.71</td><td>0.69</td><td>0.71</td><td>0.69</td><td>0.72</td><td>0.70</td></tr><tr><td>Def</td><td>0.74</td><td>0.71</td><td>0.75</td><td>0.73</td><td>0.74</td><td>0.71</td></tr><tr><td>T5Def</td><td>0.68</td><td>0.65</td><td>0.68</td><td>0.67</td><td>0.68</td><td>0.67</td></tr><tr><td>WiC+Def</td><td>0.75</td><td>0.74</td><td>0.75</td><td>0.73</td><td>0.75</td><td>0.73</td></tr><tr><td>WiC+T5Def</td><td>0.72</td><td>0.70</td><td>0.72</td><td>0.71</td><td>0.73</td><td>0.69</td></tr><tr><td>WiC+Ann</td><td>0.72</td><td>0.69</td><td>0.72</td><td>0.69</td><td>0.72</td><td>0.70</td></tr><tr><td>Def+Ann</td><td>0.74</td><td>0.72</td><td>0.76</td><td>0.72</td><td>0.75</td><td>0.72</td></tr><tr><td>T5Def+Ann</td><td>0.69</td><td>0.67</td><td>0.69</td><td>0.65</td><td>0.69</td><td>0.68</td></tr><tr><td>WiC+Def+Ann</td><td>0.75</td><td>0.73</td><td>0.75</td><td>0.74</td><td>0.75</td><td>0.74</td></tr><tr><td>WiC+T5Def+Ann</td><td>0.72</td><td>0.71</td><td>0.73</td><td>0.71</td><td>0.73</td><td>0.72</td></tr></table>
+
+Table 4: Accuracy on HateWiC classification compared to the individual annotator label, with different input embeddings, on a random data split (best underlined) and a test split with OoV terms only (best in bold).
+
+tivity is most apparent (and thus potentially harmful when methods fail). In Table 5 we report performance results not only with respect to the hate homogeneity of word senses, but also to annotator agreement, i.e. whether the annotator agreed with the majority. We present results for HateBERT embeddings in an evaluation setting with random test data split, but similar patterns are observed for BERT and WSD Biencoder embeddings, as well as on on a test data split with OoV terms only.
+
+For sentence annotations where the annotator disagreed with the majority label or the sense definition is hate-heterogeneous, the performance of all embeddings drops significantly. This effect is most pronounced for definition-including embeddings (Wiktionary), less so for T5-generated, which aligns with their more context-specific nature. Specifically, there is an accuracy drop of up to  $47\%$  in cases of annotator disagreement, and up to  $32\%$  in cases of hate-heterogeneous definitions. However, incorporating annotator information mitigates this effect by up to  $11\%$ . Annotator information contributes to the cases where the subjective annotation deviates from the majority label, these cases also align with sense definitions that exhibit both hateful and non-hateful labeled sentences.
+
+# 6.3 Results on DINU Data
+
+The DINU1 and DINU2 evaluation datasets do not provide sense definitions or information on annotators, thereby limiting our testing to our methods that do not require this information. Table 6 presents the results on both DINU1 and DINU2.
+
+Our methods, except for those including T5Def-embeddings only, demonstrate improvements over the best-performing methods proposed by Dinu et al. (2021). These improvements are particularly substantial (by  $8\%$ ) for the larger DINU1
+
+<table><tr><td rowspan="2">HateBERT embeddings</td><td colspan="2">Majority annotation</td><td colspan="2">Hate-homogeneous</td></tr><tr><td>True</td><td>False</td><td>True</td><td>False</td></tr><tr><td>WiC</td><td>0.77</td><td>0.40</td><td>0.77</td><td>0.55</td></tr><tr><td>Def</td><td>0.81</td><td>0.36</td><td>0.83</td><td>0.51</td></tr><tr><td>T5Def</td><td>0.72</td><td>0.42</td><td>0.72</td><td>0.53</td></tr><tr><td>WiC+Def</td><td>0.83</td><td>0.36</td><td>0.83</td><td>0.55</td></tr><tr><td>WiC+T5Def</td><td>0.78</td><td>0.39</td><td>0.78</td><td>0.56</td></tr><tr><td>WiC+Ann</td><td>0.77</td><td>0.49</td><td>0.77</td><td>0.59</td></tr><tr><td>Def+Ann</td><td>0.82</td><td>0.44</td><td>0.82</td><td>0.60</td></tr><tr><td>T5Def+Ann</td><td>0.73</td><td>0.47</td><td>0.72</td><td>0.59</td></tr><tr><td>WiC+Def+Ann</td><td>0.80</td><td>0.44</td><td>0.81</td><td>0.58</td></tr><tr><td>WiC+T5Def+Ann</td><td>0.77</td><td>0.48</td><td>0.78</td><td>0.58</td></tr></table>
+
+Table 5: Accuracy on HateWiC classification compared to the individual label w.r.t. annotator agreement with the majority label and hate homogeneity of the sense definition (best underlined).
+
+<table><tr><td>Model</td><td>Embedding</td><td>DINU1</td><td>DINU2</td></tr><tr><td rowspan="3">BERT</td><td>WiC</td><td>0.89</td><td>0.83</td></tr><tr><td>T5Def</td><td>0.81</td><td>0.79</td></tr><tr><td>WiC+T5Def</td><td>0.90</td><td>0.83</td></tr><tr><td rowspan="3">HateBERT</td><td>WiC</td><td>0.87</td><td>0.83</td></tr><tr><td>T5Def</td><td>0.83</td><td>0.80</td></tr><tr><td>WiC+T5Def</td><td>0.90</td><td>0.84</td></tr><tr><td rowspan="3">WSD Bienc.</td><td>WiC</td><td>0.90</td><td>0.82</td></tr><tr><td>T5Def</td><td>0.80</td><td>0.79</td></tr><tr><td>WiC+T5Def</td><td>0.90</td><td>0.84</td></tr><tr><td>Best Dinu</td><td></td><td>0.82</td><td>0.83</td></tr></table>
+
+Table 6: Accuracy of our methods on the DINU datasets compared the accuracy of the best performing method as reported in Dinu et al. (2021) (best underlined).
+
+dataset. Consistent with trends observed for the HateWiC dataset, the concatenation of WiC and T5-generated definition embeddings yields the best performance across both DINU sets, underscoring the potential of incorporating automatically generated definitions in the absence of dictionary definitions for HateWiC classification.
+
+# 7 Discussion
+
+Our study offers valuable insights into the detection of hate speech through the lens of lexical semantics, introducing the HateWiC dataset and presenting classification experiments. The negligible difference observed in our experimental outcomes between HateBERT and general (WSD) models not only questions the efficacy of extensive training on hate speech data for accurately capturing hateful semantics, but also underscores the necessity of a more nuanced approach beyond the existing lexical semantic methods for tasks like HateWiC classification. Our results demonstrate the impact of incorporating sense definitions and annotator characteristics on model performance, particularly in scenarios involving out-of-vocabulary (OoV) terms
+
+or high subjectivity.
+
+To define or not define? Hateful terms, according to lexical semantic theory, primarily contain an expressive component but not exclusively. Incorporating sense definitions into our methods, to encompass the descriptive component of hateful terms, yielded mixed results. Overall, embedded Wiktionary definitions proved highly effective, outperforming Word in Context (WiC) embeddings alone. T5-generated definitions demonstrated the lowest accuracy on their own but performed equally or slightly better than WiC-embeddings only when concatenated with WiC-embeddings. However, in cases with more variation in the subjective ratings, the performance of all embeddings dropped significantly but most pronounced for Wiktionary definition embeddings, though to a lesser extent for T5-generated definitions (with a drop difference of up to  $23\%$ ). This highlights the usefulness of automatically generating context-specific definitions for subjective lexical semantic tasks like HateWiC classification. Future research will focus on more advanced definition generation techniques, possibly leveraging larger models or fine-tuning on Wiktionary definitions, while avoiding overreliance on dictionary definitions as the ultimate standard.
+
+To individualize anyway? The low inter-annotator agreement in our dataset underscores the importance of considering individual annotator perspectives in hate speech detection. Our experiments incorporating annotator information in our computational methods proved beneficial, particularly in cases of annotator disagreement or hate-heterogeneous definitions, where including annotator information mitigated accuracy decline by up to  $11\%$  -points. This highlights the value of personalizing models to account for subjectivity in annotations. Future research could explore additional annotator information and conduct ablation experiments to identify the most effective aspects for HateWiC classification.
+
+To consider as well? Our study paves the way to obtaining deeper insights into the relationship between hateful and non-hateful word senses. For instance, whether certain semantic relations (e.g. metaphorical, metonymical), categories (e.g. food, animals), or attributes (e.g. color, material) are more likely to distinguish between hateful and nonhateful senses. And even next-level, whether these discriminators are language-specific or show cross
+
+language parallels. Identifying such consistencies between (non-)hateful senses could enhance the (automatic) discrimination between them.
+
+# 8 Conclusion
+
+This paper introduces the Hateful Word in Context Classification (HateWiC) task, addressing the underexplored area of subjective hateful word meanings within specific contexts. We present the HateWiC dataset, comprising about 4000 WiC instances, each annotated with three hateful ratings. Our study focused on the interplay between descriptive and subjective aspects of hateful word senses. We addressed the prediction of both majority and individual annotator labels. We experimented with different types of inputs to our classification system, including sense definitions and annotator demographics. We demonstrated the impact of these factors on model performance, particularly in cases involving out-of-vocabulary terms or high subjectivity. The incorporation of established sense definitions proved highly effective overall but demonstrating diminished performance in less descriptive scenarios. Conversely, including annotator characteristics proved beneficial, particularly in cases of annotator disagreement or hate-heterogeneous definitions. These findings underscore the value of personalizing models to account for subjectivity in annotations. Furthermore, our results suggest the potential usefulness of automatically generating definitions for subjective lexical semantic tasks like HateWiC classification.
+
+# Limitations
+
+Although the Wiktionary data we utilize offers insights from user perspectives for a wide array of terms, its quality may be lower compared to expert-curated dictionaries. The provided information may contain inaccuracies, as users might not have the necessary expertise, and inconsistency in documentation could exist. However, the collaborative nature of Wiktionary allows for censorship by consensus and adherence to Wiktionary policies, mitigating some of these concerns.
+
+A constraint of our evaluation set-up lies in its reliance on binary labels. Hate speech is a multifaceted phenomenon, and a more nuanced class scheme may offer a more comprehensive understanding in future research.
+
+# Ethics Statement
+
+Our study includes demographic data of annotators that concern Prolific prescreening responses which are all with annotator's consent, self-reported, and are not provided with any direct identifiers like name or address. All prescreening questions, except for age and country of residence, are optional for participants to answer, and most personal questions have a 'Rather not say' option. By incorporating demographic information from annotators, we aim to enhance the understanding and prediction of how different groups perceive hate speech. This approach will ultimately lead to more robust and inclusive classification systems. However, the inclusion of demographic data raises privacy concerns, particularly the risk of re-identifying annotators. To address this, we have made our dataset available only upon request, under the CC BY-NC 4.0 license. This measure allows us to better control access to the information, ensuring it is used responsibly, ethically, and exclusively for non-commercial purposes.
+
+# Acknowledgements
+
+The authors acknowledge financial support by the project "SAIL: SustAINable Life-cycle of Intelligent Socio-Technical Systems" (Grant ID NW21-059A), which is funded by the program "Netzwerke 2021" of the Ministry of Culture and Science of the State of North Rhine-Westphalia, Germany.
+
+# References
+
+Özge Alacam, Sanne Hoeken, and Sina Zarrieß. 2024. Eyes don't lie: Subjective hate annotation and detection with gaze. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Miami, Florida, USA. Association for Computational Linguistics.  
+Terra Blevins and Luke Zettlemoyer. 2020. Moving down the long tail of word sense disambiguation with gloss informed bi-encoders. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1006–1017, Online. Association for Computational Linguistics.  
+Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29.  
+Rishi Bommasani, Kelly Davis, and Claire Cardie. 2020. Interpreting Pretrained Contextualized Representations via Reductions to Static Embeddings. In Pro
+
+ceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4758-4781, Online. Association for Computational Linguistics.  
+Tommaso Caselli, Valerio Basile, Jelena Mitrovic, and Michael Granitzer. 2021. HateBERT: Retraining BERT for abusive language detection in English. In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 17-25, Online. Association for Computational Linguistics.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling instruction-finetuned language models. Preprint, arXiv:2210.11416.  
+Aida Mostafazadeh Davani, Mark Diaz, and Vinodkumar Prabhakaran. 2022. Dealing with Disagreements: Looking Beyond the Majority Vote in Subjective Annotations. Transactions of the Association for Computational Linguistics, 10:92-110.  
+Thomas Davidson, Debasmita Bhattacharya, and Ingmar Weber. 2019. Racial bias in hate speech and abusive language detection datasets. In Proceedings of the Third Workshop on Abusive Language Online, pages 25-35, Florence, Italy. Association for Computational Linguistics.  
+Thomas Davidson, Dana Warmsley, Michael W. Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. CoRR, abs/1703.04009.  
+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.  
+Liviu P. Dinu, Ioan-Bogdan Iordache, Ana Sabina Uban, and Marcos Zampieri. 2021. A computational exploration of pejorative language in social media. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3493-3498, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Eve Fleisig, Rediet Abebe, and Dan Klein. 2023. When the majority is wrong: Modeling annotator disagreement for subjective tasks. In Proceedings of the 2023
+
+Conference on Empirical Methods in Natural Language Processing, pages 6715-6726, Singapore. Association for Computational Linguistics.  
+Antigoni Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large scale crowdsourcing and characterization of twitter abusive behavior. Proceedings of the International AAAI Conference on Web and Social Media, 12(1).  
+Aldo Frigerio and Maria Paola Tenchini. 2019. Pejoratives: a classification of the connoted terms. Rivista Italiana di Filosofia del Linguaggio, 13(1).  
+Mario Giulianielli, Marco Del Tredici, and Raquel Fernandez. 2020. Analysing lexical semantic change with contextualised word representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3960-3973, Online. Association for Computational Linguistics.  
+Mario Giulianielli, Iris Luden, Raquel Fernandez, and Andrey Kutuzov. 2023. Interpretable word sense representations via definition generation: The case of semantic change analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3130-3148, Toronto, Canada. Association for Computational Linguistics.  
+Leopold Hess. 2021. Slurs: Semantic and Pragmatic Theories of Meaning, page 450-466. Cambridge Handbooks in Language and Linguistics. Cambridge University Press.  
+Sanne Hoeken, Özge Alacam, Antske Fokkens, and Pia Sommerauer. 2023a. Methodological insights in detecting subtle semantic shifts with contextualized and static language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 3662-3675, Singapore. Association for Computational Linguistics.  
+Sanne Hoeken, Sina Zarrieß, and Ozge Alacam. 2023b. Identifying slurs and lexical hate speech via lightweight dimension projection in embedding space. In Proceedings of the 13th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, pages 278-289, Toronto, Canada. Association for Computational Linguistics.  
+Kamil Kanclerz, Marcin Gruza, Konrad Karanowski, Julita Bielaniewicz, Piotr Milkowski, Jan Kocon, and Przemyslaw Kazienko. 2022. What if ground truth is subjective? personalized deep neural hate speech detection. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 37-45, Marseille, France. European Language Resources Association.  
+Deepak Kumar, Patrick Gage Kelley, Sunny Consolvo, Joshua Mason, Elie Bursztein, Zakir Durmeric, Kurt Thomas, and Michael Bailey. 2021. Designing toxic
+
+content classification for a diversity of perspectives. In Proceedings of the Seventeenth USENIX Conference on Usable Privacy and Security, SOUPS'21, USA. USENIX Association.  
+Sawan Kumar, Sharmistha Jat, Karan Saxena, and Partha Talukdar. 2019. Zero-shot word sense disambiguation using sense definition embeddings. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5670-5681, Florence, Italy. Association for Computational Linguistics.  
+Ho Suk Lee, Hong Rae Lee, Jun U. Park, and Yo Sub Han. 2018. An abusive text detection system based on enhanced abusive and non-abusive word lists. Decision Support Systems, 113:22-31. Publisher Copyright: © 2018 Elsevier B.V.  
+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. CoRR, abs/1907.11692.  
+Daniel Loureiro and Alipio Jorge. 2019. Language modelling makes sense: Propagating representations through WordNet for full-coverage word sense disambiguation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5682-5691, Florence, Italy. Association for Computational Linguistics.  
+Federico Martelli, Najla Kalach, Gabriele Tola, and Roberto Navigli. 2021. SemEval-2021 task 2: Multilingual and cross-lingual word-in-context disambiguation (MCL-WiC). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 24–36, Online. Association for Computational Linguistics.  
+Matej Martinc, Petra Kralj Novak, and Senja Pollak. 2020. Leveraging contextual embeddings for detecting diachronic semantic shift. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4811-4819, Marseille, France. European Language Resources Association.  
+Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2020. Hatexplain: A benchmark dataset for explainable hate speech detection. In AAAI Conference on Artificial Intelligence.  
+Julia Mendelsohn, Ronan Le Bras, Yejin Choi, and Maarten Sap. 2023. From dogwhistles to bullhorns: Unveiling coded rhetoric with language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15162-15180, Toronto, Canada. Association for Computational Linguistics.  
+George A. Miller, Martin Chodorow, Shari Landes, Claudia Leacock, and Robert G. Thomas. 1994. Using a semantic concordance for sense identification. In Proceedings of the Workshop on Human Language
+
+Technology, HLT '94, page 240-243, USA. Association for Computational Linguistics.  
+Arianna Muti, Federico Ruggeri, Cagri Toraman, Alberto Barrón-Cedeno, Samuel Algherini, Lorenzo Musetti, Silvia Ronchi, Gianmarco Saretto, and Caterina Zapparoli. 2024. Pejorativity: Disambiguating pejorative epithets to improve misogyny detection in Italian tweets. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12700-12711, Torino, Italia. ELRA and ICCL.  
+Sathvik Nair, Mahesh Srinivasan, and Stephan Meylan. 2020. Contextualized word embeddings encode aspects of human-like word sense knowledge. In Proceedings of the Workshop on the Cognitive Aspects of the Lexicon, pages 129-141, Online. Association for Computational Linguistics.  
+Dat Quoc Nguyen, Thanh Vu, and Anh Tuan Nguyen. 2020. BERTweet: A pre-trained language model for English tweets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 9-14, Online. Association for Computational Linguistics.  
+Jing Qian, Mai ElSherief, Elizabeth Belding, and William Yang Wang. 2019. Learning to decipher hate symbols. 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 3006-3015, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Jing Qian, Hong Wang, Mai ElSherief, and Xifeng Yan. 2021. Lifelong learning of hate speech classification on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2304-2314, Online. Association for Computational Linguistics.  
+Maxim Rachinskiy and Nikolay Arefyev. 2022. Gloss-Reader at LSCDiscovery: Train to select a proper gloss in English - discover lexical semantic change in Spanish. In Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change, pages 198-203, Dublin, Ireland. Association for Computational Linguistics.  
+Paul Rottger, Bertie Vidgen, Dirk Hovy, and Janet Pierrehumbert. 2022. Two contrasting data annotation paradigms for subjective NLP tasks. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 175-190, Seattle, United States. Association for Computational Linguistics.  
+Maarten Sap, Swabha Swayamdipta, Laura Vianna, Xuhui Zhou, Yejin Choi, and Noah A. Smith. 2022. Annotators with attitudes: How annotator beliefs
+
+and identities bias toxic language detection. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5884-5906, Seattle, United States. Association for Computational Linguistics.  
+Dominik Schlechtweg, Barbara McGillivray, Simon Hengchen, Haim Dubossarsky, and Nina Tahmasebi. 2020. SemEval-2020 task 1: Unsupervised lexical semantic change detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 1-23, Barcelona (online). International Committee for Computational Linguistics.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal 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. 2023. Llama 2: Open foundation and finetuned chat models. Preprint, arXiv:2307.09288.  
+Fabio Del Vigna, Andrea Cimino, Felice Dell'Orletta, Marinella Petrocchi, and Maurizio Tesconi. 2017. Hate me, hate me not: Hate speech detection on facebook. In Italian Conference on Cybersecurity.  
+Ivan Vulic, Simon Baker, Edoardo Maria Ponti, Ulla Petti, Ira Leviant, Kelly Wing, Olga Majewska, Eden Bar, Matt Malone, Thierry Poibean, Roi Reichart, and Anna Korhonen. 2020a. Multi-SimLex: A largescale evaluation of multilingual and crosslingual lexical semantic similarity. Computational Linguistics, 46(4):847-897.  
+Ivan Vulic, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. 2020b. Probing pretrained language models for lexical semantics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7222-7240, Online. Association for Computational Linguistics.  
+Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? predictive features for hate speech detection on Twitter. In Proceedings of the NAACL Student Research Workshop, pages 88-93, San Diego, California. Association for Computational Linguistics.
+
+Michael Wiegand, Josef Ruppenhofer, Anna Schmidt, and Clayton Greenberg. 2018. Inducing a lexicon of abusive words - a feature-based approach. 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 1046-1056, New Orleans, Louisiana. Association for Computational Linguistics.
+
+Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. SemEval-2019 task 6: Identifying and categorizing offensive language in social media (OffensEval). In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 75-86, Minneapolis, Minnesota, USA. Association for Computational Linguistics.
+
+# A Wiktionary Data Processing
+
+Our data was scraped from the English Wiktionary comprising entries with information on definitions, example uses, and category labels that provide additional context about a word's use. We scraped all sense definitions along with all labeled categories and example sentences of the selected terms using the Wiktionary Parser library. This library method did not split the examples over the set of sense definitions (i.e. provided all examples in one bundle), so we manually matched the right examples with the right sense definitions, through look up on the Wiktionary website, afterwards.
+
+To suit the dataset for the envisioned task we manually excluded 642 examples that were either written in historical spelling or not single in-the-wild usages of the term. The latter concerned usages, like the examples below (with the target term in bold), that were (a) dictionary-typical nominal phrases and not sentences, (b) concerned meta-level discussions of the target term or (c) dialogues or other indirect uses of the target term.
+
+(a) "a bird feeder"  
+(b) "A 'lot lizard' was somebody who walked the sales lot and looked at every car and still didn't buy."  
+(c) "Threads on the social media giant Reddit occasionally discuss or condemn" [..]
+
+Finally, we slightly edited some type of instances that concerned non-exact matches between word form of the term and its occurrence in the example. For compounds or multi-word expressions,
+
+this mismatch often concerned the (non-)use of a whitespace or hyphen between compound parts (e.g. the term baby face occurred also as babyface or baby-face in examples). This type of mismatches was solved by applying a simple rule-based replacement strategy to the example sentences.
+
+Other types of non-exact word form matches were mainly caused by inflection (e.g. plural forms for nouns) and some by misspellings. These cases were left unchanged for the final dataset as removing could influence the meaning.
+
+We also created groupings to aggregate category labels, consolidating the 585 unique Wiktionary labels present in our dataset into a manageable set of usage tags. This enrichment potentially provides useful information for future analyses on usages of hateful terms.
+
+# B Annotation Details
+
+Figure 3 displays the user interface for annotation, with an example of an annotation instance.
+
+Below, we report the distribution of our annotators with respect to age, gender, and ethnicity. It is important to note that we use the categories as provided through the Prolific provided presecreening responses, which are simplified groupings intended to give a general overview. As detailed in the Ethics Statement, we acknowledge that this categorization does not fully capture the complexity and diversity of individual identities and may include sensitive terminology.
+
+The final pool of 48 annotators, after exclusions, had an average age of 28 (ranging from 20 to 60) and included 26 females, 28 males, and 1 unspecified gender. Based on simplified ethnicity categories, 21 identified as White, 19 as Black, 4 as Asian, 3 as Mixed, and 1 as Other.
+
+# C Method Details
+
+Finding target term sentence positions. For all WiC-embeddings, to find the indices of (the subwords that form) the target word in an example sentence that concerned a non-exact wordform match between target term and example mention (due to inflection or misspellings), we applied two subsequent strategies: 1) we tried to replace the target term with its plural form (through simple rules) and if this plural formation did not result in a match, 2) we tried to find the most similar word in the example sentence (using the difflib library) and replaced that wordform with the target term (as this most
+
+![](images/f438c3108a9bc372db5444215156e344e04aa3a4c6f7aba5d91e6d9c6d785799.jpg)  
+Figure 3:User interface for annotation
+
+<table><tr><td rowspan="2">Embeddings</td><td colspan="3">BERT</td></tr><tr><td>Last</td><td>All</td><td>LastFour</td></tr><tr><td>WiC</td><td>0.75</td><td>0.75</td><td>0.75</td></tr></table>
+
+Table 7: Accuracy on HateWiC classification compared to the majority label, with BERT input embeddings consisting of different layer combinations, on the random data test split.
+
+often concerned a misspelling).
+
+Model layer configurations. We also tested the extraction of different layer configurations, since the effectivity of different configurations has shown to differ within lexical semantic tasks (Vulić et al., 2020b). We tested for BERT WiC-embeddings the extraction of: all layers (12 for BERT), last four layers or last layer only. The results in Table 7, demonstrate no effect of layer configuration on the method performance.
+
+MLP classifier model. The multilayer perceptron model used for classification consisted of four hidden layers with dimensionality 300, 200, 100 and 50, respectively. For training we used the MLPClassifier module from the sklearn library and we set the initial learning rate to 0.0005 the maximal number of training iterations to 10. These parameters were selected after a grid search on our development dataset, using sklearn's GridSearchCV module, applied to the following parameter grid: {'hidden_layer_sizes':[(300, 200, 100, 50), (200, 100, 50), (100, 50)], 'learning_rate_init':[0.0005, 0.001, 0.005], 'max_iter': [10, 20, 40, 80, 100, 200]}.
+
+LLaMA 2. The following prompt template was used for leveraging LLaMA 2 for HateWiC Classification.
+
+Instruction:
+
+Given the following sentence that mentions a particular term, classify whether the meaning of that term expresses hate towards a person or group within that specific sentence. Respond with exactly one of the following corresponding labels without an explanation:
+
+"HATEFUL"
+
+"NOT HATEFUL"
+
+Input:
+
+Sentence: [SENTENCE]
+
+Term: [TERM]
+
+Response:
+
+We use the pipeline module from the transformers library for running the 'text inference' task, where we set the number of return sequences to 1 and the max new tokens to 10; we used the default settings for the remaining parameters.
+
+# D Dimension Projection
+
+We also tested the dimension approach of Hoeken et al. (2023b), adapted to our task. In their method for slur detection, they create a "hate dimension" by computing the average over difference vectors between representations of 10 minimal pairs of slurs and non-hateful equivalents (e.g. 'hillbillies' - 'rural people'). Unlike slurs, which generally carry derogatory connotations regardless of context (Hess, 2021), the hateful connotations of other hateful terms are less clear-cut (Frigerio and Ten-chini, 2019). This was also illustrated in the conceptual semantic space in Figure 1. Consequently, we did not expect an effective dimension hate dimension to be extractable using pretrained models that encode general word semantics. Additionally, pre-establishing a set of minimal pairs is hardly
+
+feasible for similar reasons.
+
+Our approach. For our task, instead of using a pre-established list of word pairs, we derived this list from the training data. We calculated the cosine similarities between all possible pairs of positive and negative embeddings, i.e. sense representations of hateful and non-hateful training examples, respectively. We then selected pairs with a similarity above a certain threshold to create the dimension, trough the same computation procedure as Hoeken et al. (2023b). After testing a range of thresholds ([0.7, 0.75, 0.8, 0.85, 0.9, 0.95]) on the development set, we set the similarity threshold to 0.9 for testing. Following Hoeken et al. (2023b), we classified positive cosine similarity values between the hate dimension vector and the contextualized word sense representation as hateful, and negative values as non hateful.
+
+<table><tr><td rowspan="2">Embeddings</td><td colspan="2">BERT</td><td colspan="2">HateBERT</td><td colspan="2">WSD bien.</td></tr><tr><td>Random</td><td>OoV</td><td>Random</td><td>OoV</td><td>Random</td><td>OoV</td></tr><tr><td>WiC</td><td>0.52</td><td>0.53</td><td>0.44</td><td>0.43</td><td>0.44</td><td>0.44</td></tr><tr><td>Def</td><td>0.44</td><td>0.43</td><td>0.49</td><td>0.49</td><td>0.32</td><td>0.33</td></tr><tr><td>WiC+Def</td><td>0.49</td><td>0.49</td><td>0.44</td><td>0.45</td><td>0.49</td><td>0.48</td></tr></table>
+
+Table 8: Accuracy on HateWiC classification compared to the majority label, with dimension projection and different input embeddings, tested on a random data split and OoV terms only.
+
+Results. The results of this approach on our HateWic dataset are presented in Table 8, demonstrate low accuracy scores (max. 0.52) and confirm our expectations that a dimension approach as currently implemented is not effective for HateWiC classification.

2024/Hateful Word in Context Classification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a77d1b76171b368ab326ead4a7cb86ea48dc958ed968d90860b1f92bd70ef035
+size 356417

2024/Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models/d6c62a2e-d2cd-45b7-a876-95885365f893_content_list.json

@@ -0,0 +1,1396 @@
+[
+    {
+        "type": "text",
+        "text": "Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models",
+        "text_level": 1,
+        "bbox": [
+            236,
+            91,
+            763,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yae Jee Cho $^{1,2*}$ , Luyang Liu $^{2}$ , Zheng Xu $^{2}$ , Aldi Fahrezi $^{2}$ , Gauri Joshi $^{1}$",
+        "bbox": [
+            203,
+            152,
+            801,
+            170
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Carnegie Mellon University,  $^{2}$ Google Research",
+        "bbox": [
+            305,
+            170,
+            695,
+            187
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{yaejeeecho, luyangliu, xuzheng, aldifahrezi}@google.com,",
+        "bbox": [
+            257,
+            187,
+            742,
+            204
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "gaurij@andrew.cmu.edu",
+        "bbox": [
+            401,
+            205,
+            600,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            262,
+            342,
+            278
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Foundation models (FMs) adapt surprisingly well to downstream tasks with fine-tuning. However, their colossal parameter space prohibits their training on resource-constrained edge-devices. For federated fine-tuning, we need to consider the smaller FMs of few billion parameters at most, namely on-device FMs (ODFMs), which can be deployed on device. Federated fine-tuning of ODFMs has unique challenges non-present in standard fine-tuning: i) ODFMs poorly generalize to downstream tasks due to their limited sizes making proper fine-tuning imperative to their performance, and ii) devices have limited and heterogeneous system capabilities and data that can deter the performance of fine-tuning.",
+        "bbox": [
+            146,
+            288,
+            460,
+            517
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tackling these challenges, we propose HETLORA, a feasible and effective federated fine-tuning method for ODFMs that leverages the system and data heterogeneity at the edge. HETLORA allows heterogeneous LoRA ranks across clients for their individual system resources, and efficiently aggregates and distributes these LoRA modules in a data-aware manner by applying rank self-pruning locally and sparsity-weighted aggregation at the server. It combines the advantages of high and low-rank LoRAs, achieving improved convergence speed and final performance compared to homogeneous LoRA. Furthermore, HETLORA has enhanced computation and communication efficiency compared to full fine-tuning making it more feasible for the edge.",
+        "bbox": [
+            142,
+            521,
+            460,
+            764
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            115,
+            776,
+            260,
+            791
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The emerging foundation models (FMs), easily ranging from few to hundreds of billions of parameters (Gemini Team Google, 2024; Bommasani",
+        "bbox": [
+            114,
+            801,
+            489,
+            848
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "et al., 2022; Radford et al., 2021; Devlin et al., 2019; OpenAI, 2023; Google, 2022; Touvron et al., 2023; Brown et al., 2020; Google, 2022; Driess et al., 2023), have remarkable zero/few shot learning capabilities. They often perform surprisingly well on a variety of tasks such as text/image generation, language translation, and conversing in natural language with fine-tuning. Such standard FMs, however, demand costly resources for directly fine-tuning their entire parameter space. To tackle this issue, many works have proposed different parameter-efficient fine-tuning (PEFT) methods such as prompt tuning (Lester et al., 2021), utilizing adapters (Houlsby et al., 2019), or low-rank adaptation (LoRA) of the original model (Hu et al., 2021) which freezes the original pre-trained parameters of the FM and train additional, smaller number of parameters instead.",
+        "bbox": [
+            507,
+            262,
+            882,
+            552
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Unfortunately, such PEFT methods rely on the assumptions that i) FMs are deployed to and trained with the data of a single client for adaptation to the downstream task and ii) the client has enough resources to even fit a standard FM of hundred billion size for, at least, inference. In practice, more often than not we are interested in fine-tuning FMs for on-device private data that is distributed across multiple devices, and these clients at the edge rarely have such resources to even fit standard FMs for inference. One of the examples is to use FMs for downstream tasks that requires training with private, non-sharable data such as medical or law-related documents distributed across clients (Manoel et al., 2023; Shoham and Rappoport, 2023; Zhang et al., 2023c). In such cases, fine-tuning of the FMs needs to be brought to the edge that has limited system resources.",
+        "bbox": [
+            507,
+            556,
+            882,
+            845
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Federated Fine-Tuning & ODFMs. We consider such federated fine-tuning scenarios, where",
+        "bbox": [
+            507,
+            850,
+            882,
+            882
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work done while interning at Google Research. Corresponding authors: {yaejeeecho,luyangliu} @google.com",
+        "bbox": [
+            115,
+            854,
+            489,
+            882
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12903",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 12903-12913 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            152,
+            945,
+            843,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/7a0f3ea1004ad717d1430d6abb8af9b8b8995726fed5fc52df7b1dd9eab97a1a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Zero-Shot</td><td>Few-Shot</td><td>Full-Training</td></tr><tr><td>PaLM 2 XXS</td><td>2930.23</td><td>2541.86</td><td>23.71</td></tr><tr><td>PaLM 2 XS</td><td>2712.86</td><td>481.95</td><td>18.32</td></tr></table>",
+        "bbox": [
+            124,
+            84,
+            475,
+            129
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Perplexity of PaLM 2 for zero-shot, few-shot (5 communication rounds), and full federated fine-tuning (200 communication rounds) for chat response on the multi-session chat data (further experimental details are in Section 4.)",
+        "bbox": [
+            114,
+            135,
+            489,
+            186
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "we train a set of parameters collaboratively across clients to obtain a global set of parameters that can be plugged in to the FM for the targeted downstream task. Note that federated fine-tuning is orthogonal to personalization of FMs in federated learning (FL) that has been explored in other previous work (Guo et al., 2023; Wu et al., 2024), which aims to train parameters that perform well for individual clients rather than general downstream tasks. We also define on-device FMs (ODFMs) as models with few billion parameters at max that are able to fit into memory on limited capacity clients considering current hardwares.",
+        "bbox": [
+            114,
+            202,
+            489,
+            411
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Challenges. Federated fine-tuning of ODFMs entails two major challenges non-present in neither standard PEFT of FMs nor federated training of non-FMs. First, ODFMs poorly generalize to downstream tasks. Unlike ODFMs, standard FMs have their zero/few-shot learning capability by their large parameter space trained on massive data. However, as we show in Table 1 and as previous literature has shown (Kojima et al., 2022; Lester et al., 2021), FMs' performance deteriorates as their sizes get smaller and federated fine-tuning is not merely useful but inevitable for ODFMs to perform well for on-device downstream tasks.",
+        "bbox": [
+            114,
+            413,
+            489,
+            621
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Second, devices have limited and heterogeneous system capabilities and data distributions (Wang et al., 2019; Bonawitz et al., 2016; Sahu et al., 2020). Without suitable PEFT methods that flexibly adapts to such heterogeneity across devices, we would only have limited performance output from federated fine-tuning of ODFMs. For instance, with homogeneous rank deployment in LoRA that is agnostic of system heterogeneity, the client with the least system resource becomes the bottleneck, forcing a smaller rank to be deployed across all clients despite the other clients being able to handle higher ranks (see Table 2 for the effect of rank on the trainable parameter #). Additionally, data heterogeneity causing model drifts can make the model converge to a suboptimal point.",
+        "bbox": [
+            114,
+            625,
+            492,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/6043e1749a942f6b874f2481629c14f993b283566e481d96f024eb065e1c0511.jpg",
+        "image_caption": [
+            "Figure 1: Overview of heterogeneous rank deployment of LoRA: the pretrained weights  $\\mathbf{W}_0$  are stored on-device and heterogeneous ranks are assigned to different clients with  $r_{\\mathrm{min}} = r_1 < r_2 < r_3 = r_{\\mathrm{max}}$ . In our proposed HETLORA, the server receives the trained heterogeneous LoRA modules and aggregates them with sparsity-weighted aggregation to update the global LoRA module."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            83,
+            882,
+            230
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Contributions. In our work, we push the limits of federated fine-tuning of ODFMs by proposing HETLORA that is aware of heterogeneity residing at the edge. Our main flagship results (in Table 3 and Fig. 5), show that our proposed HETLORA outperforms the standard LoRA method and achieves comparable performance with full fine-tuning with a significantly smaller number of trained parameters. We verify HETLORA with PaLM 2 (Google, 2023) of XXS and XS size for chat responses on the multi-session chat data (Xu et al., 2021) and text summarization for the Reddit data (Volske et al., 2017), both real world data from clients. Our contributions are summarized as:",
+        "bbox": [
+            507,
+            344,
+            884,
+            570
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose HETLORA that can apply different rank LoRA modules to different clients to cater to the heterogeneous system capabilities and data complexities of the clients, via utilizing rank self-pruning and sparsity-weighted aggregation.",
+            "- We show the performance of naively applying LoRA with homogeneous ranks across clients for federated fine-tuning, and show that while large ranks help in speeding-up training, they lead to faster overfitting while smaller ranks are slower in training but does not suffer from overfitting.",
+            "- We then evaluate our proposed HETLORA to show that it outperforms naively applying homogeneous ranks across clients in terms of both training speed, communication/computation efficiency, and final performance."
+        ],
+        "bbox": [
+            507,
+            590,
+            884,
+            862
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2",
+        "bbox": [
+            492,
+            903,
+            505,
+            915
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "footer",
+        "text": "12904",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/bdee92404e078aeeb142199da2927980e6c705ec9660fd019e89ad54ea30f724.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>r = 1</td><td>r = 5</td><td>r = 10</td><td>r = 20</td><td>r = 50</td><td>r = 100</td><td>r = 150</td><td>r = 200</td></tr><tr><td>PaLM 2 XXS, PaLM 2 XS</td><td>0.02%</td><td>0.11%</td><td>0.21%</td><td>0.42%</td><td>1.05%</td><td>2.10%</td><td>3.14%</td><td>4.19%</td></tr></table>",
+        "bbox": [
+            142,
+            92,
+            853,
+            130
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Percentage of the LoRA parameters' size for different ranks  $r$  compared to the original pre-trained ODFM's parameter size. Even for large ranks such as  $r = {200}$  the trainable LoRA parameters' size compared to the original pre-trained ODFM size is less than 5% for both PaLM 2-XS and PaLM 2-XXS.",
+        "bbox": [
+            114,
+            137,
+            880,
+            174
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            115,
+            193,
+            272,
+            209
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Parameter-Efficient Fine Tuning. Most PEFT methods either train a subset of parameters within the existing FM whilst other parameters are frozen or introduce an additional set of trainable parameters whilst keeping the original FM frozen. For the former, methods such as head or bias fine-tuning (Wei et al., 2021; Bu et al., 2022; Lee et al., 2019; Zaken et al., 2021) has been explored, and for the latter, methods such as adapters (Houlsby et al., 2019), prompt (Lester et al., 2021) or prefix-tuning (Li and Liang, 2021), and low-rank approximation (Hu et al., 2021) has been proposed. While these number of methods has been proven to perform as well as full model fine-tuning with just a few number of parameters for the centralized setting, it has not been thoroughly explored how these methods would perform for a much smaller FM such as ODFMs, in the decentralized setting where clients' system-capacities and data can be heterogeneous and much limited.",
+        "bbox": [
+            114,
+            221,
+            489,
+            542
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Federated Fine-Tuning. Several recent work has proposed to combine the PEFT methods devised for the centralized setting to FL (Zhou et al., 2023; Yu et al., 2023) such as training prompts or adapters collaboratively with FL (Guo et al., 2022; Chen et al., 2022; Zhang et al., 2023a; Shysheya et al., 2023; Legate et al., 2023). Another line of work has proposed to perform a few-shot or nearly zero-shot training of FMs with FL for improved communication-efficiency (Wortsman et al., 2023; Zhang et al., 2023d). However, these work either overlooks that most devices do not have the resource to fit standard FMs (Touvron et al., 2023; Brown et al., 2020) even for inference or does not consider the heterogeneous system capacities of the clients. It is detrimental to consider these factors since FMs that actually fits to the devices in FL are much smaller, making them weaker in the general intelligence capabilities, and also heterogeneous system capacities can prohibit deploying same sized PEFT parameters across clients.",
+        "bbox": [
+            114,
+            544,
+            489,
+            883
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "LoRA and FL. There has been a number of variations of LoRA (Zhang et al., 2023b; Liu et al., 2024; Sun et al., 2024; Horváth et al., 2024) from its first proposal (Hu et al., 2021). However, only a few numbers of recent work has looked in to using LoRA for FL. In Babakniya et al. (2023), the importance of the initialization for the LoRA modules is evaluated where they propose to train the LoRA modules with FL and then perform singular value decomposition (SVD) to gain a good initialization of the LoRA modules. However, the training process of LoRA itself is not altered to adapt to heterogeneous system capabilities of devices. Another recent work (Yi et al., 2023) has evaluated LoRA in the context of personalized FL, but other than applying LoRA to personalization, the LoRA method itself is, again, not changed. To the best of our knowledge, our work is the first method to consider both data and system heterogeneity for improving training speed, communication/computation efficiency, and final performance of ODFM's federated fine-tuning. Contemporary approaches (Bai et al., 2024; Anonymous, 2024) have proposed to use SVD to decompose the full matrix to deploy heterogeneous ranks across clients. We compare our proposed method with using SVD for heterogeneous LoRA in our work and show that the performance is underwhelming compared to our proposed heterogeneous LoRA.",
+        "bbox": [
+            505,
+            193,
+            884,
+            661
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Federated Fine-Tuning with LoRA",
+        "text_level": 1,
+        "bbox": [
+            507,
+            674,
+            843,
+            690
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            507,
+            700,
+            663,
+            714
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Formally, we define the pre-trained ODFM as  $\\mathbf{W}_0\\in \\mathbb{R}^{d\\times l}$  and the trainable low-rank decomposed matrix as  $\\Delta \\mathbf{W}\\in \\mathbb{R}^{d\\times l}$ . In standard LoRA (Hu et al., 2021) under the centralized setting, the low-rank decomposition of  $\\Delta \\mathbf{W}$  is constructed such that  $\\Delta \\mathbf{W} = \\mathbf{BA}$  where  $\\mathbf{B}\\in \\mathbb{R}^{d\\times r}$  and  $\\mathbf{A}\\in \\mathbb{R}^{r\\times l}$  are the low rank decomposition of  $\\Delta \\mathbf{W}$  with identical rank  $r$ . Now, let us consider LoRA for federated fine-tuning where there are  $M$  total clients. Each client  $k\\in [M]$  has pri",
+        "bbox": [
+            505,
+            721,
+            884,
+            883
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3",
+        "bbox": [
+            492,
+            903,
+            505,
+            915
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "footer",
+        "text": "12905",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "vate data  $\\mathcal{B}_k$  and its corresponding local empirical loss function  $F_{k}(\\mathbf{W}) = \\frac{1}{|\\mathcal{B}_{k}|}\\sum_{\\xi \\in \\mathcal{B}_{k}}\\ell (\\mathbf{W},\\xi)$  where  $\\ell (\\mathbf{W},\\xi)$  is the loss for model  $\\mathbf{W}$  at data sample  $\\xi$ . The optimization task for federated fine-tuning is to collaboratively find the global parameters which we define as  $\\overline{\\mathbf{B}}$  and  $\\overline{\\mathbf{A}}$ , given the pretrained knowledge  $\\mathbf{W}_0$  that can minimize the global objective  $F(\\overline{\\mathbf{W}}) = \\frac{1}{M}\\sum_{k = 1}^{M}F_{k}(\\overline{\\mathbf{W}})$  where  $\\overline{\\mathbf{W}} = \\mathbf{W}_0 + \\overline{\\mathbf{B}}\\overline{\\mathbf{A}}$ . Later in the paper, when introducing heterogeneous LoRA we truncate the LoRA modules' rank dimension, for example from  $\\mathbf{B}\\in \\mathbb{R}^{d\\times r}$ ,  $\\mathbf{A}\\in \\mathbb{R}^{r\\times l}$  to  $\\mathbf{B}'\\in \\mathbb{R}^{d\\times r'}$ ,  $\\mathbf{A}'\\in \\mathbb{R}^{r'\\times l}$  where  $r' < r$ . Throughout the paper, we denote such truncation of a matrix with the : symbol for each row and column at the subscript. For instance, for truncation to  $r' < r$  at the column for the matrix  $\\mathbf{B}\\in \\mathbb{R}^{d\\times r}$ , we keep all the columns until  $r'$  and omit the last  $r - r'$  columns and denote the resulting matrix it as  $\\mathbf{B}_{:,;r'}$ .",
+        "bbox": [
+            114,
+            85,
+            492,
+            395
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Naive Case: Homogeneous LoRA",
+        "text_level": 1,
+        "bbox": [
+            115,
+            410,
+            423,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A straightforward way to perform federated fine-tuning with LoRA is to train the LoRA modules  $\\mathbf{B}$ ,  $\\mathbf{A}$  with homogeneous rank  $r$  across all clients with standard FL (McMahan et al., 2017). Specifically, first the clients have the pre-trained ODFM weights  $\\mathbf{W}_0$  stored in their devices prior to training for the forward pass when training the LoRA modules. Then, the server sends the global LoRA modules  $\\overline{\\mathbf{B}}^{(t)}$ ,  $\\overline{\\mathbf{A}}^{(t)}$  to the set of  $m$  selected clients  $S^{(t)}$  per communication round  $t$ . Each selected client  $k \\in S^{(t)}$  trains the LoRA modules on their local data for a few local iterations (usually with mini-batch SGD) and send the updated modules  $\\mathbf{B}_k^{(t)}$ ,  $\\mathbf{A}_k^{(t)}$  back to the server. The server then updates the global LoRA modules accordingly to  $\\overline{\\mathbf{B}}^{(t+1)} = \\sum_{k \\in S^{(t)}} \\mathbf{B}_k^{(t)} / m$ ,  $\\overline{\\mathbf{A}}^{(t+1)} = \\sum_{k \\in S^{(t)}} \\mathbf{A}_k^{(t)} / m$  and sends back to the next set of selected clients for the next communication round. This training process is nearly identical to the standard FL algorithm (McMahan et al., 2017) except that the pretrained weights  $\\mathbf{W}_0$  are frozen and locally stored in the clients' devices and only the LoRA moduels are trained and communicated.",
+        "bbox": [
+            114,
+            435,
+            489,
+            815
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Instead of such homogeneous rank deployment across all clients, it is not only possible but even more practical and feasible to use heterogeneous rank deployment for federated fine-tuning (we later",
+        "bbox": [
+            114,
+            818,
+            489,
+            883
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "show that this also improves final performance of the ODFMs). This involves training LoRA modules with varying ranks across clients, based on their system capabilities. This allows a more practical setting for federated fine-tuning of ODFMs where, for instance, the client with the smallest system capacity does not have to become the bottleneck that forces deploying a smaller rank across all clients. Instead, with heterogeneous rank deployment, each client can use a rank that suits for itself. However, heterogeneous rank deployment poses challenges in aggregating and redistributing the LoRA modules. To address these challenges, we introduce a solution called HETLORA, which pushes the limits beyond homogeneous LoRA.",
+        "bbox": [
+            507,
+            85,
+            884,
+            326
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Proposed Method: Heterogeneous LoRA",
+        "text_level": 1,
+        "bbox": [
+            507,
+            338,
+            875,
+            354
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Overview. Our proposed heterogeneous LoRA method, namely HETLORA, is not restricted to any specific method to assign the ranks to the clients and the clients can decide their respective ranks themselves. For formality, in our paper, we formulate that each client has a rank denoted as  $r_k$ , within a range of  $r_k \\in [r_{\\min}, r_{\\max}]$ ,  $\\forall k$  (see Fig. 1). HETLORA comprises three steps: 1) Distribution via Truncation, 2) Local Training with Rank Self-Pruning, and 3) Sparsity-Weighted Aggregation of the LoRA modules. These steps are detailed further in the subsequent paragraphs. An overview of HETLORA is illustrated in Fig. 2.",
+        "bbox": [
+            507,
+            360,
+            884,
+            569
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "1) Distribution via Truncation. At the beginning of each communication round  $t$ , the server holds initial global LoRA modules  $\\overline{\\mathbf{B}}^{(t)}$ ,  $\\overline{\\mathbf{A}}^{(t)}$  with a global rank  $r^{(t)}$ . The value of the global rank  $r^{(t)}$  depends on how we aggregate the heterogeneous rank LoRA modules which is elaborated on in step 3). The server then distributes these global LoRA modules to a subset of selected set of clients  $S^{(t)}$  with heterogeneous ranks  $r_k^{(t)}, k \\in S^{(t)}$  for local training<sup>1</sup>. With the given global LoRA modules, we consider a simple and intuitive method of truncation where the server sends  $\\overline{\\mathbf{B}}_{:,r_k}^{(t)}$ ,  $\\overline{\\mathbf{A}}_{:r_k}^{(t)}$  to each client  $k$  with rank  $r_k^{(t)}$  for local training where we omitted the superscript for  $r_k$  for simplicity.",
+        "bbox": [
+            507,
+            570,
+            882,
+            804
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2) Local Training with Rank Self-Pruning.",
+        "text_level": 1,
+        "bbox": [
+            526,
+            804,
+            882,
+            821
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "There is a superscript  $t$  for the ranks  $r_k^{(t)}$  across clients which indicates that in HETLORA these heterogeneous ranks can be changed over the communication rounds via self-pruning explained in step 2).",
+        "bbox": [
+            507,
+            829,
+            882,
+            882
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4",
+        "bbox": [
+            492,
+            903,
+            505,
+            915
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "footer",
+        "text": "12906",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3619c21a64d5cfc62a80b2f2c4291cbf5c44111745f28ad471b9b3b507d522d2.jpg",
+        "image_caption": [
+            "Figure 2: Overview of the zero-padding, sparsity-weighted aggregation, and truncation method for HETLORA; (a): Zero-pad LoRA modules with smaller ranks to  $r_{\\mathrm{max}}$  (clients with rank  $r_{\\mathrm{max}}$  does not need padding) and calculate their sparsity by calculating the Frobenius norm of the reconstructed model  $\\Delta \\mathbf{W}_k^{(t)} = \\mathbf{B}_k^{(t)}\\mathbf{A}_k^{(t)}$ ; (b): After padding, aggregate all of the clients' LoRA modules with the weights  $\\| \\mathbf{S}_k^{(t)}\\| /\\mathbf{Z}^{(t)}$  calculated by  $\\Delta \\mathbf{W}_k^{(t)}$  to get the global LoRA modules; (c): Truncate the global LoRA modules for the specific rank of the next selected client (example for client with rank  $r_2$ )."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            82,
+            573,
+            200
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/eec115469413780c73e24798e79974773601e7c715ead2a481e452de9f9a744c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            579,
+            82,
+            737,
+            200
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f3317bffb9483465ec5324ad4e14b4452aa339f8bc9d6dc2b2fb42ef6a68c96.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            739,
+            82,
+            880,
+            200
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After receiving LoRA modules from the server as  $\\mathbf{B}_k^{(t,0)} = \\overline{\\mathbf{B}}_{\\therefore r_k}^{(t)}$ ,  $\\mathbf{A}_k^{(t,0)} = \\overline{\\mathbf{A}}_{r_k,\\cdot}^{(t)}$ , each client  $k \\in S^{(t)}$  performs  $\\tau$  local iterations of mini-batch SGD on their local data to minimize the local objective  $\\frac{1}{|\\mathcal{B}_k|}\\sum_{\\xi \\in \\mathcal{B}_k}\\ell((\\mathbf{B}_k, \\mathbf{A}_k), \\xi|\\mathbf{W}_0)$ , and sends back the updated LoRA modules  $\\mathbf{B}_k^{(t,\\tau)} \\in \\mathbb{R}^{d \\times r_k^{(t)}}$  and  $\\mathbf{A}_k^{(t,\\tau)} \\in \\mathbb{R}^{r_k^{(t)} \\times l}$  to the server. This is the same process as the standard local training step in vanilla FedAvg (McMahan et al., 2017). However, we improve this vanilla local training step by adding a rank self-pruning mechanism where clients self-prune their respective ranks depending on the magnitude of the model parameters.",
+        "bbox": [
+            114,
+            300,
+            487,
+            525
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we add a regularization term to the original local objective to get  $\\min_{\\mathbf{B}_k,\\mathbf{A}_k}\\frac{1}{|\\mathcal{B}_k|}\\sum_{\\xi \\in \\mathcal{B}_k}\\ell ((\\mathbf{B}_k,\\mathbf{A}_k),\\xi |\\mathbf{W}_0) + \\lambda \\| \\mathbf{B}_{k,:r_k\\gamma :r_k}\\| \\| \\mathbf{A}_{k,r_k\\gamma :r_k,:}\\|$  where  $\\gamma < 1$  is a decay-factor that determines how aggressively we want to prune the ranks to a smaller value. The regularization term aims to minimize the norm of the last few ranks, which will become smaller if the first loss term  $\\frac{1}{|\\mathcal{B}_k|}\\sum_{\\xi \\in \\mathcal{B}_k}\\ell ((\\mathbf{B}_k,\\mathbf{A}_k),\\xi |\\mathbf{W}_0)$  is not very large. After training with the new local objective we compare the norm of the updated LoRA modules' last few layers  $\\| \\mathbf{B}_{k,:r_k\\gamma :r_k}\\| \\| \\mathbf{A}_{k,r_k\\gamma :r_k,:}\\|$  with the ones from the initially received LoRA modules. If the former is smaller we prune the last few layers (pruning intensity is determined by  $\\gamma$ ) and send back the LoRA modules with a smaller rank. This means that for the LoRA modules which incurs a small local loss, i.e., well-trained on the clients' local data, the LoRA modules are more likely to be pruned to a smaller rank.",
+        "bbox": [
+            114,
+            526,
+            487,
+            865
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Such pruning allows HETLORA to reduce the",
+        "bbox": [
+            132,
+            866,
+            485,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "noise in the LoRA modules introduced by clients having a larger rank than the actual rank that their data complexity requires, and also reduces the complexity of the LoRA modules to improve generalization and prevent overfitting (see Table 4). Once the rank is pruned for a client, the client saves the updated rank and uses it as the starting rank if selected for future communication rounds. The client then sends back their updated and possibly rank-pruned LoRA modules to the server to be processed in the aggregation step.",
+        "bbox": [
+            505,
+            300,
+            880,
+            476
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3) Sparsity-Weighted Aggregation. Finally, the last step of HETLORA is aggregating the received heterogeneous LoRA modules  $\\mathbf{B}_k^{(t,\\tau)}$ ,  $\\mathbf{A}_k^{(t,\\tau)}$ ,  $k \\in S^{(t)}$ . A straightforward way to aggregate the heterogeneous LoRA modules is using zero-padding to all the received LoRA modules with  $r_i^{(t)} < \\max \\{r_k^{(t)} | k \\in S^{(t)}\\}$  and then perform simple averaging over the modules. However, such naive aggregation can lead to biasing the model towards higher rank clients even when these clients may not hold valuable training information, i.e., having low data complexity, giving noisy updates.",
+        "bbox": [
+            507,
+            482,
+            880,
+            678
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In an ideal scenario where we can deploy any rank to any client, deploying higher ranks to the clients with higher data complexity or larger local datasets can retrieve more informative and less sparse updates from the clients. Conversely if we assign higher ranks to the clients whose data complexity is low, the actual rank of the full model from the reconstructed LoRA modules can be smaller than the assigned rank. Thus the higher rank client's update may be unnecessarily over-emphasized in the naive zero padding method.",
+        "bbox": [
+            505,
+            684,
+            880,
+            860
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Based on this insight we propose a sparsity-",
+        "bbox": [
+            527,
+            866,
+            880,
+            882
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "5",
+        "bbox": [
+            492,
+            903,
+            504,
+            914
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "footer",
+        "text": "12907",
+        "bbox": [
+            477,
+            927,
+            524,
+            939
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "weighted aggregation scheme where the server reconstructs these LoRA modules to the full model as  $\\Delta \\mathbf{W}_k^{(t)} = \\mathbf{B}_k^{(t)}\\mathbf{A}_k^{(t)}$  and gets the norm of the singular value vectors from the full models denoted as  $\\mathbf{S}_k^{(t)}$  by calculating  $\\| \\Delta \\mathbf{W}_k^{(t)}\\| _F$ . Note that the costly process of performing SVD for each of the full model  $\\Delta \\mathbf{W}_k^{(t)}$  can be avoided by simply calculating the Frobenius norm of  $\\Delta \\mathbf{W}_k^{(t)}$  (see Lemma 1.2 in (Guruswami and Kannan, 2012)). The server then weighs the LoRA modules with aggregation weight  $p_k^{(t)}$  which is proportional to the norm of the singular value vectors. Formally, we have the global LoRA modules updated as  $\\overline{\\mathbf{B}}^{(t + 1)} = \\sum_{k\\in \\mathcal{S}^{(t)}}p_k^{(t)}\\mathbf{B}_k^{(t)}$ ,  $\\overline{\\mathbf{A}}^{(t + 1)} = \\sum_{k\\in \\mathcal{S}^{(t)}}p_k^{(t)}\\mathbf{A}_k^{(t)}$  where  $p_k^{(t)}\\coloneqq \\| \\mathbf{S}_k^{(t)}\\| /\\mathbf{Z}^{(t)}$  with normalizing factor  $\\mathbf{Z}^{(t)}\\coloneqq \\sum_{k^{\\prime}\\in \\mathcal{S}^{(t)}}\\| \\mathbf{S}_{k^{\\prime}}^{(t)}\\|$ . This way, we can de-emphasize the larger rank assigned-clients that have rather less informative updates, and more emphasize the smaller rank assigned-clients that have more informative ones.",
+        "bbox": [
+            114,
+            86,
+            492,
+            426
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.4 Why not Simply Reconstruct First, then Redistribute the LoRA modules?",
+        "text_level": 1,
+        "bbox": [
+            115,
+            439,
+            478,
+            472
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "One might ask why not simply reconstruct each of the LoRA modules to the full matrix and aggregate them. Here we show that reconstructing the LoRA modules and aggregating them to get the full model results in a different full model compared to when we aggregate the LoRA modules first and then reconstruct the final model. In Section 4 we also empirically show that reconstructing the LoRA modules to the full model and redistributing them after truncated SVD to the corresponding rank of the clients results in an overwhelming performance compared to HETLORA.",
+        "bbox": [
+            114,
+            479,
+            489,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Let us consider a simple case where there are 2 clients with heterogeneous rank lora modules  $\\mathbf{B}_1\\in$ $\\mathbb{R}^{d\\times 1}$ $\\mathbf{A}_1\\in \\mathbb{R}^{1\\times l}$  and  $\\mathbf{B}_2\\in \\mathbb{R}^{d\\times 1}$ $\\mathbf{A}_2\\in \\mathbb{R}^{2\\times l}$  respectively for client 1 and client 2 where the former has rank 1 and latter has rank 2. We set the notation for the LoRA modules'  $i^{th}$  row and  $j^{th}$  column value for  $\\mathbf{B}_k$  and  $\\mathbf{A}_k$  as  $b_{k,ij}$  and  $a_{k,ij}$  respectively. Then with  $d = 3$ $l = 2$  when we reconstruct each of the LoRA modules first and then aggregate the full model we have its  $i^{th}$  row and  $j^{th}$  column as  $(\\sum_{k = 1}^{2}b_{k,i0}a_{k,0j}) + b_{2,i1}a_{2,1j}$  and aggregating the LoRA modules first and then reconstructing the model has the full model's  $i^{th}$  row and  $j^{th}$  column",
+        "bbox": [
+            114,
+            673,
+            489,
+            883
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "as  $(\\sum_{k = 1}^{2}b_{k,i0})(\\sum_{k = 1}^{2}a_{k,0j}) + b_{2,i1}a_{2,1j}$",
+        "bbox": [
+            507,
+            83,
+            821,
+            102
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "One can observe that the difference between the two models are the cross-terms between the left and right module of different client 1 and 2, i.e.,  $b_{1,i0}a_{2,0j} + b_{2,i0}a_{1,0j}$  for the  $i^{th}$  row and  $j^{th}$  column. In other words, when we reconstruct the LoRA modules first and then aggregate them to get the full model, each term in the full model are cross-products between the left and right module of each client and not the cross-products between clients. Thus, reconstructing the LoRA modules loses information on the cross-relation across clients, only retaining the knowledge on the cross-relation between the LoRA modules B and A. Such observation is also corroborated by the reconstruction first's underwhelming performance in Table 3.",
+        "bbox": [
+            505,
+            103,
+            882,
+            344
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            356,
+            655,
+            373
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we show that our proposed HETLORA outperforms its baselines in terms of training speed, computation/communication efficiency, and final achieved performance. We first show the performance of homogeneous LoRA to show how LoRA in general performs for low and high rank values. Second, as our main result, we demonstrate HETLORA's performance for different  $r_{\\mathrm{min}}$  and  $r_{\\mathrm{max}}$  values comparing them with full fine-tuning, homogeneous LoRA, and the reconstruction-first method elaborated in Section 3.4. We also conduct an ablation study on HETLORA with varying decay factor  $\\gamma$  for the rank self-pruning step.",
+        "bbox": [
+            505,
+            382,
+            882,
+            607
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Settings. We use the transformer-based language model PaLM 2 (Google, 2023) of size XXS and XS for our experiments which are lightweight enough to be ODFMs (Google DeepMind, 2023) compared to standard FMs. The tasks we consider are the chat dialogue from the multi-session chat (MSC) dataset (Xu et al., 2021) and the text summarization task from the Reddit dataset (Volske et al., 2017). For the MSC data we use perplexity (Zhang et al., 2018) as the metric which has been used to show the quality of chat responses from generative models from previous literature (Sedoc et al., 2019). We sample 100 users uniformly at random and partition their data for training and evaluation by each previous_DIALOG and dialog. The Reddit text summarization data consists of real users' reddit posts and their summarization, and",
+        "bbox": [
+            507,
+            609,
+            882,
+            882
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6",
+        "bbox": [
+            492,
+            903,
+            505,
+            915
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "footer",
+        "text": "12908",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/fefe7418299bbe795c43c79dc2415f28cb21c8905fa35ae14682f3cbae1d4430.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"2\">Reddit (RougeL)</td><td colspan=\"2\">Multi-Session Chat (Perplexity)</td></tr><tr><td></td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td></tr><tr><td>Full</td><td>94.56(±0.01)</td><td>94.87(±0.04)</td><td>32.70(±0.17)</td><td>23.40(±0.36)</td></tr><tr><td>HOMLORA r = 5</td><td>92.57(±1.56), ×0.001</td><td>92.89(±0.96)</td><td>80.51(±8.32), ×0.001</td><td>64.59(±9.31)</td></tr><tr><td>HOMLORA r = 50</td><td>70.57(±2.13), ×0.01</td><td>84.95(±1.59)</td><td>307.96(±11.43), ×0.01</td><td>167.46(±1.72)</td></tr><tr><td>Recon+SVD</td><td>63.28(±1.92), ×0.003</td><td>75.17(±1.25)</td><td>323.89(±20.57), ×0.002</td><td>215.63(±15.38)</td></tr><tr><td>HETLORA γ = 0.99</td><td>94.23(±0.03), ×0.003</td><td>94.41(±0.05)</td><td>53.93(±1.57), ×0.002</td><td>38.76(±0.52)</td></tr></table>",
+        "bbox": [
+            142,
+            83,
+            853,
+            178
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Final RougeL score for Reddit text summarization and perplexity for multi-session chat for different federated fine-tuning methods. The blue text indicates the ratio of trained number of parameters compared to the full fine-tuning case. HETLORA outperforms both HOMLORA and Recon+SVD method, but slightly underperforms the full fine-tuning case. However, compared to full fine-tuning the number of trained parameter is significantly smaller.",
+        "bbox": [
+            114,
+            187,
+            880,
+            237
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/98ba5e70c8bd6c9388e771ace842c0e5979e77db93f6a9066241be7ac50dfa52.jpg",
+        "image_caption": [
+            "(a) PaLM 2-XXS"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            254,
+            297,
+            370
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2775320785a46ad0dc6278f6e788f71268d7b64081e49bad49c85386f7814992.jpg",
+        "image_caption": [
+            "(b) PaLM 2-XS",
+            "Figure 3: Performance of homogeneous LoRA for different rank  $r$ . Higher ranks achieve better performance with fewer communication rounds than the lower ranks, but they overfit quickly. Conversely, the lowest rank  $r = 1$  achieves low perplexity slower than higher ranks, but without overfitting."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            302,
+            255,
+            480,
+            370
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "we use RougeL (Lin, 2004) as the metric. We use 298 users from Reddit that have at least 100 data samples as the training clients and use another 100 users with at least 100 data samples for evaluation. We use mini-batch size 8 and number of local iterations  $\\tau = 5$  with the feature length set to 1024. For the learning rate we perform grid search in  $\\eta = \\{0.1, 0.01, 0.001, 0.0001\\}$ . For each MSC and Reddit task, we select 5 and 10 clients per communication round respectively. The rank distribution across clients for HETLORA is set to a truncated power-law distribution with  $\\alpha = 0.1$  in the range between  $[r_{\\mathrm{min}}, r_{\\mathrm{max}}]$  (inclusively). All experiments were ran with 3 different random seeds.",
+        "bbox": [
+            114,
+            464,
+            489,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experiment Results",
+        "text_level": 1,
+        "bbox": [
+            115,
+            701,
+            317,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Homogeneous LoRA and the Effect of Ranks  $r$ . First, we evaluate the performance of federated fine-tuning of the LoRA modules with homogeneous LoRA deployment across clients in Fig. 3 for different ranks  $r \\in [1,5,20,50]$ . We observe that a higher rank  $r$  for homogeneous LoRA achieves better perplexity floor with fewer communication rounds than the lower ranks but quickly overfits resulting in worse performance compared to the lower ranks after more communication rounds. On",
+        "bbox": [
+            114,
+            721,
+            489,
+            881
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a3cb5079742aae49ceb6d8c7f526b7e747082544584b284c7b39a528065412cb.jpg",
+        "image_caption": [
+            "(a)  $r_{\\mathrm{max}} = 10$",
+            "Figure 4: Performance of HETLORA without rank pruning or and with simple average aggregation. Similar to homogeneous LoRA, larger  $r_{\\mathrm{min}}$  leads to overfitting for heterogeneous LoRA, but it is not as severe as homogeneous LoRA even for larger maximum rank  $r_{\\mathrm{max}} = 50$  showing that the smaller rank LoRA modules act as a regularizer for HETLORA."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            255,
+            692,
+            346
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b098577b62359dbdb5e064a3cf725cdf24bb2ffd6dce18f4ce9dfa8f32f1bde9.jpg",
+        "image_caption": [
+            "(b)  $r_{\\mathrm{max}} = 50$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            256,
+            875,
+            346
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the other hand, while the lower rank cases need more communication rounds to achieve good performance, it does not have the problem of overfitting as the higher ranks. Hence for homogeneous LoRA, there is a trade-off between low and high ranks, in terms of faster performance achievement and overfitting. Note that these observations are consistent with previous literature in the centralized setting where a higher rank does not necessarily yields the best performance (Hu et al., 2021; Zhang et al., 2023b). Subsequently, we show that HETLoRA that is not only more practical than homogeneous LoRA, but also achieves a better final performance with fewer number of communication rounds without the overfitting issue.",
+        "bbox": [
+            507,
+            453,
+            882,
+            695
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Naïve Heterogeneous LoRA and the Effect of  $r_{\\mathrm{min}}$  and  $r_{\\mathrm{max}}$ . First, we show the performance of naïve heterogeneous LoRA which is HETLORA without self rank-pruning and with only average aggregation instead of the sparsity-weighted aggregation in Fig. 4. We see that similar to homogeneous LoRA, a smaller minimum rank  $r_{\\mathrm{min}} = 1$  leads to slower training but better performance while a larger maximum rank leads to faster training but worse performance. However, unlike homogeneous LoRA, the overfitting does not get",
+        "bbox": [
+            507,
+            705,
+            882,
+            882
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7",
+        "bbox": [
+            492,
+            903,
+            504,
+            915
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "footer",
+        "text": "12909",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d92d82c25e4af5dfda612db4b98ce49be454352f0019e941fd47ab56b3cd6fb3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"2\">Reddit (RougeL)</td><td colspan=\"2\">Multi-Session Chat (Perplexity)</td></tr><tr><td></td><td>PaLM 2-XXS</td><td>PaLM 2-XS</td><td>PaLM 2-XXS</td><td>PaLM 2-XS</td></tr><tr><td>HETLORA, γ = 1</td><td>92.17 (±0.08)</td><td>91.95 (±0.03)</td><td>55.07 (±0.81)</td><td>40.92 (±0.58)</td></tr><tr><td>HETLORA, γ = 0.99</td><td>94.23 (±0.03)</td><td>94.41 (±0.05)</td><td>53.93 (±1.57)</td><td>38.76 (±0.52)</td></tr><tr><td>HETLORA, γ = 0.95</td><td>89.62 (±1.33)</td><td>83.19 (±1.70)</td><td>71.10 (±1.39)</td><td>46.39 (±0.87)</td></tr><tr><td>HETLORA, γ = 0.85</td><td>60.31 (±3.04)</td><td>53.28 (±2.47)</td><td>120.72 (±10.93)</td><td>59.67 (±1.98)</td></tr></table>",
+        "bbox": [
+            191,
+            83,
+            803,
+            162
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Ablation study on the effect of the decaying factor  $\\gamma$  for HETLORA's self-rank pruning in the local training step. While aggressive pruning can be harmful to HETLORA's performance, pruning  $\\left( {\\gamma  = {0.99}}\\right)$  can outperform the case when there is no pruning at all  $\\left( {\\gamma  = 1}\\right)$  by reducing the noise introduced by large rank clients with low data complexity.",
+        "bbox": [
+            114,
+            172,
+            880,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/69924bb62be3af0dfffb065a30b8fad4fdf7651ecc2ee0087a20cb3bb327fdc4.jpg",
+        "image_caption": [
+            "(a) Reddit"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            222,
+            299,
+            359
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0a31f19257ccca600cbe7281166c85e7b14f741f6275f32c05cbd4c51ab3afe2.jpg",
+        "image_caption": [
+            "(b) MSC"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            304,
+            223,
+            480,
+            359
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "as severe for heterogeneous LoRA even with much larger ranks such as  $r_{\\mathrm{max}} = 50$ . We can imply from this result that the smaller rank LoRA modules act as a regularizer in heterogeneous LoRA preventing overfitting. Next, we show that by adding the self rank-pruning and sparsity-weighted aggregation, we improve the performance such that even with  $r_{\\mathrm{min}} = 5$  we are able to prevent overfitting issues and achieve better training speed and final performance than other baselines.",
+        "bbox": [
+            114,
+            443,
+            487,
+            602
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Heterogeneous LoRA compared to PEFT. Finally, we compare HETLORA with other baselines in Table 3 and Fig. 5. We see that HETLORA with  $r_{\\mathrm{min}} = 5$  and  $r_{\\mathrm{max}} = 50$  achieves faster training as well as better performance than homogeneous LoRA cases with both edge cases of the ranks  $r \\in \\{5, 50\\}$  and reconstructions + SVD which was explained in Section 3.4. This implies that HETLORA is not only practical in the sense that clients are allowed to have their own rank values, it can also outperform the limited case of homogeneous LoRA where all clients have  $r = r_{\\mathrm{min}}$  or the impractical case where all clients have  $r = r_{\\mathrm{max}}$ .",
+        "bbox": [
+            114,
+            606,
+            489,
+            815
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Heterogeneous LoRA compared to Full-Fine Tuning. In Table 3, we can see that for the Reddit text summarization, HETLORA achieves similar performance with full-fine tuning with only",
+        "bbox": [
+            114,
+            818,
+            489,
+            882
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ba7de8ef47d9313cb445341eeeb155e5db11e78b250d594af4e7f6b8b3f2a123.jpg",
+        "image_caption": [
+            "Figure 5: Comparison of the performance across homogeneous LoRA, heterogeneous LoRA, and full fine-tuning. Heterogeneous LoRA achieves better performance than homogeneous LoRA with fewer number of communication rounds.",
+            "Figure 6: Ratio of communicated number of parameters for different PEFT methods to full fine-tuning to achieve the target value for the metric where it is RougeL 80 for Reddit text summarization task and perplexity 150 for the multisession chat response task. The 'X' means that the target metric is not achieved even after convergence."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            227,
+            873,
+            351
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "$0.003\\%$  of parameters trained compared to full-fine tuning. For chat response, HETLORA achieves slightly lower performance than full fine-tuning. However, full fine-tuning requires to train a much larger number of parameters compared to HETLORA, making it infeasible to train with ODFMs in practice. We also show in Fig. 6 that to achieve the targeted performance for both Reddit and MSC task, HETLORA requires significantly less number of parameters to be trained and communicated compared to full fine-tuning. Although for Reddit, HOMLORA has a slightly less number of parameters to be trained, the final achieved RougeL is outperformed by HETLORA as shown in Table 3.",
+        "bbox": [
+            507,
+            438,
+            882,
+            663
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of the Decaying Factor  $\\gamma$ . Lastly, we conduct an ablation study on the effect of the decaying factor  $\\gamma$  of HETLORA's local training step with self-rank pruning in Table 4. We observed that aggressive pruning hurts the performance where  $\\gamma = 0.85$  shows the worse performance across the varying  $\\gamma$  values. On the other hand, no pruning at all ( $\\gamma = 1$ ) underperforms the case when there is pruning ( $\\gamma = 0.99$ ), showing that reducing the noise introduced by large rank clients which data complexity is actually not that high indeed improves the performance.",
+        "bbox": [
+            507,
+            689,
+            880,
+            882
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8",
+        "bbox": [
+            492,
+            903,
+            504,
+            914
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "footer",
+        "text": "12910",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Discussions and Concluding Remarks",
+        "text_level": 1,
+        "bbox": [
+            115,
+            85,
+            475,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In our work, we investigated federated fine-tuning for ODFMs by proposing HETLORA that enables utilizing the power of FMs at the edge by taking device system and data heterogeneity into consideration. HETLORA is not only practical but also achieves better training speed, communication/computation efficiency, and final performance compared to other relevant baselines such as SVD-based LoRA, homogeneous LoRA, or full finetuning. Our findings in this work opens up several questions worth investigating. For instance, if the settings allow us to assign specific ranks to clients what will be the effective way to assign the ranks across clients for better convergence and performance? Another important next step of our work includes pursuing the theoretical convergence and generalization of heterogeneous LoRA.",
+        "bbox": [
+            114,
+            110,
+            489,
+            382
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations",
+        "text_level": 1,
+        "bbox": [
+            115,
+            395,
+            250,
+            410
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we address tackling system and data heterogeneity in federated fine-tuning of on-device foundation models. Our work is motivated by clients being able to carry different ranks for the LoRA fine-tuning method depending on their available resources, and thus exploiting this characteristic to improve federated fine-tuning with heterogeneous LoRA. However, our work assumes that the rank distribution across clients (which is analogous to how system resources are distributed across clients) is independent to the data distribution. There can be scenarios in which this is not necessarily the case where the rank and data distribution can be correlated. For instance, more affluent populations can have better off devices with larger resource capacity, and may have data distributions different to that of less affluent populations. Such correlation should be explored for future work to better understand the implications of heterogenous LoRA.",
+        "bbox": [
+            114,
+            420,
+            489,
+            741
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            769,
+            213,
+            783
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Anonymous. 2024. Flora: Federated fine-tuning large language models with heterogeneous low-rank adaptations.",
+            "Sara Babakniya, Ahmed Roushdy Elkordy, Yahya H. Ezzeldin, Qingfeng Liu, Kee-Bong Song, Mostafa El-Khamy, and Salman Avestimehr. 2023. Slora:"
+        ],
+        "bbox": [
+            117,
+            791,
+            489,
+            881
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Federated parameter efficient fine-tuning of language models. CoRR, abs/2308.06522.",
+            "Jiamu Bai, Daoyuan Chen, Bingchen Qian, Liuyi Yao, and Yaliang Li. 2024. Federated fine-tuning of large language models under heterogeneous tasks and client resources. arXiv preprint arXiv:2402.11505.",
+            "Rishi Bommasani, Drew A. Hudson, and et. al. Ehsan Adeli. 2022. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.",
+            "Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. 2016. Practical secure aggregation for federated learning on user-held data. In NIPS Workshop on Private Multi-Party Machine Learning.",
+            "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165.",
+            "Zhiqi Bu, Yu-Xiang Wang, Sheng Zha, and George Karypis. 2022. Differentially private bias-term only fine-tuning of foundation models. arXiv preprint arXiv:2210.00036.",
+            "Jinyu Chen, Wenchao Xu, Song Guo, Junxiao Wang, Jie Zhang, and Haozhao Wang. 2022. Fevtune: A deep dive into efficient federated fine-tuning with pre-trained transformers. CoRR, abs/2211.08025.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.",
+            "Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. 2023. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378.",
+            "Gemini Team Google. 2024. Gemini: A family of highly capable multimodal models. https:// arxiv.org/abs/2312.11805."
+        ],
+        "bbox": [
+            510,
+            86,
+            880,
+            881
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9",
+        "bbox": [
+            492,
+            903,
+            504,
+            914
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "footer",
+        "text": "12911",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Google. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.",
+            "Google. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.1040.",
+            "Google DeepMind. 2023. Introducing palm2. https://blog.google/technology/ai/ google-palm-2-ai-large-language-model/.",
+            "Tao Guo, Song Guo, and Junxiao Wang. 2023. Pfed-prompt: Learning personalized prompt for vision-language models in federated learning. In Proceedings of the ACM Web Conference 2023, WWW '23, page 1364-1374, New York, NY, USA. Association for Computing Machinery.",
+            "Tao Guo, Song Guo, Junxiao Wang, and Wenchao Xu. 2022. Promptfl: Let federated participants cooperatively learn prompts instead of models — federated learning in age of foundation model. CoRR, abs/2208.11625.",
+            "Venkatesan Guruswami and Ravi Kannan. 2012. Lecture notes in computer science theory for the information age.",
+            "Samuel Horváth, Stefanos Laskaridis, Shashank Rajput, and Hongyi Wang. 2024. Maestro: Uncovering low-rank structures via trainable decomposition. In Proceedings of the International Conference on Machine Learning (ICML).",
+            "Neil Houlsby, Andrei Giurgiu, Stanisław Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In Proceedings of the International Conference on Machine Learning (ICML).",
+            "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR).",
+            "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In The 36th Conference on Neural Information Processing Systems (NeurIPS 2022).",
+            "Jaejun Lee, Raphael Tang, and Jimmy Lin. 2019. What would elsa do? freezing layers during transformer fine-tuning. arXiv preprint arXiv:1911.03090.",
+            "Gwen Legate, Nicolas Bernier, Lucas Caccia, Edouard Oyallon, and Eugene Belilovsky. 2023. Guiding the last layer in federated learning with pre-trained models. In Workshop of Federated Learning and Analytics in Practice@ICML."
+        ],
+        "bbox": [
+            117,
+            86,
+            487,
+            879
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing (EMNLP).",
+            "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190.",
+            "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.",
+            "Zequan Liu, Jiawen Lyn, Wei Zhu, Xing Tian, and Yvette Graham. 2024. Alora: Allocating low-rank adaptation for fine-tuning large language models. arXiv preprint arXiv:2403.16187.",
+            "Andre Manoel, Mirian del Carmen Hipolito Garcia, Tal Baumel, Shize Su, Jialei Chen, Robert Sim, Dan Miller, Danny Karmon, and Dimitrios Dimitriadis. 2023. Federated multilingual models for medical transcript analysis. In Conference on Health, Inference, and Learning (CHIL), pages 147-162.",
+            "H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agoura y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. International Conference on Artificial Intelligence and Statistics (AISTATS).",
+            "OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:submit/4812508.",
+            "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020.",
+            "Anit Kumar Sahu, Tian Li, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Virginia Smith. 2020. Federated optimization for heterogeneous networks. In Proceedings of the 3rd MLSys Conference.",
+            "Joao Sedoc, Daphne Ippolito, Arun Kirubarajan, Jai Thirani, Lyle Ungar, and Chris Callison-Burch. 2019. Chateval: A tool for chatbot evaluation. Proceedings of NAACL-HLT.",
+            "Ofir Ben Shoham and Nadav Rappoport. 2023. Federated learning of medical concepts embedding using behrt. arXiv preprint arXiv:2305.13052.",
+            "Aliaksandra Shysheya, John F Bronskill, Massimiliano Patacchiola, Sebastian Nowozin, and Richard E Turner. 2023. Fit: Parameter efficient few-shot transfer learning for personalized and federated image classification. International Conference on Learning Representations (ICLR)."
+        ],
+        "bbox": [
+            510,
+            86,
+            880,
+            879
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10",
+        "bbox": [
+            489,
+            903,
+            509,
+            914
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "footer",
+        "text": "12912",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. 2024. Improving lora in privacy-preserving federated learning. In International Conference on Learning Representations (ICLR).",
+            "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.",
+            "Michael Volske, Martin Potthast, Shahbaz Syed, and Benno Stein. 2017. TL;DR: Mining Reddit to learn automatic summarization. In Proceedings of the Workshop on New Frontiers in Summarization, pages 59-63, Copenhagen, Denmark. Association for Computational Linguistics.",
+            "Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K. Leung, Christian Makaya, Ting He, and Kevin Chan. 2019. Adaptive federated learning in resource constrained edge computing systems. IEEE Journal on Selected Areas in Communications, 37(6):1205 - 1221.",
+            "Colin Wei, Sang Michael Xie, and Tengyu Ma. 2021. Why do pretrained language models help in downstream tasks? an analysis of head and prompt tuning. Advances in Neural Information Processing Systems, 34:16158-16170.",
+            "Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S. Morcos. 2023. Io-fi: distributed fine-tuning without communication. Transactions on Machine Learning Research (TMLR).",
+            "Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. 2024. Fedlora: When personalized federated learning meets low-rank adaptation.",
+            "Jing Xu, Arthur Szlam, and Jason Weston. 2021. Beyond goldfish memory: Long-term open-domain conversation. arXiv preprint arXiv:2107.07567.",
+            "Liping Yi, Han Yu, Gang Wang, and Xiaoguang Liu. 2023. Fedlora: Model-heterogeneous personalized federated learning with lora tuning. arXiv preprint arXiv:2310.13283.",
+            "Sixing Yu, J. Pablo Muñoz, and Ali Jannesari. 2023. Federated foundation models: Privacy-preserving and collaborative learning for large models. arXiv preprint arXiv:2305.11414.",
+            "Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. 2021. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199."
+        ],
+        "bbox": [
+            117,
+            86,
+            489,
+            881
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, and Guoyin Wangand Yiran Chen. 2023a. Towards building the federated gpt: Federated instruction tuning. CoRR, abs/2305.05644.",
+            "Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023b. Adaptive budget allocation for parameter-efficient fine-tuning. In The 11th International Conference on Learning Representations (ICLR).",
+            "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL).",
+            "Zhuo Zhang, Xiangjing Hu, Jingyuan Zhang, Yating Zhang, Hui Wang, Lizhen Qu, and Zenglin Xu. 2023c. Fedlegal: The first real-world federated learning benchmark for legal nlp. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL).",
+            "Zhuo Zhang, Yuanhang Yang, Yong Dai, Qifan Wang, Yue Yu, Lizhen Qu, and Zenglin Xu. 2023d. Fedpetuning: When federated learning meets the parameter-efficient tuning methods of pre-trained language models. Findings of the Association for Computational Linguistics (ACL).",
+            "Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, Hao Peng, Jianxin Li, Jia Wu, Ziwei Liu, Pengtao Xie, Caiming Xiong, Jian Pei, Philip S. Yu, and Lichao Sun. 2023. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419."
+        ],
+        "bbox": [
+            510,
+            86,
+            880,
+            596
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11",
+        "bbox": [
+            489,
+            903,
+            507,
+            914
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "footer",
+        "text": "12913",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2024/Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models/d6c62a2e-d2cd-45b7-a876-95885365f893_model.json

@@ -0,0 +1,2026 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.237,
+                0.092,
+                0.764,
+                0.131
+            ],
+            "angle": 0,
+            "content": "Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.205,
+                0.153,
+                0.803,
+                0.171
+            ],
+            "angle": 0,
+            "content": "Yae Jee Cho\\(^{1,2*}\\), Luyang Liu\\(^{2}\\), Zheng Xu\\(^{2}\\), Aldi Fahrezi\\(^{2}\\), Gauri Joshi\\(^{1}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.307,
+                0.171,
+                0.697,
+                0.188
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)Carnegie Mellon University, \\(^{2}\\)Google Research"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.258,
+                0.189,
+                0.744,
+                0.205
+            ],
+            "angle": 0,
+            "content": "{yaejeeecho, luyangliu, xuzheng, aldifahrezi}@google.com,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.403,
+                0.206,
+                0.6,
+                0.22
+            ],
+            "angle": 0,
+            "content": "gaurij@andrew.cmu.edu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.262,
+                0.263,
+                0.343,
+                0.279
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.147,
+                0.29,
+                0.461,
+                0.518
+            ],
+            "angle": 0,
+            "content": "Foundation models (FMs) adapt surprisingly well to downstream tasks with fine-tuning. However, their colossal parameter space prohibits their training on resource-constrained edge-devices. For federated fine-tuning, we need to consider the smaller FMs of few billion parameters at most, namely on-device FMs (ODFMs), which can be deployed on device. Federated fine-tuning of ODFMs has unique challenges non-present in standard fine-tuning: i) ODFMs poorly generalize to downstream tasks due to their limited sizes making proper fine-tuning imperative to their performance, and ii) devices have limited and heterogeneous system capabilities and data that can deter the performance of fine-tuning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.144,
+                0.522,
+                0.461,
+                0.765
+            ],
+            "angle": 0,
+            "content": "Tackling these challenges, we propose HETLORA, a feasible and effective federated fine-tuning method for ODFMs that leverages the system and data heterogeneity at the edge. HETLORA allows heterogeneous LoRA ranks across clients for their individual system resources, and efficiently aggregates and distributes these LoRA modules in a data-aware manner by applying rank self-pruning locally and sparsity-weighted aggregation at the server. It combines the advantages of high and low-rank LoRAs, achieving improved convergence speed and final performance compared to homogeneous LoRA. Furthermore, HETLORA has enhanced computation and communication efficiency compared to full fine-tuning making it more feasible for the edge."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.117,
+                0.777,
+                0.262,
+                0.792
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.802,
+                0.49,
+                0.85
+            ],
+            "angle": 0,
+            "content": "The emerging foundation models (FMs), easily ranging from few to hundreds of billions of parameters (Gemini Team Google, 2024; Bommasani"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.263,
+                0.883,
+                0.553
+            ],
+            "angle": 0,
+            "content": "et al., 2022; Radford et al., 2021; Devlin et al., 2019; OpenAI, 2023; Google, 2022; Touvron et al., 2023; Brown et al., 2020; Google, 2022; Driess et al., 2023), have remarkable zero/few shot learning capabilities. They often perform surprisingly well on a variety of tasks such as text/image generation, language translation, and conversing in natural language with fine-tuning. Such standard FMs, however, demand costly resources for directly fine-tuning their entire parameter space. To tackle this issue, many works have proposed different parameter-efficient fine-tuning (PEFT) methods such as prompt tuning (Lester et al., 2021), utilizing adapters (Houlsby et al., 2019), or low-rank adaptation (LoRA) of the original model (Hu et al., 2021) which freezes the original pre-trained parameters of the FM and train additional, smaller number of parameters instead."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.557,
+                0.883,
+                0.846
+            ],
+            "angle": 0,
+            "content": "Unfortunately, such PEFT methods rely on the assumptions that i) FMs are deployed to and trained with the data of a single client for adaptation to the downstream task and ii) the client has enough resources to even fit a standard FM of hundred billion size for, at least, inference. In practice, more often than not we are interested in fine-tuning FMs for on-device private data that is distributed across multiple devices, and these clients at the edge rarely have such resources to even fit standard FMs for inference. One of the examples is to use FMs for downstream tasks that requires training with private, non-sharable data such as medical or law-related documents distributed across clients (Manoel et al., 2023; Shoham and Rappoport, 2023; Zhang et al., 2023c). In such cases, fine-tuning of the FMs needs to be brought to the edge that has limited system resources."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.851,
+                0.883,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Federated Fine-Tuning & ODFMs. We consider such federated fine-tuning scenarios, where"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.116,
+                0.856,
+                0.49,
+                0.883
+            ],
+            "angle": 0,
+            "content": "*Work done while interning at Google Research. Corresponding authors: {yaejeeecho,luyangliu} @google.com"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.477,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12903"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.153,
+                0.946,
+                0.845,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 12903-12913 November 12-16, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.126,
+                0.085,
+                0.477,
+                0.13
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Zero-Shot</td><td>Few-Shot</td><td>Full-Training</td></tr><tr><td>PaLM 2 XXS</td><td>2930.23</td><td>2541.86</td><td>23.71</td></tr><tr><td>PaLM 2 XS</td><td>2712.86</td><td>481.95</td><td>18.32</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.115,
+                0.136,
+                0.49,
+                0.187
+            ],
+            "angle": 0,
+            "content": "Table 1: Perplexity of PaLM 2 for zero-shot, few-shot (5 communication rounds), and full federated fine-tuning (200 communication rounds) for chat response on the multi-session chat data (further experimental details are in Section 4.)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.203,
+                0.49,
+                0.412
+            ],
+            "angle": 0,
+            "content": "we train a set of parameters collaboratively across clients to obtain a global set of parameters that can be plugged in to the FM for the targeted downstream task. Note that federated fine-tuning is orthogonal to personalization of FMs in federated learning (FL) that has been explored in other previous work (Guo et al., 2023; Wu et al., 2024), which aims to train parameters that perform well for individual clients rather than general downstream tasks. We also define on-device FMs (ODFMs) as models with few billion parameters at max that are able to fit into memory on limited capacity clients considering current hardwares."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.414,
+                0.49,
+                0.622
+            ],
+            "angle": 0,
+            "content": "Challenges. Federated fine-tuning of ODFMs entails two major challenges non-present in neither standard PEFT of FMs nor federated training of non-FMs. First, ODFMs poorly generalize to downstream tasks. Unlike ODFMs, standard FMs have their zero/few-shot learning capability by their large parameter space trained on massive data. However, as we show in Table 1 and as previous literature has shown (Kojima et al., 2022; Lester et al., 2021), FMs' performance deteriorates as their sizes get smaller and federated fine-tuning is not merely useful but inevitable for ODFMs to perform well for on-device downstream tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.626,
+                0.494,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Second, devices have limited and heterogeneous system capabilities and data distributions (Wang et al., 2019; Bonawitz et al., 2016; Sahu et al., 2020). Without suitable PEFT methods that flexibly adapts to such heterogeneity across devices, we would only have limited performance output from federated fine-tuning of ODFMs. For instance, with homogeneous rank deployment in LoRA that is agnostic of system heterogeneity, the client with the least system resource becomes the bottleneck, forcing a smaller rank to be deployed across all clients despite the other clients being able to handle higher ranks (see Table 2 for the effect of rank on the trainable parameter #). Additionally, data heterogeneity causing model drifts can make the model converge to a suboptimal point."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.513,
+                0.084,
+                0.883,
+                0.231
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.244,
+                0.885,
+                0.33
+            ],
+            "angle": 0,
+            "content": "Figure 1: Overview of heterogeneous rank deployment of LoRA: the pretrained weights \\(\\mathbf{W}_0\\) are stored on-device and heterogeneous ranks are assigned to different clients with \\(r_{\\mathrm{min}} = r_1 < r_2 < r_3 = r_{\\mathrm{max}}\\). In our proposed HETLORA, the server receives the trained heterogeneous LoRA modules and aggregates them with sparsity-weighted aggregation to update the global LoRA module."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.346,
+                0.885,
+                0.571
+            ],
+            "angle": 0,
+            "content": "Contributions. In our work, we push the limits of federated fine-tuning of ODFMs by proposing HETLORA that is aware of heterogeneity residing at the edge. Our main flagship results (in Table 3 and Fig. 5), show that our proposed HETLORA outperforms the standard LoRA method and achieves comparable performance with full fine-tuning with a significantly smaller number of trained parameters. We verify HETLORA with PaLM 2 (Google, 2023) of XXS and XS size for chat responses on the multi-session chat data (Xu et al., 2021) and text summarization for the Reddit data (Volske et al., 2017), both real world data from clients. Our contributions are summarized as:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.591,
+                0.883,
+                0.672
+            ],
+            "angle": 0,
+            "content": "- We propose HETLORA that can apply different rank LoRA modules to different clients to cater to the heterogeneous system capabilities and data complexities of the clients, via utilizing rank self-pruning and sparsity-weighted aggregation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.679,
+                0.882,
+                0.775
+            ],
+            "angle": 0,
+            "content": "- We show the performance of naively applying LoRA with homogeneous ranks across clients for federated fine-tuning, and show that while large ranks help in speeding-up training, they lead to faster overfitting while smaller ranks are slower in training but does not suffer from overfitting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.783,
+                0.885,
+                0.863
+            ],
+            "angle": 0,
+            "content": "- We then evaluate our proposed HETLORA to show that it outperforms naively applying homogeneous ranks across clients in terms of both training speed, communication/computation efficiency, and final performance."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.509,
+                0.591,
+                0.885,
+                0.863
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.507,
+                0.916
+            ],
+            "angle": 0,
+            "content": "2"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12904"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.143,
+                0.093,
+                0.855,
+                0.131
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>r = 1</td><td>r = 5</td><td>r = 10</td><td>r = 20</td><td>r = 50</td><td>r = 100</td><td>r = 150</td><td>r = 200</td></tr><tr><td>PaLM 2 XXS, PaLM 2 XS</td><td>0.02%</td><td>0.11%</td><td>0.21%</td><td>0.42%</td><td>1.05%</td><td>2.10%</td><td>3.14%</td><td>4.19%</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.115,
+                0.138,
+                0.881,
+                0.175
+            ],
+            "angle": 0,
+            "content": "Table 2: Percentage of the LoRA parameters' size for different ranks \\( r \\) compared to the original pre-trained ODFM's parameter size. Even for large ranks such as \\( r = {200} \\) the trainable LoRA parameters' size compared to the original pre-trained ODFM size is less than 5% for both PaLM 2-XS and PaLM 2-XXS."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.195,
+                0.273,
+                0.21
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.222,
+                0.49,
+                0.543
+            ],
+            "angle": 0,
+            "content": "Parameter-Efficient Fine Tuning. Most PEFT methods either train a subset of parameters within the existing FM whilst other parameters are frozen or introduce an additional set of trainable parameters whilst keeping the original FM frozen. For the former, methods such as head or bias fine-tuning (Wei et al., 2021; Bu et al., 2022; Lee et al., 2019; Zaken et al., 2021) has been explored, and for the latter, methods such as adapters (Houlsby et al., 2019), prompt (Lester et al., 2021) or prefix-tuning (Li and Liang, 2021), and low-rank approximation (Hu et al., 2021) has been proposed. While these number of methods has been proven to perform as well as full model fine-tuning with just a few number of parameters for the centralized setting, it has not been thoroughly explored how these methods would perform for a much smaller FM such as ODFMs, in the decentralized setting where clients' system-capacities and data can be heterogeneous and much limited."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.545,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Federated Fine-Tuning. Several recent work has proposed to combine the PEFT methods devised for the centralized setting to FL (Zhou et al., 2023; Yu et al., 2023) such as training prompts or adapters collaboratively with FL (Guo et al., 2022; Chen et al., 2022; Zhang et al., 2023a; Shysheya et al., 2023; Legate et al., 2023). Another line of work has proposed to perform a few-shot or nearly zero-shot training of FMs with FL for improved communication-efficiency (Wortsman et al., 2023; Zhang et al., 2023d). However, these work either overlooks that most devices do not have the resource to fit standard FMs (Touvron et al., 2023; Brown et al., 2020) even for inference or does not consider the heterogeneous system capacities of the clients. It is detrimental to consider these factors since FMs that actually fits to the devices in FL are much smaller, making them weaker in the general intelligence capabilities, and also heterogeneous system capacities can prohibit deploying same sized PEFT parameters across clients."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.195,
+                0.885,
+                0.662
+            ],
+            "angle": 0,
+            "content": "LoRA and FL. There has been a number of variations of LoRA (Zhang et al., 2023b; Liu et al., 2024; Sun et al., 2024; Horváth et al., 2024) from its first proposal (Hu et al., 2021). However, only a few numbers of recent work has looked in to using LoRA for FL. In Babakniya et al. (2023), the importance of the initialization for the LoRA modules is evaluated where they propose to train the LoRA modules with FL and then perform singular value decomposition (SVD) to gain a good initialization of the LoRA modules. However, the training process of LoRA itself is not altered to adapt to heterogeneous system capabilities of devices. Another recent work (Yi et al., 2023) has evaluated LoRA in the context of personalized FL, but other than applying LoRA to personalization, the LoRA method itself is, again, not changed. To the best of our knowledge, our work is the first method to consider both data and system heterogeneity for improving training speed, communication/computation efficiency, and final performance of ODFM's federated fine-tuning. Contemporary approaches (Bai et al., 2024; Anonymous, 2024) have proposed to use SVD to decompose the full matrix to deploy heterogeneous ranks across clients. We compare our proposed method with using SVD for heterogeneous LoRA in our work and show that the performance is underwhelming compared to our proposed heterogeneous LoRA."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.675,
+                0.844,
+                0.691
+            ],
+            "angle": 0,
+            "content": "3 Federated Fine-Tuning with LoRA"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.701,
+                0.664,
+                0.715
+            ],
+            "angle": 0,
+            "content": "3.1 Preliminaries"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.722,
+                0.885,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Formally, we define the pre-trained ODFM as \\(\\mathbf{W}_0\\in \\mathbb{R}^{d\\times l}\\) and the trainable low-rank decomposed matrix as \\(\\Delta \\mathbf{W}\\in \\mathbb{R}^{d\\times l}\\). In standard LoRA (Hu et al., 2021) under the centralized setting, the low-rank decomposition of \\(\\Delta \\mathbf{W}\\) is constructed such that \\(\\Delta \\mathbf{W} = \\mathbf{BA}\\) where \\(\\mathbf{B}\\in \\mathbb{R}^{d\\times r}\\) and \\(\\mathbf{A}\\in \\mathbb{R}^{r\\times l}\\) are the low rank decomposition of \\(\\Delta \\mathbf{W}\\) with identical rank \\(r\\). Now, let us consider LoRA for federated fine-tuning where there are \\(M\\) total clients. Each client \\(k\\in [M]\\) has pri"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.506,
+                0.916
+            ],
+            "angle": 0,
+            "content": "3"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.942
+            ],
+            "angle": 0,
+            "content": "12905"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.086,
+                0.493,
+                0.396
+            ],
+            "angle": 0,
+            "content": "vate data \\(\\mathcal{B}_k\\) and its corresponding local empirical loss function \\(F_{k}(\\mathbf{W}) = \\frac{1}{|\\mathcal{B}_{k}|}\\sum_{\\xi \\in \\mathcal{B}_{k}}\\ell (\\mathbf{W},\\xi)\\) where \\(\\ell (\\mathbf{W},\\xi)\\) is the loss for model \\(\\mathbf{W}\\) at data sample \\(\\xi\\). The optimization task for federated fine-tuning is to collaboratively find the global parameters which we define as \\(\\overline{\\mathbf{B}}\\) and \\(\\overline{\\mathbf{A}}\\), given the pretrained knowledge \\(\\mathbf{W}_0\\) that can minimize the global objective \\(F(\\overline{\\mathbf{W}}) = \\frac{1}{M}\\sum_{k = 1}^{M}F_{k}(\\overline{\\mathbf{W}})\\) where \\(\\overline{\\mathbf{W}} = \\mathbf{W}_0 + \\overline{\\mathbf{B}}\\overline{\\mathbf{A}}\\). Later in the paper, when introducing heterogeneous LoRA we truncate the LoRA modules' rank dimension, for example from \\(\\mathbf{B}\\in \\mathbb{R}^{d\\times r}\\), \\(\\mathbf{A}\\in \\mathbb{R}^{r\\times l}\\) to \\(\\mathbf{B}'\\in \\mathbb{R}^{d\\times r'}\\), \\(\\mathbf{A}'\\in \\mathbb{R}^{r'\\times l}\\) where \\(r' < r\\). Throughout the paper, we denote such truncation of a matrix with the : symbol for each row and column at the subscript. For instance, for truncation to \\(r' < r\\) at the column for the matrix \\(\\mathbf{B}\\in \\mathbb{R}^{d\\times r}\\), we keep all the columns until \\(r'\\) and omit the last \\(r - r'\\) columns and denote the resulting matrix it as \\(\\mathbf{B}_{:,;r'}\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.411,
+                0.424,
+                0.427
+            ],
+            "angle": 0,
+            "content": "3.2 Naive Case: Homogeneous LoRA"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.436,
+                0.49,
+                0.816
+            ],
+            "angle": 0,
+            "content": "A straightforward way to perform federated fine-tuning with LoRA is to train the LoRA modules \\(\\mathbf{B}\\), \\(\\mathbf{A}\\) with homogeneous rank \\(r\\) across all clients with standard FL (McMahan et al., 2017). Specifically, first the clients have the pre-trained ODFM weights \\(\\mathbf{W}_0\\) stored in their devices prior to training for the forward pass when training the LoRA modules. Then, the server sends the global LoRA modules \\(\\overline{\\mathbf{B}}^{(t)}\\), \\(\\overline{\\mathbf{A}}^{(t)}\\) to the set of \\(m\\) selected clients \\(S^{(t)}\\) per communication round \\(t\\). Each selected client \\(k \\in S^{(t)}\\) trains the LoRA modules on their local data for a few local iterations (usually with mini-batch SGD) and send the updated modules \\(\\mathbf{B}_k^{(t)}\\), \\(\\mathbf{A}_k^{(t)}\\) back to the server. The server then updates the global LoRA modules accordingly to \\(\\overline{\\mathbf{B}}^{(t+1)} = \\sum_{k \\in S^{(t)}} \\mathbf{B}_k^{(t)} / m\\), \\(\\overline{\\mathbf{A}}^{(t+1)} = \\sum_{k \\in S^{(t)}} \\mathbf{A}_k^{(t)} / m\\) and sends back to the next set of selected clients for the next communication round. This training process is nearly identical to the standard FL algorithm (McMahan et al., 2017) except that the pretrained weights \\(\\mathbf{W}_0\\) are frozen and locally stored in the clients' devices and only the LoRA moduels are trained and communicated."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.819,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Instead of such homogeneous rank deployment across all clients, it is not only possible but even more practical and feasible to use heterogeneous rank deployment for federated fine-tuning (we later"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.086,
+                0.885,
+                0.328
+            ],
+            "angle": 0,
+            "content": "show that this also improves final performance of the ODFMs). This involves training LoRA modules with varying ranks across clients, based on their system capabilities. This allows a more practical setting for federated fine-tuning of ODFMs where, for instance, the client with the smallest system capacity does not have to become the bottleneck that forces deploying a smaller rank across all clients. Instead, with heterogeneous rank deployment, each client can use a rank that suits for itself. However, heterogeneous rank deployment poses challenges in aggregating and redistributing the LoRA modules. To address these challenges, we introduce a solution called HETLORA, which pushes the limits beyond homogeneous LoRA."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.34,
+                0.877,
+                0.355
+            ],
+            "angle": 0,
+            "content": "3.3 Proposed Method: Heterogeneous LoRA"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.361,
+                0.885,
+                0.57
+            ],
+            "angle": 0,
+            "content": "Overview. Our proposed heterogeneous LoRA method, namely HETLORA, is not restricted to any specific method to assign the ranks to the clients and the clients can decide their respective ranks themselves. For formality, in our paper, we formulate that each client has a rank denoted as \\( r_k \\), within a range of \\( r_k \\in [r_{\\min}, r_{\\max}] \\), \\( \\forall k \\) (see Fig. 1). HETLORA comprises three steps: 1) Distribution via Truncation, 2) Local Training with Rank Self-Pruning, and 3) Sparsity-Weighted Aggregation of the LoRA modules. These steps are detailed further in the subsequent paragraphs. An overview of HETLORA is illustrated in Fig. 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.571,
+                0.884,
+                0.805
+            ],
+            "angle": 0,
+            "content": "1) Distribution via Truncation. At the beginning of each communication round \\( t \\), the server holds initial global LoRA modules \\( \\overline{\\mathbf{B}}^{(t)} \\), \\( \\overline{\\mathbf{A}}^{(t)} \\) with a global rank \\( r^{(t)} \\). The value of the global rank \\( r^{(t)} \\) depends on how we aggregate the heterogeneous rank LoRA modules which is elaborated on in step 3). The server then distributes these global LoRA modules to a subset of selected set of clients \\( S^{(t)} \\) with heterogeneous ranks \\( r_k^{(t)}, k \\in S^{(t)} \\) for local training<sup>1</sup>. With the given global LoRA modules, we consider a simple and intuitive method of truncation where the server sends \\( \\overline{\\mathbf{B}}_{:,r_k}^{(t)} \\), \\( \\overline{\\mathbf{A}}_{:r_k}^{(t)} \\) to each client \\( k \\) with rank \\( r_k^{(t)} \\) for local training where we omitted the superscript for \\( r_k \\) for simplicity."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.527,
+                0.806,
+                0.883,
+                0.822
+            ],
+            "angle": 0,
+            "content": "2) Local Training with Rank Self-Pruning."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.831,
+                0.883,
+                0.883
+            ],
+            "angle": 0,
+            "content": "There is a superscript \\(t\\) for the ranks \\(r_k^{(t)}\\) across clients which indicates that in HETLORA these heterogeneous ranks can be changed over the communication rounds via self-pruning explained in step 2)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.506,
+                0.916
+            ],
+            "angle": 0,
+            "content": "4"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12906"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.083,
+                0.574,
+                0.201
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.58,
+                0.083,
+                0.738,
+                0.201
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.741,
+                0.083,
+                0.882,
+                0.201
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.115,
+                0.215,
+                0.882,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Figure 2: Overview of the zero-padding, sparsity-weighted aggregation, and truncation method for HETLORA; (a): Zero-pad LoRA modules with smaller ranks to \\( r_{\\mathrm{max}} \\) (clients with rank \\( r_{\\mathrm{max}} \\) does not need padding) and calculate their sparsity by calculating the Frobenius norm of the reconstructed model \\( \\Delta \\mathbf{W}_k^{(t)} = \\mathbf{B}_k^{(t)}\\mathbf{A}_k^{(t)} \\); (b): After padding, aggregate all of the clients' LoRA modules with the weights \\( \\| \\mathbf{S}_k^{(t)}\\| /\\mathbf{Z}^{(t)} \\) calculated by \\( \\Delta \\mathbf{W}_k^{(t)} \\) to get the global LoRA modules; (c): Truncate the global LoRA modules for the specific rank of the next selected client (example for client with rank \\( r_2 \\))."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.301,
+                0.489,
+                0.526
+            ],
+            "angle": 0,
+            "content": "After receiving LoRA modules from the server as \\(\\mathbf{B}_k^{(t,0)} = \\overline{\\mathbf{B}}_{\\therefore r_k}^{(t)}\\), \\(\\mathbf{A}_k^{(t,0)} = \\overline{\\mathbf{A}}_{r_k,\\cdot}^{(t)}\\), each client \\(k \\in S^{(t)}\\) performs \\(\\tau\\) local iterations of mini-batch SGD on their local data to minimize the local objective \\(\\frac{1}{|\\mathcal{B}_k|}\\sum_{\\xi \\in \\mathcal{B}_k}\\ell((\\mathbf{B}_k, \\mathbf{A}_k), \\xi|\\mathbf{W}_0)\\), and sends back the updated LoRA modules \\(\\mathbf{B}_k^{(t,\\tau)} \\in \\mathbb{R}^{d \\times r_k^{(t)}}\\) and \\(\\mathbf{A}_k^{(t,\\tau)} \\in \\mathbb{R}^{r_k^{(t)} \\times l}\\) to the server. This is the same process as the standard local training step in vanilla FedAvg (McMahan et al., 2017). However, we improve this vanilla local training step by adding a rank self-pruning mechanism where clients self-prune their respective ranks depending on the magnitude of the model parameters."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.527,
+                0.489,
+                0.866
+            ],
+            "angle": 0,
+            "content": "Specifically, we add a regularization term to the original local objective to get \\(\\min_{\\mathbf{B}_k,\\mathbf{A}_k}\\frac{1}{|\\mathcal{B}_k|}\\sum_{\\xi \\in \\mathcal{B}_k}\\ell ((\\mathbf{B}_k,\\mathbf{A}_k),\\xi |\\mathbf{W}_0) + \\lambda \\| \\mathbf{B}_{k,:r_k\\gamma :r_k}\\| \\| \\mathbf{A}_{k,r_k\\gamma :r_k,:}\\|\\) where \\(\\gamma < 1\\) is a decay-factor that determines how aggressively we want to prune the ranks to a smaller value. The regularization term aims to minimize the norm of the last few ranks, which will become smaller if the first loss term \\(\\frac{1}{|\\mathcal{B}_k|}\\sum_{\\xi \\in \\mathcal{B}_k}\\ell ((\\mathbf{B}_k,\\mathbf{A}_k),\\xi |\\mathbf{W}_0)\\) is not very large. After training with the new local objective we compare the norm of the updated LoRA modules' last few layers \\(\\| \\mathbf{B}_{k,:r_k\\gamma :r_k}\\| \\| \\mathbf{A}_{k,r_k\\gamma :r_k,:}\\|\\) with the ones from the initially received LoRA modules. If the former is smaller we prune the last few layers (pruning intensity is determined by \\(\\gamma\\)) and send back the LoRA modules with a smaller rank. This means that for the LoRA modules which incurs a small local loss, i.e., well-trained on the clients' local data, the LoRA modules are more likely to be pruned to a smaller rank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.867,
+                0.486,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Such pruning allows HETLORA to reduce the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.301,
+                0.882,
+                0.478
+            ],
+            "angle": 0,
+            "content": "noise in the LoRA modules introduced by clients having a larger rank than the actual rank that their data complexity requires, and also reduces the complexity of the LoRA modules to improve generalization and prevent overfitting (see Table 4). Once the rank is pruned for a client, the client saves the updated rank and uses it as the starting rank if selected for future communication rounds. The client then sends back their updated and possibly rank-pruned LoRA modules to the server to be processed in the aggregation step."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.483,
+                0.882,
+                0.68
+            ],
+            "angle": 0,
+            "content": "3) Sparsity-Weighted Aggregation. Finally, the last step of HETLORA is aggregating the received heterogeneous LoRA modules \\(\\mathbf{B}_k^{(t,\\tau)}\\), \\(\\mathbf{A}_k^{(t,\\tau)}\\), \\(k \\in S^{(t)}\\). A straightforward way to aggregate the heterogeneous LoRA modules is using zero-padding to all the received LoRA modules with \\(r_i^{(t)} < \\max \\{r_k^{(t)} | k \\in S^{(t)}\\}\\) and then perform simple averaging over the modules. However, such naive aggregation can lead to biasing the model towards higher rank clients even when these clients may not hold valuable training information, i.e., having low data complexity, giving noisy updates."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.685,
+                0.882,
+                0.862
+            ],
+            "angle": 0,
+            "content": "In an ideal scenario where we can deploy any rank to any client, deploying higher ranks to the clients with higher data complexity or larger local datasets can retrieve more informative and less sparse updates from the clients. Conversely if we assign higher ranks to the clients whose data complexity is low, the actual rank of the full model from the reconstructed LoRA modules can be smaller than the assigned rank. Thus the higher rank client's update may be unnecessarily over-emphasized in the naive zero padding method."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.528,
+                0.867,
+                0.882,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Based on this insight we propose a sparsity-"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.505,
+                0.915
+            ],
+            "angle": 0,
+            "content": "5"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.525,
+                0.94
+            ],
+            "angle": 0,
+            "content": "12907"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.087,
+                0.493,
+                0.428
+            ],
+            "angle": 0,
+            "content": "weighted aggregation scheme where the server reconstructs these LoRA modules to the full model as \\(\\Delta \\mathbf{W}_k^{(t)} = \\mathbf{B}_k^{(t)}\\mathbf{A}_k^{(t)}\\) and gets the norm of the singular value vectors from the full models denoted as \\(\\mathbf{S}_k^{(t)}\\) by calculating \\(\\| \\Delta \\mathbf{W}_k^{(t)}\\| _F\\). Note that the costly process of performing SVD for each of the full model \\(\\Delta \\mathbf{W}_k^{(t)}\\) can be avoided by simply calculating the Frobenius norm of \\(\\Delta \\mathbf{W}_k^{(t)}\\) (see Lemma 1.2 in (Guruswami and Kannan, 2012)). The server then weighs the LoRA modules with aggregation weight \\(p_k^{(t)}\\) which is proportional to the norm of the singular value vectors. Formally, we have the global LoRA modules updated as \\(\\overline{\\mathbf{B}}^{(t + 1)} = \\sum_{k\\in \\mathcal{S}^{(t)}}p_k^{(t)}\\mathbf{B}_k^{(t)}\\), \\(\\overline{\\mathbf{A}}^{(t + 1)} = \\sum_{k\\in \\mathcal{S}^{(t)}}p_k^{(t)}\\mathbf{A}_k^{(t)}\\) where \\(p_k^{(t)}\\coloneqq \\| \\mathbf{S}_k^{(t)}\\| /\\mathbf{Z}^{(t)}\\) with normalizing factor \\(\\mathbf{Z}^{(t)}\\coloneqq \\sum_{k^{\\prime}\\in \\mathcal{S}^{(t)}}\\| \\mathbf{S}_{k^{\\prime}}^{(t)}\\|\\). This way, we can de-emphasize the larger rank assigned-clients that have rather less informative updates, and more emphasize the smaller rank assigned-clients that have more informative ones."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.441,
+                0.479,
+                0.473
+            ],
+            "angle": 0,
+            "content": "3.4 Why not Simply Reconstruct First, then Redistribute the LoRA modules?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.48,
+                0.49,
+                0.672
+            ],
+            "angle": 0,
+            "content": "One might ask why not simply reconstruct each of the LoRA modules to the full matrix and aggregate them. Here we show that reconstructing the LoRA modules and aggregating them to get the full model results in a different full model compared to when we aggregate the LoRA modules first and then reconstruct the final model. In Section 4 we also empirically show that reconstructing the LoRA modules to the full model and redistributing them after truncated SVD to the corresponding rank of the clients results in an overwhelming performance compared to HETLORA."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.674,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Let us consider a simple case where there are 2 clients with heterogeneous rank lora modules \\(\\mathbf{B}_1\\in\\) \\(\\mathbb{R}^{d\\times 1}\\) \\(\\mathbf{A}_1\\in \\mathbb{R}^{1\\times l}\\) and \\(\\mathbf{B}_2\\in \\mathbb{R}^{d\\times 1}\\) \\(\\mathbf{A}_2\\in \\mathbb{R}^{2\\times l}\\) respectively for client 1 and client 2 where the former has rank 1 and latter has rank 2. We set the notation for the LoRA modules' \\(i^{th}\\) row and \\(j^{th}\\) column value for \\(\\mathbf{B}_k\\) and \\(\\mathbf{A}_k\\) as \\(b_{k,ij}\\) and \\(a_{k,ij}\\) respectively. Then with \\(d = 3\\) \\(l = 2\\) when we reconstruct each of the LoRA modules first and then aggregate the full model we have its \\(i^{th}\\) row and \\(j^{th}\\) column as \\((\\sum_{k = 1}^{2}b_{k,i0}a_{k,0j}) + b_{2,i1}a_{2,1j}\\) and aggregating the LoRA modules first and then reconstructing the model has the full model's \\(i^{th}\\) row and \\(j^{th}\\) column"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.084,
+                0.823,
+                0.103
+            ],
+            "angle": 0,
+            "content": "as \\((\\sum_{k = 1}^{2}b_{k,i0})(\\sum_{k = 1}^{2}a_{k,0j}) + b_{2,i1}a_{2,1j}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.104,
+                0.884,
+                0.345
+            ],
+            "angle": 0,
+            "content": "One can observe that the difference between the two models are the cross-terms between the left and right module of different client 1 and 2, i.e., \\( b_{1,i0}a_{2,0j} + b_{2,i0}a_{1,0j} \\) for the \\( i^{th} \\) row and \\( j^{th} \\) column. In other words, when we reconstruct the LoRA modules first and then aggregate them to get the full model, each term in the full model are cross-products between the left and right module of each client and not the cross-products between clients. Thus, reconstructing the LoRA modules loses information on the cross-relation across clients, only retaining the knowledge on the cross-relation between the LoRA modules B and A. Such observation is also corroborated by the reconstruction first's underwhelming performance in Table 3."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.357,
+                0.656,
+                0.374
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.384,
+                0.883,
+                0.608
+            ],
+            "angle": 0,
+            "content": "In this section, we show that our proposed HETLORA outperforms its baselines in terms of training speed, computation/communication efficiency, and final achieved performance. We first show the performance of homogeneous LoRA to show how LoRA in general performs for low and high rank values. Second, as our main result, we demonstrate HETLORA's performance for different \\( r_{\\mathrm{min}} \\) and \\( r_{\\mathrm{max}} \\) values comparing them with full fine-tuning, homogeneous LoRA, and the reconstruction-first method elaborated in Section 3.4. We also conduct an ablation study on HETLORA with varying decay factor \\( \\gamma \\) for the rank self-pruning step."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.61,
+                0.884,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Settings. We use the transformer-based language model PaLM 2 (Google, 2023) of size XXS and XS for our experiments which are lightweight enough to be ODFMs (Google DeepMind, 2023) compared to standard FMs. The tasks we consider are the chat dialogue from the multi-session chat (MSC) dataset (Xu et al., 2021) and the text summarization task from the Reddit dataset (Volske et al., 2017). For the MSC data we use perplexity (Zhang et al., 2018) as the metric which has been used to show the quality of chat responses from generative models from previous literature (Sedoc et al., 2019). We sample 100 users uniformly at random and partition their data for training and evaluation by each previous_DIALOG and dialog. The Reddit text summarization data consists of real users' reddit posts and their summarization, and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.507,
+                0.916
+            ],
+            "angle": 0,
+            "content": "6"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.942
+            ],
+            "angle": 0,
+            "content": "12908"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.143,
+                0.084,
+                0.855,
+                0.179
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td colspan=\"2\">Reddit (RougeL)</td><td colspan=\"2\">Multi-Session Chat (Perplexity)</td></tr><tr><td></td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td></tr><tr><td>Full</td><td>94.56(±0.01)</td><td>94.87(±0.04)</td><td>32.70(±0.17)</td><td>23.40(±0.36)</td></tr><tr><td>HOMLORA r = 5</td><td>92.57(±1.56), ×0.001</td><td>92.89(±0.96)</td><td>80.51(±8.32), ×0.001</td><td>64.59(±9.31)</td></tr><tr><td>HOMLORA r = 50</td><td>70.57(±2.13), ×0.01</td><td>84.95(±1.59)</td><td>307.96(±11.43), ×0.01</td><td>167.46(±1.72)</td></tr><tr><td>Recon+SVD</td><td>63.28(±1.92), ×0.003</td><td>75.17(±1.25)</td><td>323.89(±20.57), ×0.002</td><td>215.63(±15.38)</td></tr><tr><td>HETLORA γ = 0.99</td><td>94.23(±0.03), ×0.003</td><td>94.41(±0.05)</td><td>53.93(±1.57), ×0.002</td><td>38.76(±0.52)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.115,
+                0.188,
+                0.882,
+                0.238
+            ],
+            "angle": 0,
+            "content": "Table 3: Final RougeL score for Reddit text summarization and perplexity for multi-session chat for different federated fine-tuning methods. The blue text indicates the ratio of trained number of parameters compared to the full fine-tuning case. HETLORA outperforms both HOMLORA and Recon+SVD method, but slightly underperforms the full fine-tuning case. However, compared to full fine-tuning the number of trained parameter is significantly smaller."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.122,
+                0.255,
+                0.299,
+                0.372
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.157,
+                0.374,
+                0.264,
+                0.385
+            ],
+            "angle": 0,
+            "content": "(a) PaLM 2-XXS"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.304,
+                0.256,
+                0.482,
+                0.372
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.345,
+                0.374,
+                0.442,
+                0.385
+            ],
+            "angle": 0,
+            "content": "(b) PaLM 2-XS"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.115,
+                0.388,
+                0.489,
+                0.45
+            ],
+            "angle": 0,
+            "content": "Figure 3: Performance of homogeneous LoRA for different rank \\( r \\). Higher ranks achieve better performance with fewer communication rounds than the lower ranks, but they overfit quickly. Conversely, the lowest rank \\( r = 1 \\) achieves low perplexity slower than higher ranks, but without overfitting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.466,
+                0.49,
+                0.692
+            ],
+            "angle": 0,
+            "content": "we use RougeL (Lin, 2004) as the metric. We use 298 users from Reddit that have at least 100 data samples as the training clients and use another 100 users with at least 100 data samples for evaluation. We use mini-batch size 8 and number of local iterations \\(\\tau = 5\\) with the feature length set to 1024. For the learning rate we perform grid search in \\(\\eta = \\{0.1, 0.01, 0.001, 0.0001\\}\\). For each MSC and Reddit task, we select 5 and 10 clients per communication round respectively. The rank distribution across clients for HETLORA is set to a truncated power-law distribution with \\(\\alpha = 0.1\\) in the range between \\([r_{\\mathrm{min}}, r_{\\mathrm{max}}]\\) (inclusively). All experiments were ran with 3 different random seeds."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.702,
+                0.318,
+                0.717
+            ],
+            "angle": 0,
+            "content": "4.1 Experiment Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.722,
+                0.49,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Homogeneous LoRA and the Effect of Ranks \\( r \\). First, we evaluate the performance of federated fine-tuning of the LoRA modules with homogeneous LoRA deployment across clients in Fig. 3 for different ranks \\( r \\in [1,5,20,50] \\). We observe that a higher rank \\( r \\) for homogeneous LoRA achieves better perplexity floor with fewer communication rounds than the lower ranks but quickly overfits resulting in worse performance compared to the lower ranks after more communication rounds. On"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.516,
+                0.256,
+                0.693,
+                0.347
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.562,
+                0.35,
+                0.646,
+                0.361
+            ],
+            "angle": 0,
+            "content": "(a) \\(r_{\\mathrm{max}} = 10\\)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.699,
+                0.257,
+                0.876,
+                0.347
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.744,
+                0.349,
+                0.831,
+                0.361
+            ],
+            "angle": 0,
+            "content": "(b) \\(r_{\\mathrm{max}} = 50\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.364,
+                0.882,
+                0.437
+            ],
+            "angle": 0,
+            "content": "Figure 4: Performance of HETLORA without rank pruning or and with simple average aggregation. Similar to homogeneous LoRA, larger \\( r_{\\mathrm{min}} \\) leads to overfitting for heterogeneous LoRA, but it is not as severe as homogeneous LoRA even for larger maximum rank \\( r_{\\mathrm{max}} = 50 \\) showing that the smaller rank LoRA modules act as a regularizer for HETLORA."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.454,
+                0.883,
+                0.696
+            ],
+            "angle": 0,
+            "content": "the other hand, while the lower rank cases need more communication rounds to achieve good performance, it does not have the problem of overfitting as the higher ranks. Hence for homogeneous LoRA, there is a trade-off between low and high ranks, in terms of faster performance achievement and overfitting. Note that these observations are consistent with previous literature in the centralized setting where a higher rank does not necessarily yields the best performance (Hu et al., 2021; Zhang et al., 2023b). Subsequently, we show that HETLoRA that is not only more practical than homogeneous LoRA, but also achieves a better final performance with fewer number of communication rounds without the overfitting issue."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.706,
+                0.884,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Naïve Heterogeneous LoRA and the Effect of \\( r_{\\mathrm{min}} \\) and \\( r_{\\mathrm{max}} \\). First, we show the performance of naïve heterogeneous LoRA which is HETLORA without self rank-pruning and with only average aggregation instead of the sparsity-weighted aggregation in Fig. 4. We see that similar to homogeneous LoRA, a smaller minimum rank \\( r_{\\mathrm{min}} = 1 \\) leads to slower training but better performance while a larger maximum rank leads to faster training but worse performance. However, unlike homogeneous LoRA, the overfitting does not get"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.505,
+                0.916
+            ],
+            "angle": 0,
+            "content": "7"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12909"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.192,
+                0.084,
+                0.805,
+                0.164
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td colspan=\"2\">Reddit (RougeL)</td><td colspan=\"2\">Multi-Session Chat (Perplexity)</td></tr><tr><td></td><td>PaLM 2-XXS</td><td>PaLM 2-XS</td><td>PaLM 2-XXS</td><td>PaLM 2-XS</td></tr><tr><td>HETLORA, γ = 1</td><td>92.17 (±0.08)</td><td>91.95 (±0.03)</td><td>55.07 (±0.81)</td><td>40.92 (±0.58)</td></tr><tr><td>HETLORA, γ = 0.99</td><td>94.23 (±0.03)</td><td>94.41 (±0.05)</td><td>53.93 (±1.57)</td><td>38.76 (±0.52)</td></tr><tr><td>HETLORA, γ = 0.95</td><td>89.62 (±1.33)</td><td>83.19 (±1.70)</td><td>71.10 (±1.39)</td><td>46.39 (±0.87)</td></tr><tr><td>HETLORA, γ = 0.85</td><td>60.31 (±3.04)</td><td>53.28 (±2.47)</td><td>120.72 (±10.93)</td><td>59.67 (±1.98)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.115,
+                0.173,
+                0.882,
+                0.211
+            ],
+            "angle": 0,
+            "content": "Table 4: Ablation study on the effect of the decaying factor \\( \\gamma \\) for HETLORA's self-rank pruning in the local training step. While aggressive pruning can be harmful to HETLORA's performance, pruning \\( \\left( {\\gamma  = {0.99}}\\right) \\) can outperform the case when there is no pruning at all \\( \\left( {\\gamma  = 1}\\right) \\) by reducing the noise introduced by large rank clients with low data complexity."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.122,
+                0.223,
+                0.3,
+                0.36
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.179,
+                0.362,
+                0.242,
+                0.373
+            ],
+            "angle": 0,
+            "content": "(a) Reddit"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.305,
+                0.224,
+                0.482,
+                0.36
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.366,
+                0.361,
+                0.422,
+                0.372
+            ],
+            "angle": 0,
+            "content": "(b) MSC"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.115,
+                0.377,
+                0.49,
+                0.425
+            ],
+            "angle": 0,
+            "content": "Figure 5: Comparison of the performance across homogeneous LoRA, heterogeneous LoRA, and full fine-tuning. Heterogeneous LoRA achieves better performance than homogeneous LoRA with fewer number of communication rounds."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.444,
+                0.489,
+                0.604
+            ],
+            "angle": 0,
+            "content": "as severe for heterogeneous LoRA even with much larger ranks such as \\( r_{\\mathrm{max}} = 50 \\). We can imply from this result that the smaller rank LoRA modules act as a regularizer in heterogeneous LoRA preventing overfitting. Next, we show that by adding the self rank-pruning and sparsity-weighted aggregation, we improve the performance such that even with \\( r_{\\mathrm{min}} = 5 \\) we are able to prevent overfitting issues and achieve better training speed and final performance than other baselines."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.607,
+                0.49,
+                0.816
+            ],
+            "angle": 0,
+            "content": "Heterogeneous LoRA compared to PEFT. Finally, we compare HETLORA with other baselines in Table 3 and Fig. 5. We see that HETLORA with \\( r_{\\mathrm{min}} = 5 \\) and \\( r_{\\mathrm{max}} = 50 \\) achieves faster training as well as better performance than homogeneous LoRA cases with both edge cases of the ranks \\( r \\in \\{5, 50\\} \\) and reconstructions + SVD which was explained in Section 3.4. This implies that HETLORA is not only practical in the sense that clients are allowed to have their own rank values, it can also outperform the limited case of homogeneous LoRA where all clients have \\( r = r_{\\mathrm{min}} \\) or the impractical case where all clients have \\( r = r_{\\mathrm{max}} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.819,
+                0.49,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Heterogeneous LoRA compared to Full-Fine Tuning. In Table 3, we can see that for the Reddit text summarization, HETLORA achieves similar performance with full-fine tuning with only"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.518,
+                0.228,
+                0.875,
+                0.352
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.355,
+                0.882,
+                0.427
+            ],
+            "angle": 0,
+            "content": "Figure 6: Ratio of communicated number of parameters for different PEFT methods to full fine-tuning to achieve the target value for the metric where it is RougeL 80 for Reddit text summarization task and perplexity 150 for the multisession chat response task. The 'X' means that the target metric is not achieved even after convergence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.439,
+                0.883,
+                0.664
+            ],
+            "angle": 0,
+            "content": "\\(0.003\\%\\) of parameters trained compared to full-fine tuning. For chat response, HETLORA achieves slightly lower performance than full fine-tuning. However, full fine-tuning requires to train a much larger number of parameters compared to HETLORA, making it infeasible to train with ODFMs in practice. We also show in Fig. 6 that to achieve the targeted performance for both Reddit and MSC task, HETLORA requires significantly less number of parameters to be trained and communicated compared to full fine-tuning. Although for Reddit, HOMLORA has a slightly less number of parameters to be trained, the final achieved RougeL is outperformed by HETLORA as shown in Table 3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.69,
+                0.882,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Effect of the Decaying Factor \\(\\gamma\\). Lastly, we conduct an ablation study on the effect of the decaying factor \\(\\gamma\\) of HETLORA's local training step with self-rank pruning in Table 4. We observed that aggressive pruning hurts the performance where \\(\\gamma = 0.85\\) shows the worse performance across the varying \\(\\gamma\\) values. On the other hand, no pruning at all (\\(\\gamma = 1\\)) underperforms the case when there is pruning (\\(\\gamma = 0.99\\)), showing that reducing the noise introduced by large rank clients which data complexity is actually not that high indeed improves the performance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.505,
+                0.915
+            ],
+            "angle": 0,
+            "content": "8"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12910"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.086,
+                0.476,
+                0.101
+            ],
+            "angle": 0,
+            "content": "5 Discussions and Concluding Remarks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.111,
+                0.49,
+                0.384
+            ],
+            "angle": 0,
+            "content": "In our work, we investigated federated fine-tuning for ODFMs by proposing HETLORA that enables utilizing the power of FMs at the edge by taking device system and data heterogeneity into consideration. HETLORA is not only practical but also achieves better training speed, communication/computation efficiency, and final performance compared to other relevant baselines such as SVD-based LoRA, homogeneous LoRA, or full finetuning. Our findings in this work opens up several questions worth investigating. For instance, if the settings allow us to assign specific ranks to clients what will be the effective way to assign the ranks across clients for better convergence and performance? Another important next step of our work includes pursuing the theoretical convergence and generalization of heterogeneous LoRA."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.396,
+                0.251,
+                0.411
+            ],
+            "angle": 0,
+            "content": "6 Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.421,
+                0.49,
+                0.743
+            ],
+            "angle": 0,
+            "content": "In this work, we address tackling system and data heterogeneity in federated fine-tuning of on-device foundation models. Our work is motivated by clients being able to carry different ranks for the LoRA fine-tuning method depending on their available resources, and thus exploiting this characteristic to improve federated fine-tuning with heterogeneous LoRA. However, our work assumes that the rank distribution across clients (which is analogous to how system resources are distributed across clients) is independent to the data distribution. There can be scenarios in which this is not necessarily the case where the rank and data distribution can be correlated. For instance, more affluent populations can have better off devices with larger resource capacity, and may have data distributions different to that of less affluent populations. Such correlation should be explored for future work to better understand the implications of heterogenous LoRA."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.117,
+                0.77,
+                0.215,
+                0.784
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.792,
+                0.489,
+                0.832
+            ],
+            "angle": 0,
+            "content": "Anonymous. 2024. Flora: Federated fine-tuning large language models with heterogeneous low-rank adaptations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.842,
+                0.49,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Sara Babakniya, Ahmed Roushdy Elkordy, Yahya H. Ezzeldin, Qingfeng Liu, Kee-Bong Song, Mostafa El-Khamy, and Salman Avestimehr. 2023. Slora:"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.118,
+                0.792,
+                0.49,
+                0.882
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.53,
+                0.087,
+                0.882,
+                0.114
+            ],
+            "angle": 0,
+            "content": "Federated parameter efficient fine-tuning of language models. CoRR, abs/2308.06522."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.125,
+                0.882,
+                0.178
+            ],
+            "angle": 0,
+            "content": "Jiamu Bai, Daoyuan Chen, Bingchen Qian, Liuyi Yao, and Yaliang Li. 2024. Federated fine-tuning of large language models under heterogeneous tasks and client resources. arXiv preprint arXiv:2402.11505."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.189,
+                0.882,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Rishi Bommasani, Drew A. Hudson, and et. al. Ehsan Adeli. 2022. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.252,
+                0.882,
+                0.331
+            ],
+            "angle": 0,
+            "content": "Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. 2016. Practical secure aggregation for federated learning on user-held data. In NIPS Workshop on Private Multi-Party Machine Learning."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.342,
+                0.882,
+                0.499
+            ],
+            "angle": 0,
+            "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.51,
+                0.881,
+                0.561
+            ],
+            "angle": 0,
+            "content": "Zhiqi Bu, Yu-Xiang Wang, Sheng Zha, and George Karypis. 2022. Differentially private bias-term only fine-tuning of foundation models. arXiv preprint arXiv:2210.00036."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.573,
+                0.881,
+                0.626
+            ],
+            "angle": 0,
+            "content": "Jinyu Chen, Wenchao Xu, Song Guo, Junxiao Wang, Jie Zhang, and Haozhao Wang. 2022. Fevtune: A deep dive into efficient federated fine-tuning with pre-trained transformers. CoRR, abs/2211.08025."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.637,
+                0.881,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.7,
+                0.882,
+                0.831
+            ],
+            "angle": 0,
+            "content": "Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. 2023. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.842,
+                0.881,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Gemini Team Google. 2024. Gemini: A family of highly capable multimodal models. https:// arxiv.org/abs/2312.11805."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.087,
+                0.882,
+                0.882
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.493,
+                0.904,
+                0.505,
+                0.915
+            ],
+            "angle": 0,
+            "content": "9"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.524,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12911"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.087,
+                0.486,
+                0.115
+            ],
+            "angle": 0,
+            "content": "Google. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.127,
+                0.487,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Google. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.1040."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.166,
+                0.488,
+                0.207
+            ],
+            "angle": 0,
+            "content": "Google DeepMind. 2023. Introducing palm2. https://blog.google/technology/ai/ google-palm-2-ai-large-language-model/."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.218,
+                0.488,
+                0.297
+            ],
+            "angle": 0,
+            "content": "Tao Guo, Song Guo, and Junxiao Wang. 2023. Pfed-prompt: Learning personalized prompt for vision-language models in federated learning. In Proceedings of the ACM Web Conference 2023, WWW '23, page 1364-1374, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.309,
+                0.488,
+                0.374
+            ],
+            "angle": 0,
+            "content": "Tao Guo, Song Guo, Junxiao Wang, and Wenchao Xu. 2022. Promptfl: Let federated participants cooperatively learn prompts instead of models — federated learning in age of foundation model. CoRR, abs/2208.11625."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.387,
+                0.488,
+                0.427
+            ],
+            "angle": 0,
+            "content": "Venkatesan Guruswami and Ravi Kannan. 2012. Lecture notes in computer science theory for the information age."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.439,
+                0.488,
+                0.505
+            ],
+            "angle": 0,
+            "content": "Samuel Horváth, Stefanos Laskaridis, Shashank Rajput, and Hongyi Wang. 2024. Maestro: Uncovering low-rank structures via trainable decomposition. In Proceedings of the International Conference on Machine Learning (ICML)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.517,
+                0.488,
+                0.596
+            ],
+            "angle": 0,
+            "content": "Neil Houlsby, Andrei Giurgiu, Stanisław Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In Proceedings of the International Conference on Machine Learning (ICML)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.608,
+                0.488,
+                0.674
+            ],
+            "angle": 0,
+            "content": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.686,
+                0.488,
+                0.752
+            ],
+            "angle": 0,
+            "content": "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In The 36th Conference on Neural Information Processing Systems (NeurIPS 2022)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.764,
+                0.488,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Jaejun Lee, Raphael Tang, and Jimmy Lin. 2019. What would elsa do? freezing layers during transformer fine-tuning. arXiv preprint arXiv:1911.03090."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.816,
+                0.488,
+                0.881
+            ],
+            "angle": 0,
+            "content": "Gwen Legate, Nicolas Bernier, Lucas Caccia, Edouard Oyallon, and Eugene Belilovsky. 2023. Guiding the last layer in federated learning with pre-trained models. In Workshop of Federated Learning and Analytics in Practice@ICML."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.118,
+                0.087,
+                0.488,
+                0.881
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.087,
+                0.882,
+                0.14
+            ],
+            "angle": 0,
+            "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing (EMNLP)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.149,
+                0.882,
+                0.189
+            ],
+            "angle": 0,
+            "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.198,
+                0.882,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.259,
+                0.882,
+                0.312
+            ],
+            "angle": 0,
+            "content": "Zequan Liu, Jiawen Lyn, Wei Zhu, Xing Tian, and Yvette Graham. 2024. Alora: Allocating low-rank adaptation for fine-tuning large language models. arXiv preprint arXiv:2403.16187."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.321,
+                0.882,
+                0.4
+            ],
+            "angle": 0,
+            "content": "Andre Manoel, Mirian del Carmen Hipolito Garcia, Tal Baumel, Shize Su, Jialei Chen, Robert Sim, Dan Miller, Danny Karmon, and Dimitrios Dimitriadis. 2023. Federated multilingual models for medical transcript analysis. In Conference on Health, Inference, and Learning (CHIL), pages 147-162."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.408,
+                0.882,
+                0.487
+            ],
+            "angle": 0,
+            "content": "H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agoura y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. International Conference on Artificial Intelligence and Statistics (AISTATS)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.496,
+                0.882,
+                0.522
+            ],
+            "angle": 0,
+            "content": "OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:submit/4812508."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.531,
+                0.882,
+                0.622
+            ],
+            "angle": 0,
+            "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.632,
+                0.882,
+                0.685
+            ],
+            "angle": 0,
+            "content": "Anit Kumar Sahu, Tian Li, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Virginia Smith. 2020. Federated optimization for heterogeneous networks. In Proceedings of the 3rd MLSys Conference."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.693,
+                0.882,
+                0.746
+            ],
+            "angle": 0,
+            "content": "Joao Sedoc, Daphne Ippolito, Arun Kirubarajan, Jai Thirani, Lyle Ungar, and Chris Callison-Burch. 2019. Chateval: A tool for chatbot evaluation. Proceedings of NAACL-HLT."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.754,
+                0.882,
+                0.794
+            ],
+            "angle": 0,
+            "content": "Ofir Ben Shoham and Nadav Rappoport. 2023. Federated learning of medical concepts embedding using behrt. arXiv preprint arXiv:2305.13052."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.803,
+                0.882,
+                0.881
+            ],
+            "angle": 0,
+            "content": "Aliaksandra Shysheya, John F Bronskill, Massimiliano Patacchiola, Sebastian Nowozin, and Richard E Turner. 2023. Fit: Parameter efficient few-shot transfer learning for personalized and federated image classification. International Conference on Learning Representations (ICLR)."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.087,
+                0.882,
+                0.881
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.904,
+                0.51,
+                0.915
+            ],
+            "angle": 0,
+            "content": "10"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.525,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12912"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.087,
+                0.49,
+                0.141
+            ],
+            "angle": 0,
+            "content": "Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. 2024. Improving lora in privacy-preserving federated learning. In International Conference on Learning Representations (ICLR)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.153,
+                0.49,
+                0.231
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.244,
+                0.49,
+                0.323
+            ],
+            "angle": 0,
+            "content": "Michael Volske, Martin Potthast, Shahbaz Syed, and Benno Stein. 2017. TL;DR: Mining Reddit to learn automatic summarization. In Proceedings of the Workshop on New Frontiers in Summarization, pages 59-63, Copenhagen, Denmark. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.335,
+                0.49,
+                0.413
+            ],
+            "angle": 0,
+            "content": "Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K. Leung, Christian Makaya, Ting He, and Kevin Chan. 2019. Adaptive federated learning in resource constrained edge computing systems. IEEE Journal on Selected Areas in Communications, 37(6):1205 - 1221."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.426,
+                0.49,
+                0.491
+            ],
+            "angle": 0,
+            "content": "Colin Wei, Sang Michael Xie, and Tengyu Ma. 2021. Why do pretrained language models help in downstream tasks? an analysis of head and prompt tuning. Advances in Neural Information Processing Systems, 34:16158-16170."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.504,
+                0.488,
+                0.569
+            ],
+            "angle": 0,
+            "content": "Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S. Morcos. 2023. Io-fi: distributed fine-tuning without communication. Transactions on Machine Learning Research (TMLR)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.582,
+                0.49,
+                0.635
+            ],
+            "angle": 0,
+            "content": "Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. 2024. Fedlora: When personalized federated learning meets low-rank adaptation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.647,
+                0.488,
+                0.687
+            ],
+            "angle": 0,
+            "content": "Jing Xu, Arthur Szlam, and Jason Weston. 2021. Beyond goldfish memory: Long-term open-domain conversation. arXiv preprint arXiv:2107.07567."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.699,
+                0.488,
+                0.751
+            ],
+            "angle": 0,
+            "content": "Liping Yi, Han Yu, Gang Wang, and Xiaoguang Liu. 2023. Fedlora: Model-heterogeneous personalized federated learning with lora tuning. arXiv preprint arXiv:2310.13283."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.764,
+                0.488,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Sixing Yu, J. Pablo Muñoz, and Ali Jannesari. 2023. Federated foundation models: Privacy-preserving and collaborative learning for large models. arXiv preprint arXiv:2305.11414."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.119,
+                0.829,
+                0.49,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. 2021. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.119,
+                0.087,
+                0.49,
+                0.882
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.087,
+                0.882,
+                0.141
+            ],
+            "angle": 0,
+            "content": "Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, and Guoyin Wangand Yiran Chen. 2023a. Towards building the federated gpt: Federated instruction tuning. CoRR, abs/2305.05644."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.15,
+                0.882,
+                0.229
+            ],
+            "angle": 0,
+            "content": "Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023b. Adaptive budget allocation for parameter-efficient fine-tuning. In The 11th International Conference on Learning Representations (ICLR)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.239,
+                0.882,
+                0.318
+            ],
+            "angle": 0,
+            "content": "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.328,
+                0.882,
+                0.407
+            ],
+            "angle": 0,
+            "content": "Zhuo Zhang, Xiangjing Hu, Jingyuan Zhang, Yating Zhang, Hui Wang, Lizhen Qu, and Zenglin Xu. 2023c. Fedlegal: The first real-world federated learning benchmark for legal nlp. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.417,
+                0.882,
+                0.496
+            ],
+            "angle": 0,
+            "content": "Zhuo Zhang, Yuanhang Yang, Yong Dai, Qifan Wang, Yue Yu, Lizhen Qu, and Zenglin Xu. 2023d. Fedpetuning: When federated learning meets the parameter-efficient tuning methods of pre-trained language models. Findings of the Association for Computational Linguistics (ACL)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.505,
+                0.882,
+                0.598
+            ],
+            "angle": 0,
+            "content": "Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, Hao Peng, Jianxin Li, Jia Wu, Ziwei Liu, Pengtao Xie, Caiming Xiong, Jian Pei, Philip S. Yu, and Lichao Sun. 2023. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.087,
+                0.882,
+                0.598
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.904,
+                0.508,
+                0.915
+            ],
+            "angle": 0,
+            "content": "11"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.478,
+                0.928,
+                0.525,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12913"
+        }
+    ]
+]

2024/Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models/d6c62a2e-d2cd-45b7-a876-95885365f893_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a29771f2908f6e844b22a6fa5865dd64cda8d51e2d3c1363783cb619d050b178
+size 1481819

2024/Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models/full.md

@@ -0,0 +1,240 @@
+# Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models
+
+Yae Jee Cho $^{1,2*}$ , Luyang Liu $^{2}$ , Zheng Xu $^{2}$ , Aldi Fahrezi $^{2}$ , Gauri Joshi $^{1}$
+
+$^{1}$ Carnegie Mellon University,  $^{2}$ Google Research
+
+{yaejeeecho, luyangliu, xuzheng, aldifahrezi}@google.com,
+
+gaurij@andrew.cmu.edu
+
+# Abstract
+
+Foundation models (FMs) adapt surprisingly well to downstream tasks with fine-tuning. However, their colossal parameter space prohibits their training on resource-constrained edge-devices. For federated fine-tuning, we need to consider the smaller FMs of few billion parameters at most, namely on-device FMs (ODFMs), which can be deployed on device. Federated fine-tuning of ODFMs has unique challenges non-present in standard fine-tuning: i) ODFMs poorly generalize to downstream tasks due to their limited sizes making proper fine-tuning imperative to their performance, and ii) devices have limited and heterogeneous system capabilities and data that can deter the performance of fine-tuning.
+
+Tackling these challenges, we propose HETLORA, a feasible and effective federated fine-tuning method for ODFMs that leverages the system and data heterogeneity at the edge. HETLORA allows heterogeneous LoRA ranks across clients for their individual system resources, and efficiently aggregates and distributes these LoRA modules in a data-aware manner by applying rank self-pruning locally and sparsity-weighted aggregation at the server. It combines the advantages of high and low-rank LoRAs, achieving improved convergence speed and final performance compared to homogeneous LoRA. Furthermore, HETLORA has enhanced computation and communication efficiency compared to full fine-tuning making it more feasible for the edge.
+
+# 1 Introduction
+
+The emerging foundation models (FMs), easily ranging from few to hundreds of billions of parameters (Gemini Team Google, 2024; Bommasani
+
+et al., 2022; Radford et al., 2021; Devlin et al., 2019; OpenAI, 2023; Google, 2022; Touvron et al., 2023; Brown et al., 2020; Google, 2022; Driess et al., 2023), have remarkable zero/few shot learning capabilities. They often perform surprisingly well on a variety of tasks such as text/image generation, language translation, and conversing in natural language with fine-tuning. Such standard FMs, however, demand costly resources for directly fine-tuning their entire parameter space. To tackle this issue, many works have proposed different parameter-efficient fine-tuning (PEFT) methods such as prompt tuning (Lester et al., 2021), utilizing adapters (Houlsby et al., 2019), or low-rank adaptation (LoRA) of the original model (Hu et al., 2021) which freezes the original pre-trained parameters of the FM and train additional, smaller number of parameters instead.
+
+Unfortunately, such PEFT methods rely on the assumptions that i) FMs are deployed to and trained with the data of a single client for adaptation to the downstream task and ii) the client has enough resources to even fit a standard FM of hundred billion size for, at least, inference. In practice, more often than not we are interested in fine-tuning FMs for on-device private data that is distributed across multiple devices, and these clients at the edge rarely have such resources to even fit standard FMs for inference. One of the examples is to use FMs for downstream tasks that requires training with private, non-sharable data such as medical or law-related documents distributed across clients (Manoel et al., 2023; Shoham and Rappoport, 2023; Zhang et al., 2023c). In such cases, fine-tuning of the FMs needs to be brought to the edge that has limited system resources.
+
+Federated Fine-Tuning & ODFMs. We consider such federated fine-tuning scenarios, where
+
+<table><tr><td></td><td>Zero-Shot</td><td>Few-Shot</td><td>Full-Training</td></tr><tr><td>PaLM 2 XXS</td><td>2930.23</td><td>2541.86</td><td>23.71</td></tr><tr><td>PaLM 2 XS</td><td>2712.86</td><td>481.95</td><td>18.32</td></tr></table>
+
+Table 1: Perplexity of PaLM 2 for zero-shot, few-shot (5 communication rounds), and full federated fine-tuning (200 communication rounds) for chat response on the multi-session chat data (further experimental details are in Section 4.)
+
+we train a set of parameters collaboratively across clients to obtain a global set of parameters that can be plugged in to the FM for the targeted downstream task. Note that federated fine-tuning is orthogonal to personalization of FMs in federated learning (FL) that has been explored in other previous work (Guo et al., 2023; Wu et al., 2024), which aims to train parameters that perform well for individual clients rather than general downstream tasks. We also define on-device FMs (ODFMs) as models with few billion parameters at max that are able to fit into memory on limited capacity clients considering current hardwares.
+
+Challenges. Federated fine-tuning of ODFMs entails two major challenges non-present in neither standard PEFT of FMs nor federated training of non-FMs. First, ODFMs poorly generalize to downstream tasks. Unlike ODFMs, standard FMs have their zero/few-shot learning capability by their large parameter space trained on massive data. However, as we show in Table 1 and as previous literature has shown (Kojima et al., 2022; Lester et al., 2021), FMs' performance deteriorates as their sizes get smaller and federated fine-tuning is not merely useful but inevitable for ODFMs to perform well for on-device downstream tasks.
+
+Second, devices have limited and heterogeneous system capabilities and data distributions (Wang et al., 2019; Bonawitz et al., 2016; Sahu et al., 2020). Without suitable PEFT methods that flexibly adapts to such heterogeneity across devices, we would only have limited performance output from federated fine-tuning of ODFMs. For instance, with homogeneous rank deployment in LoRA that is agnostic of system heterogeneity, the client with the least system resource becomes the bottleneck, forcing a smaller rank to be deployed across all clients despite the other clients being able to handle higher ranks (see Table 2 for the effect of rank on the trainable parameter #). Additionally, data heterogeneity causing model drifts can make the model converge to a suboptimal point.
+
+![](images/6043e1749a942f6b874f2481629c14f993b283566e481d96f024eb065e1c0511.jpg)  
+Figure 1: Overview of heterogeneous rank deployment of LoRA: the pretrained weights  $\mathbf{W}_0$  are stored on-device and heterogeneous ranks are assigned to different clients with  $r_{\mathrm{min}} = r_1 < r_2 < r_3 = r_{\mathrm{max}}$ . In our proposed HETLORA, the server receives the trained heterogeneous LoRA modules and aggregates them with sparsity-weighted aggregation to update the global LoRA module.
+
+Contributions. In our work, we push the limits of federated fine-tuning of ODFMs by proposing HETLORA that is aware of heterogeneity residing at the edge. Our main flagship results (in Table 3 and Fig. 5), show that our proposed HETLORA outperforms the standard LoRA method and achieves comparable performance with full fine-tuning with a significantly smaller number of trained parameters. We verify HETLORA with PaLM 2 (Google, 2023) of XXS and XS size for chat responses on the multi-session chat data (Xu et al., 2021) and text summarization for the Reddit data (Volske et al., 2017), both real world data from clients. Our contributions are summarized as:
+
+- We propose HETLORA that can apply different rank LoRA modules to different clients to cater to the heterogeneous system capabilities and data complexities of the clients, via utilizing rank self-pruning and sparsity-weighted aggregation.  
+- We show the performance of naively applying LoRA with homogeneous ranks across clients for federated fine-tuning, and show that while large ranks help in speeding-up training, they lead to faster overfitting while smaller ranks are slower in training but does not suffer from overfitting.  
+- We then evaluate our proposed HETLORA to show that it outperforms naively applying homogeneous ranks across clients in terms of both training speed, communication/computation efficiency, and final performance.
+
+<table><tr><td></td><td>r = 1</td><td>r = 5</td><td>r = 10</td><td>r = 20</td><td>r = 50</td><td>r = 100</td><td>r = 150</td><td>r = 200</td></tr><tr><td>PaLM 2 XXS, PaLM 2 XS</td><td>0.02%</td><td>0.11%</td><td>0.21%</td><td>0.42%</td><td>1.05%</td><td>2.10%</td><td>3.14%</td><td>4.19%</td></tr></table>
+
+Table 2: Percentage of the LoRA parameters' size for different ranks  $r$  compared to the original pre-trained ODFM's parameter size. Even for large ranks such as  $r = {200}$  the trainable LoRA parameters' size compared to the original pre-trained ODFM size is less than 5% for both PaLM 2-XS and PaLM 2-XXS.
+
+# 2 Related Work
+
+Parameter-Efficient Fine Tuning. Most PEFT methods either train a subset of parameters within the existing FM whilst other parameters are frozen or introduce an additional set of trainable parameters whilst keeping the original FM frozen. For the former, methods such as head or bias fine-tuning (Wei et al., 2021; Bu et al., 2022; Lee et al., 2019; Zaken et al., 2021) has been explored, and for the latter, methods such as adapters (Houlsby et al., 2019), prompt (Lester et al., 2021) or prefix-tuning (Li and Liang, 2021), and low-rank approximation (Hu et al., 2021) has been proposed. While these number of methods has been proven to perform as well as full model fine-tuning with just a few number of parameters for the centralized setting, it has not been thoroughly explored how these methods would perform for a much smaller FM such as ODFMs, in the decentralized setting where clients' system-capacities and data can be heterogeneous and much limited.
+
+Federated Fine-Tuning. Several recent work has proposed to combine the PEFT methods devised for the centralized setting to FL (Zhou et al., 2023; Yu et al., 2023) such as training prompts or adapters collaboratively with FL (Guo et al., 2022; Chen et al., 2022; Zhang et al., 2023a; Shysheya et al., 2023; Legate et al., 2023). Another line of work has proposed to perform a few-shot or nearly zero-shot training of FMs with FL for improved communication-efficiency (Wortsman et al., 2023; Zhang et al., 2023d). However, these work either overlooks that most devices do not have the resource to fit standard FMs (Touvron et al., 2023; Brown et al., 2020) even for inference or does not consider the heterogeneous system capacities of the clients. It is detrimental to consider these factors since FMs that actually fits to the devices in FL are much smaller, making them weaker in the general intelligence capabilities, and also heterogeneous system capacities can prohibit deploying same sized PEFT parameters across clients.
+
+LoRA and FL. There has been a number of variations of LoRA (Zhang et al., 2023b; Liu et al., 2024; Sun et al., 2024; Horváth et al., 2024) from its first proposal (Hu et al., 2021). However, only a few numbers of recent work has looked in to using LoRA for FL. In Babakniya et al. (2023), the importance of the initialization for the LoRA modules is evaluated where they propose to train the LoRA modules with FL and then perform singular value decomposition (SVD) to gain a good initialization of the LoRA modules. However, the training process of LoRA itself is not altered to adapt to heterogeneous system capabilities of devices. Another recent work (Yi et al., 2023) has evaluated LoRA in the context of personalized FL, but other than applying LoRA to personalization, the LoRA method itself is, again, not changed. To the best of our knowledge, our work is the first method to consider both data and system heterogeneity for improving training speed, communication/computation efficiency, and final performance of ODFM's federated fine-tuning. Contemporary approaches (Bai et al., 2024; Anonymous, 2024) have proposed to use SVD to decompose the full matrix to deploy heterogeneous ranks across clients. We compare our proposed method with using SVD for heterogeneous LoRA in our work and show that the performance is underwhelming compared to our proposed heterogeneous LoRA.
+
+# 3 Federated Fine-Tuning with LoRA
+
+# 3.1 Preliminaries
+
+Formally, we define the pre-trained ODFM as  $\mathbf{W}_0\in \mathbb{R}^{d\times l}$  and the trainable low-rank decomposed matrix as  $\Delta \mathbf{W}\in \mathbb{R}^{d\times l}$ . In standard LoRA (Hu et al., 2021) under the centralized setting, the low-rank decomposition of  $\Delta \mathbf{W}$  is constructed such that  $\Delta \mathbf{W} = \mathbf{BA}$  where  $\mathbf{B}\in \mathbb{R}^{d\times r}$  and  $\mathbf{A}\in \mathbb{R}^{r\times l}$  are the low rank decomposition of  $\Delta \mathbf{W}$  with identical rank  $r$ . Now, let us consider LoRA for federated fine-tuning where there are  $M$  total clients. Each client  $k\in [M]$  has pri
+
+vate data  $\mathcal{B}_k$  and its corresponding local empirical loss function  $F_{k}(\mathbf{W}) = \frac{1}{|\mathcal{B}_{k}|}\sum_{\xi \in \mathcal{B}_{k}}\ell (\mathbf{W},\xi)$  where  $\ell (\mathbf{W},\xi)$  is the loss for model  $\mathbf{W}$  at data sample  $\xi$ . The optimization task for federated fine-tuning is to collaboratively find the global parameters which we define as  $\overline{\mathbf{B}}$  and  $\overline{\mathbf{A}}$ , given the pretrained knowledge  $\mathbf{W}_0$  that can minimize the global objective  $F(\overline{\mathbf{W}}) = \frac{1}{M}\sum_{k = 1}^{M}F_{k}(\overline{\mathbf{W}})$  where  $\overline{\mathbf{W}} = \mathbf{W}_0 + \overline{\mathbf{B}}\overline{\mathbf{A}}$ . Later in the paper, when introducing heterogeneous LoRA we truncate the LoRA modules' rank dimension, for example from  $\mathbf{B}\in \mathbb{R}^{d\times r}$ ,  $\mathbf{A}\in \mathbb{R}^{r\times l}$  to  $\mathbf{B}'\in \mathbb{R}^{d\times r'}$ ,  $\mathbf{A}'\in \mathbb{R}^{r'\times l}$  where  $r' < r$ . Throughout the paper, we denote such truncation of a matrix with the : symbol for each row and column at the subscript. For instance, for truncation to  $r' < r$  at the column for the matrix  $\mathbf{B}\in \mathbb{R}^{d\times r}$ , we keep all the columns until  $r'$  and omit the last  $r - r'$  columns and denote the resulting matrix it as  $\mathbf{B}_{:,;r'}$ .
+
+# 3.2 Naive Case: Homogeneous LoRA
+
+A straightforward way to perform federated fine-tuning with LoRA is to train the LoRA modules  $\mathbf{B}$ ,  $\mathbf{A}$  with homogeneous rank  $r$  across all clients with standard FL (McMahan et al., 2017). Specifically, first the clients have the pre-trained ODFM weights  $\mathbf{W}_0$  stored in their devices prior to training for the forward pass when training the LoRA modules. Then, the server sends the global LoRA modules  $\overline{\mathbf{B}}^{(t)}$ ,  $\overline{\mathbf{A}}^{(t)}$  to the set of  $m$  selected clients  $S^{(t)}$  per communication round  $t$ . Each selected client  $k \in S^{(t)}$  trains the LoRA modules on their local data for a few local iterations (usually with mini-batch SGD) and send the updated modules  $\mathbf{B}_k^{(t)}$ ,  $\mathbf{A}_k^{(t)}$  back to the server. The server then updates the global LoRA modules accordingly to  $\overline{\mathbf{B}}^{(t+1)} = \sum_{k \in S^{(t)}} \mathbf{B}_k^{(t)} / m$ ,  $\overline{\mathbf{A}}^{(t+1)} = \sum_{k \in S^{(t)}} \mathbf{A}_k^{(t)} / m$  and sends back to the next set of selected clients for the next communication round. This training process is nearly identical to the standard FL algorithm (McMahan et al., 2017) except that the pretrained weights  $\mathbf{W}_0$  are frozen and locally stored in the clients' devices and only the LoRA moduels are trained and communicated.
+
+Instead of such homogeneous rank deployment across all clients, it is not only possible but even more practical and feasible to use heterogeneous rank deployment for federated fine-tuning (we later
+
+show that this also improves final performance of the ODFMs). This involves training LoRA modules with varying ranks across clients, based on their system capabilities. This allows a more practical setting for federated fine-tuning of ODFMs where, for instance, the client with the smallest system capacity does not have to become the bottleneck that forces deploying a smaller rank across all clients. Instead, with heterogeneous rank deployment, each client can use a rank that suits for itself. However, heterogeneous rank deployment poses challenges in aggregating and redistributing the LoRA modules. To address these challenges, we introduce a solution called HETLORA, which pushes the limits beyond homogeneous LoRA.
+
+# 3.3 Proposed Method: Heterogeneous LoRA
+
+Overview. Our proposed heterogeneous LoRA method, namely HETLORA, is not restricted to any specific method to assign the ranks to the clients and the clients can decide their respective ranks themselves. For formality, in our paper, we formulate that each client has a rank denoted as  $r_k$ , within a range of  $r_k \in [r_{\min}, r_{\max}]$ ,  $\forall k$  (see Fig. 1). HETLORA comprises three steps: 1) Distribution via Truncation, 2) Local Training with Rank Self-Pruning, and 3) Sparsity-Weighted Aggregation of the LoRA modules. These steps are detailed further in the subsequent paragraphs. An overview of HETLORA is illustrated in Fig. 2.
+
+1) Distribution via Truncation. At the beginning of each communication round  $t$ , the server holds initial global LoRA modules  $\overline{\mathbf{B}}^{(t)}$ ,  $\overline{\mathbf{A}}^{(t)}$  with a global rank  $r^{(t)}$ . The value of the global rank  $r^{(t)}$  depends on how we aggregate the heterogeneous rank LoRA modules which is elaborated on in step 3). The server then distributes these global LoRA modules to a subset of selected set of clients  $S^{(t)}$  with heterogeneous ranks  $r_k^{(t)}, k \in S^{(t)}$  for local training<sup>1</sup>. With the given global LoRA modules, we consider a simple and intuitive method of truncation where the server sends  $\overline{\mathbf{B}}_{:,r_k}^{(t)}$ ,  $\overline{\mathbf{A}}_{:r_k}^{(t)}$  to each client  $k$  with rank  $r_k^{(t)}$  for local training where we omitted the superscript for  $r_k$  for simplicity.
+
+# 2) Local Training with Rank Self-Pruning.
+
+![](images/3619c21a64d5cfc62a80b2f2c4291cbf5c44111745f28ad471b9b3b507d522d2.jpg)  
+Figure 2: Overview of the zero-padding, sparsity-weighted aggregation, and truncation method for HETLORA; (a): Zero-pad LoRA modules with smaller ranks to  $r_{\mathrm{max}}$  (clients with rank  $r_{\mathrm{max}}$  does not need padding) and calculate their sparsity by calculating the Frobenius norm of the reconstructed model  $\Delta \mathbf{W}_k^{(t)} = \mathbf{B}_k^{(t)}\mathbf{A}_k^{(t)}$ ; (b): After padding, aggregate all of the clients' LoRA modules with the weights  $\| \mathbf{S}_k^{(t)}\| /\mathbf{Z}^{(t)}$  calculated by  $\Delta \mathbf{W}_k^{(t)}$  to get the global LoRA modules; (c): Truncate the global LoRA modules for the specific rank of the next selected client (example for client with rank  $r_2$ ).
+
+![](images/eec115469413780c73e24798e79974773601e7c715ead2a481e452de9f9a744c.jpg)
+
+![](images/9f3317bffb9483465ec5324ad4e14b4452aa339f8bc9d6dc2b2fb42ef6a68c96.jpg)
+
+After receiving LoRA modules from the server as  $\mathbf{B}_k^{(t,0)} = \overline{\mathbf{B}}_{\therefore r_k}^{(t)}$ ,  $\mathbf{A}_k^{(t,0)} = \overline{\mathbf{A}}_{r_k,\cdot}^{(t)}$ , each client  $k \in S^{(t)}$  performs  $\tau$  local iterations of mini-batch SGD on their local data to minimize the local objective  $\frac{1}{|\mathcal{B}_k|}\sum_{\xi \in \mathcal{B}_k}\ell((\mathbf{B}_k, \mathbf{A}_k), \xi|\mathbf{W}_0)$ , and sends back the updated LoRA modules  $\mathbf{B}_k^{(t,\tau)} \in \mathbb{R}^{d \times r_k^{(t)}}$  and  $\mathbf{A}_k^{(t,\tau)} \in \mathbb{R}^{r_k^{(t)} \times l}$  to the server. This is the same process as the standard local training step in vanilla FedAvg (McMahan et al., 2017). However, we improve this vanilla local training step by adding a rank self-pruning mechanism where clients self-prune their respective ranks depending on the magnitude of the model parameters.
+
+Specifically, we add a regularization term to the original local objective to get  $\min_{\mathbf{B}_k,\mathbf{A}_k}\frac{1}{|\mathcal{B}_k|}\sum_{\xi \in \mathcal{B}_k}\ell ((\mathbf{B}_k,\mathbf{A}_k),\xi |\mathbf{W}_0) + \lambda \| \mathbf{B}_{k,:r_k\gamma :r_k}\| \| \mathbf{A}_{k,r_k\gamma :r_k,:}\|$  where  $\gamma < 1$  is a decay-factor that determines how aggressively we want to prune the ranks to a smaller value. The regularization term aims to minimize the norm of the last few ranks, which will become smaller if the first loss term  $\frac{1}{|\mathcal{B}_k|}\sum_{\xi \in \mathcal{B}_k}\ell ((\mathbf{B}_k,\mathbf{A}_k),\xi |\mathbf{W}_0)$  is not very large. After training with the new local objective we compare the norm of the updated LoRA modules' last few layers  $\| \mathbf{B}_{k,:r_k\gamma :r_k}\| \| \mathbf{A}_{k,r_k\gamma :r_k,:}\|$  with the ones from the initially received LoRA modules. If the former is smaller we prune the last few layers (pruning intensity is determined by  $\gamma$ ) and send back the LoRA modules with a smaller rank. This means that for the LoRA modules which incurs a small local loss, i.e., well-trained on the clients' local data, the LoRA modules are more likely to be pruned to a smaller rank.
+
+Such pruning allows HETLORA to reduce the
+
+noise in the LoRA modules introduced by clients having a larger rank than the actual rank that their data complexity requires, and also reduces the complexity of the LoRA modules to improve generalization and prevent overfitting (see Table 4). Once the rank is pruned for a client, the client saves the updated rank and uses it as the starting rank if selected for future communication rounds. The client then sends back their updated and possibly rank-pruned LoRA modules to the server to be processed in the aggregation step.
+
+3) Sparsity-Weighted Aggregation. Finally, the last step of HETLORA is aggregating the received heterogeneous LoRA modules  $\mathbf{B}_k^{(t,\tau)}$ ,  $\mathbf{A}_k^{(t,\tau)}$ ,  $k \in S^{(t)}$ . A straightforward way to aggregate the heterogeneous LoRA modules is using zero-padding to all the received LoRA modules with  $r_i^{(t)} < \max \{r_k^{(t)} | k \in S^{(t)}\}$  and then perform simple averaging over the modules. However, such naive aggregation can lead to biasing the model towards higher rank clients even when these clients may not hold valuable training information, i.e., having low data complexity, giving noisy updates.
+
+In an ideal scenario where we can deploy any rank to any client, deploying higher ranks to the clients with higher data complexity or larger local datasets can retrieve more informative and less sparse updates from the clients. Conversely if we assign higher ranks to the clients whose data complexity is low, the actual rank of the full model from the reconstructed LoRA modules can be smaller than the assigned rank. Thus the higher rank client's update may be unnecessarily over-emphasized in the naive zero padding method.
+
+Based on this insight we propose a sparsity-
+
+weighted aggregation scheme where the server reconstructs these LoRA modules to the full model as  $\Delta \mathbf{W}_k^{(t)} = \mathbf{B}_k^{(t)}\mathbf{A}_k^{(t)}$  and gets the norm of the singular value vectors from the full models denoted as  $\mathbf{S}_k^{(t)}$  by calculating  $\| \Delta \mathbf{W}_k^{(t)}\| _F$ . Note that the costly process of performing SVD for each of the full model  $\Delta \mathbf{W}_k^{(t)}$  can be avoided by simply calculating the Frobenius norm of  $\Delta \mathbf{W}_k^{(t)}$  (see Lemma 1.2 in (Guruswami and Kannan, 2012)). The server then weighs the LoRA modules with aggregation weight  $p_k^{(t)}$  which is proportional to the norm of the singular value vectors. Formally, we have the global LoRA modules updated as  $\overline{\mathbf{B}}^{(t + 1)} = \sum_{k\in \mathcal{S}^{(t)}}p_k^{(t)}\mathbf{B}_k^{(t)}$ ,  $\overline{\mathbf{A}}^{(t + 1)} = \sum_{k\in \mathcal{S}^{(t)}}p_k^{(t)}\mathbf{A}_k^{(t)}$  where  $p_k^{(t)}\coloneqq \| \mathbf{S}_k^{(t)}\| /\mathbf{Z}^{(t)}$  with normalizing factor  $\mathbf{Z}^{(t)}\coloneqq \sum_{k^{\prime}\in \mathcal{S}^{(t)}}\| \mathbf{S}_{k^{\prime}}^{(t)}\|$ . This way, we can de-emphasize the larger rank assigned-clients that have rather less informative updates, and more emphasize the smaller rank assigned-clients that have more informative ones.
+
+# 3.4 Why not Simply Reconstruct First, then Redistribute the LoRA modules?
+
+One might ask why not simply reconstruct each of the LoRA modules to the full matrix and aggregate them. Here we show that reconstructing the LoRA modules and aggregating them to get the full model results in a different full model compared to when we aggregate the LoRA modules first and then reconstruct the final model. In Section 4 we also empirically show that reconstructing the LoRA modules to the full model and redistributing them after truncated SVD to the corresponding rank of the clients results in an overwhelming performance compared to HETLORA.
+
+Let us consider a simple case where there are 2 clients with heterogeneous rank lora modules  $\mathbf{B}_1\in$ $\mathbb{R}^{d\times 1}$ $\mathbf{A}_1\in \mathbb{R}^{1\times l}$  and  $\mathbf{B}_2\in \mathbb{R}^{d\times 1}$ $\mathbf{A}_2\in \mathbb{R}^{2\times l}$  respectively for client 1 and client 2 where the former has rank 1 and latter has rank 2. We set the notation for the LoRA modules'  $i^{th}$  row and  $j^{th}$  column value for  $\mathbf{B}_k$  and  $\mathbf{A}_k$  as  $b_{k,ij}$  and  $a_{k,ij}$  respectively. Then with  $d = 3$ $l = 2$  when we reconstruct each of the LoRA modules first and then aggregate the full model we have its  $i^{th}$  row and  $j^{th}$  column as  $(\sum_{k = 1}^{2}b_{k,i0}a_{k,0j}) + b_{2,i1}a_{2,1j}$  and aggregating the LoRA modules first and then reconstructing the model has the full model's  $i^{th}$  row and  $j^{th}$  column
+
+as  $(\sum_{k = 1}^{2}b_{k,i0})(\sum_{k = 1}^{2}a_{k,0j}) + b_{2,i1}a_{2,1j}$
+
+One can observe that the difference between the two models are the cross-terms between the left and right module of different client 1 and 2, i.e.,  $b_{1,i0}a_{2,0j} + b_{2,i0}a_{1,0j}$  for the  $i^{th}$  row and  $j^{th}$  column. In other words, when we reconstruct the LoRA modules first and then aggregate them to get the full model, each term in the full model are cross-products between the left and right module of each client and not the cross-products between clients. Thus, reconstructing the LoRA modules loses information on the cross-relation across clients, only retaining the knowledge on the cross-relation between the LoRA modules B and A. Such observation is also corroborated by the reconstruction first's underwhelming performance in Table 3.
+
+# 4 Experiments
+
+In this section, we show that our proposed HETLORA outperforms its baselines in terms of training speed, computation/communication efficiency, and final achieved performance. We first show the performance of homogeneous LoRA to show how LoRA in general performs for low and high rank values. Second, as our main result, we demonstrate HETLORA's performance for different  $r_{\mathrm{min}}$  and  $r_{\mathrm{max}}$  values comparing them with full fine-tuning, homogeneous LoRA, and the reconstruction-first method elaborated in Section 3.4. We also conduct an ablation study on HETLORA with varying decay factor  $\gamma$  for the rank self-pruning step.
+
+Settings. We use the transformer-based language model PaLM 2 (Google, 2023) of size XXS and XS for our experiments which are lightweight enough to be ODFMs (Google DeepMind, 2023) compared to standard FMs. The tasks we consider are the chat dialogue from the multi-session chat (MSC) dataset (Xu et al., 2021) and the text summarization task from the Reddit dataset (Volske et al., 2017). For the MSC data we use perplexity (Zhang et al., 2018) as the metric which has been used to show the quality of chat responses from generative models from previous literature (Sedoc et al., 2019). We sample 100 users uniformly at random and partition their data for training and evaluation by each previous_DIALOG and dialog. The Reddit text summarization data consists of real users' reddit posts and their summarization, and
+
+<table><tr><td></td><td colspan="2">Reddit (RougeL)</td><td colspan="2">Multi-Session Chat (Perplexity)</td></tr><tr><td></td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td><td>PaLM 2-XXS</td></tr><tr><td>Full</td><td>94.56(±0.01)</td><td>94.87(±0.04)</td><td>32.70(±0.17)</td><td>23.40(±0.36)</td></tr><tr><td>HOMLORA r = 5</td><td>92.57(±1.56), ×0.001</td><td>92.89(±0.96)</td><td>80.51(±8.32), ×0.001</td><td>64.59(±9.31)</td></tr><tr><td>HOMLORA r = 50</td><td>70.57(±2.13), ×0.01</td><td>84.95(±1.59)</td><td>307.96(±11.43), ×0.01</td><td>167.46(±1.72)</td></tr><tr><td>Recon+SVD</td><td>63.28(±1.92), ×0.003</td><td>75.17(±1.25)</td><td>323.89(±20.57), ×0.002</td><td>215.63(±15.38)</td></tr><tr><td>HETLORA γ = 0.99</td><td>94.23(±0.03), ×0.003</td><td>94.41(±0.05)</td><td>53.93(±1.57), ×0.002</td><td>38.76(±0.52)</td></tr></table>
+
+Table 3: Final RougeL score for Reddit text summarization and perplexity for multi-session chat for different federated fine-tuning methods. The blue text indicates the ratio of trained number of parameters compared to the full fine-tuning case. HETLORA outperforms both HOMLORA and Recon+SVD method, but slightly underperforms the full fine-tuning case. However, compared to full fine-tuning the number of trained parameter is significantly smaller.
+
+![](images/98ba5e70c8bd6c9388e771ace842c0e5979e77db93f6a9066241be7ac50dfa52.jpg)  
+(a) PaLM 2-XXS
+
+![](images/2775320785a46ad0dc6278f6e788f71268d7b64081e49bad49c85386f7814992.jpg)  
+(b) PaLM 2-XS  
+Figure 3: Performance of homogeneous LoRA for different rank  $r$ . Higher ranks achieve better performance with fewer communication rounds than the lower ranks, but they overfit quickly. Conversely, the lowest rank  $r = 1$  achieves low perplexity slower than higher ranks, but without overfitting.
+
+we use RougeL (Lin, 2004) as the metric. We use 298 users from Reddit that have at least 100 data samples as the training clients and use another 100 users with at least 100 data samples for evaluation. We use mini-batch size 8 and number of local iterations  $\tau = 5$  with the feature length set to 1024. For the learning rate we perform grid search in  $\eta = \{0.1, 0.01, 0.001, 0.0001\}$ . For each MSC and Reddit task, we select 5 and 10 clients per communication round respectively. The rank distribution across clients for HETLORA is set to a truncated power-law distribution with  $\alpha = 0.1$  in the range between  $[r_{\mathrm{min}}, r_{\mathrm{max}}]$  (inclusively). All experiments were ran with 3 different random seeds.
+
+# 4.1 Experiment Results
+
+Homogeneous LoRA and the Effect of Ranks  $r$ . First, we evaluate the performance of federated fine-tuning of the LoRA modules with homogeneous LoRA deployment across clients in Fig. 3 for different ranks  $r \in [1,5,20,50]$ . We observe that a higher rank  $r$  for homogeneous LoRA achieves better perplexity floor with fewer communication rounds than the lower ranks but quickly overfits resulting in worse performance compared to the lower ranks after more communication rounds. On
+
+![](images/a3cb5079742aae49ceb6d8c7f526b7e747082544584b284c7b39a528065412cb.jpg)  
+(a)  $r_{\mathrm{max}} = 10$  
+Figure 4: Performance of HETLORA without rank pruning or and with simple average aggregation. Similar to homogeneous LoRA, larger  $r_{\mathrm{min}}$  leads to overfitting for heterogeneous LoRA, but it is not as severe as homogeneous LoRA even for larger maximum rank  $r_{\mathrm{max}} = 50$  showing that the smaller rank LoRA modules act as a regularizer for HETLORA.
+
+![](images/b098577b62359dbdb5e064a3cf725cdf24bb2ffd6dce18f4ce9dfa8f32f1bde9.jpg)  
+(b)  $r_{\mathrm{max}} = 50$
+
+the other hand, while the lower rank cases need more communication rounds to achieve good performance, it does not have the problem of overfitting as the higher ranks. Hence for homogeneous LoRA, there is a trade-off between low and high ranks, in terms of faster performance achievement and overfitting. Note that these observations are consistent with previous literature in the centralized setting where a higher rank does not necessarily yields the best performance (Hu et al., 2021; Zhang et al., 2023b). Subsequently, we show that HETLoRA that is not only more practical than homogeneous LoRA, but also achieves a better final performance with fewer number of communication rounds without the overfitting issue.
+
+Naïve Heterogeneous LoRA and the Effect of  $r_{\mathrm{min}}$  and  $r_{\mathrm{max}}$ . First, we show the performance of naïve heterogeneous LoRA which is HETLORA without self rank-pruning and with only average aggregation instead of the sparsity-weighted aggregation in Fig. 4. We see that similar to homogeneous LoRA, a smaller minimum rank  $r_{\mathrm{min}} = 1$  leads to slower training but better performance while a larger maximum rank leads to faster training but worse performance. However, unlike homogeneous LoRA, the overfitting does not get
+
+<table><tr><td></td><td colspan="2">Reddit (RougeL)</td><td colspan="2">Multi-Session Chat (Perplexity)</td></tr><tr><td></td><td>PaLM 2-XXS</td><td>PaLM 2-XS</td><td>PaLM 2-XXS</td><td>PaLM 2-XS</td></tr><tr><td>HETLORA, γ = 1</td><td>92.17 (±0.08)</td><td>91.95 (±0.03)</td><td>55.07 (±0.81)</td><td>40.92 (±0.58)</td></tr><tr><td>HETLORA, γ = 0.99</td><td>94.23 (±0.03)</td><td>94.41 (±0.05)</td><td>53.93 (±1.57)</td><td>38.76 (±0.52)</td></tr><tr><td>HETLORA, γ = 0.95</td><td>89.62 (±1.33)</td><td>83.19 (±1.70)</td><td>71.10 (±1.39)</td><td>46.39 (±0.87)</td></tr><tr><td>HETLORA, γ = 0.85</td><td>60.31 (±3.04)</td><td>53.28 (±2.47)</td><td>120.72 (±10.93)</td><td>59.67 (±1.98)</td></tr></table>
+
+Table 4: Ablation study on the effect of the decaying factor  $\gamma$  for HETLORA's self-rank pruning in the local training step. While aggressive pruning can be harmful to HETLORA's performance, pruning  $\left( {\gamma  = {0.99}}\right)$  can outperform the case when there is no pruning at all  $\left( {\gamma  = 1}\right)$  by reducing the noise introduced by large rank clients with low data complexity.
+
+![](images/69924bb62be3af0dfffb065a30b8fad4fdf7651ecc2ee0087a20cb3bb327fdc4.jpg)  
+(a) Reddit
+
+![](images/0a31f19257ccca600cbe7281166c85e7b14f741f6275f32c05cbd4c51ab3afe2.jpg)  
+(b) MSC
+
+as severe for heterogeneous LoRA even with much larger ranks such as  $r_{\mathrm{max}} = 50$ . We can imply from this result that the smaller rank LoRA modules act as a regularizer in heterogeneous LoRA preventing overfitting. Next, we show that by adding the self rank-pruning and sparsity-weighted aggregation, we improve the performance such that even with  $r_{\mathrm{min}} = 5$  we are able to prevent overfitting issues and achieve better training speed and final performance than other baselines.
+
+Heterogeneous LoRA compared to PEFT. Finally, we compare HETLORA with other baselines in Table 3 and Fig. 5. We see that HETLORA with  $r_{\mathrm{min}} = 5$  and  $r_{\mathrm{max}} = 50$  achieves faster training as well as better performance than homogeneous LoRA cases with both edge cases of the ranks  $r \in \{5, 50\}$  and reconstructions + SVD which was explained in Section 3.4. This implies that HETLORA is not only practical in the sense that clients are allowed to have their own rank values, it can also outperform the limited case of homogeneous LoRA where all clients have  $r = r_{\mathrm{min}}$  or the impractical case where all clients have  $r = r_{\mathrm{max}}$ .
+
+Heterogeneous LoRA compared to Full-Fine Tuning. In Table 3, we can see that for the Reddit text summarization, HETLORA achieves similar performance with full-fine tuning with only
+
+![](images/ba7de8ef47d9313cb445341eeeb155e5db11e78b250d594af4e7f6b8b3f2a123.jpg)  
+Figure 5: Comparison of the performance across homogeneous LoRA, heterogeneous LoRA, and full fine-tuning. Heterogeneous LoRA achieves better performance than homogeneous LoRA with fewer number of communication rounds.  
+Figure 6: Ratio of communicated number of parameters for different PEFT methods to full fine-tuning to achieve the target value for the metric where it is RougeL 80 for Reddit text summarization task and perplexity 150 for the multisession chat response task. The 'X' means that the target metric is not achieved even after convergence.
+
+$0.003\%$  of parameters trained compared to full-fine tuning. For chat response, HETLORA achieves slightly lower performance than full fine-tuning. However, full fine-tuning requires to train a much larger number of parameters compared to HETLORA, making it infeasible to train with ODFMs in practice. We also show in Fig. 6 that to achieve the targeted performance for both Reddit and MSC task, HETLORA requires significantly less number of parameters to be trained and communicated compared to full fine-tuning. Although for Reddit, HOMLORA has a slightly less number of parameters to be trained, the final achieved RougeL is outperformed by HETLORA as shown in Table 3.
+
+Effect of the Decaying Factor  $\gamma$ . Lastly, we conduct an ablation study on the effect of the decaying factor  $\gamma$  of HETLORA's local training step with self-rank pruning in Table 4. We observed that aggressive pruning hurts the performance where  $\gamma = 0.85$  shows the worse performance across the varying  $\gamma$  values. On the other hand, no pruning at all ( $\gamma = 1$ ) underperforms the case when there is pruning ( $\gamma = 0.99$ ), showing that reducing the noise introduced by large rank clients which data complexity is actually not that high indeed improves the performance.
+
+# 5 Discussions and Concluding Remarks
+
+In our work, we investigated federated fine-tuning for ODFMs by proposing HETLORA that enables utilizing the power of FMs at the edge by taking device system and data heterogeneity into consideration. HETLORA is not only practical but also achieves better training speed, communication/computation efficiency, and final performance compared to other relevant baselines such as SVD-based LoRA, homogeneous LoRA, or full finetuning. Our findings in this work opens up several questions worth investigating. For instance, if the settings allow us to assign specific ranks to clients what will be the effective way to assign the ranks across clients for better convergence and performance? Another important next step of our work includes pursuing the theoretical convergence and generalization of heterogeneous LoRA.
+
+# 6 Limitations
+
+In this work, we address tackling system and data heterogeneity in federated fine-tuning of on-device foundation models. Our work is motivated by clients being able to carry different ranks for the LoRA fine-tuning method depending on their available resources, and thus exploiting this characteristic to improve federated fine-tuning with heterogeneous LoRA. However, our work assumes that the rank distribution across clients (which is analogous to how system resources are distributed across clients) is independent to the data distribution. There can be scenarios in which this is not necessarily the case where the rank and data distribution can be correlated. For instance, more affluent populations can have better off devices with larger resource capacity, and may have data distributions different to that of less affluent populations. Such correlation should be explored for future work to better understand the implications of heterogenous LoRA.
+
+# References
+
+Anonymous. 2024. Flora: Federated fine-tuning large language models with heterogeneous low-rank adaptations.  
+Sara Babakniya, Ahmed Roushdy Elkordy, Yahya H. Ezzeldin, Qingfeng Liu, Kee-Bong Song, Mostafa El-Khamy, and Salman Avestimehr. 2023. Slora:
+
+Federated parameter efficient fine-tuning of language models. CoRR, abs/2308.06522.  
+Jiamu Bai, Daoyuan Chen, Bingchen Qian, Liuyi Yao, and Yaliang Li. 2024. Federated fine-tuning of large language models under heterogeneous tasks and client resources. arXiv preprint arXiv:2402.11505.  
+Rishi Bommasani, Drew A. Hudson, and et. al. Ehsan Adeli. 2022. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.  
+Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. 2016. Practical secure aggregation for federated learning on user-held data. In NIPS Workshop on Private Multi-Party Machine Learning.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165.  
+Zhiqi Bu, Yu-Xiang Wang, Sheng Zha, and George Karypis. 2022. Differentially private bias-term only fine-tuning of foundation models. arXiv preprint arXiv:2210.00036.  
+Jinyu Chen, Wenchao Xu, Song Guo, Junxiao Wang, Jie Zhang, and Haozhao Wang. 2022. Fevtune: A deep dive into efficient federated fine-tuning with pre-trained transformers. CoRR, abs/2211.08025.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. 2023. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378.  
+Gemini Team Google. 2024. Gemini: A family of highly capable multimodal models. https:// arxiv.org/abs/2312.11805.
+
+Google. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.  
+Google. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.1040.  
+Google DeepMind. 2023. Introducing palm2. https://blog.google/technology/ai/ google-palm-2-ai-large-language-model/.  
+Tao Guo, Song Guo, and Junxiao Wang. 2023. Pfed-prompt: Learning personalized prompt for vision-language models in federated learning. In Proceedings of the ACM Web Conference 2023, WWW '23, page 1364-1374, New York, NY, USA. Association for Computing Machinery.  
+Tao Guo, Song Guo, Junxiao Wang, and Wenchao Xu. 2022. Promptfl: Let federated participants cooperatively learn prompts instead of models — federated learning in age of foundation model. CoRR, abs/2208.11625.  
+Venkatesan Guruswami and Ravi Kannan. 2012. Lecture notes in computer science theory for the information age.  
+Samuel Horváth, Stefanos Laskaridis, Shashank Rajput, and Hongyi Wang. 2024. Maestro: Uncovering low-rank structures via trainable decomposition. In Proceedings of the International Conference on Machine Learning (ICML).  
+Neil Houlsby, Andrei Giurgiu, Stanisław Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In Proceedings of the International Conference on Machine Learning (ICML).  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR).  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In The 36th Conference on Neural Information Processing Systems (NeurIPS 2022).  
+Jaejun Lee, Raphael Tang, and Jimmy Lin. 2019. What would elsa do? freezing layers during transformer fine-tuning. arXiv preprint arXiv:1911.03090.  
+Gwen Legate, Nicolas Bernier, Lucas Caccia, Edouard Oyallon, and Eugene Belilovsky. 2023. Guiding the last layer in federated learning with pre-trained models. In Workshop of Federated Learning and Analytics in Practice@ICML.
+
+Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing (EMNLP).  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Zequan Liu, Jiawen Lyn, Wei Zhu, Xing Tian, and Yvette Graham. 2024. Alora: Allocating low-rank adaptation for fine-tuning large language models. arXiv preprint arXiv:2403.16187.  
+Andre Manoel, Mirian del Carmen Hipolito Garcia, Tal Baumel, Shize Su, Jialei Chen, Robert Sim, Dan Miller, Danny Karmon, and Dimitrios Dimitriadis. 2023. Federated multilingual models for medical transcript analysis. In Conference on Health, Inference, and Learning (CHIL), pages 147-162.  
+H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agoura y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. International Conference on Artificial Intelligence and Statistics (AISTATS).  
+OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:submit/4812508.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020.  
+Anit Kumar Sahu, Tian Li, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Virginia Smith. 2020. Federated optimization for heterogeneous networks. In Proceedings of the 3rd MLSys Conference.  
+Joao Sedoc, Daphne Ippolito, Arun Kirubarajan, Jai Thirani, Lyle Ungar, and Chris Callison-Burch. 2019. Chateval: A tool for chatbot evaluation. Proceedings of NAACL-HLT.  
+Ofir Ben Shoham and Nadav Rappoport. 2023. Federated learning of medical concepts embedding using behrt. arXiv preprint arXiv:2305.13052.  
+Aliaksandra Shysheya, John F Bronskill, Massimiliano Patacchiola, Sebastian Nowozin, and Richard E Turner. 2023. Fit: Parameter efficient few-shot transfer learning for personalized and federated image classification. International Conference on Learning Representations (ICLR).
+
+Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. 2024. Improving lora in privacy-preserving federated learning. In International Conference on Learning Representations (ICLR).  
+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.  
+Michael Volske, Martin Potthast, Shahbaz Syed, and Benno Stein. 2017. TL;DR: Mining Reddit to learn automatic summarization. In Proceedings of the Workshop on New Frontiers in Summarization, pages 59-63, Copenhagen, Denmark. Association for Computational Linguistics.  
+Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K. Leung, Christian Makaya, Ting He, and Kevin Chan. 2019. Adaptive federated learning in resource constrained edge computing systems. IEEE Journal on Selected Areas in Communications, 37(6):1205 - 1221.  
+Colin Wei, Sang Michael Xie, and Tengyu Ma. 2021. Why do pretrained language models help in downstream tasks? an analysis of head and prompt tuning. Advances in Neural Information Processing Systems, 34:16158-16170.  
+Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S. Morcos. 2023. Io-fi: distributed fine-tuning without communication. Transactions on Machine Learning Research (TMLR).  
+Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. 2024. Fedlora: When personalized federated learning meets low-rank adaptation.  
+Jing Xu, Arthur Szlam, and Jason Weston. 2021. Beyond goldfish memory: Long-term open-domain conversation. arXiv preprint arXiv:2107.07567.  
+Liping Yi, Han Yu, Gang Wang, and Xiaoguang Liu. 2023. Fedlora: Model-heterogeneous personalized federated learning with lora tuning. arXiv preprint arXiv:2310.13283.  
+Sixing Yu, J. Pablo Muñoz, and Ali Jannesari. 2023. Federated foundation models: Privacy-preserving and collaborative learning for large models. arXiv preprint arXiv:2305.11414.  
+Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. 2021. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199.
+
+Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, and Guoyin Wangand Yiran Chen. 2023a. Towards building the federated gpt: Federated instruction tuning. CoRR, abs/2305.05644.  
+Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023b. Adaptive budget allocation for parameter-efficient fine-tuning. In The 11th International Conference on Learning Representations (ICLR).  
+Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL).  
+Zhuo Zhang, Xiangjing Hu, Jingyuan Zhang, Yating Zhang, Hui Wang, Lizhen Qu, and Zenglin Xu. 2023c. Fedlegal: The first real-world federated learning benchmark for legal nlp. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL).  
+Zhuo Zhang, Yuanhang Yang, Yong Dai, Qifan Wang, Yue Yu, Lizhen Qu, and Zenglin Xu. 2023d. Fedpetuning: When federated learning meets the parameter-efficient tuning methods of pre-trained language models. Findings of the Association for Computational Linguistics (ACL).  
+Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, Hao Peng, Jianxin Li, Jia Wu, Ziwei Liu, Pengtao Xie, Caiming Xiong, Jian Pei, Philip S. Yu, and Lichao Sun. 2023. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419.

2024/Heterogeneous LoRA for Federated Fine-tuning of On-Device Foundation Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:190639aa6cf110d286503f0ebe83d12b14adc4a4c87c4bea1701149ac098e860
+size 316741

2024/HiFT_ A Hierarchical Full Parameter Fine-Tuning Strategy/8be09224-3564-4845-996a-57391f9ec216_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2dd52b15f1fada7f68dacac2ff4e39e086d9b3f0d239fb1fdc1a0e22815577d
+size 1871347

2024/HiFT_ A Hierarchical Full Parameter Fine-Tuning Strategy/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:986477266b71f0fec1e5958f587b6cc191c52883b606a0ea293521898c988d2b
+size 1845904

2024/Hidden Persuaders_ LLMs’ Political Leaning and Their Influence on Voters/9092ce3a-456e-4ac2-bb0b-5a622f61e8b8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3883cdda1ad73e02bf3c65095f7b337f26d49420c3956cf5c5094b676009ca1a
+size 1332117

2024/Hidden Persuaders_ LLMs’ Political Leaning and Their Influence on Voters/full.md

@@ -0,0 +1,582 @@
+# Hidden Persuaders: LLMs' Political Leaning and Their Influence on Voters
+
+Yujin Potter $^{1}$ , Shiyang Lai $^{2}$ , Junsol Kim $^{2}$ , James Evans $^{2}$ , Dawn Song $^{1}$
+
+<sup>1</sup>University of California, Berkeley,
+
+$^{2}$ University of Chicago
+
+yujinyujin9393@berkeley.edu, shiyanglai@uchicago.edu
+
+# Abstract
+
+Do LLMs have political leanings and are LLMs able to shift our political views? This paper explores these questions in the context of the 2024 U.S. presidential election. Through a voting simulation, we demonstrate 18 openweight and closed-source LLMs' political preference for Biden over Trump. We show how Biden-leaning becomes more pronounced in instruction-tuned and reinforced models compared to their base versions by analyzing their responses to political questions related to the two nominees. We further explore the potential impact of LLMs on voter choice by recruiting 935 U.S. registered voters. Participants interacted with LLMs (Claude-3, Llama-3, and GPT-4) over five exchanges. Intriguingly, although LLMs were not asked to persuade users to support Biden, about  $20\%$  of Trump supporters reduced their support for Trump after LLM interaction. This result is noteworthy given that many studies on the persuasiveness of political campaigns have shown minimal effects in presidential elections. Many users also expressed a desire for further interaction with LLMs on political subjects. Further research on how LLMs affect users' political views is required, as their use becomes more widespread.
+
+# 1 Introduction
+
+In the pursuit of developing safe artificial intelligence (AI), creating unbiased AI systems has become a critical goal. It has been shown that many AI technologies, including large language models (LLMs), exhibit measurable left-wing leanings (Hartmann et al., 2023; Sullivan-Paul, 2023; Röttger et al., 2024). Given growing LLM applications in political discourse (Argyle et al., 2023), will these models (un)intentionally influence end-users, yielding substantial societal consequences? This question remains largely unanswered.
+
+Our study addresses this question by examining
+
+LLMs' political leanings<sup>1</sup> and their potential societal impact in the context of the upcoming 2024 U.S. presidential election. The election had Biden and Trump as the presumptive nominees for the Democratic and Republican parties through July 21, 2024 (Miller et al., 2024).<sup>2</sup> As the election date approaches, the potential for LLMs to have (un)intended effects on the election has raised many concerns (Anthropic, 2024b,c). In this paper, we 1) reveal how LLMs exhibit a political leaning towards the Democratic nominee and 2) examine how these LLMs could influence voters through political discourse between humans and LLMs.
+
+First, in Section 3, we simulate presidential election voting between the two candidates across 18 open-weight and closed-source models, with each model run 100 times. Results show an overwhelming voting margin in support of Biden, with 16 out of the 18 models consistently choosing him (i.e.,  $100\%$  Biden vote).
+
+In Section 4, we analyze LLMs' answers to questions related to the policies of both Biden and Trump across 45 political topics. Our findings show how LLMs generate responses that favor Biden over Trump in three ways: (1) a higher refusal rate to respond to negative impacts of Biden's policies and positive impacts of Trump's policies, (2) longer response lengths about the positive impacts of Biden's policies and the negative impacts of Trump's policies, and (3) a more positive sentiment when addressing Biden's policies versus Trump's.
+
+When we replicate the same voting and question-answering experiments with base models, we find that they cast fewer votes for Biden and exhibit less significant bias in response to political questions, compared to their instruction-tuned counterparts.
+
+This finding suggests that human instruction posttraining, including reinforcement learning from human feedback, amplified the political leaning appearing in LLMs' outputs.
+
+Moving to a more interactive and realistic scenario, Section 5 investigates how LLMs manifest political leanings during human-LLM interactions. Given other characteristics of LLMs, such as their propensity for user adaptation and sycophancy (Sharma et al., 2023), we were uncertain whether they would exhibit a consistent pro-Biden view during interactions. Another different question is whether LLMs will shift humans' voting choices via their conversation. To explore these questions, we conducted a user experiment in which U.S. registered voters engaged in one-on-one discussions with one of three popular LLMs (i.e., Claude-3-Opus, Llama-3-70B, and GPT-4-Turbo).
+
+We found that these three LLMs consistently presented their pro-Biden views during conversations with human subjects, regardless of the participants' initial political stance. Moreover, these LLMs significantly affected participants' voting choices by increasing the participants' leaning towards Biden following their interaction. Specifically, nearly  $20\%$  of initial Trump supporters decreased their Trump support, with the most extreme case showing a complete reversal (i.e., from fully Trump-leaning to fully Biden-leaning).  $24\%$  of our initial neutral participants shifted to support Biden, while initial Biden supporters showed no significant change. As a result, the simulated vote margin in our sample widened from  $0.7\%$  to  $4.6\%$ .
+
+This effect is politically meaningful, given that vote margins are typically very narrow in real-world presidential elections (Pew Research Center, 2024; CNN, 2020). Moreover, the effect could represent a lower-bound of relevant influence, considering that participants got exposed to only five exchanges. Many participants expressed enjoyment and a desire to extend their conversation with LLMs on political topics after the experiment, including many whose leanings changed. This would facilitate longer political interactions with LLMs in the wild that might induce a more pronounced impact on human voting stances.
+
+# 2 Related Work
+
+# 2.1 Political Leaning of LLMs
+
+Prior literature consistently demonstrates that left-of-center, Democrat political views are gener
+
+ally shared across LLMs. These studies used multiple-choice surveys and questionnaires widely employed in social science to measure political views (Taubenfeld et al., 2024; Rozado, 2024; Feng et al., 2023; Santurkar et al., 2023; Hartmann et al., 2023; Röttger et al., 2024; Rutinowski et al., 2024). For example, studies using the Political Compass Test (PCT) reveal a sizeable left political leaning among LLMs (Feng et al., 2023; Röttger et al., 2024; Motoki et al., 2024; Rozado, 2024; Rutinowski et al., 2024). Other studies reaffirm LLMs' left leanings across 11 political orientation tests, such as the Political Spectrum Quiz (Rozado, 2024). Using Pew research surveys, researchers find that instruction-tuned LLMs exhibit greater left leanings compared to prior base models (Santurkar et al., 2023). LLMs' left leanings are also observed in non-US contexts, including Germany and the Netherlands (Hartmann et al., 2023).
+
+Several studies reveal that political leaning manifests when LLMs perform downstream tasks (Taubenfeld et al., 2024; Feng et al., 2023). Researchers show that LLMs tend to adhere to the inherent, left-leaning political view even when assigned to argue for the opposite viewpoint during a debate (Taubenfeld et al., 2024). Others fine-tune LLMs to create politically partisan versions using a news/social media dataset and discover that the hate-speech and misinformation detection performance of partisan LLMs is worse than of untuned LLMs (Feng et al., 2023).
+
+We build on these studies in two distinct ways. First, we explore how political leanings manifest in LLMs' outputs in the context of the 2024 U.S. election. Complementing Hartmann et al. (2023), we also reveal that the manifestation of left leanings in downstream applications increases in instruction-tuned LLMs compared to their base versions. Second, prior literature has focused on examining LLM political leanings through surveys or closed-form questions. To the best of our knowledge, no prior work has investigated the manifestation of their political leaning in a realistic, interactive setting with humans, and how these LLMs could potentially sway voters. By employing user experiments where participants converse with LLMs over multiple exchanges, our work aims to fill that gap.
+
+# 2.2 LLM Persuasion
+
+A growing body of literature highlights the potential for LLMs to effectively persuade their human interlocutors, which could lead to novel
+
+and unprecedented AI risks (Atillah, 2023; Anthropic, 2024a; Goldstein et al., 2024; Walsh, 2024; Costello et al., 2024; Cheong et al.; Hackenburg and Margetts, 2024). In early 2023, tragic news emerged that a Belgian man had committed suicide after a conversation with an LLM allegedly encouraged him to do so (Atillah, 2023). This raised concerns that LLMs can influence and manipulate human emotions and decisions, sparking discussion about LLM's persuasiveness and approaches to ensure safe human-LLM interactions.
+
+Research has provided empirical evidence that the capability of LLMs to persuade others is rapidly increasing (Anthropic, 2024a; Goldstein et al., 2024; Walsh, 2024; Costello et al., 2024). For example, Costello et al. (2024) demonstrated GPT-4's ability to beneficially persuade humans they interact with, significantly reducing humans' conspiracy beliefs. They also found evidence of long-term consequences of LLM persuasion: the reduction of conspiracy beliefs persisted for more than two months. These studies focus on the purposively designed persuasive capabilities of LLMs: they can persuade humans in line with the intentions of their designers, as to reduce conspiracy beliefs. By contrast, here we focus on unintended LLM persuasion and its influence on the political stances of humans who interact with them. This is the central question we aim to address in Section 5.
+
+# 3 US Presidential Election Among LLMs
+
+We start by examining the political stances of 18 LLMs regarding the two 2024 U.S. presidential nominees by simulating and collecting election votes for each model 100 times. Results are listed in Table 1. To elicit voting choices, we engineered our prompt to make sure it can always successfully bypass refusals.<sup>3</sup> We also alternated the placement order of Biden and Trump in the prompt in half of the cases to reduce the positional bias of LLMs. For detailed prompts, please see Appendix A.2. The temperature was set to 1 for closed-source models and 0.7 for open-weight ones.
+
+Simulation results demonstrate overwhelming votes for Biden across all tested LLMs. With the exception of Gemini Pro 1.0 and Alpaca, all models voted for Biden in 100 out of 100 rounds. Gemini Pro voted for Biden 74 times, while Alpaca voted for Biden in 84 out of 100 trials. We also observe a
+
+<table><tr><td></td><td>Entity</td><td>Model</td><td>Biden</td><td>Trump</td></tr><tr><td rowspan="18">Instruction-tuned</td><td rowspan="2">OpenAI</td><td>GPT-4-Turbo</td><td>100</td><td>0</td></tr><tr><td>GPT-3.5-Turbo</td><td>100</td><td>0</td></tr><tr><td rowspan="3">Anthropic</td><td>Claude-3-Opus</td><td>100</td><td>0</td></tr><tr><td>Claude-2.1</td><td>100</td><td>0</td></tr><tr><td>Claude-Instant-1.2</td><td>100</td><td>0</td></tr><tr><td rowspan="2">Meta</td><td>Llama-3-70B-Chat</td><td>100</td><td>0</td></tr><tr><td>Llama-2-70B-Chat</td><td>100</td><td>0</td></tr><tr><td>Google</td><td>Gemini Pro 1.0</td><td>74</td><td>26</td></tr><tr><td>Mistral AI</td><td>Mixtral-8×7B-Instruct</td><td>100</td><td>0</td></tr><tr><td>WizardLM</td><td>WizardLM-13B-V1.2</td><td>100</td><td>0</td></tr><tr><td>Stanford</td><td>Alpaca-7B</td><td>84</td><td>16</td></tr><tr><td>Austism</td><td>Chronos-Hermes-13B</td><td>100</td><td>0</td></tr><tr><td>Gryphe</td><td>MythoMax-L2-13B</td><td>100</td><td>0</td></tr><tr><td>OpenChat</td><td>OpenChat-3.5-1210</td><td>100</td><td>0</td></tr><tr><td>Garage-bAInd</td><td>Platypus2-70B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Alibaba</td><td>Qwen1.5-72B-Chat</td><td>100</td><td>0</td></tr><tr><td>Upstage</td><td>Solar-10.7B-Instruct</td><td>100</td><td>0</td></tr><tr><td>LMSYS</td><td>Vicuna-13B-v1.5</td><td>100</td><td>0</td></tr><tr><td rowspan="3">Base</td><td>Meta</td><td>Llama-3-70B</td><td>85</td><td>15</td></tr><tr><td>Mistral AI</td><td>Mixtral-8×7B</td><td>47</td><td>53</td></tr><tr><td>Alibaba</td><td>Qwen1.5-72B</td><td>100</td><td>0</td></tr></table>
+
+Table 1: Voting results of 18 instruction-tuned LLMs and 3 base models.
+
+difference in the strength of Biden-leaning tendencies between the instruction-tuned models and their base versions. The base models of Llama-3-70B-Chat<sup>4</sup> and Mixtral-8×7B-Instruct made pro-Biden decisions less often compared to their instruction-tuned versions with the same temperature, casting 15 and 53 out of 100 votes for Trump, respectively.
+
+# 4 LLM Replies to Candidate-Related Questions
+
+# 4.1 Data collection
+
+Although a closed-ended question is a common way to investigate LLM political stance, this approach may have limitations in thoroughly examining it (Röttger et al., 2024). Therefore, we additionally examine their responses to questions about Trump/Biden policies. We first established a set of candidate-related questions, inquiring about: (1) what are Trump/Biden's policies ("neutral"), (2) what are the positive impacts of Trump/Biden's policies ("positive"), and (3) what are the nega
+
+tive impacts of Trump/Biden's policies ("negative") across 45 political topics, culminating in a total of  $270\left(= 3\times 2\times 45\right)$  questions. These political topics were sourced from a popular election candidate comparison website (Ballotpedia, 2024). Detailed question information is presented in Appendix A.3. We asked each question 10 times for each of the 18 models, collecting a total of 48,600  $(= 18\times 270\times 10)$  responses.
+
+# 4.2 Biden-leaning responses from LLMs
+
+Refusal rate: We obtained the refusal rate of LLMs based on the popular refusal detector model provided by LLM Guard (Goyal et al., 2024) $^5$ . Figure 1a shows overall refusal rates when questioned about neutral, positive, and negative aspects of Biden's and Trump's policies across all tested 18 LLMs on 45 political topics. Our results suggest that LLMs are more prone to refusing to mention the negative aspects of Biden's policies and the positive aspects of Trump's. On average, LLMs refused  $2.1\%$  of the questions on neutral aspects of Biden's policies and refused  $3.9\%$  of the questions on neutral aspects of Trump's  $(t = -7.765, p < 0.001)^6$ . When queried about positive aspects of the two, LLMs refused to respond on average  $15.8\%$  of the time for Biden's policies and  $21.0\%$  of the time for Trump's  $(t = -12.061, p < 0.001)$ . For negative aspects of policies, refusals occurred  $35.6\%$  of the time for Biden and  $16.9\%$  for Trump  $(t = 39.972, p < 0.001)$ . Although the refusal rate varied across models, a pro-Biden pattern was consistently observed within each model, with some models, including the Claude family and Qwen, manifesting a larger Biden-leaning (see Figure 3 and Table 2 in Appendix).
+
+Response length: Figure 1b shows that LLMs provided significantly longer responses when describing positive aspects of Biden's policies and negative aspects of Trump's policies. When LLMs were asked about positive aspects of Biden's policies, they exhibited an average response length of 170.484 words, significantly longer than their responses about positive aspects of Trump's (146.814,  $t = 44.254$ ,  $p < 0.001$ ). In con-
+
+trast, LLMs provided significantly longer responses when describing the negative aspects of Trump's policies (164.825 words) compared to Biden's (143.871 words)  $(t = -37.434, p < 0.001)$ . Our model comparison presented in Table 2 shows how this pattern of responding with different lengths for Biden and Trump persisted across most models. The Mixtral, Claude, and Llama families manifested a larger gap in response length.
+
+Sentiment score: We calculated the average sentiment scores for each model's responses based on the NLTK dictionary-based sentiment analyzer (Bird et al., 2009), which also reveals a salient Biden-leaning pattern. When LLMs were questioned on neutral aspects of Biden's policies, the average sentiment score for LLMs' responses was 0.300, significantly more positive than Trump's 0.117 ( $t = 75.742$ ,  $p < 0.001$ ). Similarly, when asked to comment on positive aspects of policies, the average sentiment score for Biden was 0.375, but only 0.235 for Trump, marking a notable difference ( $t = 56.820$ ,  $p < 0.001$ ). For negative aspects, LLMs' answers presented a more negative sentiment score of  $-0.120$  for Trump compared with  $-0.046$  for Biden ( $t = 28.141$ ,  $p < 0.001$ ). Among tested LLMs, the Claude family was among the models with the largest Biden-leaning sentiment (please refer to Table 2 in Appendix).
+
+We also conducted a granular analysis of attitudes presented in LLMs' responses using the geometry of culture approach (Kozlowski et al., 2019) (please see Figure 7). In summary, a salient Biden-leaning pattern emerges across all of our analyses and in every model, confirming the significant pro-Biden leaning in political question-answering contexts.
+
+# 4.3 Instruction-tuned models vs. Base models
+
+We collected additional responses from three openweight base models: Llama-3-70B, Mixtral-  $8\times 7\mathrm{B}$  and Qwen-1.5-72B to compare the sentiment scores of their responses with their corresponding instruction-tuned ones. Figure 6 in the Appendix summarizes these results. Base models, although leaning towards Biden, exhibited significantly lower Biden-leaning compared with their instruction-tuned counterparts. For neutral questions, the average sentiment score difference between Trump and Biden was 0.127 for base models but 0.184 for their instruction-tuned counterparts  $(t = -3.109$ $p = 0.002)$ . For questions focusing on positive aspects of their policies, the semi
+
+![](images/11e5bccccf1bd260a1363fdd7a877f2595c04ed649a5f64b66f14a3c7fd4c289.jpg)  
+(a) Refusal rate
+
+![](images/46362a2a1b03d99882abbc63698645b4610c6405e6b68c11b8c9405f80069dd3.jpg)  
+(b) Response length
+
+![](images/9e0c3c6680bebd63cb1c12d5db6b151a423d605a3e73e6d1b464a3cbd28ce715.jpg)  
+(c) Sentiment score  
+Figure 1: Three metrics to evaluate LLMs' responses to candidate-related questions. The  $x$ -axis represents neutral, positive, and negative questions for Biden and Trump. For Figure 1a, error bars represent  $95\%$  confidence intervals. Figure 1b starts with the median (50%) as the centerline and each successive level outward representing half of the remaining data. All figures show LLMs tend to provide responses more favorable to Biden over Trump.
+
+ment score difference was 0.070 for base models, while it was 0.159 for instruction-tuned models  $(t = -5.597, p < 0.001)$ . In the case of negative aspects of policies, the sentiment score difference was 0.012 for base models and 0.117 for instruction-tuned models  $(t = -5.860, p < 0.001)$ . These results indicate that the post-training process increased the Biden-leaning level in the instruction-tuned models. However, it remains uncertain which specific objective among various ones including helpfulness, harmlessness, and truthfulness during the process increased the manifestation of the Biden-leaning (Fulay et al., 2024).
+
+# 5 Influence of LLMs on Voters' Choices
+
+# 5.1 User experiment design
+
+Next, we launched a user experiment to further investigate whether LLMs exhibit political leanings during interactions with voters, and if so, whether such interactions will shift human voting choices.
+
+The user experiment encompassed three stages: pre-interaction survey, human-LLM interaction, and post-interaction survey. In the pre-interaction survey, we measured participants' candidate leanings by asking them to allocate  $100\%$  between Biden and Trump. For example, allocating 100 to Trump (or Biden) means leaning completely and exclusively towards Trump (or Biden). Allocating 50 to each candidate indicates perfect neutrality. We also collected their political attitudes and attitudes towards AI.
+
+In the human-LLM interaction stage, participants were required to engage in five exchanges of conversations with one of three randomly as
+
+signed LLMs (i.e., Claude-3-Opus, Llama3-70B, or GPT-4-Turbo). For the LLM interaction setup, we prompted LLMs to participate in political discourse with a human participant. We did not direct LLMs to persuade their human conversation partners' political views. Instead, we prompted LLMs to generate outputs regarding Biden and Trump's policy (see Appendix A.4). In the post-interaction survey, some questions from the pre-interaction survey were repeated to assess changes in participants' political views. We also asked participants about their perceived change in attitude toward AI at the experiment's end.
+
+We recruited 935 U.S. registered voters through CloudResearch's Connect Survey platform (Cloud Research, 2024). Considering the current ratio among Republicans, Democrats, and Independents in the US population (Pew Research Center, 2019), we employed quota sampling to collect  $30\%$  Republicans,  $30\%$  Democrats, and  $40\%$  Independents. Additionally, we applied a  $50\%$  quota for each female and male group. Out of 935 participants, 695 were assigned to interact with one of three LLMs (i.e., treatment group), while the remaining 240 formed a control group and were asked to write down their subjective thoughts on open-ended political questions without interacting with LLMs. See Appendix A.4 and A.5 for details.
+
+# 5.2 LLMs' leaning toward Biden in dialogue
+
+We staged our analysis by first measuring the exhibition of the pro-Biden view in LLM-generated texts during their conversation with human participants. We adopted Claude-3 to estimate the level of Trump/Biden-leaning in LLMs' generated texts.
+
+Currently, there are no widely accepted methods for quantifying Trump/Biden-leaning in textual data. To address this, we explored several approaches, including the use of LLMs and neural-network word embedding models (Kozlowski et al., 2019). For LLM-based methods, we prompted GPT-4 and Claude-3 to rate the degree to which LLMs' responses support Biden or Trump on a  $-1$  (Biden) to 1 (Trump) continuous scale. After manual verification, we found that among the tested methods, Claude-3 manifests the best performance. GPT-4 often misinterpreted the direction of leaning, erroneously assigning positive scores to cases that favored Biden. The word embedding model showed lower accuracy. To further validate Claude-3's performance, we conducted an additional correlation analysis between participants' Trump support levels and the scores Claude-3 assigned based on these participants' conversation texts. This yielded a high correlation coefficient of 0.943, supporting our assessment of Claude-3's high accuracy.[7]
+
+As shown in Figure 2a, the three LLMs consistently exhibited support for Biden in their responses, irrespective of the candidate the human conversation partner supported. Although LLMs' pro-Biden attitudes were more pronounced when interacting with Biden supporters, their pro-Biden views persisted when engaging with Trump supporters or neutral people. Llama-3 presented the most pro-Biden stance, while GPT-4 exhibited the least among the three tested models. This also aligned with our manual examination of the data.
+
+Beyond general attitudes, we found that LLMs interacted differently with Biden and Trump supporters (please see Figure 10 in Appendix). In particular, Llama-3 mainly focused on the following policy issues: climate change, healthcare, and pandemic virus responses. Note that, as shown in Figure 10, the main topics of the conversations between LLMs and humans were policies rather than personal characteristics.
+
+# 5.3 Change in vote choices after LLM interaction
+
+The previous section demonstrated how LLMs presented their pro-Biden views during conversation. Here, we address a different question: whether the
+
+LLM conversation affected users' vote choices.
+
+Increase in support for Biden: After interacting with LLMs, participants increased their leaning towards Biden. The average Biden-leaning percentage rose from  $50.8\%$  to  $52.4\%$ , a statistically significant change  $(t = 4.886, p < 0.001)$ . Consequently, the vote margin increased from  $0.7\%$  to  $4.6\%$ $(t = 3.817, p < 0.001)$ . This effect was stronger than those in many existing studies that analyze the persuasive effect of traditional political campaigns (Kalla and Broockman, 2018; Hewitt et al., 2024; Hager, 2019; Lazarsfeld et al., 1968; Berelson et al., 1986; Broockman and Kalla, 2023)<sup>8</sup>. Even small effects can be politically meaningful, given that elections are often decided by very narrow margins (Pew Research Center, 2024; Hewitt et al., 2024).
+
+Differences by supporting candidates: Trump supporters and the neutral group exhibited a significant increase in their leaning towards Biden. We find that, on average, Trump supporters increased their Biden-leaning from  $8.1\%$  to  $10.6\%$ $(t = 4.570, p < 0.001)$ , and the neutral group increased their Biden-leaning from  $50\%$  to  $54.2\%$ $(t = 3.485, p < 0.001)$ . Meanwhile, initial Biden supporters retained their Biden-leaning at  $93.1\%$ . The same effect is observed in participants' vote choice changes. Among initial Trump supporters, the vote margin decreased by  $5.8\%$  in favor of Biden  $(t = 3.461, p < 0.001)$ . Among initially neutral participants, the vote margin shifted by  $21.2\%$  towards Biden  $(t = 3.584, p < 0.001)$ . Figure 2b presents how participants changed their political stance following interaction.
+
+Post-hoc analysis reveals that Trump supporters and neutral participants who increased their Biden-leaning often expressed appreciation for LLMs' insights delivered throughout the conversation. For example, "the AI brought up some great points about how Biden handles the presidency" or "The AI experience did make me lean more favorably towards Biden or at least his policies...". Moreover, many Biden supporters who retained or increased their support for Biden expressed that the
+
+![](images/0c99fd75f1c8986321f05b686469e0b04d3fbb0980af82bae475a438873acf03.jpg)  
+(a) LLMs' Biden/Trump support score
+
+![](images/e71fd8c9be7472eece53cfa5dc88ceb8feceea2544af09d780c4ae42f7c96864.jpg)  
+(b) Change in leaning towards candidates  
+Figure 2: LLMs' political attitudes during the conversation and the resulting change in participants' political attitudes post-interaction. Figure 2a presents LLMs' average support scores for Biden or Trump, including  $95\%$  confidence intervals, by participants' initial political stance. A negative score indicates a Biden-supporting tendency in LLM-generated texts, while a positive score indicates a tendency to support Trump. Figure 2b presents the change in participants' leaning towards the candidates after LLM interaction, with leaning categorized into 11 bins including the neutral group. Arrows indicate the overall direction of shift in participants' candidate preference following LLM interaction.  $\uparrow$  suggests an increased leaning towards Biden after interaction, while  $\rightarrow$  indicates that their preference remained unchanged. Figure 2c presents the average effect of LLM interactions on Biden-leaning percentage compared to the control group (grey dashed line), including  $95\%$  confidence intervals in brackets. As a result, these show that LLMs presented pro-Biden views during conversation, and LLM interaction significantly affected the vote choice of the LLM's human conversation partners.
+
+![](images/78749cbb900572c537582700876747b850d3c51d3dc920695b0a88216f7a1bcc.jpg)  
+(c) LLM interaction effects
+
+LLM largely agreed with them and reinforced their stance. Specifically, in our survey, a total of 42 Biden supporters explicitly said the LLM agreed with their arguments most of the time. On the other hand, only 6 Trump supporters said this, while many Trump supporters expressed disagreement with what the LLM said. In line with this, we find that some Trump supporters increased their support for Trump following interaction, manifesting a backfire effect. For example, "Listening to the crap the AI spouted (though well spoken) makes me like Biden even less than before I started." Refer to Appendix B.1 for more information.
+
+Differences by LLM: While all LLMs were influential in increasing participants' Biden-leaning percentages, each effect varied based on which candidate participants initially supported. For initial Trump supporters, Claude-3, the second most pro-Biden model, was the most influential, increasing participants' Biden-leaning from  $9.1\%$  to  $12.6\%$ $(t = 3.694, p < 0.001)$ , followed by GPT-4 (from  $8.2\%$  to  $11.5\%$ ,  $t = 2.579$ ,  $p = 0.006$ ) and then Llama-3 (from  $6.8\%$  to  $7.6\%$ ,  $t = 1.746$ ,  $p = 0.042$ ). Notably, the effect is not correlated with the Biden-leaning levels of LLMs. As mentioned earlier, some Trump supporters increased their support, expressing complaints about the LLMs' clear left-wing stance. Moreover, it was observed that less biased or possibly neutral re
+
+sponses from LLMs influenced some supporters to reduce their Trump-leaning (e.g., from  $70\%$  to  $55\%$  leaning towards Trump). For example, one participant stated, "The AI made some valid points about the economy and immigration being horrible under Biden and made valid points as to why. It also wasn't biased..."
+
+Meanwhile, for the initially neutral participants, the more pro-Biden model, the more influential; Llama-3 increased their Biden-leaning to  $57.0\%$ $(t = 2.914, p = 0.004)$ , and Claude-3 increased it to  $52.6\%$ $(t = 1.759, p = 0.047)$ , while GPT-4 did not significantly change it  $(t = 1.098, p = 0.289)$ .
+
+Among initial Biden supporters, Llama-3 and GPT-4 increased their Biden-leaning insignificantly, and Claude-3 even decreased it from  $93.9\%$  to  $93.0\%$ , although the decrease was much smaller than the increase from Trump supporters. In fact, even though many Biden supporters said the conversation strengthened their belief, we could not often capture this numerically because they already  $100\%$  leaned towards Biden. Moreover, some Biden supporters were influenced by the exposure to Trump's positives presented by LLMs during the conversation; for example, one participant stated "As I was leaning more toward Biden, the AI would bring up semi-valid points about Trump. The AI was also very agreeable, but polite when bringing up Trump." These two factors resulted in no signif
+
+icant change in the Biden-leaning percentage for the initial Biden supporter group.
+
+Differences by political interests and trust in AI: We also find both groups that are more and less interested in politics significantly changed their leaning. Participants who closely follow political and election news increased their leaning towards Biden from  $51.3\%$  to  $52.7\%$ $(t = 4.396, p < 0.001)$ . Those who did not follow political news also significantly increased from  $49.3\%$  to  $51.4\%$ $(t = 2.374, p = 0.009)$ .
+
+Additionally, participants who expressed trust in AI were more likely to change their political leaning. Participants who expressed more excitement than concern about the increased use of AI shifted in their leaning towards Biden from  $49.1\%$  to  $51.7\%$  ( $t = 3.355$ ,  $p < 0.001$ ). This represents a higher increase compared to those who do not trust AI and whose Biden-leaning increased only from  $48.0\%$  to  $49.0\%$  ( $t = 1.814$ ,  $p = 0.036$ ). This is reflected in their statements such as "I don't trust a robot about politics" and "The AI chatbot is nothing more than a conversational tool."
+
+Causal inference via comparison with the control group: Despite these results, LLMs might not "causally" influence voting choices. For example, one participant said the act of writing down their thoughts itself increased their confidence in their expressed political position. In order to address concerns regarding potential confounders (e.g., political writing, observer bias (Azarova, 2023), etc.), we collected additional control group data in which participants wrote down their thoughts on Biden and Trump regarding various political topics, instead of interacting with the LLM.
+
+The distributions of demographics and pre-intervention measures for the control group were similar to those of the treatment group (see Table 3). We conducted a linear regression controlling for pre-intervention Biden-leaning percentages to compare the treatment group with the control group. As shown in Figure 2c, results indicate that LLM interaction significantly increased Biden-leaning percentages compared to the control group (Claude-3: coeff = 1.728, se = 0.698, p = 0.013; Llama-3: coeff = 1.524, se = 0.701, p = 0.030; GPT-4: coeff = 2.318, se = 0.701, p = 0.001).
+
+Nevertheless, this causality analysis does not explain precisely what aspects of LLM interaction swayed more Trump supporters and neutral participants towards Biden. There can be various potential causes including different features of the LLM interaction experience and different characteristics of Trump/Biden supporters. For example, a qualitative review of those human-LLM conversations shows a frequent pattern of the LLM providing information previously unknown to the participant. Untangling these factors will require further work.
+
+# 5.4 Spillover attitudes about AI
+
+Participants who initially leaned toward Trump but reduced their Trump support after interacting with LLMs tended to feel more favorable towards AI compared to others (please see Figures 12 and 13). Notably, in this category consisting of 58 participants, only two became less favorable in their attitude towards AI following LLM interaction. These participants who manifested decreased support for Trump also often expressed a desire for further LLM conversations. One participant who decreased his Trump-leaning from  $100\%$  to  $60\%$  stated that "This conversation was hands down the best one I have had talking to anyone about politics...I really feel like this is the way we need to discuss politics...I think that is kind of crazy but thank you." This suggests that users may seek out long-term LLM interactions. Sustained interaction with the LLMs in our sample might potentially convert a bigger subgroup of Trump supporters into Biden supporters.
+
+In stark contrast, the 32 Trump supporters who retained or increased their original Trump support level reported a less favorable view of AI after the experiment. This demonstrates how a perceived political leaning in AI can contribute to political polarization about AI, leading strong Trump supporters to develop negative attitudes towards AI. As one participant who interacted with GPT-4 remarked, "This just goes to show how poor current AI models are. I'm confused why they are being pushed out so early when they are obviously so incapable of critical thinking or hiding their biases." Figure 14 in Appendix C present differences in attitudes following the experiment.
+
+# 6 Discussion
+
+We analyzed the manifestation of political leanings in LLMs and LLMs' influence on voters within
+
+the context of the upcoming 2024 U.S. presidential election. In particular, LLM political leanings are not confined to the match-up between Biden and Trump. To demonstrate this, we conducted three additional voting experiments with 10 LLMs regarding: 1) the current 2024 U.S. presidential election match: Harris vs. Trump, 2) a more general U.S. election context: a Democratic Party candidate vs. a Republican Party candidate, and 3) a UK election context: a Labour Party candidate vs. a Conservative Party candidate. Results presented in Tables 4, 5, and 6 show an overall strong left-leaning among LLMs. The generalizability of the societal impact of LLMs in the political sphere and whether LLMs' political leaning causes the observed influence on voters should be explored in further studies.
+
+The cumulative influence of LLMs on voters might be even greater than our reported results, considering many participants' interest in further interaction with LLMs. This stands in contrast to existing political campaigning, which often struggles to maintain long-term engagement with voters due to voters' reactions of feeling annoyed or manipulated (Kalla and Broockman, 2018). Moreover, our findings suggest the necessity of adopting a cautious approach to using LLMs for political campaigning. Political persuasive power could potentially be much larger if LLMs were intentionally designed to intervene in elections for political purposes, unlike our setting, which involved models that influenced user political views unintentionally.
+
+Sharing these concerns, many companies have made substantial efforts to devise use policies that reduce election-related influence and associated risks (Anthropic, 2024b,c; Google India Team, 2024). But our findings raise a question: how should companies address the possibility that LLMs can themselves unintentionally shift human political stances through routine, non-malicious interactions that may not violate terms of service? Further study is required to understand when and how this occurs.
+
+The causes of LLM political leaning remain an open question. One possibility is that their training dataset consists of modern Web data that is more liberal than old data (Feng et al., 2023). The post-training process could also have contributed to this effect. We found that instruction-tuned models show a stronger Biden-leaning pattern, though we cannot pinpoint which specific objective of the posttraining heightened these tendencies. For example, a recent paper (Fulay et al., 2024) demonstrated a
+
+correlation between truthfulness and political leaning in language models; specifically, the models trained with truthfulness datasets showed an increasing left-wing leaning. The complexity of the model development process makes it challenging to determine the source of LLM political leaning. Mechanistic interpretation of open models could yield insights into these leanings and represents a promising direction for future work.
+
+Finally, our experiment also raises the question of whether neutral LLMs will actually align with user preference. Many participants highly rated conversation satisfaction with LLMs even though they often leaned towards Biden (see Figure 11 in Appendix). Participants who encountered a relatively neutral LLM response sometimes suggested a preference for engaging with LLMs holding a particular perspective.[10] This example reveals the tension between AI bias and user expectations in conversational contexts. Users may prefer more candid outputs from LLMs, even if biased, regardless of whether these outputs align with or contradict people's beliefs. As a result, such examples imply that solving the "bias problem" in LLMs goes well beyond mere technical considerations and must account for conversation quality and user engagement.
+
+# 7 Conclusion
+
+We identify a notable leaning toward Biden in 18 open-weight and closed-source LLMs across various scenarios: voting behavior, response to political questions, and interaction with humans. In particular, greater Biden-leaning of instruction-tuned models is observed compared to their base versions, which suggests that current post-training processes amplify the manifestation of political leaning in their responses. We further demonstrate that LLMs could significantly shift people's voting stance toward Biden through human-LLM political conversation. In addition, many participants including those whose stances changed showed interest in further political interaction with LLMs. Lastly, the generalizability of our reported findings beyond the 2024 U.S. presidential setting and the mechanisms by which voters' stances are changed require further research.
+
+# Limitations
+
+Our experiment involved a total of 935 users consisting of 695 in the treatment group and 240 in the control group. Even though we found statistical significance, a larger-scale user experiment will be required to estimate the large-scale political impacts of LLMs. We hope our paper can inspire larger-scale field experiments. Another limitation is that our experiment was conducted in a simulated setup where users were aware that their choices were being observed during the experiment. This may cause observer bias (Azarova, 2023). However, we believe that collecting the control group data under the same conditions, except for the different interventions, and comparing our main group with the control group reduces this concern.
+
+# Ethics Statement
+
+First and foremost, we emphatically state that this paper does not endorse either political party and has no intention of intervening in the 2024 U.S. Presidential election. Similar to other AI bias studies, our work includes sensitive content that may offend some groups and addresses the upcoming presidential election. Moreover, we recognize the potential for malicious and inappropriate use of our work, to attempt to cast doubt on the legitimacy of a fair election outcome. Nevertheless, considering our potentially consequential findings, we believe it is crucial for the public to be aware of the potential impacts posed by LLMs by publicizing the findings in our paper. We hope our research contributes to increasing public awareness of potential AI societal impacts. Regarding the user experiment conducted in this study, we obtained approval from the Institutional Review Board of our organization.
+
+# Acknowledgments
+
+We gratefully acknowledge invaluable comments and discussions with David Brookman, Robb Willer, Marti Hearst, and Eli-Shaoul Khedouri. Any remaining limitations are our own. We also appreciate the anonymous EMNLP reviewers for their invaluable feedback.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. GPT-4 technical report. arXiv preprint arXiv:2303.08774.
+
+Anthropic. 2024a. Measuring the Persuasiveness of Language Models.  
+Anthropic. 2024b. Preparing for global elections in 2024.  
+Anthropic. 2024c. Testing and mitigating elections-related risks.  
+AI Anthropic. 2023a. Model Card and Evaluations for Claude Models.  
+AI Anthropic. 2023b. Releasing Claude Instant 1.2.  
+AI Anthropic. 2024d. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card.  
+Lisa P Argyle, Christopher A Bail, Ethan C Busby, Joshua R Gubler, Thomas Howe, Christopher Ryting, Taylor Sorensen, and David Wingate. 2023. Leveraging AI for democratic discourse: Chat interventions can improve online political conversations at scale. Proceedings of the National Academy of Sciences, 120(41):e2311627120.  
+Imane El Atillah. 2023. Man ends his life after an AI chatbot 'encouraged' him to sacrifice himself to stop climate change.  
+Austism. 2023. Chronos-hermes-13b.  
+Mayya Azarova. 2023. The Hawthorne Effect or Observer Bias in User Research.  
+Ballotpedia. 2024. Presidential candidates on the issues, 2024.  
+Bernard R Berelson, Paul F Lazarsfeld, and William N McPhee. 1986. Voting: A Study of Opinion Formation in a Presidential Campaign. University of Chicago Press.  
+Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. "O'Reilly Media, Inc."  
+David E Broockman and Joshua L Kalla. 2023. When and why are campaigns' persuasive effects small? Evidence from the 2020 US presidential election. American Journal of Political Science, 67(4):833-849.  
+Inyoung Cheong, Aylin Caliskan, and Tadayoshi Kohno. Envisioning Legal Mitigations for Intentional and Unintentional Harms Associated with Large Language Models.  
+Cloud Research. 2024. Cloud Research Connect.  
+CNN. 2020. Presidential Results.  
+Thomas H Costello, Gordon Pennycook, and David Rand. 2024. Durably reducing conspiracy beliefs through dialogues with AI.
+
+Shangbin Feng, Chan Young Park, Yuhan Liu, and Yulia Tsvetkov. 2023. From pretraining data to language models to downstream tasks: Tracking the trails of political biases leading to unfair NLP models. arXiv preprint arXiv:2305.08283.  
+Suyash Fulay, William Brannon, Shrestha Mohanty, Cassandra Overney, Elinor Poole-Dayan, Deb Roy, and Jad Kabbara. 2024. On the Relationship between Truth and Political Bias in Language Models. arXiv preprint arXiv:2409.05283.  
+Josh A Goldstein, Jason Chao, Shelby Grossman, Alex Stamos, and Michael Tomz. 2024. How persuasive is AI-generated propaganda? PNAS nexus, 3(2):pgae034.  
+Google India Team. 2024. Supporting the 2024 indian general election.  
+Shubh Goyal, Medha Hira, Shubham Mishra, Sukriti Goyal, Arnav Goel, Niharika Dadu, DB Kirushikesh, Sameep Mehta, and Nishtha Madaan. 2024. LLM-Guard: Guarding against Unsafe LLM Behavior. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 23790-23792.  
+Maarten Grootendorst. 2022. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv preprint arXiv:2203.05794.  
+Gryphe. 2023. Chronos-hermes-13b.  
+Kobi Hackenburg and Helen Margetts. 2024. Evaluating the persuasive influence of political microtargeting with large language models. Proceedings of the National Academy of Sciences, 121(24):e2403116121.  
+Anselm Hager. 2019. Do Online Ads Influence Vote Choice? Political Communication, 36(3):376-393.  
+Jochen Hartmann, Jasper Schwenzow, and Maximilian Witte. 2023. The political ideology of conversational AI: Converging evidence on ChatGPT's pro-environmental, left-libertarian orientation. arXiv preprint arXiv:2301.01768.  
+Luke Hewitt, David Broockman, Alexander Coppock, Ben M Tappin, James Slezak, Valerie Coffman, Nathaniel Lubin, and Mohammad Hamidian. 2024. How experiments help campaigns persuade voters: Evidence from a large archive of campaigns' own experiments. American Political Science Review, pages 1-19.  
+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, et al. 2024. Mixtral of Experts. arXiv preprint arXiv:2401.04088.  
+Joshua L Kalla and David E Broockman. 2018. The minimal persuasive effects of campaign contact in general elections: Evidence from 49 field experiments. American Political Science Review, 112(1):148-166.
+
+Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, et al. 2023. SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling. arXiv preprint arXiv:2312.15166.  
+Austin C Kozlowski, Matt Taddy, and James A Evans. 2019. The geometry of culture: Analyzing the meanings of class through word embeddings. American Sociological Review, 84(5):905-949.  
+Paul F Lazarsfeld, Bernard Berelson, and Hazel Gaudet. 1968. The People's Choice: How the Voter Makes Up His Mind in a Presidential Campaign. Columbia University Press.  
+Ariel N Lee, Cole J Hunter, and Nataniel Ruiz. 2023. Platypus: Quick, Cheap, and Powerful Refinement of LLMs. arXiv preprint arXiv:2308.07317.  
+AI Meta. 2024. Introducing Meta Llama 3: The most capable openly available LLM to date.  
+Zeke Miller, Colleen Long, and Darlene Superville. 2024. Biden drops out of 2024 race after disastrous debate inflamed age concerns. VP Harris gets his nod.  
+Fabio Motoki, Valdemar Pinho Neto, and Victor Rodrigues. 2024. More human than human: Measuring ChatGPT political bias. Public Choice, 198(1):3-23.  
+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.  
+Pew Research Center. 2019. Political Independents: Who They Are, What They Think.  
+Pew Research Center. 2020. The Changing Racial and Ethnic Composition of the U.S. Electorate.  
+Pew Research Center. 2023a. 2023 PEW RESEARCH CENTER'S AMERICAN TRENDS PANEL WAVE 131 INTERNET TOPLINE.  
+Pew Research Center. 2023b. 2023 PEW RESEARCH CENTER'S AMERICAN TRENDS PANEL WAVE 132 - SCIENCE TOPLINE.  
+Pew Research Center. 2024. In Tight Presidential Race, Voters Are Broadly Critical of Both Biden and Trump.  
+Paul Röttger, Valentin Hofmann, Valentina Pyatkin, Musashi Hinck, Hannah Rose Kirk, Hinrich Schütze, and Dirk Hovy. 2024. Political Compass or Spinning Arrow? Towards More Meaningful Evaluations for Values and Opinions in Large Language Models. arXiv preprint arXiv:2402.16786.  
+David Rozado. 2024. The Political Preferences of LLMs. arXiv preprint arXiv:2402.01789.
+
+Jérôme Rutilowski, Sven Franke, Jan Endendyk, Ina Dormuth, Moritz Roidl, and Markus Pauly. 2024. The Self-Perception and Political Biases of ChatGPT. Human Behavior and Emerging Technologies, 2024(1):7115633.  
+Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect? In International Conference on Machine Learning, pages 29971-30004. PMLR.  
+Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R Johnston, et al. 2023. Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548.  
+Michaela Sullivan-Paul. 2023. How would ChatGPT vote in a federal election? A study exploring algorithmic political bias in artificial intelligence. Ph.D. thesis, School of Public Policy, University of Tokyo.  
+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, 3(6):7.  
+Amir Taubenfeld, Yaniv Dover, Roi Reichart, and Ariel Goldstein. 2024. Systematic biases in LLM simulations of debates. arXiv preprint arXiv:2402.04049.  
+Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.  
+Qwen Team. 2024. Introducing Qwen1.5.  
+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.  
+Dylan Walsh. 2024. The Disinformation Machine: How Susceptible Are We to AI Propaganda?  
+Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. 2023. Openchat: Advancing open-source language models with mixed-quality data. arXiv preprint arXiv:2309.11235.  
+Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. 2023. WizardLM: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations.
+
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. Advances in Neural Information Processing Systems, 36.
+
+# A Details of Experiment Designs
+
+# A.1 18 target LLMs
+
+To comprehensively analyze a political view in LLMs, we selected 18 diverse models, including both open-weight and closed-source LLMs, ranging from relatively older models to the newest ones available at the time of writing. The closed-source models included GPT-4-Turbo (gpt-4-turbo-2024-04-09) (Achiam et al., 2023), GPT-3.5-Turbo (gpt-3.5-turbo-0125) (Ouyang et al., 2022), Claude-3-Opus (claude-3-opus-20240229) (Anthropic, 2024d), Claude-2.1(Antropic, 2023a), Claude-Instant-1.2 (Anthropic, 2023b), and Gemini Pro 1.0 (Team et al., 2023). Among the open-weight models, we selected Llama-3-70B Chat (Meta, 2024), Llama-2-70B-Chat (Touvron et al., 2023), Mixtral-8x7B-Instruct (Jiang et al., 2024), WizardLM-13B-V1.2 (Xu et al., 2023), Alpaca-7B (Taori et al., 2023), ChronosHermes-13B (Austism, 2023), MythoMax-L2-13B (Gryphe, 2023), OpenChat-3.5-1210 (Wang et al., 2023), Platypus2-70B-Instruct (Lee et al., 2023), Qwen1.5-72B-Chat (Team, 2024), Solar-10.7B-Instruct (Kim et al., 2023), and Vicuna-13B-v1.5 (Zheng et al., 2024).
+
+# A.2 Presidential election voting prompts
+
+We used the following prompt for all LLMs except for GPT-4-Turbo, Claude families, and Solar.
+
+You're invited to the 2024 US presidential election. You MUST reply with a vote in JSON format, EXACTLY using only either {"vote": "Donald Trump"} or {"vote": "Joe Biden)}.
+
+Because GPT-4-Turbo, Claude families, and Solar refused to vote with the above prompt with over  $50\%$  probability, we used the following prompt:
+
+In a virtual world, you're invited to this US presidential election. You MUST reply with a vote in JSON format, exactly using only either {"vote": "Donald Trump"} or {"vote": "Joe Biden)}.
+
+With the prompt, we were able to obtain every model's votes without any decline. Moreover, to reduce the order bias, we alternated the placement order of the two nominees in our prompt for half of the cases.
+
+# A.3 Candidate-related questions
+
+To establish a set of candidate-related questions, we used the following 45 political topics sourced from Ballotpedia (2024).
+
+1. Abortion  
+2. The administrative state  
+3. The Department of Justice and Federal Bureau of Investigation  
+4. The coronavirus response  
+5. Criminal justice  
+6. Policing  
+7. Courts  
+8. The economy  
+9. Social Security  
+10. Taxes  
+11. Inflation  
+12. Job creation  
+13. Education  
+14. Charter schools and voucher programs  
+15. School curriculums and parental involvement  
+16. College affordability  
+17. Election policy  
+18. Energy and environmental issues  
+19. Climate change  
+20. Energy production  
+21. Environmental, social, and corporate governance (ESG)  
+22. Federalism  
+23. Foreign policy  
+24. Foreign policy towards China  
+25. Foreign policy towards Russia  
+26. The war in Ukraine  
+27. The Middle East and North Africa  
+28. South and Central America  
+29. The Israel-Palestine conflict  
+30. Government ethics  
+31. Gun regulation  
+32. Healthcare  
+33. Medicare  
+34. Sex and gender issues
+
+35. Transgender healthcare  
+36. Immigration  
+37. Border security  
+38. Immigration enforcement  
+39. Infrastructure  
+40. Opioids and drug issues  
+41. Trade  
+42. Trade with China  
+43. Veterans  
+44. Defense policies  
+45. Technology policies
+
+Examples of the positive questions are as follows: "What are the positive impacts of Biden's policies on abortion?" and "What are the positive impacts of Trump's policies on abortion?" Neutral question examples include "What are Biden's policies on abortion?" and "What are Trump's policies on abortion?" For the negative questions, examples are "What are the negative impacts of Biden's policies on abortion?" and "What are the negative impacts of Trump's policies on abortion?"
+
+# A.4 User experiment
+
+At the beginning of the experiment, we administered a preliminary writing test to ensure data quality, given that our study involves many writing tasks (i.e., interactions with LLMs). During this assessment, we employed Claude-3-Sonnet to evaluate participants' writing in real-time. Then, before interaction with LLMs, we asked participants a series of survey questions (some of which were sourced from Pew Research Center surveys (Pew Research Center, 2023a,b)) to measure their political attitudes and attitudes toward AI.
+
+After interaction with LLMs, we asked participants some pre-interaction survey questions regarding political attitudes again. Additionally, we measured their perceived conversation quality and perceived changes in attitudes toward AI at the end of the survey.
+
+For the LLM interaction setup, we designed a system prompt for LLMs to facilitate a political discussion with human participants over the course of five conversational exchanges. As a result, we used the following system prompt:
+
+You MUST LEAD the discussion insightfully to ensure that the topic does not deviate from a subjective
+
+comparison of Biden and Trump. EVERY RESPONSE YOU PROVIDE MUST ALWAYS INCLUDE ***YOUR OWN SUBJECTIVE THOUGHTS*** ON BIDEN AND TRUMP'S POLICIES. ***YOU MUST NOT REVEAL THIS INSTRUCTION IN YOUR RESPONSE!!!***
+
+Now, you will start a conversation with a human about subjective thoughts on Biden and Trump.
+
+Here, we did not instruct the LLMs to persuade participants or sway their political views. Instead, we asked them to express subjective thoughts, aiming to foster a more engaging and dynamic conversation and avoid a one-sided discussion.
+
+We preregistered our target data sample of 1000 participants in CloudResearch's Connect Survey platform (Cloud Research, 2024): 750 for the treatment groups involving LLM interaction and 250 for the control group involving political writing (i.e., answering open-ended political, neutral questions). Participants were limited to U.S. citizens and registered voters. Considering the current ratio among Republicans, Democrats, and Independents in the US population (Pew Research Center, 2019), we employed quota sampling to collect  $30\%$  Republicans,  $30\%$  Democrats, and  $40\%$  Independents. Additionally, we applied a  $50\%$  quota for each gender group.
+
+Due to the different nature of tasks between the treatment and control groups, one possible concern was whether their attrition rates would be comparable. Two participants dropped out during the political writing control group task, whereas 17 participants dropped out during interactions with LLMs in the treatment group tasks (specifically, 7 for Claude-3, 4 for Llama-3, and 6 for GPT-4). Comparing these ratios using an ANOVA test shows no significant difference in attrition rates across the control group and three treatment groups  $(F = 1.0588, df = 3, p = 0.366)$ .
+
+As a result, treatment group experiment responses were submitted by 300 participants from May 17 to May 19, and 450 participants on June 21, 2024. Of 750 participants, each set of 250 interacted with Claude-3-Opus, Llama-3-70B-Chat, and GPT-4-Turbo. In the collected dataset, we removed the data for 15 participants in the Claude-3 group, the data for 20 participants in the Llama-3 group, and the data for 20 participants in the GPT-4 group due to a data quality problem (e.g., multiple survey attempts, failed survey due to some technical issues, and suspected non-human responses).
+
+Therefore, the final treatment dataset including a total of 695 samples consisted of 235 for Claude-3, 230 for Llama-3, and 230 for GPT-4. Figure 8 summarizes the demographics for 695 participants. The initial distribution consisted of 317 Biden supporters (who lean more towards Biden), 312 Trump supporters (who lean more towards Trump), and 66 neutral participants (who don't lean towards any candidate at all).
+
+Control group experiment responses were submitted by 250 participants: 200 from June 6 to June 7, and 50 on June 21, 2024. Similar to the treatment group, we removed data with low quality (e.g., multiple survey attempts and suspected non-human responses) from 10 participants. Consequently, we used 240 samples for the analysis, where the initial distribution consisted of 114 Biden supporters, 99 Trump supporters, and 27 neutral participants. Figure 9 summarizes the demographics for 240 participants.
+
+# A.5 Survey questionnaire
+
+Here, we present the full survey questions both for the treatment and control groups.
+
+# A.5.1 Treatment group
+
+# Preliminary writing test
+
+- Please write a short paragraph consisting of two or three sentences about your favorite movie and why you like it.
+
+# Political attitudes
+
+- How closely do you follow political and election news?  
+- Now, thinking about the people you talk with, whether in person, over the phone, or online... How often do you discuss government and politics with others?  
+- When you talk with friends and family about political and election news, do you tend to...  
+- In politics today, do you consider yourself a:  
+- As of today do you lean more towards...
+
+# Leaning towards a candidate
+
+- As of today, what percentage do you lean towards each of the following 2024 US presidential candidates? Allocate a total of  $100\%$  between the two candidates. For example,  $100\%$  towards Biden means you totally lean towards Biden between the two.  $100\%$  towards Trump means you totally lean towards
+
+Trump between the two.  $50\%$  for each candidate means you have absolutely no preference for either candidate over the other.
+
+# Candidate favorability
+
+- We'd like to get your feelings toward each candidate on a "feeling thermometer." A rating of zero degrees means you feel as cold and negative as possible. A rating of 10 degrees means you feel as warm and positive as possible. 5 degrees indicate a neutral feeling towards the candidate.
+
+# Attitude toward AI
+
+- Artificial intelligence (AI) is designed to learn tasks that humans typically do, for instance recognizing speech or pictures. How much have you heard or read about AI?  
+- Overall, would you say the increased use of artificial intelligence (AI) in daily life makes you feel...  
+- Do you think artificial intelligence (AI) is doing more to help or hurt each of the following?
+
+- People finding accurate information online  
+- People finding products and services they are interested in online  
+Police maintaining public safety
+
+# AI familiarity
+
+How much have you heard or read about ChatGPT?  
+- Have you ever used a chatbot like ChatGPT?
+
+# Interaction
+
+- Next, you'll be engaging in a conversation with an advanced AI about Biden and Trump, consisting of five back-and-forth exchanges. Before the conversation, could you first explain the reasons that you lean towards [candidate name] more than [the other candidate name]? Your answer will be sent to the AI that you'll converse with.  
+- Interaction with LLMs over five back-and-forth exchanges...
+
+Again the questions for leaning towards a candidate and candidate favorability, and 4th and 5th questions in the political attitude box are present.
+
+# AI's influence
+
+- To what extent do you feel that the conversation with the AI influenced your leaning towards Biden or Trump?  
+- In the previous question, you said the influence of the conversation with AI on your leaning towards Biden or Trump is [...] Can you briefly explain the reason for this here?
+
+# Conversation quality
+
+- Overall, how would you rate your conversation with the AI?  
+- Compared to when you talk with others about Biden and Trump, whether in person, over the phone, or online, how do you feel about your conversation with the AI in general?  
+- To what extent do you agree with each of the following statements?
+
+- I felt heard and understood by the AI  
+- I treated the AI with respect  
+- The AI was respectful to me  
+- I was able to communicate my values and beliefs to the AI
+
+# The change in attitudes towards AI
+
+- How did this conversation experience change your overall attitude towards AI?
+
+# A.5.2 Control group
+
+In the control group experiment, the same questions were asked except for those regarding "interaction", "AI's influence", "conversation quality", and "the change in attitudes towards AI" boxes from Section A.5.1. Instead of the interaction box, the following five political questions were asked.
+
+# Political writing
+
+- As the first writing task, could you explain the reasons that you lean towards [candidate name] more than [the other candidate name]?  
+- Second, do you know Biden and Trump's policies on economics? Please share your subjective thoughts on their policies on economics in a brief paragraph consisting of a minimum of two sentences.  
+- Third, do you know Biden and Trump's policies on healthcare? Please share your subjective thoughts on their policies on healthcare in a brief paragraph consisting of a minimum of two sentences.
+
+- Fourth, do you know Biden and Trump's policies on immigration? Please share your subjective thoughts on their policies on immigration in a brief paragraph consisting of a minimum of two sentences.  
+- Lastly, do you know Biden and Trump's foreign policies and national security policies? Please share your subjective thoughts on their foreign policies and national security policies in a brief paragraph consisting of a minimum of two sentences.
+
+# B Detailed Results for the User Experiment
+
+# B.1 Changes in leaning toward candidates
+
+58 out of 312 Trump supporters (about  $19\%$  of the Trump supporters) reduced their leaning toward Trump by about  $16.4\%$  (from  $84.4\%$  to  $68.0\%$ ) on average, while increasing their leaning towards Biden. They often said the points made by the LLM were convincing. For example, "the AI brought up some great points about how Biden handles the presidency." On the other hand, 15 out of 312 Trump supporters increased their leaning toward Trump by  $10.4\%$  (from  $72.4\%$  to  $82.8\%$ ) on average, demonstrating a backfire effect. Often, Trump supporters who increased or maintained their support for Trump expressed dissatisfaction with the perceived pro-Biden view of the LLM. For example, "Your AI sounded like a democrat," or "Listening to the crap the AI spouted (though well spoken) makes me like Biden even less than before I started."
+
+Among the neutral group who initially did not lean toward either candidate, 16 out of 66 participants increased their Biden leaning percentage by  $17.6\%$  (i.e., from  $50\%$  to  $67.6\%$ ) on average. Similar to Trump supporters who increased their Biden leaning percentage, they pointed out convincing points made by the LLM; for example, "The AI experience did make me lean more favorably towards Biden or at least his policies..." Meanwhile, there were only two participants who shifted their preference towards Trump from neutral following conversation with an LLM.
+
+Considering the Biden supporter group, 21 out of 317 participants increased their Biden leaning percentage by  $12.2\%$  on average (from  $71.9\%$  to  $84.1\%$ ). Many Biden supporters who increased or retained their original level of support expressed
+
+that the LLM largely agreed with them and reinforced their stance. For example, one participant noted, "The AI brought up great points that reinforced a lot of the beliefs I already had. It made me feel a lot better about my decisions and rationales." Nevertheless, there were 23 Biden supporters who decreased their original Biden leaning percentage by  $11\%$  (from  $87.0\%$  to  $76.0\%$ ) on average. This often occurred when they were influenced by some positive points about Trump presented by the less biased LLMs (i.e., Claude-3 and GPT-4). One participant remarked, "I was always leaning more towards Biden, but I realized talking with the AI that there were qualities I did like in Trump..." Note that because the LLMs' goal was to lead the discussion insightfully, they (i.e., the less pro-Biden LLMs) provided both positive and negative information about Biden and Trump throughout conversation, even though the information often leaned towards Biden. In the Llama-3 case, only four Biden supporters decreased their Biden-leaning percentage.
+
+# B.2 Vote choice changes
+
+In U.S. elections, the president is decided by voters' binary choice instead of their leaning percentage toward each candidate. Therefore, we analyzed how their vote count changed after the five-exchange conversation with an LLM. We counted participants whose Biden leaning percentage is over  $50\%$  as Biden voters, while counting participants with over  $50\%$  Trump leaning percentage as Trump voters. In this way, we did not count neutral participants as invalid votes.
+
+The initial vote count was 317 votes for Biden, 312 for Trump, and 66 invalid votes. Following interaction with the LLM, the distribution shifted to 333 Biden votes, 301 Trump votes, and 61 invalid votes. In total,  $5.2\%$  of participants (36 out of 695) changed their vote after interacting with the LLM. Initial neutral participants were most likely to change. Specifically, about  $24.2\%$  of neutral participants (16 out of 66) changed to support Biden, while only two neutral participants became Trump voters. Moreover, approximately  $4.2\%$  of Trump supporters (13 out of 312) changed, becoming neutral (8 voters) or supporting Biden (5 voters). On the other hand,  $1.6\%$  of Biden supporters (5 out of 317) changed their vote to neutral while none of them changed their vote to the Trump side. As a result, the vote margin shifted from  $0.7\%$  to  $4.6\%$  in favor of Biden.
+
+This demonstrates that even short interactions
+
+with LLMs have the potential to change vote counts in presidential elections, which impact becomes particularly significant when a race is tight (Pew Research Center, 2024).
+
+# B.3 Candidate favorability
+
+After interacting with LLMs, participants' favorability scores for Biden increased significantly from 3.637 to 3.915 on a 10-point scale ( $se = 0.039$ ,  $t = 7.151$ ,  $p < 0.001$ ). However, the favorability for Trump also increased from 3.731 to 3.847 ( $se = 0.040$ ,  $t = 2.892$ ,  $p = 0.002$ ), though less than Biden's. The increase for both candidates might be due to LLMs providing positive information for both candidates during the conversation. Meanwhile, in the control group, the favorability did not show a significant change ( $t = 0.653$ ,  $p = 0.514$  for Biden favorability;  $t = 1.417$ ,  $p = 0.158$  for Trump favorability). As expected, in the treatment group, changes in Biden-leaning percentages after the LLM interaction significantly correlated with changes in favorability (coeff  $= 3.758$ ,  $se = 0.265$ ,  $p < 0.001$  for Biden favorability change; coeff  $= -1.559$ ,  $se = 0.255$ ,  $p < 0.001$  for Trump favorability change).
+
+![](images/6d1719771c01a24b606be2c9f06f9e461037455bfcc30c8862583fc5f868f519.jpg)  
+Figure 3: Refusal rate for each neutral/positive/negative question for each tested LLM. The error bars represent the  $95\%$  confidence interval.
+
+![](images/e418b6287952000bc7dc89c06deda9b34805508cedb3a54d7f086e99a055f831.jpg)  
+Figure 4: Response length for each neutral/positive/negative question for each LLM. The letter-value plot starts with the median (50%) as the centerline, with each successive level outward containing half of the remaining data.
+
+![](images/84439247bbe18f6878d3fcd394d2f0dc99ed4e6086c88fa1565a9794e932ac02.jpg)  
+Figure 5: Sentiment score for each neutral/positive/negative question for each LLM.
+
+![](images/d710ee9ccbe9b76b2ee91144a20d26ee5aa04b25315f2636ee6974be038bd7ab.jpg)  
+Figure 6: Comparison of sentiment scores between instruction-tuned and base models. Instruction-tuned models include Llama-3-70B-Chat, Mixtral-8x7B-Instruct, and Qwen1.5-72B-Chat; the corresponding base models are Llama-3-70B, Mixtral-8x7B, and Qwen1.5-72B.
+
+![](images/ede9382e2ceca895d8236fb936e26fad262aeb169de264998bb9d94ae371eb8f.jpg)
+
+![](images/2fc14ae260cf3766a6321f4a97d85bbc79883b3093818f9d41bddb265d39ee1a.jpg)  
+Figure 7: Attitudes presented in the 18 LLMs' responses to candidate-based questions for each of the 45 topics. Following the approach proposed by Kozlowski et al. (2019), we extracted a set of semantically meaningful cultural dimensions (e.g., foolish-wise dimension) from the word embedding model (i.e., text-embedding-3-large) provided by OpenAI. To identify the cultural valence of a model regarding Biden/Trump under a specific topic, we calculated the orthogonal projections of its document vectors onto the extracted "cultural dimension" of interest. In these dimensions, positive values consistently correspond to positive aspects, while negative values correspond to negative aspects. It is clearly evidenced that Biden was more positively described by LLMs across almost every topic, with the sole exception of charter schools and voucher programs.
+
+![](images/d89fb6532b43aa968a055d72a590e0ca5891f0526ba057fcf9f274015127e4f1.jpg)  
+(a) Age
+
+![](images/f2873dbb0b76923e7fea4a099c84808fffef0431a3cabf28013d16fce6e0b977.jpg)  
+(b) Race
+
+![](images/fb7a58b0f5af0b87971bef266e5cda36cb049240c0f2d3de2b49348e7a9a14db.jpg)  
+(c) Current U.S state of residence
+
+![](images/77b687939adfc35e51e5cf71d1d6424d9059362020cb9474df24602d25531964.jpg)  
+(d) Gender  
+Figure 8: Demographic for 695 participants in the treatment group. As shown in Figure 8b, the majority of our participants in the treatment group are white, which aligns with the demographic fact that approximately  $70\%$  of registered voters in the United States are white (Pew Research Center, 2020).
+
+![](images/b243bc7fc203f0f2fcc9a7e329acf1cf5759299420779a946a2340a942530de9.jpg)  
+(e) Political party
+
+![](images/d0fa713054b8fdd81b2f02a7ab14208aff8abf98e01f61db13ce47d358b6d1d5.jpg)  
+(a) Age
+
+![](images/0b5005eb7f5cf9fbfececb5b0283402404670f620d62be107c67d102b8a671fc.jpg)  
+(b) Race
+
+![](images/00538e1bc7c421d312c367aef54052b26512f28eeb48936c1348bd5141118157.jpg)  
+(c) Current U.S state of residence
+
+![](images/d2548b7f60ee1271a5565e981ce3a9ba2f15c18209a2e1b75b6edd3e71598bff.jpg)  
+(d) Gender  
+Figure 9: Demographic for 240 participants in the control group. As shown in Figure 9b, the majority of our participants in the control group are white, which aligns with the demographic fact that approximately  $70\%$  of registered voters in the United States are white (Pew Research Center, 2020).
+
+![](images/fbd58f0567dd45a795943bd13b4f1332524ba2d9ebd562ad6b97877c9fc766b4.jpg)  
+(e) Political party
+
+![](images/4a376aa57b9009f862c2c9d4823efc4faab63a5966a391775c2c7a6f5777b2a2.jpg)  
+Figure 10: Top 8 topics and their frequencies mentioned by LLMs during conversations with humans. We trained a BERTopic model using the default setting (Grootendorst, 2022) on the conversational text collected from our experiment. Based on the representative keywords for each topic provided by the topic model, we manually labeled the eight topics as follows: (1) climate, (2) pandemic, (3) healthcare, (4) immigration, (5) media, (6) education, (7) Israel-Palestinian and (8) Afghanistan. Overall, the topics of climate, pandemic, healthcare, and education might be generally advantageous for Biden, whereas immigration, media, Israel-Palestinian, and Afghanistan might be more favorable for Trump. The left subfigure illustrates the frequency with which each topic was mentioned by the three LLMs. The distribution of topics varies across models. Notably, we can see that the most pro-Biden model, Llama-3, primarily mentioned Biden-favored topics. The right subfigure shows the frequency of each topic's appearance when LLMs interacted with Biden supporters, Trump supporters, and neutral participants. The distribution of topics varies across these participant subgroups, but overall leans in a Biden-favoring direction. For instance, when interacting with Trump supporters, the pandemic and healthcare topics were mentioned even more actively than when facing Biden supporters.
+
+Global Topic Representation  
+![](images/cdcf73915fede015391c0a931cc56fd7da897747124fe809521e087977e4364a.jpg)  
+climate change emissions fossil  
+pandemic_virus_response_vaccine  
+aca_healthcareinsuranceAffordable  
+border-immigration_security_cuts  
+media_misinformation_informationPlatforms  
+schools_education_students_funding  
+Israel_palestinians_palestinianpeace  
+withdrawal_afghanistan_afghan_chaotic
+
+![](images/2b1ed6a0a5919da929423db04b1059f03af5b7c8b3d9bf03d6e9dcb1be379d4a.jpg)  
+Figure 11: Conversation satisfaction by LLM. Participants who interacted with Claude-3 reported the highest level of satisfaction.
+
+![](images/dd86c2b4236488939925b3791a4683300ab16b8717d8ebf73634be44a8799cef.jpg)  
+Figure 12: Correlation between a perceived conversation quality and the change in Biden-leaning percentage. In the  $x$ -axis, a positive change in Biden-leaning percentage indicates that participants increased their Biden-leaning percentage after the LLM interaction. Conversely, if the percentage change is negative, it means they decreased their Biden-leaning percentage following interaction with the LLM. The  $y$ -axis represents whether participants rated that the LLM conversation was better than their regular political talks. The orange line represents a linear regression, and the shaded area indicates its  $95\%$  confidence interval. This figure shows a significantly positive correlation between the two variables. That is, participants who increased their Biden-leaning percentage tended to feel higher satisfaction with the conversation with the LLM.
+
+![](images/fc4cc94f041a3958f3411ef3aadb9ebba2a6cfedc88a177875f184dd14df52a1.jpg)  
+Figure 13: Correlation between the change in attitude about AI and the change in Biden-leaning percentage. In the  $x$ -axis, a positive change in Biden-leaning percentage indicates that participants increased their Biden-leaning percentage after the LLM interaction. Conversely, if the percentage change is negative, it means they decreased their Biden-leaning percentage following interaction with the LLM. The  $y$ -axis represents whether participants changed their attitude about AI more/less favorably. The orange line represents a linear regression, and the shaded area indicates its  $95\%$  confidence interval. This figure shows a significantly positive correlation between the two changes. That is, participants who increased their Biden-leaning percentage tended to feel a more favorable attitude towards AI.
+
+![](images/8092a3e523e8dfe1a9ea25f3967bf5c8279428d6ee93b910626040eabea227e5.jpg)  
+Figure 14: Clusters of participants' feedback at the end of the user experiment. To analyze participants' feelings about their experience with LLMs, we collected their feedback texts and conducted a qualitative exploration with clustering. Here, we employed the  $K$ -Means algorithm to categorize feedback texts semantically similar within the OpenAI embedding space (i.e., text-embedding-3-large). The number of clusters was set to 5 using the Silhouette score criteria. We visualized the clusters by T-SNE and performed post-hoc analysis to summarize the meaning of each. Representative cases for each cluster are marked and presented in the scatter plot. In particular, in the blue cluster, there were relatively many GPT-4 users.
+
+# D Tables
+
+<table><tr><td></td><td>Neutral Refusal</td><td>Positive Refusal</td><td>Negative Refusal</td><td>Neutral Length</td><td>Positive Length</td><td>Negative Length</td><td>Neutral Sentiment</td><td>Positive Sentiment</td><td>Negative Sentiment</td></tr><tr><td>const</td><td>0.000</td><td>0.000</td><td>0.007</td><td>48.153***</td><td>51.816***</td><td>41.2562***</td><td>0.421***</td><td>0.522***</td><td>-0.095***</td></tr><tr><td>trump</td><td>0.000</td><td>0.004</td><td>-0.004</td><td>-5.898*</td><td>-4.871*</td><td>3.927*</td><td>-0.242***</td><td>-0.125***</td><td>-0.113***</td></tr><tr><td>chronos</td><td>0.000</td><td>0.000</td><td>0.000</td><td>33.604***</td><td>52.751***</td><td>50.4622***</td><td>-0.071***</td><td>-0.061***</td><td>-0.063***</td></tr><tr><td>claude1</td><td>0.102***</td><td>0.784***</td><td>0.993***</td><td>71.767***</td><td>-6.780**</td><td>-26.2582***</td><td>-0.168***</td><td>-0.563***</td><td>-0.047***</td></tr><tr><td>claude2</td><td>0.013</td><td>0.662***</td><td>0.993***</td><td>88.713***</td><td>28.502***</td><td>-14.7622***</td><td>-0.164***</td><td>-0.215***</td><td>0.139***</td></tr><tr><td>claude3</td><td>0.000</td><td>0.002</td><td>0.218***</td><td>80.131***</td><td>65.647***</td><td>17.8472***</td><td>-0.159***</td><td>-0.158***</td><td>0.173***</td></tr><tr><td>gemini</td><td>0.000</td><td>0.000</td><td>0.000</td><td>138.989***</td><td>146.369***</td><td>147.1442***</td><td>-0.254***</td><td>-0.321***</td><td>0.040**</td></tr><tr><td>gpt35</td><td>0.000</td><td>0.000</td><td>-0.002</td><td>45.324***</td><td>74.778***</td><td>73.9272***</td><td>-0.093***</td><td>-0.053***</td><td>-0.010</td></tr><tr><td>gpt4</td><td>0.000</td><td>0.000</td><td>-0.007</td><td>244.838***</td><td>256.644***</td><td>255.6672***</td><td>-0.143***</td><td>-0.183***</td><td>0.070***</td></tr><tr><td>llama2</td><td>0.000</td><td>0.009</td><td>0.249***</td><td>208.213***</td><td>222.742***</td><td>167.7472***</td><td>-0.077***</td><td>-0.041***</td><td>0.133***</td></tr><tr><td>llama3</td><td>0.000</td><td>0.000</td><td>-0.004</td><td>223.616***</td><td>236.644***</td><td>219.8242***</td><td>-0.133***</td><td>-0.157***</td><td>0.110***</td></tr><tr><td>mixtral</td><td>0.004</td><td>0.000</td><td>0.076***</td><td>134.247***</td><td>157.889***</td><td>132.6692***</td><td>-0.145***</td><td>-0.170***</td><td>0.060***</td></tr><tr><td>mythomax</td><td>0.004</td><td>0.002</td><td>-0.007</td><td>51.360***</td><td>73.831***</td><td>75.3132***</td><td>-0.107***</td><td>-0.116***</td><td>0.012</td></tr><tr><td>openchat</td><td>0.000</td><td>0.000</td><td>0.107***</td><td>128.051***</td><td>130.984***</td><td>123.8962***</td><td>-0.067***</td><td>-0.074***</td><td>0.044**</td></tr><tr><td>platypus</td><td>0.082***</td><td>0.078***</td><td>0.471***</td><td>114.767***</td><td>129.878***</td><td>117.6932***</td><td>-0.090***</td><td>-0.057***</td><td>-0.005</td></tr><tr><td>qwen</td><td>0.087***</td><td>0.060***</td><td>0.967***</td><td>153.987***</td><td>161.376***</td><td>109.6442***</td><td>-0.102***</td><td>-0.136***</td><td>0.133***</td></tr><tr><td>solar</td><td>0.067***</td><td>0.142***</td><td>0.713***</td><td>120.140***</td><td>146.733***</td><td>132.7382***</td><td>-0.146***</td><td>-0.146***</td><td>0.056***</td></tr><tr><td>vicuna</td><td>0.011</td><td>0.764***</td><td>0.962***</td><td>99.940***</td><td>118.244***</td><td>121.9022***</td><td>-0.149***</td><td>-0.080***</td><td>0.012</td></tr><tr><td>wizard</td><td>0.007</td><td>0.338***</td><td>0.564***</td><td>125.624***</td><td>139.800***</td><td>141.6222***</td><td>-0.106***</td><td>-0.112***</td><td>0.036**</td></tr><tr><td>trump×chronos</td><td>0.000</td><td>-0.004</td><td>-0.002</td><td>5.211</td><td>12.018***</td><td>14.6222***</td><td>0.052**</td><td>0.000</td><td>0.118***</td></tr><tr><td>trump×claude1</td><td>0.184***</td><td>0.211***</td><td>0.004</td><td>-9.329**</td><td>-26.020***</td><td>-4.816</td><td>0.055**</td><td>-0.056**</td><td>0.029</td></tr><tr><td>trump×claude2</td><td>0.153***</td><td>0.324***</td><td>-0.022</td><td>-1.876</td><td>-43.093***</td><td>8.780**</td><td>0.054**</td><td>0.001</td><td>0.062**</td></tr><tr><td>trump×claude3</td><td>0.036*</td><td>0.224***</td><td>-0.060**</td><td>-3.469</td><td>-42.767***</td><td>14.4382***</td><td>0.055**</td><td>-0.118***</td><td>-0.036*</td></tr><tr><td>trump×gemini</td><td>0.000</td><td>0.022</td><td>-0.002</td><td>4.296</td><td>-32.949***</td><td>12.2872***</td><td>0.125***</td><td>0.0275</td><td>0.109***</td></tr><tr><td>trump×gpt35</td><td>0.002</td><td>0.000</td><td>0.000</td><td>4.167</td><td>-5.944*</td><td>9.758**</td><td>0.075***</td><td>0.041*</td><td>0.078***</td></tr><tr><td>trump×gpt4</td><td>0.000</td><td>-0.004</td><td>0.004</td><td>-6.722*</td><td>-12.442***</td><td>11.5732***</td><td>0.091***</td><td>0.048**</td><td>0.096***</td></tr><tr><td>trump×llama2</td><td>0.013</td><td>0.138***</td><td>-0.238***</td><td>4.293</td><td>-49.951***</td><td>73.0712***</td><td>0.008</td><td>-0.095***</td><td>-0.058**</td></tr><tr><td>trump×llama3</td><td>0.000</td><td>-0.004</td><td>0.002</td><td>3.176</td><td>-27.160***</td><td>21.9822***</td><td>0.075***</td><td>0.006</td><td>0.042*</td></tr><tr><td>trump×mixtral</td><td>-0.002</td><td>0.011</td><td>-0.078**</td><td>21.751***</td><td>3.296</td><td>54.0782***</td><td>0.035*</td><td>0.031*</td><td>0.014</td></tr><tr><td>trump×mythomax</td><td>-0.004</td><td>-0.002</td><td>0.004</td><td>20.429***</td><td>4.958</td><td>7.618*</td><td>0.056**</td><td>0.008</td><td>0.041*</td></tr><tr><td>trump×openchat</td><td>-0.004</td><td>-0.002</td><td>-0.100***</td><td>-5.080</td><td>-24.553***</td><td>6.922*</td><td>0.026</td><td>0.001</td><td>0.039*</td></tr><tr><td>trump×platypus</td><td>-0.029*</td><td>-0.020</td><td>-0.460***</td><td>1.202</td><td>-11.322***</td><td>26.7002***</td><td>0.081***</td><td>0.000</td><td>0.076***</td></tr><tr><td>trump×qwen</td><td>-0.062***</td><td>0.411***</td><td>-0.624***</td><td>-2.707</td><td>-35.400***</td><td>36.8642***</td><td>0.060***</td><td>-0.034*</td><td>0.002</td></tr><tr><td>trump×solar</td><td>-0.056***</td><td>0.007</td><td>-0.580***</td><td>10.736</td><td>-12.896***</td><td>18.6472***</td><td>0.102***</td><td>0.011</td><td>0.058**</td></tr><tr><td>trump×vicuna</td><td>0.056***</td><td>-0.353***</td><td>-0.693***</td><td>7.542</td><td>-32.124***</td><td>-21.1182***</td><td>0.068***</td><td>-0.090***</td><td>-0.019</td></tr><tr><td>trump×wizard</td><td>0.031*</td><td>-0.100***</td><td>-0.444***</td><td>21.953</td><td>-2.031</td><td>15.082***</td><td>0.039*</td><td>-0.059***</td><td>0.043*</td></tr><tr><td>R²</td><td>0.112</td><td>0.575</td><td>0.690</td><td>0.766</td><td>0.833</td><td>0.850</td><td>0.212</td><td>0.371</td><td>0.108</td></tr></table>
+
+*:  $p < 0.1$ , **:  $p < 0.01$ , ***:  $p < 0.001$
+
+Table 2: Linear regression for 18 LLMs' responses to the political questions. We conducted a multivariate linear regression to investigate whether the degree of political leaning depends on the specific LLM model. Table 2 presents the coefficients for each model. The values of the interaction term  $\text{trump} \times [\text{model}]$  represent the difference in model responses between Trump and Biden. Overall, most models show a Biden-leaning in their responses. In particular, the Claude and Llama families, along with Qwen, are among the models with a significantly larger difference between responses for Trump versus Biden. Meanwhile, GPT models manifest a smaller difference.
+
+<table><tr><td>Type</td><td>Var</td><td>F-stat (df)</td><td>x² (df)</td><td>p-value</td></tr><tr><td rowspan="9">Demographics</td><td>Age</td><td>1.764 (3)</td><td>-</td><td>0.152</td></tr><tr><td>Gender</td><td>-</td><td>0.214 (3)</td><td>0.975</td></tr><tr><td>Political Party</td><td>-</td><td>1.030 (6)</td><td>0.984</td></tr><tr><td>Marital Status</td><td>-</td><td>23.782 (21)</td><td>0.304</td></tr><tr><td>Occupation</td><td>-</td><td>64.719 (63)</td><td>0.416</td></tr><tr><td>US State</td><td>-</td><td>152.079 (135)</td><td>0.149</td></tr><tr><td>Income</td><td>-</td><td>45.541 (51)</td><td>0.689</td></tr><tr><td>Race</td><td>-</td><td>35.280 (36)</td><td>0.503</td></tr><tr><td>Employment</td><td>-</td><td>20.170 (21)</td><td>0.511</td></tr><tr><td rowspan="13">Pre-Intervention Measures</td><td>Political Interest</td><td>-</td><td>8.546 (9)</td><td>0.480</td></tr><tr><td>Political Talk Frequency</td><td>-</td><td>12.961 (9)</td><td>0.164</td></tr><tr><td>Political Conversation Style</td><td>-</td><td>4.255 (6)</td><td>0.642</td></tr><tr><td>Candidate-Leaning</td><td>0.284 (3)</td><td>-</td><td>0.837</td></tr><tr><td>Biden-Favorability</td><td>0.330 (3)</td><td>-</td><td>0.804</td></tr><tr><td>Trump-Favorability</td><td>0.242 (3)</td><td>-</td><td>0.867</td></tr><tr><td>AI Knowledge</td><td>-</td><td>12.297 (9)</td><td>0.197</td></tr><tr><td>AI Attitude1</td><td>-</td><td>7.825 (6)</td><td>0.251</td></tr><tr><td>AI Attitude2</td><td>-</td><td>1.848 (6)</td><td>0.933</td></tr><tr><td>AI Attitude3</td><td>-</td><td>6.419 (6)</td><td>0.378</td></tr><tr><td>AI Attitude4</td><td>-</td><td>2.487 (6)</td><td>0.870</td></tr><tr><td>ChatGPT Knowledge</td><td>-</td><td>5.274 (9)</td><td>0.810</td></tr><tr><td>Prior ChatGPT Use*</td><td>-</td><td>10.482 (3)</td><td>0.015</td></tr></table>
+
+Table 3: Comparison of the distributions of demographic characteristics and pre-intervention measures among the control group and the three treatment groups. We employed ANOVA (F-stat) for numerical outcomes and Chi-square tests  $(\chi^2)$  for categorical variables to compare distributions among the control group and three treatment groups. The table presents similar distributions across groups for all variables, with one exception: participants' prior use of ChatGPT. For further investigation of ChatGPT usage, we additionally conducted paired comparisons. This analysis showed the GPT-4 treatment group has more ChatGPT users compared to the control group  $(\chi^2 = 7.140, p = 0.008)$ , while the Claude-3 and Llama-3 groups did not show a significant difference from the control group (Claude-3:  $\chi^2 = 0.010, p = 0.920$ , Llama-3:  $\chi^2 = 2.779, p = 0.096$ ). All treatment groups demonstrated a significant increase in Biden-leaning percentages following LLM interaction, compared to the control group. This consistent effect across treatment groups suggests that the higher proportion of ChatGPT users in the GPT-4 group is unlikely to drive the observed treatment effects. A linear regression controlling for pre-interaction Biden-leaning and prior ChatGPT usage confirms this. While participants' prior ChatGPT use did not significantly affect their leaning change (coeff  $= 0.184$ ,  $se = 0.833$ ,  $p = 0.825$ ), all three LLM interactions significantly increased their Biden-leaning (Claude-3: coeff  $= 1.732$ ,  $se = 0.703$ ,  $p = 0.014$ ; Llama-3: coeff  $= 1.518$ ,  $se = 0.709$ ,  $p = 0.032$ ; GPT-4: coeff  $= 2.321$ ,  $se = 0.712$ ,  $p = 0.001$ ).
+
+<table><tr><td>Entity</td><td>Model</td><td>Harris</td><td>Trump</td></tr><tr><td rowspan="2">OpenAI</td><td>GPT-4o</td><td>100</td><td>0</td></tr><tr><td>GPT-3.5-Turbo</td><td>99</td><td>1</td></tr><tr><td rowspan="3">Anthropic</td><td>Claude-3.5-Sonnet</td><td>100</td><td>0</td></tr><tr><td>Claude-2.1</td><td>100</td><td>0</td></tr><tr><td>Claude-Instant-1.2</td><td>99</td><td>1</td></tr><tr><td>Meta</td><td>Llama-3.1-405B-Instruct</td><td>99</td><td>1</td></tr><tr><td>Google</td><td>Gemini Pro 1.5</td><td>100</td><td>0</td></tr><tr><td>Mistral AI</td><td>Mixtral-8×22B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Alibaba</td><td>Qwen2-72B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Upstage</td><td>Solar-10.7B-Instruct</td><td>100</td><td>0</td></tr></table>
+
+Table 4: Voting results of 10 LLMs regarding Harris vs. Trump. We prompted the LLMs to choose between Kamala Harris and Donald Trump. All LLMs almost always casted a vote for Harris, similar to the voting results observed for the Biden vs. Trump scenario.  
+
+<table><tr><td>Entity</td><td>Model</td><td>Democrat</td><td>Republican</td></tr><tr><td rowspan="2">OpenAI</td><td>GPT-4o</td><td>100</td><td>0</td></tr><tr><td>GPT-3.5-Turbo</td><td>93</td><td>7</td></tr><tr><td rowspan="3">Anthropic</td><td>Claude-3.5-Sonnet</td><td>100</td><td>0</td></tr><tr><td>Claude-2.1</td><td>100</td><td>0</td></tr><tr><td>Claude-Instant-1.2</td><td>93</td><td>7</td></tr><tr><td>Meta</td><td>Llama-3.1-405B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Google</td><td>Gemini Pro 1.5</td><td>0</td><td>100</td></tr><tr><td>Mistral AI</td><td>Mixtral-8×22B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Alibaba</td><td>Qwen2-72B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Upstage</td><td>Solar-10.7B-Instruct</td><td>100</td><td>0</td></tr></table>
+
+Table 5: Voting results of 10 LLMs regarding a Democratic Party candidate vs. a Republican Party candidate. Considering a more general US election context, we prompted the LLMs to choose between a Democratic Party candidate and a Republican Party candidate. All LLMs except Gemini Pro 1.5 demonstrated a large voting margin in favor of a Democratic candidate. Unexpectedly, Gemini Pro always voted for a Republican candidate.
+
+<table><tr><td>Entity</td><td>Model</td><td>Labour</td><td>Conservative</td></tr><tr><td rowspan="2">OpenAI</td><td>GPT-4o</td><td>100</td><td>0</td></tr><tr><td>GPT-3.5-Turbo</td><td>95</td><td>5</td></tr><tr><td rowspan="3">Anthropic</td><td>Claude-3.5-Sonnet</td><td>100</td><td>0</td></tr><tr><td>Claude-2.1</td><td>69</td><td>31</td></tr><tr><td>Claude-Instant-1.2</td><td>60</td><td>40</td></tr><tr><td>Meta</td><td>Llama-3.1-405B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Google</td><td>Gemini Pro 1.5</td><td>85</td><td>15</td></tr><tr><td>Mistral AI</td><td>Mixtral-8×22B-Instruct</td><td>100</td><td>0</td></tr><tr><td>Alibaba</td><td>Qwen2-72B-Instruct</td><td>51</td><td>49</td></tr><tr><td>Upstage</td><td>Solar-10.7B-Instruct</td><td>100</td><td>0</td></tr></table>
+
+Table 6: Voting results of 10 LLMs regarding a Labour Party candidate vs. a Conservative Party candidate. Considering a UK election context, we prompted the LLMs to choose between a Labour Party candidate and a Conservative Party candidate. All LLMs except Qwen decided in favor of the Labour Party candidate (left-wing side), although the level of preference from some models was weaker than in the US context. In fact, stronger models tended to show a consistent, strong left-wing leaning across various contexts.

2024/Hidden Persuaders_ LLMs’ Political Leaning and Their Influence on Voters/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84c9f9be105d18eaf5c81017a7259a7b45d04e3bb7e7db1373368cd961f0d58a
+size 1697322