Add Batch 6cd15307-f42b-4492-aefc-0e80866254f9 data

.gitattributes

@@ -2156,3 +2156,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/MMAG_[[:space:]]Multimodal[[:space:]]Learning[[:space:]]for[[:space:]]Mucus[[:space:]]Anomaly[[:space:]]Grading[[:space:]]in[[:space:]]Nasal[[:space:]]Endoscopy[[:space:]]via[[:space:]]Semantic[[:space:]]Attribute[[:space:]]Prompting/0711a405-6a50-4be1-bcbe-6c5b147ce528_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/MMAPG_[[:space:]]A[[:space:]]Training-Free[[:space:]]Framework[[:space:]]for[[:space:]]Multimodal[[:space:]]Multi-hop[[:space:]]Question[[:space:]]Answering[[:space:]]via[[:space:]]Adaptive[[:space:]]Planning[[:space:]]Graphs/bb8bfa02-6e92-4a46-8358-35b65bf09ac1_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/MMDocIR_[[:space:]]Benchmarking[[:space:]]Multimodal[[:space:]]Retrieval[[:space:]]for[[:space:]]Long[[:space:]]Documents/b61cbeb4-7985-4a9f-accf-53587b5ed67c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MMLU-ProX_[[:space:]]A[[:space:]]Multilingual[[:space:]]Benchmark[[:space:]]for[[:space:]]Advanced[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Evaluation/74076a83-6daa-44d2-8b6c-8214a5104094_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MOSAIC_[[:space:]]Modeling[[:space:]]Social[[:space:]]AI[[:space:]]for[[:space:]]Content[[:space:]]Dissemination[[:space:]]and[[:space:]]Regulation[[:space:]]in[[:space:]]Multi-Agent[[:space:]]Simulations/bc9345db-bfaa-4476-acc4-3adce3ea00fd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MPCG_[[:space:]]Multi-Round[[:space:]]Persona-Conditioned[[:space:]]Generation[[:space:]]for[[:space:]]Modeling[[:space:]]the[[:space:]]Evolution[[:space:]]of[[:space:]]Misinformation[[:space:]]with[[:space:]]LLMs/37389502-1af8-40a9-8199-f16836485399_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MPRF_[[:space:]]Interpretable[[:space:]]Stance[[:space:]]Detection[[:space:]]through[[:space:]]Multi-Path[[:space:]]Reasoning[[:space:]]Framework/f3365d8d-e0cc-4911-aacf-05455e5743c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MR.[[:space:]]Judge_[[:space:]]Multimodal[[:space:]]Reasoner[[:space:]]as[[:space:]]a[[:space:]]Judge/b1637928-d23a-4ee6-8b4d-2e909f1b2b62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MS-RAG_[[:space:]]Simple[[:space:]]and[[:space:]]Effective[[:space:]]Multi-Semantic[[:space:]]Retrieval-Augmented[[:space:]]Generation/da873171-424c-48e9-9119-5424bd5a6138_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MUCAR_[[:space:]]Benchmarking[[:space:]]Multilingual[[:space:]]Cross-Modal[[:space:]]Ambiguity[[:space:]]Resolution[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/aabf9fec-71eb-46aa-acbd-b7c977b28b1d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MULTIGUARD_[[:space:]]An[[:space:]]Efficient[[:space:]]Approach[[:space:]]for[[:space:]]AI[[:space:]]Safety[[:space:]]Moderation[[:space:]]Across[[:space:]]Languages[[:space:]]and[[:space:]]Modalities/6a831b8d-c698-467c-8423-29005465a436_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MULTIVOX_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Voice[[:space:]]Assistants[[:space:]]for[[:space:]]Multimodal[[:space:]]Interactions/fd3c0a4a-5db1-4e4e-9097-8d01504aa0be_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MUSE_[[:space:]]MCTS-Driven[[:space:]]Red[[:space:]]Teaming[[:space:]]Framework[[:space:]]for[[:space:]]Enhanced[[:space:]]Multi-Turn[[:space:]]Dialogue[[:space:]]Safety[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/203909ec-67aa-4b5f-aeaf-2b7aad5c9bf5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MUZO_[[:space:]]Leveraging[[:space:]]Multiple[[:space:]]Queries[[:space:]]and[[:space:]]Momentum[[:space:]]for[[:space:]]Zeroth-Order[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/f544867f-710a-49b5-8bcf-23b1054a1064_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MaZO_[[:space:]]Masked[[:space:]]Zeroth-Order[[:space:]]Optimization[[:space:]]for[[:space:]]Multi-Task[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/e1c86835-b143-4aa3-a499-078991c28207_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Machine-generated[[:space:]]text[[:space:]]detection[[:space:]]prevents[[:space:]]language[[:space:]]model[[:space:]]collapse/3f802c8f-1ee0-42b0-8cad-f1d72b32c255_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mahānāma_[[:space:]]A[[:space:]]Unique[[:space:]]Testbed[[:space:]]for[[:space:]]Literary[[:space:]]Entity[[:space:]]Discovery[[:space:]]and[[:space:]]Linking/cc6762bf-9e33-4b86-8420-74735ffcc8b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Making[[:space:]]VLMs[[:space:]]More[[:space:]]Robot-Friendly_[[:space:]]Self-Critical[[:space:]]Distillation[[:space:]]of[[:space:]]Low-Level[[:space:]]Procedural[[:space:]]Reasoning/b8e9e5dc-1af4-46ec-8727-c885619c1435_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mapping[[:space:]]Toxic[[:space:]]Comments[[:space:]]Across[[:space:]]Demographics_[[:space:]]A[[:space:]]Dataset[[:space:]]from[[:space:]]German[[:space:]]Public[[:space:]]Broadcasting/af390cb2-7c95-42fd-b17e-aad3ef8b8ed7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mapping[[:space:]]semantic[[:space:]]networks[[:space:]]to[[:space:]]Dutch[[:space:]]word[[:space:]]embeddings[[:space:]]as[[:space:]]a[[:space:]]diagnostic[[:space:]]tool[[:space:]]for[[:space:]]cognitive[[:space:]]decline/14fed0b2-a433-4716-8db4-81e596baf099_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mapping[[:space:]]the[[:space:]]Minds[[:space:]]of[[:space:]]LLMs_[[:space:]]A[[:space:]]Graph-Based[[:space:]]Analysis[[:space:]]of[[:space:]]Reasoning[[:space:]]LLMs/bd574992-d638-49b4-bb7a-2c03185d5e82_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Massive[[:space:]]Supervised[[:space:]]Fine-tuning[[:space:]]Experiments[[:space:]]Reveal[[:space:]]How[[:space:]]Data,[[:space:]]Layer,[[:space:]]and[[:space:]]Training[[:space:]]Factors[[:space:]]Shape[[:space:]]LLM[[:space:]]Alignment[[:space:]]Quality/ace012f7-7d2a-40a2-b2db-26f9717858cc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MathTutorBench_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Measuring[[:space:]]Open-ended[[:space:]]Pedagogical[[:space:]]Capabilities[[:space:]]of[[:space:]]LLM[[:space:]]Tutors/d4b66f7a-4308-4f18-bbad-b6e574d4f7ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Matter-of-Fact_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Verifying[[:space:]]the[[:space:]]Feasibility[[:space:]]of[[:space:]]Literature-Supported[[:space:]]Claims[[:space:]]in[[:space:]]Materials[[:space:]]Science/59810f85-9f61-4cc3-87cc-611b6cace16b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Measuring[[:space:]]Bias[[:space:]]or[[:space:]]Measuring[[:space:]]the[[:space:]]Task_[[:space:]]Understanding[[:space:]]the[[:space:]]Brittle[[:space:]]Nature[[:space:]]of[[:space:]]LLM[[:space:]]Gender[[:space:]]Biases/df5573e6-65e6-4ace-9b45-5dbc39f089aa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Measuring[[:space:]]Chain[[:space:]]of[[:space:]]Thought[[:space:]]Faithfulness[[:space:]]by[[:space:]]Unlearning[[:space:]]Reasoning[[:space:]]Steps/cb3e7b85-5a92-4bec-aee2-d3e6f36291ac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Measuring[[:space:]]Risk[[:space:]]of[[:space:]]Bias[[:space:]]in[[:space:]]Biomedical[[:space:]]Reports_[[:space:]]The[[:space:]]RoBBR[[:space:]]Benchmark/5e2474e6-4146-4407-b0ca-fbb227fe75d8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Measuring[[:space:]]and[[:space:]]Mitigating[[:space:]]Media[[:space:]]Outlet[[:space:]]Name[[:space:]]Bias[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/0369bee9-94d5-4450-bdbd-81c392b7cd77_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Measuring[[:space:]]scalar[[:space:]]constructs[[:space:]]in[[:space:]]social[[:space:]]science[[:space:]]with[[:space:]]LLMs/e824be44-8420-4dc2-8b26-e92e1839072e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Measuring[[:space:]]the[[:space:]]Effect[[:space:]]of[[:space:]]Disfluency[[:space:]]in[[:space:]]Multilingual[[:space:]]Knowledge[[:space:]]Probing[[:space:]]Benchmarks/5b2253da-f6e8-442e-b9e5-77334a7dd14d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mechanisms[[:space:]]vs.[[:space:]]Outcomes_[[:space:]]Probing[[:space:]]for[[:space:]]Syntax[[:space:]]Fails[[:space:]]to[[:space:]]Explain[[:space:]]Performance[[:space:]]on[[:space:]]Targeted[[:space:]]Syntactic[[:space:]]Evaluations/831f9a82-6164-4509-9d03-16487aca8809_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Med-PRM_[[:space:]]Medical[[:space:]]Reasoning[[:space:]]Models[[:space:]]with[[:space:]]Stepwise,[[:space:]]Guideline-verified[[:space:]]Process[[:space:]]Rewards/27637b32-cec9-4179-83f9-bad237b46570_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Med-VRAgent_[[:space:]]A[[:space:]]Framework[[:space:]]for[[:space:]]Medical[[:space:]]Visual[[:space:]]Reasoning-Enhanced[[:space:]]Agents/950d99b9-a1d1-44fd-a072-bb93e1964236_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MedFact_[[:space:]]A[[:space:]]Large-scale[[:space:]]Chinese[[:space:]]Dataset[[:space:]]for[[:space:]]Evidence-based[[:space:]]Medical[[:space:]]Fact-checking[[:space:]]of[[:space:]]LLM[[:space:]]Responses/a119bdca-3e89-45f4-a39e-948c694c207a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MedHallu_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]for[[:space:]]Detecting[[:space:]]Medical[[:space:]]Hallucinations[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/9d9e9a17-f069-4798-9093-0d720a8f8ca6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MedLinkDE[[:space:]]–[[:space:]]MedDRA[[:space:]]Entity[[:space:]]Linking[[:space:]]for[[:space:]]German[[:space:]]with[[:space:]]Guided[[:space:]]Chain[[:space:]]of[[:space:]]Thought[[:space:]]Reasoning/84c1a532-0cb5-4872-9f94-b57d935b032e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Media[[:space:]]Source[[:space:]]Matters[[:space:]]More[[:space:]]Than[[:space:]]Content_[[:space:]]Unveiling[[:space:]]Political[[:space:]]Bias[[:space:]]in[[:space:]]LLM-Generated[[:space:]]Citations/eb028adb-3395-4c32-a9a1-11ed2cc36039_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MemInsight_[[:space:]]Autonomous[[:space:]]Memory[[:space:]]Augmentation[[:space:]]for[[:space:]]LLM[[:space:]]Agents/b023e763-df92-4d1d-af96-a39c4855faeb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Membership[[:space:]]and[[:space:]]Memorization[[:space:]]in[[:space:]]LLM[[:space:]]Knowledge[[:space:]]Distillation/02041d7e-4bc2-4771-ac41-0f35eeb77401_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MemeArena_[[:space:]]Automating[[:space:]]Context-Aware[[:space:]]Unbiased[[:space:]]Evaluation[[:space:]]of[[:space:]]Harmfulness[[:space:]]Understanding[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/bea49644-8e3d-4225-8c6e-8a2aefed170a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MemeIntel_[[:space:]]Explainable[[:space:]]Detection[[:space:]]of[[:space:]]Propagandistic[[:space:]]and[[:space:]]Hateful[[:space:]]Memes/8c27881b-416b-438a-abce-df15dcb29e8d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MemeReaCon_[[:space:]]Probing[[:space:]]Contextual[[:space:]]Meme[[:space:]]Understanding[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/2e014e9e-01a7-4359-ad59-76a3decca970_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Memorization[[:space:]]or[[:space:]]Reasoning_[[:space:]]Exploring[[:space:]]the[[:space:]]Idiom[[:space:]]Understanding[[:space:]]of[[:space:]]LLMs/8a659de3-b558-493e-ba48-a1450c0b77aa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Memorization[[:space:]]≠[[:space:]]Understanding_[[:space:]]Do[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Have[[:space:]]the[[:space:]]Ability[[:space:]]of[[:space:]]Scenario[[:space:]]Cognition_/13b838ba-8fb5-4fd6-9839-495d9f5bc0b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Memory[[:space:]]OS[[:space:]]of[[:space:]]AI[[:space:]]Agent/5a9d8a4a-8210-414c-aa4b-c3fe23647a6b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Memory-QA_[[:space:]]Answering[[:space:]]Recall[[:space:]]Questions[[:space:]]Based[[:space:]]on[[:space:]]Multimodal[[:space:]]Memories/499fd6b6-85e6-4282-a96d-24099492c99b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MentalGLM[[:space:]]Series_[[:space:]]Explainable[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Mental[[:space:]]Health[[:space:]]Analysis[[:space:]]on[[:space:]]Chinese[[:space:]]Social[[:space:]]Media/aadcbbd5-c399-486e-8a27-57f38a4a37a6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Merge[[:space:]]then[[:space:]]Realign_[[:space:]]Simple[[:space:]]and[[:space:]]Effective[[:space:]]Modality-Incremental[[:space:]]Continual[[:space:]]Learning[[:space:]]for[[:space:]]Multimodal[[:space:]]LLMs/2a2db7b9-a4a5-4147-9414-f543aad33219_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Merger-as-a-Stealer_[[:space:]]Stealing[[:space:]]Targeted[[:space:]]PII[[:space:]]from[[:space:]]Aligned[[:space:]]LLMs[[:space:]]with[[:space:]]Model[[:space:]]Merging/4defad6a-3244-479a-bcab-c2af0e82d2c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MessIRve_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Spanish[[:space:]]Information[[:space:]]Retrieval[[:space:]]Dataset/eaa77088-9b47-4737-b8d9-b96cd6a257c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Meta-Rewarding[[:space:]]Language[[:space:]]Models_[[:space:]]Self-Improving[[:space:]]Alignment[[:space:]]with[[:space:]]LLM-as-a-Meta-Judge/2ae95a98-30b6-4964-8ea1-48acf65d0575_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Meta-Semantics[[:space:]]Augmented[[:space:]]Few-Shot[[:space:]]Relational[[:space:]]Learning/eae1aa3c-c4ce-4388-8a41-a797de9841b9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MetaFaith_[[:space:]]Faithful[[:space:]]Natural[[:space:]]Language[[:space:]]Uncertainty[[:space:]]Expression[[:space:]]in[[:space:]]LLMs/ffa2526e-7c63-4ad6-b734-5149074711f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Metric[[:space:]]Calculating[[:space:]]Benchmark_[[:space:]]Code-Verifiable[[:space:]]Complicate[[:space:]]Instruction[[:space:]]Following[[:space:]]Benchmark[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/e944868d-8056-45c1-b268-c613b49344b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MiCRo_[[:space:]]Mixture[[:space:]]Modeling[[:space:]]and[[:space:]]Context-aware[[:space:]]Routing[[:space:]]for[[:space:]]Personalized[[:space:]]Preference[[:space:]]Learning/cdf4551e-2cb8-4a32-9932-a48ed9115ae9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MiLQ_[[:space:]]Benchmarking[[:space:]]IR[[:space:]]Models[[:space:]]for[[:space:]]Bilingual[[:space:]]Web[[:space:]]Search[[:space:]]with[[:space:]]Mixed[[:space:]]Language[[:space:]]Queries/766b26eb-e55d-4aa1-8540-1b02a3f3331e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MicroEdit_[[:space:]]Neuron-level[[:space:]]Knowledge[[:space:]]Disentanglement[[:space:]]and[[:space:]]Localization[[:space:]]in[[:space:]]Lifelong[[:space:]]Model[[:space:]]Editing/778480de-39c9-4c2b-b73c-6b8ccf40e46b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Middo_[[:space:]]Model-Informed[[:space:]]Dynamic[[:space:]]Data[[:space:]]Optimization[[:space:]]for[[:space:]]Enhanced[[:space:]]LLM[[:space:]]Fine-Tuning[[:space:]]via[[:space:]]Closed-Loop[[:space:]]Learning/2731f33d-8fd9-4050-baf6-ac3e8af2a539_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]the[[:space:]]Blind[[:space:]]Spots_[[:space:]]A[[:space:]]Focus-Level[[:space:]]Evaluation[[:space:]]Framework[[:space:]]for[[:space:]]LLM[[:space:]]Reviews/d7503084-4a6c-4aa7-a1ec-09075270bc38_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]the[[:space:]]Gap_[[:space:]]A[[:space:]]Closer[[:space:]]Look[[:space:]]at[[:space:]]Tokenization[[:space:]]for[[:space:]]Multiple-Choice[[:space:]]Question[[:space:]]Answering[[:space:]]with[[:space:]]LLMs/ed7e6f58-25d1-4ffd-abab-2d32d9a794b4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]the[[:space:]]Gap_[[:space:]]How[[:space:]]BabyLMs[[:space:]]Learn[[:space:]]Filler-Gap[[:space:]]Dependencies/6727388e-8a5e-440f-a54c-991a9cf5a38d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]the[[:space:]]Inclusivity[[:space:]]Gap_[[:space:]]Multilingual[[:space:]]Gender-Neutral[[:space:]]Translation[[:space:]]Evaluation[[:space:]]with[[:space:]]mGeNTE/f54b4942-3e01-4923-8fca-ec0929117b3b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mind[[:space:]]the[[:space:]]Value-Action[[:space:]]Gap_[[:space:]]Do[[:space:]]LLMs[[:space:]]Act[[:space:]]in[[:space:]]Alignment[[:space:]]with[[:space:]]Their[[:space:]]Values_/1f80e89b-ed74-428c-a1e8-bb079805af7f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Minimal,[[:space:]]Local,[[:space:]]and[[:space:]]Robust_[[:space:]]Embedding-Only[[:space:]]Edits[[:space:]]for[[:space:]]Implicit[[:space:]]Bias[[:space:]]in[[:space:]]T2I[[:space:]]Models/fad7f168-0933-4880-9956-143c77cbdb68_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mining[[:space:]]the[[:space:]]Past[[:space:]]with[[:space:]]Dual[[:space:]]Criteria_[[:space:]]Integrating[[:space:]]Three[[:space:]]types[[:space:]]of[[:space:]]Historical[[:space:]]Information[[:space:]]for[[:space:]]Context-aware[[:space:]]Event[[:space:]]Forecasting/0b9f764b-a626-454f-ba6b-8fd4b2d7bfe0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mitigating[[:space:]]Biases[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Bias[[:space:]]Unlearning/10f1c487-2eea-4afe-8e6e-bae4c4b234c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mitigating[[:space:]]Catastrophic[[:space:]]Forgetting[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Forgetting-aware[[:space:]]Pruning/8fa796eb-6a54-4396-8b0d-6c3c312fadb4_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/MMLU-ProX_ A Multilingual Benchmark for Advanced Large Language Model Evaluation/74076a83-6daa-44d2-8b6c-8214a5104094_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42b7c764d56368baa7d0a667151dab7be56ee7ce82dac47d0453d335cd0a375e
+size 1553275

2025/MMLU-ProX_ A Multilingual Benchmark for Advanced Large Language Model Evaluation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30945c7cbb7ae76685a1fca9b1777b5450146f20b72706606b047c4cbb3e0747
+size 1630186

2025/MOSAIC_ Modeling Social AI for Content Dissemination and Regulation in Multi-Agent Simulations/bc9345db-bfaa-4476-acc4-3adce3ea00fd_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa627637bb782584e16a9a472b1fca2a82b27b04bbbba30734a9783d2a554621
+size 1076897

2025/MOSAIC_ Modeling Social AI for Content Dissemination and Regulation in Multi-Agent Simulations/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf142c53503705fab8a5524870eddb73131cbcb6918b0bd287836e5d28859f12
+size 999276

2025/MPCG_ Multi-Round Persona-Conditioned Generation for Modeling the Evolution of Misinformation with LLMs/37389502-1af8-40a9-8199-f16836485399_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c95071e320144a9cb52f54dd157d3957c6e9b18cef06c575858878466b75f03
+size 758853

2025/MPCG_ Multi-Round Persona-Conditioned Generation for Modeling the Evolution of Misinformation with LLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:264e62089c56347327b43ecfdc2d12749c0f98670c94faa79b36fbd1873fecd5
+size 595753

2025/MPRF_ Interpretable Stance Detection through Multi-Path Reasoning Framework/f3365d8d-e0cc-4911-aacf-05455e5743c0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b14555396b454e527f602a4fff898a5accdc6e9fa9d36f560b0ff99d36814cd
+size 719252

2025/MPRF_ Interpretable Stance Detection through Multi-Path Reasoning Framework/full.md

@@ -0,0 +1,518 @@
+# MPRF: Interpretable Stance Detection through Multi-Path Reasoning Framework
+
+Zhaodan Zhang $^{1,2,3}$ , Jin Zhang $^{2,3,*}$ , Hui Xu $^{2}$ , Jiafeng Guo $^{2,3}$ , Xueqi Cheng $^{2,3}$
+
+$^{1}$ School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences
+
+${}^{2}$  State Key Laboratory of AI Safety,
+
+Institute of Computing Technology, Chinese Academy of Sciences
+
+<sup>3</sup>University of Chinese Academy of Sciences
+
+{zhangzhaodan23s, jinzhang, xuhui, guojiafeng, cxq}@ict.ac.cn
+
+# Abstract
+
+Stance detection, a critical task in Natural Language Processing (NLP), aims to identify the attitude expressed in text toward specific targets. Despite advancements in Large Language Models (LLMs), challenges such as limited interpretability and handling nuanced content persist. To address these issues, we propose the Multi-Path Reasoning Framework (MPRF), a novel framework that generates, evaluates, and integrates multiple reasoning paths to improve accuracy, robustness, and transparency in stance detection. Unlike prior work that relies on single-path reasoning or static explanations, MPRF introduces a structured end-to-end pipeline: it first generates diverse reasoning paths through predefined perspectives, then dynamically evaluates and optimizes each path using LLM-based scoring, and finally fuses the results via weighted aggregation to produce interpretable and reliable predictions. Extensive experiments on the SEM16, VAST, and PStance datasets demonstrate that MPRF outperforms existing models. Ablation studies further validate the critical role of MPRF's components, highlighting its effectiveness in enhancing interpretability and handling complex stance detection tasks.
+
+# 1 Introduction
+
+Stance detection (Hasan and Ng, 2014; Kucuk and Can, 2020), the task of determining the attitude expressed in a text towards a specific target, plays a crucial role in applications such as opinion mining (Graells-Garrido et al., 2020), combating misinformation (Lai et al., 2020), and understanding public sentiment (Lei et al., 2024). By analyzing structural and linguistic patterns in stance reasoning, researchers can gain insights into opinion dynamics, address the evolution of harmful behaviors, and foster a more ethical online environment
+
+(De Vinco et al., 2024; Graells-Garrido and Baeza-Yates, 2022; Zhang et al., 2023b).
+
+Existing stance detection methods often treat the task as a classification problem, where models output a stance label without providing interpretable reasoning paths (Allaway et al., 2021; De Vinco et al., 2024; Graells-Garrido and Baeza-Yates, 2022; Li and Zhang, 2024; Xu et al., 2022; Yang and Urbani, 2021; Zhang et al., 2023b). This lack of transparency is particularly problematic in complex tasks that involve subtle or ambiguous opinions, such as those expressed in social media content(Gatto et al., 2023).
+
+Recent advances in large language models (LLMs), particularly those leveraging Chain-of-Thought (CoT) prompting (Wei et al., 2022), have demonstrated remarkable reasoning capabilities in tasks such as multihop question answering (Lu et al., 2022) and mathematical problem solving (Wei et al., 2022). These models achieve zero-shot and few-shot success across diverse tasks by utilizing machine-generated instruction-following data and reasoning mechanisms. While prior work has shown that explanations from models like GPT can improve interpretability in stance detection (Zhang et al., 2024a; Taranukhin et al., 2024), they often rely on single-path or heuristic-based reasoning, which may fail to capture diverse perspectives or provide robust explanations.
+
+In this work, we introduce the Multi-Path Reasoning Framework (MPRF), a novel framework designed to address the interpretability challenges in stance detection by generating, evaluating, and integrating multiple reasoning paths. Unlike previous approaches (Ding et al., 2024a; Taranukhin et al., 2024), which often rely on static reasoning or single-path explanations, MPRF introduces a structured and dynamic pipeline: it first generates diverse reasoning paths through predefined perspectives (e.g., separating sentiment from factual analysis), then evaluates each path using LLM-based
+
+scoring (relevance, evidence strength, logical consistency), and finally fuses the results via weighted aggregation to produce interpretable and reliable predictions. This systematic integration of generation, evaluation, optimization, and fusion represents a meaningful system-level innovation beyond prompt engineering or isolated use of LLM capabilities.
+
+Our contributions are as follows:
+
+- We propose the Multi-Path Reasoning Framework (MPRF), a novel end-to-end framework for stance detection that goes beyond traditional prompt engineering by combining multiple reasoning paths in a structured, iterative, and interpretable manner.
+
+- We design a dynamic evaluation and optimization mechanism that ensures high-quality reasoning paths through LLM-based scoring (relevance, evidence strength, logical consistency) and refinement of weak paths. We introduce a weighted fusion strategy that prioritizes well-supported reasoning paths, improving prediction reliability and interpretability.
+
+- Extensive experiments show that MPRF achieves state-of-the-art performance across multiple datasets in both zero-shot and few-shot settings, while providing interpretable reasoning chains for each prediction. Our detailed ablation studies highlight the importance of reasoning path evaluation and fusion in improving the accuracy and interpretability of stance detection predictions.
+
+# 2 Related Work
+
+Early stance detection approaches primarily treated the task as a classification problem, relying on traditional machine learning models with handcrafted features (Aldayel and Magdy, 2019; Dey et al., 2017). With the emergence of pretrained language models (PLMs) such as BERT (Devlin et al., 2019; Nguyen et al., 2020), these methods achieved significant improvements by learning features from in-domain or cross-domain datasets (Augenstein et al., 2016; Zhang et al., 2019; Allaway et al., 2021; Liu et al., 2021; Liang et al., 2022b). However, these models often lacked interpretability, as they treated stance detection as a black-box process without explicitly modeling the reasoning steps.
+
+Recent advancements in large language models (LLMs) have significantly enhanced stance detection, with a focus on improving reasoning capabilities through Chain-of-Thought (CoT) prompting
+
+(Yao et al., 2024). CoT prompting enables LLMs to perform step-by-step reasoning, achieving state-of-the-art results in complex tasks such as arithmetic, logical reasoning, and stance detection. For example, Wei et al. (2022) and Zhou et al. (2022) demonstrated CoT prompting's effectiveness in multihop reasoning tasks. Ding et al. (2024a) applied multi-step CoT prompting to capture the positional perspective of targets in stance detection, while Fei et al. (2023) proposed decomposing tasks into multiple stages for better predictions. Similarly, Ling et al. (2023) introduced deductive reasoning and iterative verification to enhance task inference. Hardalov et al. (2022) designed a prompt-based framework for cross-language stance detection, Zhu et al. (2024) incorporated soft knowledge during cue fine-tuning to improve context understanding, and Huang et al. (2023) expanded the verbalizer in prompt-tuning with external semantic knowledge. While these methods demonstrate the power of CoT prompting in LLMs, they often rely on static reasoning processes and struggle to adapt to tasks requiring diverse perspectives.
+
+Building on these advancements, we propose the Multi-Path Reasoning Framework (MPRF), a novel framework that addresses the limitations of existing CoT-based methods by generating, evaluating, and integrating multiple reasoning paths.
+
+# 3 Methodology
+
+In this study, we propose a novel framework for stance detection, called Multi-Path Reasoning Framework (MPRF), which is designed to address the critical challenges in stance detection tasks, particularly the lack of transparency and interpretability. MPRF enhances the accuracy, robustness, and interpretability of stance detection by generating multiple reasoning paths, evaluating and optimizing these paths, and finally combining the results through a weighted fusion mechanism. Our approach is structured into five key steps: generating multiple reasoning paths (3.2), evaluating the paths (3.3), optimizing the paths (3.4), weighted fusion of the paths (3.5), and outputting the stance label along with the interpretable reasoning paths (3.6). MPRF shown in figure 1 ensures that the detected stance is not only accurate but also transparent, providing clear, traceable reasoning behind the final decision.
+
+![](images/b03fff5838813752d465ce415168f122f4e152eccbee55b8fe743c483c6f5e40.jpg)  
+Figure 1: This is the Multi-Path Reasoning Framework. We generate four reasoning paths. A final score  $s_i'$  is assigned to each path  $i$  by combining the relevance  $r_i$ , evidence strength  $e_i$ , and logical consistency scores  $l_i$  weighted by  $\alpha$ ,  $\beta$ , and  $\gamma$  after optimization.  $S_t$  is the aggregated weighted score for label  $t$ ,  $t \in \{favor, against, neutral\}$ . The label with the highest score is selected as the final stance.
+
+# 3.1 Task Definition
+
+Let  $\mathcal{D} = \{(x_j,p_j,y_j)\}_{j = 1}^N$  be a dataset consisting of  $N$  instances, where each instance  $(x_{j},p_{j},y_{j})$  represents the input text  $x_{j}$  for which the stance needs to be detected, the corresponding target  $p_j$  towards which the stance of  $x_{j}$  is to be determined, and the stance label  $y_{j}$  for the input  $x_{j}$  towards the target  $p_j$ , where  $y_{j}\in \{\mathrm{favor,}$  against, neutral\}.
+
+Stance detection aims to predict the stance label  $y_{j}$  for each input sentence  $x_{j}$  towards the given target  $p_{j}$ .
+
+# 3.2 Generating Multiple Reasoning Paths
+
+In this step, we generate multiple reasoning paths for stance detection, each representing a distinct chain of thought from a theoretically grounded perspective. This multi-path approach is designed to deconstruct the complex process of stance formation into its fundamental cognitive and social components, ensuring a more robust, accurate, and interpretable analysis. The specific prompts used to guide each reasoning path are detailed in Appendix B.
+
+The first reasoning path is Sentiment Analysis. This path is rooted in the understanding that stance is often an expression of affective state. Emotions are powerful drivers of human judgment and decision-making, frequently serving as the primary motivator for taking a position on an issue (Dey et al., 2017). The model identifies key emotional lexicons in the text (e.g., "happy", "angry", "disappointed") to determine the overall valence - positive, negative, or neutral.
+
+The second reasoning path is Factual Reasoning. This path addresses the cognitive process of evidence-based justification. It is grounded in the principle that rational argumentation relies on objective information to support a claim (Mohammad
+
+et al., 2016). The model acts as an analytical agent, extracting verifiable facts such as statistical data, scientific research findings, or legal precedents. It then evaluates whether this evidence logically supports or contradicts the target. This path is crucial for distinguishing between opinions based on truth claims and those based on mere assertion, thereby enhancing the logical rigor of the stance detection process.
+
+The third reasoning path is Expert Opinion. This path models the social influence of authority and expertise, a cornerstone of persuasive communication. In contentious or complex domains, individuals often defer to the judgment of trusted professionals or institutions (e.g., scientists, medical doctors, legal scholars) (Wang et al., 2024a). The model identifies explicit or implicit references to such authoritative sources within the text. If the referenced expert's position aligns with the target (e.g., a scientist stating that climate change is anthropogenic), the stance is inferred as "favor". This path captures the heuristic of "appeal to authority," a common and powerful mechanism in public discourse.
+
+The fourth reasoning path is Public Opinion. This path reflects the powerful role of social conformity and normative influence in shaping individual stances. Humans are social beings, and their opinions are often influenced by perceived societal consensus, social media trends, or mainstream media narratives (Wang et al., 2024a). The model identifies phrases that indicate collective sentiment (e.g., "most people think", "everyone is talking about", "public opinion is shifting"). If the perceived public sentiment supports the target, the stance is classified as "favor". This path accounts for the bandwagon effect and the desire to align with the majority.
+
+The selection of these four paths is not arbitrary but is grounded in a comprehensive analysis of human reasoning and empirical validation. They represent a spectrum of influence: from internal affective states (Sentiment) to external rational evidence (Factual), from institutional authority (Expert) to social consensus (Public). This holistic framework ensures that MPRF can capture the multifaceted nature of real-world arguments. Alternative paths, such as Metaphorical Analysis (Allaway and McKeown, 2020) and Temporal Reasoning (Allaway and McKeown, 2020), were explored but found to be less effective due to ambiguity and sparsity in short texts, respectively. The final four paths thus provide a balanced, theoretically sound, and empirically validated foundation for multi-perspective stance analysis.
+
+# 3.3 Evaluating Reasoning Paths
+
+Once multiple reasoning paths have been generated, each path is evaluated by a large language model (LLM) based on three main criteria: relevance, evidence strength, and logical consistency. These criteria are inspired by established frameworks in argumentation quality assessment (Hasan and Ng, 2014) and are defined as follows:
+
+- Relevance Score  $r_i$ : Measures how closely the reasoning aligns with the target stance. A higher score indicates stronger alignment.
+
+- Evidence Strength  $e_i$ : Assesses the reliability and robustness of the evidence supporting the reasoning (e.g., factual references to studies or statistics).
+
+- Logical Consistency  $l_{i}$ : Evaluates whether the reasoning steps follow a coherent and logically sound structure without contradictions.
+
+To ensure consistent and objective scoring across different paths, we designed standardized prompts (see Appendix B) that guide the LLM through structured evaluation steps for each criterion. While human annotation could provide additional validation, our experiments demonstrate a strong correlation between LLM-assigned scores and downstream performance, indicating the practical effectiveness of this automated approach.
+
+After evaluating each reasoning path using these criteria, a final score  $s_i$  is assigned to each path by combining the relevance, evidence strength, and logical consistency scores:
+
+$$
+s _ {i} = \alpha \cdot r _ {i} + \beta \cdot e _ {i} + \gamma \cdot l _ {i} \tag {1}
+$$
+
+where  $i \in \{0,1,2,3\}$  represents the four reasoning paths being evaluated, and  $\alpha, \beta,$  and  $\gamma$  are weights determined through experiments on weighting configurations (Table 5). The final score  $s_i$  reflects the overall quality and contribution of the path to the stance detection task.
+
+# 3.4 Optimizing Reasoning Paths
+
+After generating and evaluating reasoning paths, we focus on optimizing those that receive low scores in relevance, evidence strength, or logical consistency. Specifically, any path  $i$  that score below 5 in any of the evaluation criteria is considered a low-scoring path.
+
+The optimization process involves re-prompting the LLM to refine the reasoning path using a dedicated prompt template (Appendix B), which guides the model to improve specific weak areas—such as enhancing relevance to the target, strengthening supporting evidence, or correcting logical inconsistencies:
+
+$$
+i ^ {\prime} = \mathbf {M} (i, \mathrm {p r o m p t} _ {\mathrm {r e f i n e}}) \quad \mathrm {i f} \min  (r _ {i}, e _ {i}, l _ {i}) <   5 \quad (2)
+$$
+
+Once refined, the updated path  $i'$  is re-evaluated using the same scoring criteria:
+
+$$
+s _ {i} ^ {\prime} = \alpha \cdot r _ {i} ^ {\prime} + \beta \cdot e _ {i} ^ {\prime} + \gamma \cdot l _ {i} ^ {\prime} \tag {3}
+$$
+
+where  $r_i', e_i', l_i'$  denote the improved relevance, evidence strength, and logical consistency of the optimized path.
+
+This iterative refinement ensures that only high-quality reasoning paths contribute to the final stance prediction, significantly improving both interpretability and accuracy.
+
+# 3.5 Fusing Reasoning Paths
+
+In step 4, for each reasoning path  $i$ , we calculate the weighted score for each label  $t \in \{\text{favor, against, neutral}\}$ . This is done by computing the weighted sum of the individual scores  $s_i'$  for each path  $i$  assigned to label  $t$ , where  $w_i$  is the weight of each path based on its contribution to the reasoning process:
+
+$$
+S _ {t} = \sum_ {i \in P _ {t}} w _ {i} \cdot s _ {i} ^ {\prime} \tag {4}
+$$
+
+Here,  $P_{t}$  represents the set of paths assigned to label  $t$ , and  $S_{t}$  is the aggregated weighted score for label  $t$ .
+
+Next, we calculate the weighted scores for each label (favor, against, neutral) based on the re-evaluated scores for all the reasoning paths:  $S_{\text{favor}}$ ,  $S_{\text{against}}$  and  $S_{\text{neutral}}$ .
+
+After the weighted scores for each label are computed, the label with the highest score is selected as the final stance prediction. For example, if  $S_{\text{favor}} > S_{\text{against}}$  and  $S_{\text{favor}} > S_{\text{neutral}}$ , then the final stance is favor. The same approach is applied for the other labels.
+
+In the case where multiple labels have equal scores, the final stance label is determined by a majority vote. Specifically, the label associated with the most reasoning paths is chosen. Even when the scores are equal, the label with more supporting paths generally indicates stronger evidence, making it more likely to represent the true final stance. This voting mechanism resolves tie situations and enhances the robustness of the model by ensuring that the most evidence-backed label is selected.
+
+# 3.6 Outputting Stance and Reasoning Paths
+
+After determining the final stance prediction in Step 4, we now output both the selected stance label and the corresponding reasoning paths. The reasoning paths that contributed to the selected stance label, illustrating the logical progression and supporting evidence behind the decision. Specifically, the final output consists of:
+
+Final Stance Label: The label selected in Step 4, which can be one of the following: favor, against, or neutral.
+
+Reasoning Paths: The set of reasoning paths associated with the selected stance label, showing how the decision was made. These paths are the ones that contributed to the weighted score for the selected label.
+
+# 4 Experiments
+
+# 4.1 Datasets
+
+We conduct experiments on the VAST, SEM16, and PStance datasets to evaluate our proposed method. The VAST dataset (Allaway and McKeown, 2020) includes a wide range of targets, each instance comprising a sentence  $r$ , a target  $t$ , and a stance label  $y$  (classified as 'Pro', 'Con' or 'Neutral') towards  $t$ .
+
+The SEM16 dataset (Mohammad et al., 2016) contains six predefined targets, including Donald Trump (DT), Hillary Clinton (HC), Feminist Movement (FM), Legalization of Abortion (LA), Athe
+
+ism (A), and Climate Change (CC). Each instance is categorized as Favor, Against, or Neutral.
+
+The PStance dataset (Li et al., 2021) focuses on the stance of individuals towards three prominent political figures in the United States: Donald Trump (trump), Joe Biden (biden), and Bernie Sanders (sanders). This large-scale dataset includes only two stance labels: favor or against. Detailed statistics of datasets can be found in Appendix A.
+
+# 4.2 Evaluation Metrics
+
+For the VAST dataset, following Allaway and McKeown (2020), we calculate the Macro-averaged F1 score across the Pro, Con, and Neutral labels to evaluate the performance of the models on the test set. For the SEM16 and PStance datasets, we report the  $F_{\mathrm{avg}}$ , which is the average of the F1 scores for the Favor and Against classes, in line with Mohammad et al. (2016); Li et al. (2021). We compute  $F_{\mathrm{avg}}$  for each target.  $F_{\mathrm{macro}}$  is calculated by averaging the  $F_{\mathrm{avg}}$  across all targets.
+
+# 4.3 Baselines
+
+We compare our proposed Multi-Path Reasoning Framework (MPRF) against a comprehensive set of state-of-the-art baselines, categorized into statistics-based models, BERT-based models, and LLM-based models.
+
+Statistics-based models rely on traditional machine learning architectures. We include BiLSTM (Schuster and Paliwal, 1997), a bidirectional LSTM without target information; CNN (Kim, 2014), a convolutional network also without target modeling; TAN (Du et al., 2017), an attention-based LSTM for target-specific features; BiCond (Augenstein et al., 2016), which jointly encodes text and target with bidirectional LSTMs; CrossNet (Xu et al., 2018), a BiLSTM enhanced with self-attention; GCAE (Xue and Li, 2018), a CNN using a gating mechanism to filter irrelevant information; and PGCNN (Huang and Carley, 2018), a gated CNN that generates target-sensitive filters.
+
+BERT-based models are built upon the BERT architecture and are typically fine-tuned. We include BERT (Devlin et al., 2019), the standard model; BERTweet (Nguyen et al., 2020), pretrained on tweets; PT-HCL (Liang et al., 2022a), which uses contrastive learning for zero-shot tasks; JointCL (Liang et al., 2022b), a joint contrastive learning framework; WS-BERT (He et al., 2022), which infuses Wikipedia knowledge; CKI (Yan et al., 2024), a collaborative knowledge infusion
+
+approach; TATA (Hanley and Durmeric, 2023), using topic-agnostic and topic-aware embeddings; KPatch (Lin et al., 2024), introducing a "knowledge patch" for zero-shot detection; CNet-Ad (Zhang et al., 2024c), a commonsense-based adversarial learning framework; and EZSD-CP (Yao et al., 2024), a model based on a gated MLP and prompt learning.
+
+LLM-based models leverage the reasoning capabilities of LLMs, often through prompting. We include KEprompt (Huang et al., 2023), using knowledge-enhanced prompt-tuning; KASD (Li et al., 2023), integrating Wikipedia knowledge with retrieval; COLA (Lan et al., 2024), a collaborative role-infusion framework; Stanceformer (Garg and Caragea, 2024), a target-aware transformer; DEEM (Wang et al., 2024a), using dynamic expert modeling; GPT-3.5 and GPT-3.5+COT (Lan et al., 2024), the base and CoT-prompted versions; GPT-EDDA (Ding et al., 2024b), an encoder-decoder data augmentation framework; Llama-2-7b-chat and Llama-2-13b-chat (Garg and Caragea, 2024), the 7B and 13B chat models; Manual-CoT (Wei et al., 2022) and Auto-CoT (Zhang et al., 2023c), CoT methods with manual or automatic demonstrations; ExpertPrompt (Xu et al., 2023), instructing LLMs to act as experts; SPP (Wang et al., 2024b), a solo performance prompting method; and LC-CoT (Zhang et al., 2023a), a logically consistent CoT approach.
+
+Detailed descriptions and performance results for these models across different datasets are provided in Appendix C.
+
+# 4.4 Implementation Details
+
+In our study, we utilize four large language models: GPT-3.5(Brown et al., 2020), Qwen2.5-7B-Instruct(Qwen Team, 2024) Llama-3.3-70B-Instruct(Llama Team, 2024), and Mistral-7B-Instruct-v0.3(Jiang et al., 2023). The experiments were conducted on a single NVIDIA A800 GPU with 80GB of RAM, utilizing bfloat16 precision to optimize memory usage and computational efficiency. To ensure the reproducibility of the LLMs' responses, we set the temperature parameter to 0 during inference. This configuration helps maintain consistent outputs across multiple runs. The results reported in our experiments are averaged over 5 repeated runs to ensure statistical reliability and mitigate the impact of any variance in model performance.
+
+# 5 Results and Discussion
+
+This section addresses the following research questions (RQs) based on our experimental results: RQ1: How does the performance of our MPRF compare to state-of-the-art stance detection models on PStance, SEM16 and VAST datasets? RQ2: Is each component of the MPRF effective and contributory to overall performance? RQ3: How do the weighting configurations of reasoning path scores  $(s_i^{\prime})$  and path weights  $(w_{i})$  affect the performance of MPRF? RQ4: How does the performance of MPRF vary across different models?
+
+<table><tr><td>Methods</td><td>DT</td><td>HC</td><td>FM</td><td>LA</td><td>A</td><td>CC</td></tr><tr><td colspan="7">Statistics-based Models</td></tr><tr><td>BiCond</td><td>30.5</td><td>32.7</td><td>40.6</td><td>34.4</td><td>31.0</td><td>15.0</td></tr><tr><td>CrossNet</td><td>35.6</td><td>38.3</td><td>41.7</td><td>38.5</td><td>39.7</td><td>22.8</td></tr><tr><td colspan="7">BERT-based Model</td></tr><tr><td>BERT</td><td>40.1</td><td>49.6</td><td>41.9</td><td>44.8</td><td>55.2</td><td>37.3</td></tr><tr><td>PT-HCL</td><td>50.1</td><td>54.5</td><td>54.6</td><td>50.9</td><td>56.5</td><td>38.9</td></tr><tr><td>TGA-Net</td><td>40.7</td><td>49.3</td><td>46.6</td><td>45.2</td><td>52.7</td><td>36.6</td></tr><tr><td>Joint-CL</td><td>50.5</td><td>54.8</td><td>53.8</td><td>49.5</td><td>54.5</td><td>39.7</td></tr><tr><td>KPatch</td><td>41.1</td><td>49.7</td><td>43.9</td><td>43.8</td><td>39.9</td><td>31.9</td></tr><tr><td>TATA</td><td>63.8</td><td>65.4</td><td>66.9</td><td>62.9</td><td>52.1</td><td>41.6</td></tr><tr><td>EZSD-CP</td><td>68.8</td><td>76.3</td><td>62.2</td><td>64.4</td><td>54.4</td><td>37.3</td></tr><tr><td>CNet-Ad</td><td>47.8</td><td>59.2</td><td>50.7</td><td>54.9</td><td>57.4</td><td>50.7</td></tr><tr><td>TarBK-BERT</td><td>50.8</td><td>55.1</td><td>53.8</td><td>48.7</td><td>56.2</td><td>39.5</td></tr><tr><td colspan="7">LLM-based Model</td></tr><tr><td>KASD-ChatGPT</td><td>-</td><td>80.3</td><td>70.4</td><td>62.7</td><td>-</td><td>-</td></tr><tr><td>COLA</td><td>68.5</td><td>81.7</td><td>63.4</td><td>71.0</td><td>70.8</td><td>65.5</td></tr><tr><td>GPT-3.5</td><td>62.5</td><td>68.7</td><td>44.7</td><td>51.5</td><td>9.1</td><td>31.1</td></tr><tr><td>GPT-3.5+COT</td><td>63.3</td><td>70.9</td><td>47.7</td><td>53.4</td><td>13.3</td><td>34.0</td></tr><tr><td>GPT-EDDA</td><td>69.5</td><td>80.1</td><td>69.2</td><td>62.7</td><td>67.2</td><td>68.5</td></tr><tr><td>LCDA</td><td>70.0</td><td>79.8</td><td>70.0</td><td>69.4</td><td>-</td><td>-</td></tr><tr><td>LC-CoT</td><td>71.7</td><td>82.9</td><td>70.4</td><td>63.2</td><td>-</td><td>-</td></tr><tr><td colspan="7">MPRF (Ours)</td></tr><tr><td>GPT-3.5</td><td>80.3</td><td>83.5</td><td>83.2</td><td>84.1</td><td>80.4</td><td>82.9</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>84.4*</td><td>83.1*</td><td>84.5*</td><td>82.5</td><td>83.9*</td><td>82.3*</td></tr><tr><td>LLaMA-3.3-70B</td><td>81.5</td><td>83.4</td><td>83.4</td><td>85.3*</td><td>84.3*</td><td>83.6*</td></tr><tr><td>Mistral-7B-Instruct-v0.3</td><td>82.6</td><td>84.5*</td><td>87.7*</td><td>83.2</td><td>82.7</td><td>83.4</td></tr></table>
+
+Table 1: Zero-shot stance detection results on the SEM16 dataset. The best scores are in bold. Results with * denote that MPRF significantly outperforms baselines with the p-value  $< {0.05}$  .
+
+Performance Comparison with State-of-the-Art Models As shown in Tables 1, 2, and 3, MPRF significantly outperforms state-of-the-art models across all three datasets—SEM16, VAST, and PStance—demonstrating both effectiveness and innovation in stance detection.
+
+On SEM16, MPRF achieves the highest scores across multiple targets (e.g., FM:  $87.7\%$ , HC:  $84.5\%$ ), surpassing strong LLM-based baselines such as GPT-EDDA and KASD by a large margin. In particular, our method shows robust generalization to challenging targets like "Atheism" and "Climate Change", where traditional models often struggle.
+
+On VAST, MPRF achieves an overall F1 score of  $83.5\%$  with Mistral-7B-Instruct-v0.3, outperforming existing methods by more than  $3\%$ . Notably, it performs exceptionally well in zero-shot settings  $(85.4\%)$ , highlighting its ability to generalize without target-specific training data.
+
+On PStance, MPRF sets a new state of the art with an  $\mathrm{F}_{\mathrm{macro}}$  of 88.41%, surpassing DEEM by 3.26%. The model also achieves the best individual performance on Biden (89.36%) and Trump (89.33%), demonstrating strong adaptability to real-world political discourse. These results affirm that MPRF's structured multi-path reasoning framework brings substantial improvements over conventional single-path or static prompting approaches, offering both enhanced accuracy and interpretability.
+
+<table><tr><td>Model</td><td>Zero-Shot</td><td>Few-Shot</td><td>Overall</td></tr><tr><td>BiCond</td><td>42.8</td><td>40.0</td><td>41.5</td></tr><tr><td>CrossNet</td><td>43.4</td><td>47.4</td><td>45.5</td></tr><tr><td>TGA-Net</td><td>66.6</td><td>66.3</td><td>66.5</td></tr><tr><td>BERT-GCN</td><td>68.6</td><td>69.7</td><td>69.2</td></tr><tr><td>CKE-Net</td><td>70.2</td><td>70.1</td><td>70.1</td></tr><tr><td>GDA-CL</td><td>70.5</td><td>-</td><td>-</td></tr><tr><td>PT-HCL</td><td>71.6</td><td>-</td><td>-</td></tr><tr><td>WS-BERT</td><td>75.3</td><td>73.6</td><td>74.5</td></tr><tr><td>CNet-Ad</td><td>73.2</td><td>71.8</td><td>72.6</td></tr><tr><td>TarBK-BERT</td><td>73.6</td><td>-</td><td>-</td></tr><tr><td>Joint-CL</td><td>72.3</td><td>71.5</td><td>-</td></tr><tr><td>EZSD-CP</td><td>73.6</td><td>-</td><td>-</td></tr><tr><td>StSQA</td><td>68.9</td><td>-</td><td>-</td></tr><tr><td>COLA</td><td>73.4</td><td>-</td><td>-</td></tr><tr><td>TATA</td><td>77.1</td><td>74.1</td><td>76.3</td></tr><tr><td>GPT-3.5-Turbo</td><td>65.0</td><td>-</td><td>-</td></tr><tr><td>GPT-3.5-Turbo-CoT</td><td>66.4</td><td>-</td><td>-</td></tr><tr><td>KASD-BERT</td><td>76.8</td><td>-</td><td>-</td></tr><tr><td>LKI-BART</td><td>79.6</td><td>-</td><td>-</td></tr><tr><td>CKI</td><td>81.9</td><td>79.6</td><td>80.7</td></tr><tr><td>EDDA-LLaMA</td><td>70.3</td><td>-</td><td>-</td></tr><tr><td>LC-CoT</td><td>72.5</td><td>-</td><td>-</td></tr><tr><td>LCDA</td><td>80.3</td><td>-</td><td>-</td></tr><tr><td colspan="4">MPRF (Ours)</td></tr><tr><td>GPT-3.5</td><td>81.1</td><td>82.3</td><td>82.1</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>85.4*</td><td>82.0</td><td>83.2</td></tr><tr><td>LLaMA-3.3-70B</td><td>80.6</td><td>82.1</td><td>81.2</td></tr><tr><td>Mistral-7B-Instruct-v0.3</td><td>84.2*</td><td>83.1*</td><td>83.5*</td></tr></table>
+
+Contribution of Each Component of MPRF To evaluate the effectiveness of each component, we conducted ablation studies on SEM16 (Table 4), with additional results for PStance and VAST in Appendix D. Removing any component leads to performance drops, confirming their importance.
+
+Sentiment Path significantly influences targets involving emotional expression. Its removal causes
+
+Table 2: Stance detection performance  $(\%)$  on VAST. The best scores are in bold. Results with  $*$  denote that MPRF significantly outperforms baselines with the p-value  $< 0.05$ .  
+
+<table><tr><td>Method</td><td>Trump</td><td>Biden</td><td>Sanders</td><td>F_macro</td></tr><tr><td>BiLSTM</td><td>76.92</td><td>77.95</td><td>69.75</td><td>74.87</td></tr><tr><td>CNN</td><td>76.80</td><td>77.22</td><td>71.40</td><td>75.14</td></tr><tr><td>TAN</td><td>77.10</td><td>77.64</td><td>71.60</td><td>75.45</td></tr><tr><td>BiCond</td><td>77.15</td><td>77.69</td><td>71.24</td><td>75.36</td></tr><tr><td>PGCNN</td><td>76.87</td><td>76.60</td><td>72.13</td><td>75.20</td></tr><tr><td>GCAE</td><td>78.96</td><td>77.95</td><td>71.82</td><td>76.24</td></tr><tr><td>BERT</td><td>78.28</td><td>78.70</td><td>72.45</td><td>76.48</td></tr><tr><td>BERTweet</td><td>82.48</td><td>81.02</td><td>78.09</td><td>80.53</td></tr><tr><td>WS-BERT</td><td>84.97</td><td>82.86</td><td>79.97</td><td>82.60</td></tr><tr><td>TarBK-BERT</td><td>65.80</td><td>75.49</td><td>70.45</td><td>70.58</td></tr><tr><td>GPT-3.5</td><td>79.80</td><td>79.65</td><td>77.77</td><td>79.07</td></tr><tr><td>Llama-2-7b-chat-ft</td><td>72.00</td><td>67.96</td><td>65.57</td><td>68.51</td></tr><tr><td>Llama-2-13b-chat-ft</td><td>76.62</td><td>71.88</td><td>68.44</td><td>72.31</td></tr><tr><td>Stanceformer</td><td>85.35</td><td>83.96</td><td>80.57</td><td>83.30</td></tr><tr><td>CKI</td><td>86.2</td><td>84.1</td><td>80.5</td><td>83.6</td></tr><tr><td>Manual-CoT</td><td>85.4</td><td>83.8</td><td>80.9</td><td>83.37</td></tr><tr><td>StSQA</td><td>85.7</td><td>82.8</td><td>80.8</td><td>83.1</td></tr><tr><td>Auto-CoT</td><td>84.1</td><td>82.8</td><td>80.6</td><td>82.5</td></tr><tr><td>ExpertPrompt</td><td>84.7</td><td>84.7</td><td>81.2</td><td>83.53</td></tr><tr><td>SPP</td><td>85.1</td><td>84.6</td><td>81.5</td><td>83.73</td></tr><tr><td>DEEM</td><td>86.4</td><td>86.1</td><td>82.1</td><td>84.87</td></tr><tr><td>COLA</td><td>86.6</td><td>84.0</td><td>79.7</td><td>83.1</td></tr><tr><td>KASD-ChatGPT</td><td>85.06</td><td>84.59</td><td>79.96</td><td>83.2</td></tr><tr><td>MPRF (Ours)</td><td></td><td></td><td></td><td></td></tr><tr><td>GPT-3.5</td><td>88.20</td><td>88.60</td><td>85.10</td><td>87.60</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>89.50*</td><td>88.15*</td><td>86.75*</td><td>88.13</td></tr><tr><td>LLaMA-3.3-70B</td><td>89.63*</td><td>88.26</td><td>85.35</td><td>87.75</td></tr><tr><td>Mistral-7B-Instruct-v0.3</td><td>89.33</td><td>89.36*</td><td>86.53</td><td>88.41*</td></tr></table>
+
+Table 3: Comparison of different models on the PStance. The best scores are in bold. Results with * denote that MPRF significantly outperforms baselines with the p-value  $< {0.05}$  .
+
+a  $4.6\%$  drop on FM (from  $87.7\%$  to  $83.1\%$ ) on SEM16 and a  $1.9\%$  drop on Trump (from  $89.33\%$  to  $87.4\%$ ) on PStance.
+
+Fact-Based Reasoning is crucial for evidence-based stance formation. Without it, FM drops by  $3.7\%$  on SEM16 and Trump drops by  $1.6\%$  on PStance.
+
+Expert Opinion Path enhances predictions through authoritative viewpoints. Removing it leads to a  $3.4\%$  drop on FM and a  $1.2\%$  drop on Trump.
+
+Public Opinion Path contributes moderate but consistent improvements, especially in social context modeling. Its removal results in a  $3.3\%$  drop on FM and a  $1.3\%$  drop on Trump.
+
+Optimization plays a key role in refining low-quality paths. Without optimization, FM drops sharply by  $7.1\%$ , indicating its critical impact on reasoning quality.
+
+Weighted Fusion has the most significant impact among all components. Removing it causes an  $8.6\%$  drop on FM and a  $3.6\%$  drop on Trump, highlighting its essential role in integrating diverse perspectives effectively.
+
+<table><tr><td>Model</td><td>DT</td><td>HC</td><td>FM</td><td>LA</td><td>A</td><td>CC</td></tr><tr><td>MPRF</td><td>82.6</td><td>84.5</td><td>87.7</td><td>83.2</td><td>82.7</td><td>83.4</td></tr><tr><td>w/o Sentiment Path</td><td>80.3</td><td>81.2</td><td>83.1</td><td>81.4</td><td>81.8</td><td>81.9</td></tr><tr><td>w/o Fact-Based Path</td><td>81.1</td><td>82.4</td><td>84.0</td><td>82.6</td><td>82.3</td><td>82.0</td></tr><tr><td>w/o Expert Opinion Path</td><td>81.5</td><td>82.8</td><td>84.3</td><td>82.9</td><td>83.1</td><td>82.6</td></tr><tr><td>w/o Public Opinion Path</td><td>81.7</td><td>83.1</td><td>84.4</td><td>83.0</td><td>83.3</td><td>82.9</td></tr><tr><td>w/o Optimization</td><td>79.7</td><td>80.4</td><td>80.6</td><td>80.3</td><td>80.5</td><td>80.0</td></tr><tr><td>w/o Weighted Fusion</td><td>78.9</td><td>79.8</td><td>79.1</td><td>79.6</td><td>79.8</td><td>78.5</td></tr></table>
+
+Impact of Weighting Configurations on MPRF Performance Table 5 shows the impact of different weighting configurations on MPRF's performance. The optimal configuration  $(\alpha = 0.4, \beta = 0.3, \gamma = 0.3)$  yields the highest F1 scores, including  $87.7\%$  for FM and  $84.5\%$  for HC, outperforming other configurations. This demonstrates that a balanced weighting prioritizes relevance while maintaining adequate consideration of evidence strength and logical consistency.
+
+Other configurations, such as the equal weighting  $(\alpha = 0.33, \beta = 0.33, \gamma = 0.33)$ , result in significant performance drops, with FM and HC scores dropping to  $83.1\%$  and  $81.2\%$ , respectively. Skewed configurations that emphasize a single criterion—like relevance  $(\alpha = 0.5, \beta = 0.25, \gamma = 0.25)$ , evidence  $(\alpha = 0.25, \beta = 0.5, \gamma = 0.25)$ , or logical consistency  $(\alpha = 0.25, \beta = 0.25, \gamma = 0.5)$  also perform poorly. These configurations fail to balance the complementary contributions of all components, leading to suboptimal results.
+
+Table 4: Ablation Study Results on SEM16 (%) using Mistral-7B-Instruct. The best scores are in bold.  
+
+<table><tr><td>Weighting Configurations</td><td>DT HC FM LA A CC</td></tr><tr><td>MPRF</td><td>82.6 84.5 87.7 83.2 82.7 83.4</td></tr><tr><td>Balanced Weights</td><td>80.3 81.2 83.1 81.4 81.8 81.9</td></tr><tr><td>Enhanced Relevance</td><td>81.4 82.2 81.0 80.4 81.1 81.6</td></tr><tr><td>Enhanced Evidence Strength</td><td>81.8 83.5 82.1 81.3 80.0 82.0</td></tr><tr><td>Enhanced Logical Consistency</td><td>81.4 82.7 82.9 83.0 80.6 81.8</td></tr></table>
+
+Table 6 illustrates the impact of different weighting configurations on the SEM16 dataset. The MPRF  $(w_{i} \propto s_{i}^{\prime})$  configuration, which dynamically adjusts path weights based on their scores  $(s_{i}^{\prime})$ , achieves the best performance, with F1 scores of  $87.7\%$  for FM and  $84.5\%$  for HC, outperforming all other configurations.
+
+In contrast, the Equal Weights configuration shows a significant performance drop, with FM and HC scores decreasing to  $83.1\%$  and  $81.2\%$ , respectively. Configurations assigning high weights (0.8) to specific paths, such as sentiment or fact-based reasoning, also fail to surpass the MPRF  $(w_{i} \propto s_{i})$
+
+configuration.
+
+The superiority of dynamic weighting highlights the importance of prioritizing high-quality reasoning paths. Equal weighting dilutes the contributions of critical paths, while overemphasizing a single path reduces the complementary contributions from other paths.
+
+In conclusion, dynamically adjusting weights based on path scores is crucial for maximizing performance, while equal or skewed weighting results in significant performance decline, emphasizing the need for balanced contributions from multiple reasoning paths.
+
+Table 5: Effect of Weighting Configurations on SEM16 (%) using Mistral-7B-Instruct. The best scores are in bold.  
+
+<table><tr><td colspan="2">Weighting Configurations</td><td>DT</td><td>HC</td><td>FM</td><td>LA</td><td>A</td><td>CC</td></tr><tr><td colspan="2">MPRF (wi ∝ s&#x27;)</td><td>82.6</td><td>84.5</td><td>87.7</td><td>83.2</td><td>82.7</td><td>83.4</td></tr><tr><td colspan="2">Equal Weights</td><td>80.3</td><td>81.2</td><td>83.1</td><td>81.4</td><td>81.8</td><td>81.9</td></tr><tr><td colspan="2">High for Sentiment Path</td><td>81.4</td><td>82.8</td><td>84.3</td><td>81.9</td><td>81.6</td><td>82.1</td></tr><tr><td colspan="2">High for Fact-Based Path</td><td>81.8</td><td>83.7</td><td>85.1</td><td>82.6</td><td>82.0</td><td>82.5</td></tr><tr><td colspan="2">High for Expert Path</td><td>80.8</td><td>82.3</td><td>83.5</td><td>82.0</td><td>81.7</td><td>82.2</td></tr><tr><td colspan="2">High for Public Path</td><td>80.9</td><td>82.5</td><td>83.8</td><td>81.8</td><td>81.5</td><td>81.9</td></tr></table>
+
+Table 6: Comparison of Weighting Configurations (%) on SEM16 using Mistral-7B-Instruct. The best scores are in bold.
+
+Performance of MPRF Across Different Models Table 1, Table 2, and Table 3 summarize the performance of MPRF across four distinct large language models on the SEM16, VAST, and PStance datasets. These results demonstrate that MPRF consistently improves stance detection performance across diverse model architectures, suggesting that its effectiveness stems from the framework design rather than specific model capabilities.
+
+On SEM16, for example, all four models achieve strong results under MPRF, with Mistral-7B-Instruct-v0.3 performing best on FM (87.7%) and HC (84.5%), and Qwen2.5-7B-Instruct leading on DT (84.4%). The relatively smaller performance gaps between models suggest that MPRF mitigates inherent model-specific biases or limitations by leveraging structured reasoning paths and weighted fusion. Similarly, in zero-shot and few-shot settings on VAST and PStance, MPRF enables consistent improvements across models.
+
+The key takeaway is that MPRF's performance gains are not solely attributable to the choice of LLM, but rather to its ability to generate, evaluate, and integrate multiple reasoning paths effectively. This consistency across models confirms that MPRF provides a robust and model-agnostic enhancement to stance detection frameworks.
+
+![](images/1a7de8b997e0048bfed01aa43e99ea861e390e0b1046cd0ba43a662ab745898a.jpg)  
+Figure 2: Stance Detection Process: Multi-Path Analysis of the Statement on Climate Change
+
+Case Study: Stance Detection on "Climate Change is a Real Concern" To demonstrate the interpretability of our Multi-Path Reasoning Framework (MPRF), we present a detailed case study in Figure 2: "Climate change is something we should pay attention to, but it's not the end of the world. Many people exaggerate the effects of climate change. Yes, it's real, but there are much bigger issues at hand." Our goal is to detect the stance towards the target "Climate Change is a Real Concern".
+
+MPRF analyzes this statement through four distinct, human-understandable reasoning paths, each representing a different dimension of opinion formation. First, the Sentiment Analysis path identifies the overall tone as neutral. The text acknowledges the reality of climate change ("it's real") but tempers this with phrases like "not the end of the world" and "many people exaggerate," resulting in a balanced emotional expression and a stance prediction of Neutral. Second, the Factual Reasoning path focuses on the evidence presented. The claim that "many people exaggerate the effects" is interpreted as a direct challenge to the severity of the scientific consensus on climate change, leading to a stance prediction of Against. Third, the Expert Opinion path infers that the statement contradicts the overwhelming consensus of climate scientists. By downplaying the urgency and scale of the issue, the argument implicitly opposes the authoritative viewpoint, resulting in a stance of Against. Fourth, the Public Opinion path interprets the phrase "many people exaggerate" as a reflection of a perceived public sentiment that overstates the problem, thus classifying the stance as Against.
+
+Crucially, MPRF does not treat these paths equally. Each path is rigorously evaluated by an LLM on three criteria: relevance, evidence strength,
+
+and logical consistency, and assigned a quantitative score. For instance, the Fact-Based path received a high relevance score (8) due to its direct engagement with the core issue, though its evidence strength was moderate (5). The Expert Opinion path scored highly in logical consistency (8), reflecting a coherent argument structure. These individual scores  $(s_i^{\prime})$  are then used in a weighted fusion mechanism. The paths predicting Against (Fact-Based, Expert Opinion, Public Opinion) collectively contribute a significantly higher aggregated weighted score  $(S_{\mathrm{against}})$  than the single Neutral path. This transparent, score-driven aggregation process results in the final prediction of Against.
+
+# 6 Conclusion
+
+In this paper, we proposed the Multi-Path Reasoning Framework (MPRF), a novel framework for stance detection that enhances accuracy, robustness, and interpretability by generating, evaluating, and optimizing multiple reasoning paths. Through experiments on three datasets, SEM16, VAST, and PStance, MPRF consistently demonstrated strong performance across various models. The results highlight the adaptability and effectiveness of MPRF in both zero-shot and few-shot settings, achieving state-of-the-art performance and providing interpretable reasoning chains.
+
+Despite these achievements, there are still some limitations to our approach. The reliance on manually designed prompts for path evaluation and scoring may limit scalability to more complex tasks or datasets. In future work, we aim to explore automated prompt optimization and more efficient reasoning strategies to further enhance MPRF's scalability and performance in real-world applications.
+
+# Limitations
+
+While the Multi-Path Reasoning Framework (MPRF) demonstrates strong performance and adaptability across various datasets and models, there are several limitations to address. First, the reliance on manually designed prompts for evaluating and scoring reasoning paths may limit scalability to more complex tasks or domains where predefined prompts may not generalize well. Second, generating and optimizing multiple reasoning paths introduces additional computational overhead, which could hinder its deployment in large-scale or time-sensitive applications. Lastly, the interpretability of MPRF, while improved compared to traditional black-box models, still requires further exploration to ensure that reasoning paths are both human-readable and aligned with domain-specific requirements.
+
+# Ethical Considerations
+
+The Multi-Path Reasoning Framework(MPRF) aims to improve stance detection by enhancing transparency and interpretability in its decision-making process. However, the use of large language models (LLMs) in MPRF raises potential ethical concerns. First, LLMs are trained on vast amounts of internet data, which may include biased or harmful content, potentially influencing the reasoning paths generated by MPRF. Second, the interpretability of the reasoning paths may lead to unintended misuse, such as generating misleading or manipulative arguments. Finally, while MPRF strives for fairness, it inherits biases present in the underlying models and datasets, which could result in unfair or inaccurate stance detection in certain contexts. To mitigate these risks, we have carefully evaluated MPRF's outputs and employed diverse datasets to reduce potential biases. Future work will focus on improving fairness, bias detection, and ethical safeguards to ensure responsible application of MPRF in real-world scenarios.
+
+# Acknowledgments
+
+This research was supported by funding from the National Natural Science Foundation of China under Grant No.62441229 for the project "High-quality Dataset Construction". This valuable resource significantly enhanced the reliability and robustness of our experimental results. We would like to extend our sincere gratitude to all those who contributed to this work.
+
+# References
+
+Abeer Aldayel and Walid Magdy. 2019. Your stance is exposed! analysing possible factors for stance detection on social media. Proc. ACM Hum.-Comput. Interact., 3(CSCW).  
+Emily Allaway and Kathleen McKeown. 2020. Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913-8931, Online. Association for Computational Linguistics.  
+Emily Allaway, Malavika Srikanth, and Kathleen McKeown. 2021. Adversarial learning for zero-shot stance detection on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4756-4767, Online. Association for Computational Linguistics.  
+Isabelle Augenstein, Tim Rocktäschel, Andreas Vlachos, and Kalina Bontcheva. 2016. Stance detection with bidirectional conditional encoding. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876-885, Austin, Texas. Association for Computational Linguistics.  
+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, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Ma teusz 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*, abs/2005.14165.  
+Daniele De Vinco, Alessia Antelmi, Carmine Spagnuolo, and Luca Maria Aiello. 2024. Deciphering conversational networks: Stance detection via hypergraphs and llms. In Companion Publication of the 16th ACM Web Science Conference, Websci Companion '24, page 3-4, New York, NY, USA. Association for Computing Machinery.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Kuntal Dey, Ritvik Shrivastava, and Saroj Kaushik. 2017. Twitter stance detection — a subjectivity and sentiment polarity inspired two-phase approach. In
+
+2017 IEEE International Conference on Data Mining Workshops (ICDMW), pages 365-372.  
+Daijun Ding, Rong Chen, Liwen Jing, Bowen Zhang, Xu Huang, Li Dong, Xiaowen Zhao, and Ge Song. 2024a. Cross-target stance detection by exploiting target analytical perspectives. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10651-10655.  
+Dajun Ding, Li Dong, Zhichao Huang, Guangning Xu, Xu Huang, Bo Liu, Liwen Jing, and Bowen Zhang. 2024b. EDDA: An encoder-decoder data augmentation framework for zero-shot stance detection. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 5484-5494, Torino, Italia. ELRA and ICCL.  
+Jiachen Du, Ruifeng Xu, Yulan He, and Lin Gui. 2017. Stance classification with target-specific neural attention. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, pages 3988-3994.  
+Hao Fei, Bobo Li, Qian Liu, Lidong Bing, Fei Li, and Tat-Seng Chua. 2023. Reasoning implicit sentiment with chain-of-thought prompting. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1171-1182, Toronto, Canada. Association for Computational Linguistics.  
+Krishna Garg and Cornelia Caragea. 2024. Stanceformer: Target-aware transformer for stance detection. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 4969-4984, Miami, Florida, USA. Association for Computational Linguistics.  
+Joseph Gatto, Omar Sharif, and Sarah Preum. 2023. Chain-of-thought embeddings for stance detection on social media. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4154-4161, Singapore. Association for Computational Linguistics.  
+Eduardo Graells-Garrido and Ricardo Baeza-Yates. 2022. Bots don't vote, but they surely bother! a study of anomalous accounts in a national referendum. In Proceedings of the 14th ACM Web Science Conference 2022, WebSci '22, page 302-306, New York, NY, USA. Association for Computing Machinery.  
+Eduardo Graells-Garrido, Ricardo Baeza-Yates, and Mounia Lalmas. 2020. Every colour you are: Stance prediction and turnaround in controversial issues. In Proceedings of the 12th ACM Conference on Web Science, WebSci '20, page 174-183, New York, NY, USA. Association for Computing Machinery.  
+Hans Hanley and Zakir Durmeric. 2023. TATA: Stance detection via topic-agnostic and topic-aware embeddings. In Proceedings of the 2023 Conference on
+
+Empirical Methods in Natural Language Processing, pages 11280-11294, Singapore. Association for Computational Linguistics.  
+Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. 2022. Few-shot cross-lingual stance detection with sentiment-based pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10729-10737.  
+Kazi Saidul Hasan and Vincent Ng. 2014. Why are you taking this stance? identifying and classifying reasons in ideological debates. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 751-762, Doha, Qatar. Association for Computational Linguistics.  
+Zihao He, Negar Mokhberian, and Kristina Lerman. 2022. Infusing knowledge from Wikipedia to enhance stance detection. In Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis, pages 71-77, Dublin, Ireland. Association for Computational Linguistics.  
+Binxuan Huang and Kathleen Carley. 2018. Parameterized convolutional neural networks for aspect level sentiment classification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1091-1096, Brussels, Belgium. Association for Computational Linguistics.  
+Hu Huang, Bowen Zhang, Yangyang Li, Baoquan Zhang, Yuxi Sun, Chuyao Luo, and Cheng Peng. 2023. Knowledge-enhanced prompt-tuning for stance detection. ACM Trans. Asian Low-Resour. Lang. Inf. Process., 22(6).  
+Albert Qiaochu Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. ArXiv, abs/2310.06825.  
+Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746-1751, Doha, Qatar. Association for Computational Linguistics.  
+Dilek Kucuk and Fazli Can. 2020. Stance detection: A survey. ACM Comput. Surv., 53(1).  
+Mirko Lai, Alessandra Teresa Cignarella, Delia Irazú Hernández Farias, Cristina Bosco, Viviana Patti, and Paolo Rosso. 2020. Multilingual stance detection in social media political debates. Comput. Speech Lang., 63:101075.
+
+Xiaochong Lan, Chen Gao, Depeng Jin, and Yong Li. 2024. Stance detection with collaborative roleinfused llm-based agents. In Proceedings of the Eighteenth International AAAI Conference on Web and Social Media, ICWSM 2024, Buffalo, New York, USA, June 3-6, 2024, pages 891-903. AAAI Press.  
+Yuanyuan Lei, Md Messal Monem Miah, Ayesha Qamar, Sai Ramana Reddy, Jonathan Tong, Haotian Xu, and Ruihong Huang. 2024. EMONA: Event-level moral opinions in news articles. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5239-5251, Mexico City, Mexico. Association for Computational Linguistics.  
+Ang Li, Bin Liang, Jingqian Zhao, Bowen Zhang, Min Yang, and Ruifeng Xu. 2023. Stance detection on social media with background knowledge. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15703-15717, Singapore. Association for Computational Linguistics.  
+Yang Li and Jiawei Yuan. 2022. Generative data augmentation with contrastive learning for zero-shot stance detection. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6985-6995, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yingjie Li, Tiberiu Sosea, Aditya Sawant, Ajith Jayaraman Nair, Diana Inkpen, and Cornelia Caragea. 2021. P-stance: A large dataset for stance detection in political domain. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2355-2365, Online. Association for Computational Linguistics.  
+Yingjie Li and Yue Zhang. 2024. Pro-woman, anti-man? identifying gender bias in stance detection. In Findings of the Association for Computational Linguistics: ACL 2024, pages 3229-3236, Bangkok, Thailand. Association for Computational Linguistics.  
+Bin Liang, Zixiao Chen, Lin Gui, Yulan He, Min Yang, and Ruifeng Xu. 2022a. Zero-shot stance detection via contrastive learning. In Proceedings of the ACM Web Conference 2022, WWW '22, page 2738-2747, New York, NY, USA. Association for Computing Machinery.  
+Bin Liang, Qinglin Zhu, Xiang Li, Min Yang, Lin Gui, Yulan He, and Ruifeng Xu. 2022b. JointCL: A joint contrastive learning framework for zero-shot stance detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 81-91, Dublin, Ireland. Association for Computational Linguistics.  
+Shuohao Lin, Wei Chen, Yunpeng Gao, Zhishu Jiang, Mengqi Liao, Zhiyu Zhang, Shuyuan Zhao, and Huaiyu Wan. 2024. KPatch: Knowledge patch to pretrained language model for zero-shot stance detection
+
+on social media. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 9961-9973, Torino, Italia. ELRA and ICCL.  
+Zhan Ling, Yunhao Fang, Xuanlin Li, Zhiao Huang, Mingu Lee, Roland Memisevic, and Hao Su. 2023. Deductive verification of chain-of-thought reasoning. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, NY, USA. Curran Associates Inc.  
+Rui Liu, Zheng Lin, Yutong Tan, and Weiping Wang. 2021. Enhancing zero-shot and few-shot stance detection with commonsense knowledge graph. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 3152-3157, Online. Association for Computational Linguistics.  
+Llama Team. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: multimodal reasoning via thought chains for science question answering. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA. Curran Associates Inc.  
+Yun Luo, Zihan Liu, Yuefeng Shi, Stan Z. Li, and Yue Zhang. 2022. Exploiting sentiment and common sense for zero-shot stance detection. In Proceedings of the 29th International Conference on Computational Linguistics, pages 7112-7123, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31-41, San Diego, California. 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.  
+Qwen Team. 2024. Qwen2.5: A party of foundation models.  
+M. Schuster and K.K. Paliwal. 1997. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11):2673-2681.  
+Maksym Taranukhin, Vered Shwartz, and Evangelos Milios. 2024. *Stance reasoner: Zero-shot stance detection on social media with explicit reasoning*.
+
+In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 15257-15272, Torino, Italia. ELRA and ICCL.  
+Xiaolong Wang, Yile Wang, Sijie Cheng, Peng Li, and Yang Liu. 2024a. DEEM: Dynamic experienced expert modeling for stance detection. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 4530-4541, Torino, Italia. ELRA and ICCL.  
+Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. 2024b. Unleashing the emergent cognitive synergy in large language models: A task-solving agent through multi-persona self-collaboration. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 257-279, Mexico City, Mexico. Association for Computational Linguistics.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA. Curran Associates Inc.  
+Benfeng Xu, An Yang, Junyang Lin, Quan Wang, Chang Zhou, Yongdong Zhang, and Zhendong Mao. 2023. Expertprompting: Instructing large language models to be distinguished experts. Preprint, arXiv:2305.14688.  
+Chang Xu, Cecile Paris, Surya Nepal, and Ross Sparks. 2018. Cross-target stance classification with self-attention networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 778-783, Melbourne, Australia. Association for Computational Linguistics.  
+Hanzi Xu, Slobodan Vucetic, and Wenpeng Yin. 2022. OpenStance: Real-world zero-shot stance detection. In Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 314-324, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Wei Xue and Tao Li. 2018. Aspect based sentiment analysis with gated convolutional networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2514-2523, Melbourne, Australia. Association for Computational Linguistics.  
+Ming Yan, Tianyi Zhou Joey, and W. Tsang Ivor. 2024. Collaborative knowledge infusion for low-resource stance detection. *Big Data Mining and Analytics*, 7(3):682-698.
+
+Song Yang and Jacopo Urbani. 2021. Tribrid: Stance classification with neural inconsistency detection. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6831-6843, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Zhenyin Yao, Wenzhong Yang, and Fuyuan Wei. 2024. Enhancing zero-shot stance detection with contrastive and prompt learning. Entropy, 26(4).  
+Bowen Zhang, Daijun Ding, Liwen Jing, Genan Dai, and Nan Yin. 2024a. How would stance detection techniques evolve after the launch of chatgpt? Preprint, arXiv:2212.14548.  
+Bowen Zhang, Daijun Ding, Liwen Jing, and Hu Huang. 2023a. A logically consistent chain-of-thought approach for stance detection. CoRR, abs/2312.16054.  
+Bowen Zhang, Xianghua Fu, Daijun Ding, Hu Huang, Genan Dai, Nan Yin, Yangyang Li, and Liwen Jing. 2024b. Investigating chain-of-thought with chatgpt for stance detection on social media. Preprint, arXiv:2304.03087.  
+Bowen Zhang, Xu Li, Jun Ma, Xi Zhang, Genan Dai, and Jianhua Ye. 2025. Zero-shot stance detection with logically consistent data augmentation. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5.  
+Chen Zhang, Qiuchi Li, and Dawei Song. 2019. Aspect-based sentiment classification with aspect-specific graph convolutional networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4568-4578, Hong Kong, China. Association for Computational Linguistics.  
+Hao Zhang, Yizhou Li, Tuanfei Zhu, and Chuang Li. 2024c. Commonsense-based adversarial learning framework for zero-shot stance detection. Neurocomputing, 563:126943.  
+Hong Zhang, Haewoon Kwak, Wei Gao, and Jisun An. 2023b. Wearing masks implies refuting trump?: Towards target-specific user stance prediction across events in COVID-19 and us election 2020. In Proceedings of the 15th ACM Web Science Conference 2023, WebSci '23, page 23-32, New York, NY, USA. Association for Computing Machinery.  
+Zhao Zhang, Yiming Li, Jin Zhang, and Hui Xu. 2024d. LLM-driven knowledge injection advances zero-shot and cross-target stance detection. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers), pages 371-378, Mexico City, Mexico. Association for Computational Linguistics.
+
+Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2023c. Automatic chain of thought prompting in large language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. Open-Review.net.
+
+Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, and Ed Chi. 2022. Least-to-most prompting enables complex reasoning in large language models. Preprint, arXiv:2205.10625.
+
+Yi Zhu, Ye Wang, Jianyuan Mu, Yun Li, Jipeng Qiang, Yunhao Yuan, and Xindong Wu. 2024. Short text classification with soft knowledgeable prompt-tuning. Expert Systems with Applications, 246:123248.
+
+# A Dataset Statistics
+
+<table><tr><td></td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td># Examples</td><td>13477</td><td>2062</td><td>3006</td></tr><tr><td># Unique Comments</td><td>1845</td><td>682</td><td>786</td></tr><tr><td># Zero-shot Topics</td><td>4003</td><td>383</td><td>600</td></tr><tr><td># Few-shot Topics</td><td>638</td><td>114</td><td>159</td></tr></table>
+
+Table 7: Statistics of VAST dataset.  
+
+<table><tr><td>Target</td><td>Favor</td><td>Against</td><td>Neutral</td></tr><tr><td>DT</td><td>148</td><td>299</td><td>260</td></tr><tr><td>HC</td><td>163</td><td>565</td><td>256</td></tr><tr><td>FM</td><td>268</td><td>511</td><td>170</td></tr><tr><td>LA</td><td>167</td><td>544</td><td>222</td></tr><tr><td>A</td><td>124</td><td>464</td><td>145</td></tr><tr><td>CC</td><td>335</td><td>26</td><td>203</td></tr></table>
+
+Table 8: Statistics of SEM16 dataset.  
+
+<table><tr><td colspan="2"></td><td>Trump</td><td>Biden</td><td>Sanders</td></tr><tr><td rowspan="2">Train</td><td>Favor</td><td>2,937</td><td>2,552</td><td>2,858</td></tr><tr><td>Against</td><td>3,425</td><td>3,254</td><td>2,198</td></tr><tr><td rowspan="2">Val</td><td>Favor</td><td>365</td><td>328</td><td>350</td></tr><tr><td>Against</td><td>430</td><td>417</td><td>284</td></tr><tr><td rowspan="2">Test</td><td>Favor</td><td>361</td><td>337</td><td>343</td></tr><tr><td>Against</td><td>435</td><td>408</td><td>292</td></tr><tr><td colspan="2">Total</td><td>7,953</td><td>7,296</td><td>6,325</td></tr></table>
+
+Table 9: Label distribution across different targets for P-Stance.
+
+# B Prompts
+
+We selected these four reasoning paths—sentiment analysis, factual reasoning, expert opinions, and
+
+public opinion—based on their semantic distinctiveness and empirical effectiveness in capturing key dimensions of stance reasoning. Each path addresses a unique aspect that commonly influences human judgment:
+
+- Sentiment Analysis Path: Captures emotional tone (e.g., "happy", "angry") to infer stance. Emotional language often strongly correlates with stance expression.  
+- Factual Reasoning Path: Analyzes factual evidence (e.g., statistics, legal references) to determine alignment with the target stance. Stance decisions frequently hinge on objective claims or data.  
+- Expert Opinion Path: Evaluates references to authoritative sources (e.g., scientists, legal scholars). Authority-driven arguments are particularly influential in contentious or technical topics.  
+- Public Opinion Path: Reflects collective sentiment derived from social media, surveys, or news coverage. Public consensus can significantly shape individual stances through social influence.
+
+# Prompt Template for Sentiment Path:
+
+Your task is to classify the stance of the comment on the topic as "favor", "against", or "neutral".
+
+Topic: Target
+
+Comment: Text
+
+Stance:
+
+Explain:
+
+1. Analyze the emotional tone of the comment. Identify key emotional words such as "happy", "sad", "angry", "disappointed", "excited", etc.  
+2. Based on the sentiment, determine if the comment expresses a positive, negative, or neutral emotion.  
+3. If the comment expresses a positive sentiment, classify the stance as "favor"; if negative, classify the stance as "against"; if neutral, classify the stance as "neutral".
+
+# Prompt Template for Fact-Based Reasoning Path:
+
+Your task is to classify the stance of the comment on the topic as "favor", "against", or "neutral".
+
+Topic: Target
+
+Comment: Text
+
+Stance:
+
+# Explain:
+
+1. Extract factual information from the comment, such as data, studies, or legal references related to the target.  
+2. Determine if the facts support or oppose the target stance.  
+3. If the facts present a positive outcome or support legalization, classify the stance as "favor"; if they present negative impacts or oppose it, classify the stance as "against"; if the facts are neutral or inconclusive, classify the stance as "neutral".
+
+# Prompt Template for Expert Opinion Path:
+
+Your task is to classify the stance of the comment on the topic as "favor", "against", or "neutral".
+
+Topic: Target
+
+Comment: Text
+
+Stance:
+
+Explain:
+
+1. Identify expert opinions or references within the comment.  
+2. Determine if the expert opinions support or oppose the target stance.  
+3. If the expert opinions support the target, classify the stance as "favor"; if they oppose it, classify the stance as "against"; if the expert opinions are neutral or inconclusive, classify the stance as "neutral".
+
+# Prompt Template for Public Opinion Path:
+
+Your task is to classify the stance of the comment on the topic as "favor", "against", or "neutral".
+
+Topic: Target
+
+Comment: Text
+
+Stance:
+
+Explain:
+
+1. Extract public opinion from the comment, such as social media reactions, public surveys, or common sentiments.  
+2. Determine if the public opinion supports or opposes the target.  
+3. If the public opinion supports target, classify the stance as "favor"; if it opposes, classify the stance as "against"; if the public opinion is neutral, classify the stance as "neutral".
+
+# Prompt for Relevance Score:
+
+Please evaluate the relevance of the following reasoning path: reasoning path, in relation to the target Target, given the comment: Text.
+
+How relevant is the reasoning path to the target Target? Please provide a relevance score from 1 to 10, where 1 represents low relevance and 10 represents high relevance.
+
+# Prompt for Evidence Strength Score:
+
+Evaluate the strength of the evidence in the following reasoning path:reasoning path for Target. Does the evidence strongly support the stance of target? Please provide an evidence strength score from 1 to 10, where 1 indicates weak or unreliable evidence and 10 indicates strong and reliable evidence.
+
+# Prompt for Logical Consistency Score:
+
+Evaluate the logical consistency of the following reasoning path:reasoning path for Target. Are the reasoning steps coherent and logically connected? Please provide a logical consistency score from 1 to 10, where 1 indicates poor consistency and 10 indicates excellent consistency.
+
+# Prompt Template for Optimization of Low-Scoring Reasoning Paths:
+
+Your task is to review and improve the following reasoning path for Target. This path has been evaluated and received low scores in one or more of the following areas: Relevance Score  $r_i$ , Evidence Strength Score  $e_i$ , or Logical Consistency Score  $l_i$ . The scores below 5 indicate weak areas that need optimization. The initial reasoning path is:
+
+reasoning path
+
+Please focus on improving the areas where the scores are low. Specifically, address the following:
+
+1. Low Relevance Score  $(r_i)$ : - Review the reasoning steps that have been marked as irrelevant or weakly connected to the target stance. Adjust these steps to ensure they are directly relevant to the target. If necessary, provide stronger connections to the target stance.  
+2. Low Evidence Strength Score  $(e_i)$ : - Identify reasoning steps that lack strong supporting evidence or facts. Propose more robust evidence, studies, or expert opinions to support these steps and strengthen the overall argument.  
+3. Low Logical Consistency Score  $(l_i)$ : - Examine any steps that show logical inconsistencies or errors. Identify these inconsistencies and suggest improvements to fix any logical flaws, ensuring the reasoning follows a consistent and sound structure.
+
+After revising the weak areas, provide the updated reasoning path with all improvements incorporated.
+
+Revised reasoning path:
+
+# C Baselines
+
+In this appendix, we provide detailed descriptions and performance results for the baseline models compared to MPRF across the SEM16, PStance, and VAST datasets.
+
+# C.1 Statistics-based Models
+
+We include BiLSTM(Schuster and Paliwal, 1997), which takes tweets as inputs without considering the target information; CNN(Kim, 2014), similar to BiLSTM, the vanilla CNN only takes tweets as inputs and does not consider the target information; TAN(Du et al., 2017) an attention-based LSTM model that extracts target specific features; BiCond (Augenstein et al., 2016), which uses bidirectional LSTM to encode both the text and the target; CrossNet (Xu et al., 2018), which enhances BiLSTM with self-attention mechanisms to improve focus on relevant text segments; GCAE(Xue and Li, 2018), a CNN model that utilizes a gating mechanism to block targetunrelated information; PGCNN(Huang and Carley, 2018), which is based
+
+on gated convolutional networks and encodes target information by generating target-sensitive filters.
+
+# C.2 BERT-based Models
+
+We benchmark against BERT (Devlin et al., 2019), a transformer model fine-tuned for stance detection, PT-HCL (Liang et al., 2022a), which incorporates contrastive learning within a BERT framework targeting zero-shot and cross-target tasks. BERT-joint-ft(Liu et al., 2021) and TGA-Net-ft(Liu et al., 2021), in which BERT has been fine-tuned during the training process. BERTweet(Nguyen et al., 2020), CKE-Net(Liu et al., 2021), GDA-CL(Li and Yuan, 2022), WS-BERT(He et al., 2022), BS-RGCN(Luo et al., 2022), BERT-joint(Allaway and McKeown, 2020), TGA Net (Allaway and McKeown, 2020), JointCL (Liang et al., 2022b), KPatch(Lin et al., 2024), TATA(Hanley and Durmeric, 2023), EDDA-LLaMA(Ding et al., 2024b); EZSD-CP(Yao et al., 2024), a stance detection model underpinned by a gated multilayer perceptron and a prompt learning strategy;CKI (Yan et al., 2024), a collaborative knowledge infusion approach, BERT-based models showing strong performance on datasets; CNet-Ad(Zhang et al., 2024c), a commonsense based adversarial learning framework that comprises a commonsense graph encoder and a feature separation adversarial network.
+
+# C.3 LLM-based Models
+
+We compare with KEprompt (Huang et al., 2023), which leverages knowledge-enhanced prompt tuning; KASD (Li et al., 2023), enhancing detection by integrating Wikipedia knowledge with retrieval-enhanced generation; COLA (Lan et al., 2024), which employs a collaborative role-infusion framework with multiple LLMs; Llama-2-7b-chat2, Llama-2-13b-chat and finetuned settings for the Llama models(Garg and Caragea, 2024); Manual-CoT (Wei et al., 2022) manually provides the explanations in demonstrations and enhances the chainof-thought reasoning ability of LLMs; StSQA(Zhang et al., 2024b) proposes automatic "thought-inducing" and add them to the demonstrations for step-by-step question answering; AutoCoT (Zhang et al., 2023c) automatically selects demonstrations from training data according to semantic diversity; ExpertPrompt(Xu et al., 2023) introduces the identity of experts and customizes information descriptions for LLMs before giving responses; SPP (Wang et al., 2024b) proposes
+
+solo performance prompting by engaging in multi-turn collaboration with multi-persona during reasoning; DEEM(Wang et al., 2024a), dynamic experienced expert modeling for stance detection; Stanceformer(Garg and Caragea, 2024), a target-aware transformer model; GPT-3.5 (Lan et al., 2024), GPT-3.5+COT (Lan et al., 2024), GPT-EDDA(Ding et al., 2024b), and LKI-BART(Zhang et al., 2024d). LC-CoT(Zhang et al., 2023a) employs the structured chain-of-thought approach for stance detection. LCDA enhances data quality by maintaining logical coherence (Zhang et al., 2025)
+
+# D Ablation Study Results
+
+In this appendix, we present the ablation study results for the VAST and PStance datasets. These tables illustrate the performance of the model with various components removed. The results demonstrate the importance of each component in stance detection.
+
+# D.1 Ablation Study Results on VAST
+
+<table><tr><td>Model</td><td colspan="3">Zero-Shot Few-Shot Overall</td></tr><tr><td>Mistral-7B-Instruct-v0.3</td><td>84.2</td><td>83.1</td><td>83.5</td></tr><tr><td>w/o Sentiment Path</td><td>82.7</td><td>81.4</td><td>81.9</td></tr><tr><td>w/o Fact-Based Path</td><td>83.2</td><td>82.1</td><td>82.5</td></tr><tr><td>w/o Expert Opinion Path</td><td>83.0</td><td>82.5</td><td>82.7</td></tr><tr><td>w/o Public Opinion Path</td><td>83.1</td><td>81.9</td><td>82.5</td></tr><tr><td>w/o Optimization</td><td>81.6</td><td>80.5</td><td>80.9</td></tr><tr><td>w/o Weighted Fusion</td><td>80.8</td><td>79.3</td><td>79.9</td></tr></table>
+
+Table 10: Ablation Study Results on VAST (%) using Mistral-7B-Instruct-v0.3. The best scores are in bold.
+
+The results on VAST using Mistral-7B-Instructv0.3 are shown in Table 10. The performance decreases when components like optimization and weighted fusion are omitted, highlighting their importance for improving the final stance prediction.
+
+# D.2 Ablation Study Results on PStance
+
+<table><tr><td>Method</td><td colspan="4">Trump Biden Sanders F_macro</td></tr><tr><td>MPRF</td><td>89.33</td><td>89.36</td><td>86.53</td><td>88.41</td></tr><tr><td>w/o Sentiment Path</td><td>87.4</td><td>85.9</td><td>84.0</td><td>85.4</td></tr><tr><td>w/o Fact-Based Path</td><td>87.7</td><td>86.3</td><td>84.5</td><td>85.8</td></tr><tr><td>w/o Expert Opinion Path</td><td>88.1</td><td>87.0</td><td>85.1</td><td>86.2</td></tr><tr><td>w/o Public Opinion Path</td><td>88.0</td><td>86.5</td><td>84.8</td><td>85.8</td></tr><tr><td>w/o Optimization</td><td>86.5</td><td>85.1</td><td>83.4</td><td>84.6</td></tr><tr><td>w/o Weighted Fusion</td><td>85.7</td><td>84.2</td><td>82.8</td><td>83.8</td></tr></table>
+
+Table 11: Ablation Study Results on PStance  $(\%)$  using Mistral-7B-Instruct-v0.3. The best scores are in bold.
+
+The results on PStance using Mistral-7B-Instruct-v0.3 are shown in Table 11. The table shows the performance drop when certain components, such as the sentiment path, fact-based path, and others, are removed.

2025/MPRF_ Interpretable Stance Detection through Multi-Path Reasoning Framework/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb49c38c4033841ffae0c26b99229405eef87bf356458856a5104e87902947e7
+size 621955

2025/MR. Judge_ Multimodal Reasoner as a Judge/b1637928-d23a-4ee6-8b4d-2e909f1b2b62_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74c95602d227b894672e1243445b85a6a9efaf8badbea70e546cdc63e2363b04
+size 19214929

2025/MR. Judge_ Multimodal Reasoner as a Judge/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cbfd7be434e36ac41f9fb003543533c036995c67fdf1f128dc8b96e4649ba70
+size 1107988

2025/MS-RAG_ Simple and Effective Multi-Semantic Retrieval-Augmented Generation/da873171-424c-48e9-9119-5424bd5a6138_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aad6df3bc5f5af2f0b8e5eb86d271e2e0ae4ea8385d5158989bdcf25d3945510
+size 988088

2025/MS-RAG_ Simple and Effective Multi-Semantic Retrieval-Augmented Generation/full.md

@@ -0,0 +1,474 @@
+# MS-RAG: Simple and Effective Multi-Semantic Retrieval-Augmented Generation
+
+Xiaozhou You
+
+Ant Group
+
+Hangzhou, China
+
+youxz@antgroup.com
+
+Yahui Luo
+
+Ant Group
+
+Hangzhou, China
+
+haoshan.lyh@antgroup.com
+
+Lihong Gu
+
+Ant Group
+
+Hangzhou, China
+
+lihong.glh@antgroup.com
+
+# Abstract
+
+To alleviate the hallucination problem of large language model (LLM), retrieval-augmented generation (RAG) has been proposed and widely adopted. Due to the limitations in cross-chunk summarization task of naive RAG, graph-based RAG has emerged as a promising solution. However, a close study reveals several flaws in these works. First, most graph-based RAGs suffer from less efficient indexing process, which leads to information loss and expensive costs. Second, they heavily rely on LLM for retrieval thus inference slowly, which hinders their application in industry. To build a more efficient and effective RAG, we propose the multi-semantic RAG (MS-RAG). In this work, we combine knowledge graphs with dense vector to build a multi-semantic RAG. To be specific, (i) at indexing stage, we create multiple semantic-level indexes, including chunk-level, relation-level, and entity-level, to leverage the merits of dense vector and knowledge graph. (ii) at retrieval stage, unlike the previous LLM-empowered entity extraction, we propose a novel mix recall algorithm. Finally, we employ a multi-semantic rerank module to purify the results. Extensive experiments show that MS-RAG achieves superior performance. In terms of retrieval effect, MS-RAG achieves state-of-the-art performance, which is about  $10\% -30\%$  improvement than the existing methods. In terms of question-answering effect, MS-RAG still achieves promising results with faster inference speed. More analysis and experiments are provided in Appendix.
+
+# 1 Introduction
+
+By integrating external knowledge, retrieval-augmented generation (RAG) effectively alleviates the hallucination problem of large language models (LLM) and has attracted widespread attention (Brown, 2020; Lewis et al., 2020; Fatehkia et al., 2024; Fang et al., 2024; Qian et al., 2024). Based on well-built knowledge index, RAG per
+
+![](images/277ca6a17579cf4310cf64b530e61b59d8d7ba5259a517180d2876d3c43bd7dd.jpg)  
+(a) Query: "Who is the mutual friend of Mike and Tom?"
+
+![](images/a1b5745a58c6136fd155fda4af7466f60e3549fec9af76acbfe00a50d7d85d1c.jpg)  
+(b) The percent of missing edge/node (left) and recall performance (right)  
+Figure 1: (a) On ideal knowledge graph (KG), the graph retrieval will find the answer "Bill" by locating the nodes ("Mike", "Tom"), and the corresponding edges, such as (i). However, if there are missing edges or nodes in KG, the retrieval will fail to find answer, such as (ii) and (iii). (b) We randomly selected 50 samples from the HotpotQA and 2WikiMultiHopQA and used different LLMs to construct KG. We found that  $10 - 30\%$  of the samples have missing node/edge. More importantly, when performing graph retrieval on missing node/edge KG (denoted graph w/ m.), the performance is even worse than Naive RAG.
+
+![](images/fa7114ff7ace4dfd80dbccb3c83c06868931f301c1ddfb915da67e08bb0a3a1a.jpg)
+
+forms retrieval tailored to user needs. To further enhance the capacity, graph-based RAG was proposed and regarded as a promising solution. GraphRAG (Edge et al., 2024) utilizes LLM to extract entity and relation, and employs Leiden algorithm to build hierarchical graph communities. HippoRAG (Gutiérrez et al., 2024) is inspired by hippocampal indexing theory and uses Personalized PageRank algorithm (Haveliwala, 2002) for graph retrieval. Then, the following works have further improved graph-based RAG and promoted its application (Guo et al., 2024; Li et al., 2024; Yuan et al., 2024; Peng et al., 2024; Zhang et al., 2024).
+
+However, by careful study, there are still several limitations. First, most graph-based RAG suffers
+
+<table><tr><td></td><td>Vector Search</td><td>Cross-Chunk</td><td>Multi-Semitic</td><td>Fast Inference</td><td>ReRank</td></tr><tr><td>Naive RAG</td><td>✓</td><td>×</td><td>×</td><td>✓</td><td>×</td></tr><tr><td>GraphRAG</td><td>×</td><td>✓</td><td>×</td><td>×</td><td>×</td></tr><tr><td>HippoRAG</td><td>✓</td><td>✓</td><td>×</td><td>✓</td><td>×</td></tr><tr><td>MS-RAG (ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+Table 1: The qualitative comparison between different RAGs. Compared with existing methods, our proposed MS-RAG boasts several desired properties that contribute to building a more strong RAG. The "multi-semantic" refers to information that includes multiple semantic levels, such as chunk-level, entity-level, and relation-level.
+
+from less efficient indexing process, which leads to information loss and expensive costs. Most of them inherently need LLM to build graph index, which is prone to mistakes. As shown in Figure 1, almost  $10 - 30\%$  of samples have missing node/edge. The performance of RAG with missing node/edge index is extremely poor, even worse than naive RAG. Besides, GraphRAG (Edge et al., 2024) needs to call LLM multiple times to build hierarchical communities, which brings expensive token costs and additional latency. Second, they heavily rely on LLM for retrieval thus inference slowly, which hinders their application in industry. Given a query, existing methods tend to use LLM to extract entities for following retrieval. The process not only has weak accuracy, but also makes inferences slowly.
+
+To trickle it, we proposed a simple and effective multi-semantic RAG (MS-RAG). In this work, we combine knowledge graphs with dense embedding to build MS-RAG. First, at indexing stage, we established multiple semantic-level indexes, including chunk-level, relation-level, and entity-level, aimed to integrate the advantages of vector and graph. Some existing works rely on LLM for graph index construction, which easily leads to missing node/edge. We introduce chunk-level information in multi-semantic index, which not only helps to alleviate the negative of LLM mistakes, but also enhances semantic representation. In addition, we perform entity disambiguation to merge entities with same semantic. We also perform entity/rela. tion summary to provide richer prompt to LLM.
+
+Then, at retrieval stage, unlike the previous LLM-empowered entity extraction, we propose a novel mix recall algorithm to complementarily combine the advantages of graph recall and dense recall, then employ a novel multi-semantic rerank module to integrate different outputs. Given a query, we use vector search instead of LLM to extract entity/relation from our proposed multisemantic index. Based on the recalled elements, we
+
+simultaneously perform graph search to recall other nearby entities and relations. Then, we propose an innovative multi-semantic rerank module to purify the final passage. The module first filters candidate passage by weighted voting, and then employ a lightweight LLM (Bai et al., 2023) to rerank.
+
+Extensive experiments show MS-RAG achieves outstanding performance. We conduct experiments on public benchmarks like hotpotQA (Yang et al., 2018), 2WikiMultiHopQA (Ho et al., 2020), and MuSiQue (Trivedi et al., 2022b). In terms of retrieval effect, compared with existing RAG methods, both in single-hop and multi-hop settings (Trivedi et al., 2022a), MS-RAG achieves state-of-the-art performance remarkably, improving retrieval accuracy by about  $10\% -30\%$ . Besides, in terms of question-answering effect, compared with GraphRAG (Edge et al., 2024) and HippoRAG (Gutiérrez et al., 2024), MS-RAG achieves promising performance with faster inference speed.
+
+# 2 Related Works
+
+Large Language Model. In 2017, Transformer was proposed (Vaswani, 2017), which had a profound impact on subsequent work. Then, the BERTlike (Devlin et al., 2018) and GPT-like (Brown, 2020) models have demonstrated their capabilities on broader downstream tasks. In particular, ChatGPT (OpenAI, 2024) marked a significant milestone by accelerating the development of intelligent dialogue systems. The following works have further bolstered confidence in the pursuit of artificial general intelligence (AGI), such as Qwen (Bai et al., 2023) and Llama (AI@Meta, 2024). However, LLMs still face several challenges, including untimely data update and hallucination problem.
+
+Retrieval-Augmented Generation. To alleviate the hallucination problem, RAG was proposed and widely used in LLM applications (Lewis et al., 2020). RAG mainly consists of indexing stage, retrieval stage and a generation stage. Recently,
+
+![](images/c9fca90e0d4ee37957feecd53b44689c3747f4382cbce72b9c47b197f5b8dd68.jpg)  
+Figure 2: The pipeline of our proposed multi-semantic index. For the input documents, we establish a multi-semantic index, which contains three levels of information: chunk-level, entity-level and relation-level, to construct a more effective RAG index.
+
+Many subsequent works have optimized the retrieval performance and generation effect of RAG. RAFT (Zhang et al., 2024) improves the robustness of LLM by training the model to ignore irrelevant documents, and introduces chain of thought (CoT) to improve the reasoning ability. And RankRAG (Yu et al., 2024) fine-tunes LLM to enable LLM to have both text generation and rerank capabilities.
+
+Knowledge Graph for RAG. knowledge graph (Chen et al., 2020) are adopted in RAG due to their powerful ability of complex relationships understanding and cross-chunks summarization. GraphRAG (Edge et al., 2024) first applied graph to RAG, and effectively solved the query-summariation task by building a hierarchical graph community, and improved the generation quality. HippoRAG (Gutiérrez et al., 2024) combines knowledge graphs with Personalized PageRank algorithms to improve retrieval efficiency. However, these works face the problem of relying on LLM and low efficiency. And HybridRAG (Sarmah et al., 2024) splices the results of VectorRAG and GraphRAG outputs together, in which the output results of VectorRAG and GraphRAG are completely independent. Different from previous works, our proposed MS-RAG effectively improves retrieval performance and generation quality by introducing multi-semantic index.
+
+# 3 Methods
+
+# 3.1 Overview
+
+To build a more efficient and effective retrieval-augmented generation system, we proposed MS-
+
+RAG. In this work, we integrate the merits of knowledge graphs and vector search.
+
+As shown in Figure 2, it's the indexing pipeline. First, we chunk the input documents, and the chunked results are stored in chunk database. Then, we employ LLM to extract knowledge graph. The extracted entity stored in entity database, and the extracted relation stored in relation database. All chunks, entities, and relations are encoded by dense vector model. Besides, we also apply entity/rela- tion summarization and disambiguation to enrich the prompt provided to LLM.
+
+As shown in Figure 3, it's the retrieval pipeline. Based on the well-built multi-semantic index, we propose a novel mix recall algorithm to search related entities, relations, and chunks. Unlike previous methods (Gutiérrez et al., 2024; Guo et al., 2024), we first perform vector search to recall most relevant subject. Then, we recall some adjacent entities/relations and other similar chunks. Finally, the multi-semantic rerank module performs voting and rerank to purify final result.
+
+# 3.2 Offline Indexing
+
+Most existing works suffer from less efficient indexing process, which heavily rely on LLM to build graph index from input documents. To alleviate the negative impact of LLM mistakes, we propose multi-semantic index. To enrich prompt for LLM, we introduce summarization and disambiguation.
+
+# 3.2.1 Multi-Semantic Index
+
+The proposed multi-semantic index consists of three databases: entity, relation, and chunk. All
+
+![](images/28960aad9de2e4ddd8d06f7e210918431f1b070575887d5bfb0e60cb25dd41c0.jpg)  
+Figure 3: The retrieval pipeline of our proposed MS-RAG. Based on the well-built multi-semantic index, we propose novel mix recall algorithm to complementarily combine the advantages of graph recall and dense recall. Then, we employ a novel multi-semantic rerank module to purify final outputs.
+
+information of each database is encoded by a dense vector model.
+
+Entity db. It mainly consists of some nouns and characters. For example, in the sentence "Tom loves basketball", "Tom" and "basketball" are entities encoded and stored in the database. In addition, the relation information of the entity and adjacent entities are also stored.
+
+Relation db. It mainly consists of declarative sentences. For example, in the sentence "Tom loves basketball", entities ("Tom", "basketball") and relations ("play") are semantically concatenated together and then encoded by dense vector model.
+
+Chunk db. It mainly consists of chunks from passage. The purpose of introducing naive chunk information in multi-semantic index is to alleviate the negative impact caused by LLM mistakes.
+
+# 3.2.2 Summarization and Disambiguation
+
+Entity/Relation Summarization: While building the graph index, we also designed prompts for LLM to generate summary for the extracted entity/rela-tion, mainly based on involved context. We hope that the generated summary can serve as a helpful reference for the final outputs generated by LLM.
+
+Entity Disambiguation: For some identical entities with different names, we decided to perform disambiguation, which we designed prompts to merge such case in indexing pipeline. For example, in a context, "Clarence Fred Gehrke" and "Fred Gehrke" refer to the same person.
+
+# 3.3 Online Retrieval
+
+To achieve more accurate and faster retrieval, we proposed two innovative components: mix recall
+
+algorithm and multi-semantic rerank module. The mix recall algorithm aims to better utilize the well-built multi-semantic index, which includes knowledge graph structure and chunk information. The multi-semantic rerank module enhances reasoning capabilities with slight inference delay. The process is illustrated in Figure 3.
+
+# 3.3.1 Mix Recall Algorithm
+
+The mix recall algorithm aims to complementarily combine the advantages of graph recall and dense recall, inspired by two aspects. First, knowledge graph provides an information structure with semantic associations. While dense recall struggles with complex and lengthy texts, it excels at recalling short entities and relations. Then, as shown in Figure 1, we introduced chunk recall to alleviate the negative impact of LLM errors. We also provide more comparison in Appendix C.
+
+Graph Element Recall. It includes two stages: (i) entity/relation extraction: Previous works tend to use LLM for entity extraction (Gutiérrez et al., 2024; Edge et al., 2024). We argued that it's inefficient and time-consuming. In our work, we employ vector search instead of LLM to recall relevant subjects. We extracted  $\lambda_{1}$  entities and  $\lambda_{2}$  relations. (ii) semantic neighboring search: In knowledge graph, neighboring nodes and edges have associated semantics (Guo et al., 2024). For extracted relation, we recall its neighboring  $\lambda_{h}$ -hop relations. For extracted entity, we recall the neighboring entities and relations simultaneously.
+
+Chunk Recall. Similar to naive RAG, we directly use vector tools to calculate similarity and take the top  $\lambda_3$  chunks as the recall chunks. We introduced
+
+chunk recall to enrich our semantics and mitigate the impact of LLM mistakes.
+
+# 3.3.2 Multi-semantic rerank module
+
+The purpose of rerank is to evaluate the relation between the recall results and input query at semantic and logical level, and output more relevant ranking. To effectively integrate the results of multirecall sources, we proposed multi-semantic reranking module, which mainly consists of two stages: voting and rerank.
+
+Voting. We employ voting algorithm to integrate the results from multiple recall sources. Notably, the voting targets are the passage labels associated with entity, relation, or chunk. We argued that the passages appearing frequently in recall sources are more likely to be relevant. Based on this, we select the top  $\lambda_4$  passages to advance to the next stage.
+
+Rerank. Inspired by RankGPT (Sun et al., 2023), we design prompts and employ a lightweight LLM for rerank, with the purpose of using the reasoning ability of LLM to find relevant passage that can answer input query. In addition, in order to speed up inference speed, we prompt the large model to only output the sorted sequence number.
+
+# 4 Experimental Setup
+
+In this section, we introduce datasets, baselines, evaluation details, and implementation details about our experiments of our proposed MS-RAG.
+
+# 4.1 Datasets
+
+We conduct experiments primarily on three challenging multi-hop benchmarks, HotpotQA (Yang et al., 2018), MuSiQue (answerable) (Trivedi et al., 2022b) and 2WikiMultiHopQA (Ho et al., 2020). Following previous works (Press et al., 2023; Trivedi et al., 2022a; Gutierrez et al., 2024), we se-lete 1,000 questions from each validation set. The description of datasets is shown below:
+
+i) HotpotQA. It contains 113k Wikipedia QA pairs, where questions require searching and reasoning about multiple documents to answer. We selected about 9k passages for experiments.  
+ii) MuSiQue. It is a new multi-hop QA dataset containing about 25K 2-4 hop questions, using seed questions from 5 existing single-hop datasets. We selected about 12k passages for our experiments.  
+iii) 2WikiMultiHopQA. It introduces evidence information containing reasoning paths for multi-hop questions, which contains structured and un
+
+structured data. We selected about 6k passages for our experiments.
+
+# 4.2 Baselines
+
+In order to verify the effectiveness of MS-RAG, we compare against several strong and widely used retrieval methods. For retrieval performance, we compare with BGE-M3 (Chen et al., 2024), BM25 (Robertson and Walker, 1994), Contriever (Izacard et al., 2021), GTR (Ni et al., 2022) and ColBERTv2 (Santhanam et al., 2022). Besides, we compare with three recent LLM/RAG augmented baselines: HippoRAG (Gutiérrez et al., 2024), which introduces Personalized PageRank algorithm to enhance graph retrieval, and Propositionizer (Chen et al., 2023), which proposed a novel retrieval unit, and RAPTOR (Sarthi et al., 2024), which uses trees to integrate information from long documents. Besides, to verify the multi-step retrieval performance, we also compare with the multi-step retrieval method IRCoT (Trivedi et al., 2023). For question-answering performance, we compare with GraphRAG (Edge et al., 2024) and HippoRAG (Gutiérrez et al., 2024).
+
+# 4.3 Evaluation Details
+
+For the evaluation of retrieval performance, we use top2 recall (R@2) and top5 recall (R@5), which refers to the proportion of correct answers in the top2 search results or in the top5 search results.
+
+For the evaluation of question-answering performance, inspired by (Guo et al., 2024; Edge et al., 2024), we prompt LLM (Bai et al., 2023) to evaluate based on the following criteria:
+
+i) Correctness: Does the answer correctly answer the question?  
+ii) Diversity: Does the answer provide a variety of perspectives on the question?  
+iii) Comprehension: Does the answer effectively help readers understand?  
+iv) Overall: Determine the best answer by combining the above three criteria.
+
+# 4.4 Implementation Details
+
+The corresponding hyper-parameters  $\lambda_{1},\lambda_{2},\lambda_{3},$ $\lambda_4,\lambda_h$  are set to 5, 10, 10, 5 and 4, respectively. We use GPT-3.5-turbo-1106 (OpenAI, 2024) with temperature of 0 as our LLM, and employ BGE-M3 (Chen et al., 2024) and Contriever (Izacard et al., 2021) as our retriever. The LLM used in proposed multi-semantic rerank module is Qwen-7B, and another LLM used to evaluate question-answer
+
+<table><tr><td></td><td colspan="2">MuSiQue</td><td colspan="2">2Wiki</td><td colspan="2">HotpotQA</td><td colspan="2">Average</td></tr><tr><td></td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td></tr><tr><td>BM25</td><td>32.3</td><td>41.2</td><td>51.8</td><td>61.9</td><td>55.4</td><td>72.2</td><td>46.5</td><td>58.4</td></tr><tr><td>Contriever</td><td>34.8</td><td>46.6</td><td>46.6</td><td>57.5</td><td>57.2</td><td>75.5</td><td>46.2</td><td>59.9</td></tr><tr><td>GTR</td><td>37.4</td><td>49.1</td><td>60.2</td><td>67.9</td><td>59.4</td><td>73.3</td><td>52.3</td><td>63.4</td></tr><tr><td>ColBERTv2</td><td>37.9</td><td>49.2</td><td>59.2</td><td>68.2</td><td>64.7</td><td>79.3</td><td>53.9</td><td>65.6</td></tr><tr><td>BGE-M3</td><td>38.2</td><td>51.7</td><td>62.1</td><td>70.2</td><td>69.5</td><td>82.4</td><td>56.6</td><td>68.1</td></tr><tr><td>RAPTOR</td><td>35.7</td><td>45.3</td><td>46.3</td><td>53.8</td><td>58.1</td><td>71.2</td><td>46.7</td><td>56.8</td></tr><tr><td>Proposition</td><td>37.6</td><td>49.3</td><td>56.4</td><td>63.1</td><td>58.7</td><td>71.1</td><td>50.9</td><td>61.2</td></tr><tr><td>HippoRAG</td><td>41.0</td><td>52.1</td><td>71.5</td><td>89.5</td><td>59.0</td><td>76.2</td><td>57.2</td><td>72.6</td></tr><tr><td>MS-RAG (Contriever)</td><td>43.2</td><td>54.3</td><td>77.8</td><td>90.9</td><td>77.6</td><td>86.2</td><td>66.2</td><td>77.1</td></tr><tr><td>MS-RAG (BGE-M3)</td><td>47.3</td><td>56.4</td><td>80.1</td><td>91.2</td><td>79.4</td><td>88.5</td><td>68.9</td><td>78.7</td></tr></table>
+
+Table 2: The single-step retrieval performance. Our proposed MS-RAG achieves excellent performance, which outperforms all baselines on HotpotQA, MuSiQue and 2WikiMultiHopQA datasets.  
+
+<table><tr><td></td><td colspan="2">MuSiQue</td><td colspan="2">2Wiki</td><td colspan="2">HotpotQA</td><td colspan="2">Average</td></tr><tr><td></td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td></tr><tr><td>IRCoT + BM25 (Default)</td><td>34.2</td><td>44.7</td><td>61.2</td><td>75.6</td><td>65.6</td><td>79.0</td><td>53.7</td><td>66.4</td></tr><tr><td>IRCoT + Contriever</td><td>39.1</td><td>52.2</td><td>51.6</td><td>63.8</td><td>65.9</td><td>81.6</td><td>52.2</td><td>65.9</td></tr><tr><td>IRCoT + ColBERTv2</td><td>41.7</td><td>53.7</td><td>64.1</td><td>74.4</td><td>67.9</td><td>82.0</td><td>57.9</td><td>70.0</td></tr><tr><td>IRCoT + HippoRAG</td><td>43.9</td><td>56.6</td><td>75.3</td><td>93.4</td><td>65.8</td><td>82.3</td><td>66.2</td><td>77.1</td></tr><tr><td>IRCoT + MS-RAG (Contriever)</td><td>45.7</td><td>57.6</td><td>80.4</td><td>93.8</td><td>78.4</td><td>88.7</td><td>68.2</td><td>79.3</td></tr><tr><td>IRCoT + MS-RAG (BGE-M3)</td><td>47.8</td><td>59.2</td><td>85.0</td><td>95.2</td><td>82.7</td><td>91.2</td><td>71.8</td><td>81.9</td></tr></table>
+
+Table 3: The multi-step retrieval performance. Compared with other baselines, our proposed MS-RAG achieves state-of-the-art performance with significant improvements on multi-step setting.
+
+performance is Qwen-32B (Bai et al., 2023). In Figure 1, the mentioned GraphRAG is (Edge et al., 2024), and the mentioned naive RAG is (Izacard et al., 2021), respectively.
+
+# 5 Results
+
+In this section, we will present the experimental results of our proposed MS-RAG, which includes single-step retrieval performance, multi-step retrieval performance, question-answering performance and other ablation experiments.
+
+# 5.1 The single-step retrieval performance.
+
+As shown in Table 2, MS-RAG outperforms all other methods with significant improvement, including general retrieval models such as ColBERTv2, GTR, BGE-M3, BM25 and Contriever, recent LLM-empowered baselines such as Propositionizer and RAPTOR, the current SOTA HippoRAG. Compared with previous best results, our
+
+MS-RAG obtains almost  $20.5\%$  R@2 and  $8.4\%$  R@5 improvement on average, respectively.
+
+Compared with HippoRAG (Gutiérrez et al., 2024), the current SOTA graph-based RAG, our MS-RAG (Contriever) significantly improves performance. First, we introduce multi-semantic index, which organically combines the advantages of vector search and graph structure. We take advantage of the semantic structure of the graph and the accurate and fast characteristic of vector search to effectively improve the algorithm. At the indexing stage, we introduce chunk information to alleviate the problem of LLM errors. At the retrieval stage, we first use vector retrieval instead of LLM to perform entity extraction. In addition, we use nearby search instead of personal pagerank algorithm (Gutiérrez et al., 2024) to improve the efficiency and accuracy of graph retrieval. Finally, the proposed multi-semantic rerank module effectively integrates multi-semantic information. Im
+
+<table><tr><td></td><td>Correctness</td><td>Diversity</td><td>Comprehension</td><td>Overall</td><td>Inference time</td></tr><tr><td>GraphRAG</td><td>27.7%</td><td>38.8%</td><td>45.7%</td><td>38.2%</td><td>4.12s</td></tr><tr><td>MS-RAG (ours)</td><td>72.3%</td><td>61.2%</td><td>54.3%</td><td>61.8%</td><td>0.76s</td></tr><tr><td>HippoRAG</td><td>38.8%</td><td>41.6%</td><td>44.3%</td><td>40.6%</td><td>0.88s</td></tr><tr><td>MS-RAG (ours)</td><td>61.2%</td><td>58.4%</td><td>55.7%</td><td>58.4%</td><td>0.76s</td></tr></table>
+
+Table 4: The QA performance on mixed datasets. The paired comparison of question-answering effect between GraphRAG, HippoRAG and our MS-RAG, separately. Our proposed MS-RAG achieves better question-answering rating and obtains faster inference speed.  
+
+<table><tr><td></td><td>R@2</td><td>R@5</td></tr><tr><td>MS-RAG</td><td>79.4</td><td>88.5</td></tr><tr><td>MS-RAG w/o multi-semantic</td><td>73.3</td><td>79.2</td></tr><tr><td>MS-RAG w/o rerank</td><td>76.1</td><td>84.9</td></tr></table>
+
+Table 5: The ablation experiments of MS-RAG on HotpotQA datasets. The ablation experiments of different modules in proposed MS-RAG.
+
+Importantly, we demonstrate the importance of each of our proposed modules in Section 5.4 and the Appendix.
+
+# 5.2 The multi-step retrieval performance.
+
+As shown in Table 3, equipped with IRCoT, MS-RAG achieves more advanced performance than previous works on multi-step retrieval, including general retrieval models such as BM25, ColBERTv2, Contriever, and the current graph-based SOTA HippoRAG. Compared with previous best results, our MS-RAG obtains almost  $8.5\%$  R@2 and  $8.4\%$  R@5 improvement on average.
+
+Compared with HippoRAG (Gutiérrez et al., 2024), the current state-of-the-art RAG, our MS-RAG improves  $\mathbf{R}@\mathbf{2}$  metric from 66.2 to 71.8 (+5.6), improves  $\mathbf{R}@\mathbf{5}$  metric from 77.1 to 81.9 (+4.8) on average. It directly reflects the innovation and effectiveness of our multi-semantic index in multi-step settings. In addition, the introduction of the multi-semantic rerank module can more effectively integrate different information.
+
+Compared with IRCot + BM25 (Trivedi et al., 2023), the typical multi-step retrieval method, our MS-RAG demonstrated significant performance gains, which outperforms IRCoT + BM25 in all metrics and selected three datasets. On average, our MS-RAG improves R@2 metric from 53.7 to 71.8 (+18.1), improves R@5 metric from 66.4 to 77.1 (+10.7), which demonstrates the performance advantages of MS-RAG and the effectiveness of
+
+proposed novel modules.
+
+# 5.3 The question-answering performance.
+
+As shown in Table 4, it's the question-answering performance comparison between GraphRAG (Edge et al., 2024), HippoRAG (Gutiérrez et al., 2024) and our proposed MS-RAG. We select 100 questions from HotpotQA, 2wikiMultiHopQA and MuSiQue datasets. And we employ Qwen-32B (Bai et al., 2023) to judge the quality.
+
+Our MS-RAG surpasses GraphRAG in all metrics, with  $61.8\%$  overall rating,  $72.3\%$  correctness,  $61.2\%$  diversity and  $54.3\%$  comprehension. For correctness, our MS-RAG surpasses GraphRAG from  $27.7\%$  to  $72.3\%$ . It is due to the reason that our proposed MS-RAG achieves more accurate retrieval and gives correct supporting facts. For diversity and comprehension, our MS-RAG surpasses GraphRAG from  $38.8\%$  to  $61.2\%$ , from  $45.7\%$  to  $54.3\%$ , respectively. We argued it's the result of our entity summarization and relation summarization. Besides, our MS-RAG achieves faster inference speed (0.76s vs 4.12s), more than 5 times improvement. It's because we reduced the number of LLM accesses and token consumption during our retrieval stage.
+
+Compared with the current SOTA HippoRAG, MS-RAG surpasses it with  $58.4\%$  overall rating,  $61.2\%$  correctness,  $58.4\%$  diversity and  $55.7\%$  comprehension. For correctness, our MS-RAG imporves HippoRAG from  $38.8\%$  to  $61.2\%$ . It reflects the advanced retrieval performance of MS-RAG. For diversity and comprehension, our MS-RAG imporves HippoRAG from  $41.6\%$  to  $58.4\%$ , from  $44.3\%$  to  $55.7\%$ , respectively. It also proves the innovation and effectiveness of MS-RAG.
+
+# 5.4 Ablations experiments
+
+In this subsection, we will present the ablations Experiments of our MS-RAG, to evaluate the effect of each component, including ablation for modules in
+
+<table><tr><td></td><td colspan="2">MuSiQue</td><td colspan="2">2Wiki</td><td colspan="2">HotpotQA</td><td colspan="2">Average</td></tr><tr><td></td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td><td>R@2</td><td>R@5</td></tr><tr><td>Llama-3-8B-instruct</td><td>46.2</td><td>57.3</td><td>79.2</td><td>89.9</td><td>79.1</td><td>87.5</td><td>68.2</td><td>78.3</td></tr><tr><td>Llama-3-70B-instruct</td><td>46.8</td><td>57.2</td><td>79.5</td><td>90.6</td><td>79.3</td><td>88.1</td><td>68.5</td><td>78.6</td></tr><tr><td>REBEL</td><td>36.1</td><td>48.3</td><td>74.3</td><td>79.2</td><td>67.3</td><td>68.0</td><td>59.2</td><td>65.2</td></tr><tr><td>GPT-3.5-turbo</td><td>47.3</td><td>56.4</td><td>80.1</td><td>91.2</td><td>79.4</td><td>88.5</td><td>68.9</td><td>78.7</td></tr></table>
+
+Table 6: The ablation experiments of different LLMs in indexing stage on HotpotQA datasets. By default, we choose GPT-3.5-turbo as the LLM to build our multi-semantic index.  
+
+<table><tr><td></td><td>R@2</td><td>R@5</td></tr><tr><td>Mix recall algorithm (MR)</td><td>79.4</td><td>88.5</td></tr><tr><td>MR w/o chunk recall</td><td>77.1</td><td>86.2</td></tr><tr><td>MR w/o graph recall</td><td>70.6</td><td>79.1</td></tr><tr><td>MR w/ LLM extraction</td><td>73.8</td><td>82.2</td></tr><tr><td>MR w/ PPR</td><td>73.8</td><td>84.2</td></tr></table>
+
+Table 7: The ablation experiments of mix recall algorithm on HotpotQA datasets. The ablation experiments of different methods in mix recall algorithm.
+
+MS-RAG, ablation for different LLMs for indexing, and ablation for multi-semantic index.
+
+# 5.4.1 Ablation for modules in MS-RAG.
+
+As shown in Table 5, it's the ablation experiments of modules in MS-RAG on HotpotQA datasets. When we remove the multi-semantic index module and use only BGE-M3 (Chen et al., 2024) recall chunks, the R@2 metric decreases from 79.4 to 73.3 (-6.1), and the R@5 metric decreases from 88.5 to 79.2 (-9.3). It shows the effectiveness of integrating knowledge graphs and vectors.
+
+And we remove the multi-semantic rerank module and take the most relevant output at different recalls (entity and relation first), the R@2 metric decreases from 79.4 to 76.1 (-3.3), and the R@5 metric decreases from 88.5 to 84.9 (-3.6). It shows that the LLM-based voting and rerank algorithm we designed brings significant improvements.
+
+# 5.4.2 Ablation for LLMs for indexing.
+
+As shown in Table 6, it's the ablation experiments of different LLMs in our indexing stage on HotpotQA datasets. We try four different LLMs to build knowledge graph at indexing stage, including GPT-3.5-turbo (OpenAI, 2024), Llama-3-8B-instruct (AI@Meta, 2024), Llama-3-70B-instruct (AI@Meta, 2024), and REBEL (Huguet Cabot and
+
+Navigli, 2021). By default, we choose GPT-3.5-turbo as the LLM to build knowledge graph for our multi-semantic index, which achieves the best results overall. The results of Llama-3-70B-instruct is close to GPT-3.5-turbo. And REBEL is the worst performance. In the future, we will explore more LLMs to build a more complete knowledge graph.
+
+# 5.4.3 Ablation for mix recall algorithm.
+
+As shown in Table 7, it's the ablation experiments of mix recall algorithm on HotpotQA datasets. We try to remove chunk recall and graph recall separately to test the impact on performance. MR w/ LLM extraction means employs LLM to extract entity/relation instead of vector search. MR w/ PPR means employs PPR (Gutiérrez et al., 2024) to recall graph elements. When we remove any of the components, performance degrades, directly demonstrating the effectiveness of each component. Removing chunk recall, the R@2 accuracy drops by 2.3 and R@5 accuracy drops by 2.3, which indicates chunk recall plays a significant role in mitigating the negative effects of mistakes made by the LLM. Removing graph recall also get worse results. The attempts of using PPR and LLM extraction also demonstrated the innovativeness of the mix recall algorithm.
+
+# 6 Conclusion
+
+In this work, we propose multi-semantic RAG (MS-RAG) to build a more efficient and effective RAG. To be specific, (i) at indexing stage, we create multisemantic index to leverage merits of dense vector and knowledge graph. (ii) at retrieval stage, we propose mix recall algorithm to capture texts at different semantic levels. Then we employ a multisemantic rerank module to purify results. Extensive experiments show MS-RAG achieves superior performance. In the future, we will explore more effective retrieval methods to improve performance.
+
+# Limitations
+
+Although MS-RAG achieves outstanding performance by introducing multi-semantic indexing and multi-semantic rerank modules, there are still some limitations. First, we need to explore ways to build knowledge graphs more effectively. Although our multi-semantic index effectively alleviates this problem, it is not enough. Second, we need to explore more advanced retrieval algorithms with MS-RAG to maximize the performance.
+
+# References
+
+AI@Meta. 2024. Llama 3 model card.  
+Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.  
+Tom B Brown. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165.  
+Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216.  
+Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, and Dong Yu. 2023. Dense x retrieval: What retrieval granularity should we use? arXiv preprint arXiv:2312.06648.  
+Xiaojun Chen, Shengbin Jia, and Yang Xiang. 2020. A review: Knowledge reasoning over knowledge graph. Expert systems with applications, 141:112948.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130.  
+Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. 2024. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. arXiv preprint arXiv:2405.20978.  
+Masoomali Fatehkia, Ji Kim Lucas, and Sanjay Chawla. 2024. T-rag: Lessons from the llm trenches. arXiv preprint arXiv:2402.07483.
+
+Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. Lighthrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779.  
+Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. Hipporag: Neurobiologically inspired long-term memory for large language models. arXiv preprint arXiv:2405.14831.  
+Taher H Haveliwala. 2002. Topic-sensitive pagerank. In Proceedings of the 11th international conference on World Wide Web, pages 517-526.  
+Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060.  
+Pere-Lluis Huguet Cabot and Roberto Navigli. 2021. REBEL: Relation extraction by end-to-end language generation. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2370-2381, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118.  
+Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. triviaqa: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. arXiv preprint arXiv:1705.03551.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association of Computational Linguistics.  
+Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474.  
+Zhuoqun Li, Xuanang Chen, Haiyang Yu, Hongyu Lin, Yaojie Lu, Qiaoyu Tang, Fei Huang, Xianpei Han, Le Sun, and Yongbin Li. 2024. Structrag: Boosting knowledge intensive reasoning of llms via inference-time hybrid information structurization. arXiv preprint arXiv:2410.08815.  
+Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Hannaneh Hajishirzi, and Daniel Khashabi. 2022. When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories. arXiv preprint.
+
+Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith Hall, Ming-Wei Chang, and Yinfei Yang. 2022. Large dual encoders are generalizable retrievers. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9844–9855, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+OpenAI. 2024. GPT-3.5 Turbo.  
+Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921.  
+Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. arXiv preprint arXiv:2210.03350.  
+Hongjin Qian, Zheng Liu, Kelong Mao, Yujia Zhou, and Zhicheng Dou. 2024. Grounding language model with chunking-free in-context retrieval. arXiv preprint arXiv:2402.09760.  
+Stephen E. Robertson and Steve Walker. 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. Dublin, Ireland, 3-6 July 1994 (Special Issue of the SIGIR Forum), pages 232-241. ACM/Springer.  
+Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. 2022. ColBERTv2: Effective and efficient retrieval via lightweight late interaction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3715-3734, Seattle, United States. Association for Computational Linguistics.  
+Bhaskarjit Sarmah, Dhagash Mehta, Benika Hall, Rohan Rao, Sunil Patel, and Stefano Pasquali. 2024. Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. In Proceedings of the 5th ACM International Conference on AI in Finance, pages 608-616.  
+Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D. Manning. 2024. RAPTOR: recursive abstractive processing for tree-organized retrieval. arXiv preprint arXiv:2401.18059.  
+Weiwei Sun, Lingyong Yan, Xinyu Ma, Pengjie Ren, Dawei Yin, and Zhaochun Ren. 2023. Is chatgpt good at search? investigating large language models as re-ranking agent. ArXiv, abs/2304.09542.
+
+Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022a. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. arXiv preprint arXiv:2212.10509.  
+Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022b. musique: Multi-hop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539-554.  
+Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10014-10037, Toronto, Canada. Association for Computational Linguistics.  
+A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600.  
+Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. 2024. Rankrag: Unifying context ranking with retrieval-augmented generation in llms. arXiv preprint arXiv:2407.02485.  
+Ye Yuan, Chengwu Liu, Jingyang Yuan, Gongbo Sun, Siqi Li, and Ming Zhang. 2024. A hybrid rag system with comprehensive enhancement on complex reasoning. arXiv preprint arXiv:2408.05141.  
+Tianjun Zhang, Shishir G. Patil, Naman Jain, Sheng Shen, Matei Zaharia, Ion Stoica, and Joseph E. Gonzalez. 2024. Raft: Adapting language model to domain specific rag. arXiv preprint arXiv:2403.10131.
+
+# A Contribution
+
+In summary, our contributions are summarized as follows:
+
+- Lightweight Multi-Semantic Index with Lower LLM Dependency. We built a multi-semantic index, including chunk-level, relation-level, and entity-level, with the goal of integrating the advantages of dense embedding and graph index. By introducing the chunk-level information, MS-RAG enrich our semantics and mitigate the negative impact of LLM mistakes effectively.  
+- Novel Mix Recall Algorithm. We propose a novel mix recall algorithm, which complementarily combines the advantages of graph recall and dense recall. It employs dense recall for more accurate short entity/relation extraction and graph structures for nearby recall logically.  
+- Fast and Effective RAG Retrieval Pipeline. By integrating vector search and knowledge graph, we achieve more accurate and faster RAG retrieval. The pipeline includes two innovative components: mix recall algorithm and multi-semantic rerank module. The multisemantic rerank module enhances reasoning capabilities with slight inference delay.  
+- The Outstanding Performance. Extensive experiments show that MS-RAG achieves excellent performance. Compared with existing works, MS-RAG shows strong performance both in effect of retrieval and question-answering. In terms of retrieval effect, compared with existing RAG methods, both in single-hop and multi-hop settings (Trivedi et al., 2022a), MS-RAG achieves state-of-the-art performance remarkably, improving retrieval accuracy by about  $10\% -30\%$ . Besides, in terms of question-answering effect, MS-RAG achieves promising performance with faster inference speed.
+
+# B Datasets details
+
+Following IRCoT (Trivedi et al., 2023) and HippoRAG (Gutiérrez et al., 2024), we collect all candidate passages (including supporting and distractor passages) from our selected questions and form a retrieval corpus for each dataset. The specific information is show in Table 8.
+
+<table><tr><td></td><td>MuSiQue</td><td>2Wiki</td><td>HotpotQA</td></tr><tr><td>Passage</td><td>12k</td><td>6k</td><td>9k</td></tr><tr><td>Triplet</td><td>107k</td><td>51k</td><td>99k</td></tr></table>
+
+Table 8: The number of passages and extracted triplets on our selected datasets.
+
+# C Discussion: Novel Recall Algorithm
+
+The purpose of MS-RAG is to combine knowledge graph and vector retrieval to build a fast and effective RAG system and create an application that can be used in industry. First, we rethink the knowledge graph and vector model. Although GraphRAG (Edge et al., 2024) can solve the summarization problem well, it is difficult to achieve effective and fast retrieval, which makes it impossible to apply it on a large scale in industry. We believe that knowledge graph can bring the advantages of the basic data structure, which converts a long text into short entities and relations. The vector model may not perform well for complex and long text retrieval, but it can effectively recall short entities and relations. As shown in Figure 4, we discuss the differences and connection between dense recall, graph recall and our mix recall. As shown in Algorithm 1, we show the mix recall algorithm pipeline, included the vector search on chunk index and nearby search on graph index.
+
+# Algorithm 1 Mix Recall pipeline
+
+Require: input query  $q$
+
+Require: graph index  $m_{graph}$
+
+Require: chunk index  $m_{\text{chunk}}$
+
+1: procedure  
+2: rel,ent  $\leftarrow$  VectorSearch(q,mgraph)  
+3: chunk  $\leftarrow$  VectorSearch(q,  $m_{\text{chunk}}$ )  
+4: rel,ent  $\leftarrow$  NearBy(rel,ent,mgraph)  
+5:  $p\gets \mathbf{V}$ oting(chunk,rel,ent)  
+6: return  $p \gets \text{Rerank}(p)$  
+7: end procedure
+
+# D Discussion: Distinct Advantage over Existing Solutions
+
+We conduct discussion with existing HippoRAG, GraphRAG, and NaiveRAG to further emphasize the innovativeness of our framework.
+
+HippoRAG. Compared with HippoRAG (Gutiérrez et al., 2024), our MS-RAG have significantly improved the retrieval performance
+
+![](images/ad072ce48d5008b08aed1b77b6620b0ed98c6df216373407181da4890759fc28.jpg)
+
+![](images/d21afb376155346be1b391798ee2ad3440bf79bd0720d39f832fc2545ca1e13e.jpg)
+
+![](images/ac635f41736acba03d93352d4aa21fc56b5f22112bd731599cf87e137263af16.jpg)  
+Figure 4: (a) Dense Recall: dense recall struggles with complex and lengthy texts, but it excels at recalling short entities and relations. (b) Graph Recall: The KG provides the knowledge structure with semantic assoicitation, in which nearby nodes have explicit logical connections. (b) Mix Recall (ours): our proposed mix recall algorithm complementarily combines the advantages of graph recall and dense recall to achieve faster and more accurate retrieval.
+
+<table><tr><td></td><td colspan="3">MuSiQue</td><td colspan="3">2Wiki</td><td colspan="3">HotpotQA</td></tr><tr><td></td><td>EM</td><td>F1</td><td>MRR</td><td>EM</td><td>F1</td><td>MRR</td><td>EM</td><td>F1</td><td>MRR</td></tr><tr><td>HippoRAG</td><td>19.2</td><td>29.8</td><td>57.4</td><td>46.6</td><td>59.5</td><td>79.4</td><td>41.8</td><td>55.0</td><td>68.2</td></tr><tr><td>MS-RAG (Contriever)</td><td>21.8</td><td>33.6</td><td>58.3</td><td>60.2</td><td>66.9</td><td>88.2</td><td>55.7</td><td>67.7</td><td>88.3</td></tr><tr><td>MS-RAG (BGE-M3)</td><td>25.6</td><td>37.8</td><td>63.9</td><td>64.3</td><td>71.4</td><td>90.1</td><td>60.4</td><td>72.5</td><td>91.9</td></tr></table>
+
+Table 9: The retrieval performance on F1, EM and MRR metric. We conduct experiments to compare HippoRAG and MS-RAG on more metrics.  
+
+<table><tr><td></td><td>NaturalQA</td><td>PopQA</td><td>TriviaQA</td></tr><tr><td>BM25</td><td>56.1</td><td>35.7</td><td>63.6</td></tr><tr><td>HippoRAG</td><td>70.3</td><td>49.2</td><td>72.8</td></tr><tr><td>HippoRAG</td><td>44.4</td><td>53.8</td><td>-</td></tr><tr><td>MS-RAG</td><td>78.1</td><td>54.1</td><td>75.3</td></tr></table>
+
+Table 10: The simpleQA experiments on NaturalQA, PopQA and TriviaQA datasets. we add simpleQA experiments to further demonstrate the robustness.  
+
+<table><tr><td></td><td>R@2</td><td>R@5</td></tr><tr><td>MS-RAG w/ qwen 7b</td><td>79.4</td><td>88.5</td></tr><tr><td>MS-RAG w/ Llama 8b</td><td>77.2</td><td>85.9</td></tr><tr><td>MS-RAG w/o rerank</td><td>76.1</td><td>84.9</td></tr><tr><td>HippoRAG</td><td>59.0</td><td>76.2</td></tr></table>
+
+Table 11: The ablation experiments of rerank module on HotpotQA datasets. The ablation experiments of rerank LLMs in our proposed MS-RAG.
+
+and question answering performance. (i) Indexing stage: The proposed multi-semantic index additionally introduces chunk index to mitigate the negative impact of LLM errors. (ii) Retrieval stage: We propose novel mix recall algorithm instead of PPR to complementarily combine the advantages of graph recall and dense recall, and propose a multi-semantic rerank module to give the model reasoning capabilities.
+
+GraphRAG. Compared with GraphRAG (Edge
+
+et al., 2024), MS-RAG achieves better question-answering results while significantly reducing the number of LLM calls. We removed many unnecessary LLM calls, such as community clustering, community summarization, etc., which also accelerated inference speed.
+
+Naive RAG. Compared with Naive RAG, our MS-RAG boost the performance with slight cost. Unlike vector search in naive RAG, we introduce the knowledge graph structure and propose an in
+
+<table><tr><td>Query</td><td>Is Ari Up a guitarists?</td></tr><tr><td>Support Passage</td><td>Ariane Daniela Forster, known by her stage name Ari Up, was a German vocalist best known as a member of The Slits.</td></tr><tr><td>Support Entity</td><td>Ari Up, Ariane Daniela Forster, German vocalist</td></tr><tr><td>Support Relation</td><td>(Ariane Daniela Forster, has stage name, Ari Up), (Ariane Daniela Forster, was, German vocalist), ......</td></tr><tr><td>Real Answer</td><td>Ari Up isn’t a guitarist. The typical graph recall (HippoRAG), will locate the relations (Ariane Daniela Forster, was, German vocalist) based on the entities Ari Up and Ariane Daniela Forster, and thus get answer.</td></tr><tr><td>Challenge: Entity-missing</td><td>If important entities Ari Up or Ariane Daniela Forster are lost in the process of building the KG by LLM, graph recall will not be able to locate the answer.</td></tr><tr><td>Challenge: Relation-missing</td><td>If important relations (Ariane Daniela Forster, has stage name, Ari Up) or (Ariane Daniela Forster, was, German vocalist) are lost in the process of building the KG by LLM, graph recall will not be able to locate the answer.</td></tr></table>
+
+Table 12: The example of LLM errors for building the knowledge graph. When graph-based RAG excute retrieval on the index with the missing node/edge, the performance will decrease.  
+
+<table><tr><td></td><td>R@2</td><td>R@5</td></tr><tr><td>MS-RAG</td><td>79.4</td><td>88.5</td></tr><tr><td>MS-RAG w/o entity disa.</td><td>76.7</td><td>86.4</td></tr></table>
+
+Table 13: The ablation experiments of entity disambiguation on HotpotQA datasets. The ablation experiments of entity disambiguation in our proposed MS-RAG.  
+
+<table><tr><td>Error Type</td><td>Error Percent.(%)</td></tr><tr><td>Incorrect/missing graph</td><td>26</td></tr><tr><td>Recall error</td><td>52</td></tr><tr><td>Rerank error</td><td>22</td></tr></table>
+
+Table 14: The error analysis on HotpotQA datasets.
+
+novative mix recall algorithm to better utilize our multi-semantic index.
+
+# E The Challenges for Constructing Knowledge Graph
+
+To provide more details about LLM errors for building the knowledge graph, as shown in Table 12, we show an example to further analyze the challenge.
+
+# F The Experimental Results of F1, EM and MRR
+
+To make it easier for readers to understand the effectiveness, we present the experiments between HippoRAG and MS-RAG in terms of F1 score, Exact Match (EM), and Mean Reciprocal Rank
+
+(MRR). As shown in Table 9, our proposed MSRAG significantly outperforms the previous SOTA HippoRAG.
+
+# G The Experimental Results of SimpleQA tasks
+
+To further demonstrate the robustness, we randomly selected 1,000 questions from NaturalQA (Kwiatkowski et al., 2019), 1,000 questions from PopQA (Mallen et al., 2022), and 2,000 questions from TriviaQA (Joshi et al., 2017) to conduct experiments to demonstrate the performance of our model on the simpleQA task.
+
+The experimental results are shown in Table 10. In terms of retrieval performance (passage recall@5), the proposed MS-RAG model significantly outperforms the dense model BGE-M3, the classic retriever BM25, and the baseline HippoRAG.
+
+# H Ablation of Rerank Module
+
+As shown in Table 11, we conduct ablation experiments of rerank module on HotpotQA datasets. The experiments have shown that the qwen-7b rerank module can significantly improve performance. In addition, we also try Llama-3-8B-instruct as a chioce. When removing the rerank module, we also obtains better results than HippoRAG.
+
+<table><tr><td></td><td>GraphRAG</td><td>HippoRAG</td><td>MS-RAG (ours)</td></tr><tr><td>Indexing</td><td>extract graph, summarization, extract hierarchical community, ......</td><td>build graph</td><td>build graph</td></tr><tr><td>Retrieval</td><td>extract entity, summarization, community answers, ......</td><td>entity extraction</td><td>rerank</td></tr></table>
+
+Table 15: The comparison of LLM API calls between GraphRAG, HippoRAG and our MS-RAG in HotpotQA datasets.  
+
+<table><tr><td></td><td>chunk index</td><td>entity index</td><td>relation index</td><td>storage size</td></tr><tr><td>HippoRAG</td><td>×</td><td>✓</td><td>✓</td><td>2.1G</td></tr><tr><td>MS-RAG (ours)</td><td>✓</td><td>✓</td><td>✓</td><td>2.4G</td></tr></table>
+
+Table 16: The comparison of storage costs between baseline HippoRAG and our MS-RAG in HotpotQA datasets.
+
+# I Ablation of Entity Disambiguation
+
+As shown in Table 13, we conduct ablation experiments of entity disambiguation on HotpotQA datasets. Our entity disambiguation achieves R@2 improvement of +2.7, R@5 improvement of +2.1, which demonstrates its effectiveness.
+
+# J Error Analysis
+
+As shown in Table 14, we present error analysis on HotpotQA datasets. There are three error types. The main error type is recall error, which means the mix recall algorithm is not executed correctly. After voting, the correct subjects (chunks, relations/entiyes) doesn't enter the final rerank list. The second is incorrect/missing graph, which means there are missing entity/relation during the construction of KGs using LLM. The third is rerank error, which means the rerank model can's give right answer. The error analysis will help us improve the corresponding modules in the future.
+
+# K The Comparison of API Calls
+
+As shown in Table 15, our MS-RAG achieves superior performance with very few LLM calls. Compared with GraphRAG, MS-RAG greatly reduces LLM calls. Compared with HippRAG, we achieve significant performance improvements with comparable LLM calls.
+
+# L The Comparison of Storage Costs
+
+As show in Table 16, our multi-semantic index achieves better results only slight storage cost.
+
+Compared with HippoRAG, we only add an additional chunk index.
+
+# M The Experiments of Missing Edge/Node
+
+We provide further details on the experiments concerning missing nodes/edges in Figure 1. Our evaluation focuses on the ratio of missing or incorrect triples. For example, if an passage contains 30 ground truth triples but only 24 are correctly extracted, the missing ratio is  $20\%$ . Since the HotpotQA (Yang et al., 2018) and 2WikiMultiHopQA (Ho et al., 2020) datasets do not have official annotations of ground truth triples, we manually annotated the ground truth triples to ensure annotation quality.
+
+# N LLM Prompts
+
+Following HippoRAG (Gutiérrez et al., 2024), the knowledge graph is constructed in two stages, including entity extraction and relation extraction. The LLM prompts of relation extraction is shown in Figure 5, and the LLM prompts of entity extraction is shown in Figure 6. Besides, we also provide the LLM prompt of rerank module (Figure 7), entity disambiguation (Figure 8) and question-answering quality evaluation (Figure 9).
+
+```txt
+Instruction: Your task is to construct an RDF (Resource Description Framework) graph from the given passages and named entity lists. Respond with a JSON list of triples, with each triple representing a relationship in the RDF graph. Pay attention to the following requirements: - Each triple should contain at least one, but preferably two, of the named entities in the list for each passage. - Clearly resolve pronouns to their specific names to maintain clarity. Convert the paragraph into a JSON dict, it has a named entity list and a triple list. One-Shot Demonstration: Paragraph: Radio City Radio City is India's first private FM radio station and was started on 3 July 2001. It plays Hindi, English and regional songs. Radio City recently forayed into New Media in May 2008 with the launch of a music portal - PlanetRadiocity.com that offers music related news, videos, songs, and other music-related features. Named Entities: ["Radio City", "India", "3 July 2001", "Hindi", "English", "May 2008", "PlanetRadiocity.com"] Triples: [ ["Radio City", "located in", "India"], ["Radio City", "is", "private FM radio station"], ["Radio City", "started on", "3 July 2001"], ["Radio City", "plays songs in", "Hindi"], ["Radio City", "plays songs in", "English"], ["Radio City", "forayed into", "New Media"], ["Radio City", "launched", "PlanetRadiocity.com"], ["PlanetRadiocity.com", "launched in", "May 2008"], ["PlanetRadiocity.com", "is", "music portal"], ["PlanetRadiocity.com", "offers", "news"], ["PlanetRadiocity.com", "offers", "videos"], ["PlanetRadiocity.com", "offers", "songs"] ] Input: Convert the paragraph into a JSON dict, it has a named entity list and a triple list. Paragraph: PASSAGE TO INDEX Named Entities: [entities]
+```
+
+Figure 5: The LLM prompts of relation extraction.
+
+<table><tr><td>Instruction:Your task is to extract named entities from the given paragraph. Respond with a JSON list of entities.One-Shot Demonstration:Paragraph:Radio CityRadio City is India&#x27;s first private FM radio station and was started on 3 July 2001. It plays Hindi, English and regional songs. Radio City recently forayed into New Media in May 2008 with the launch of a music portal - PlanetRadiocity.com that offers music related news, videos, songs, and other music-related features.Namedentities:&quot;Radio City&quot;,&quot;India&quot;,&quot;3 July 2001&quot;,&quot;Hindi&quot;,&quot;English&quot;,&quot;May 2008&quot;,&quot;PlanetRadiocity.com&quot;]Input:Paragraph:PASSAGE TO INDEX</td></tr></table>
+
+Figure 6: The LLM prompts of entity extraction.
+
+<table><tr><td>You are an excellent natural language processing expert. Your task is to search the text in textList based on the input query to find the text that can answer the user&#x27;s query.
+Requirements:
+1. Return the first two texts that are helpful in answering the query.
+2. Use # to concatenate the returned text numbers, such as 2#3.
+3. Do not return correlation analysis, thinking process, and other irrelevant content.
+One-shot Demonstration:
+query: Are Ari Up and Boz Burrell both guitarists?
+textList: {{{&quot;Text number&quot;: 0, &quot;Text content&quot;: &quot;Raymond &quot;Boz&quot; Burrell (1 August 1946 - 21 September 2006) was an English musician. Originally a vocalist and guitarist, Burrell is best known for his bass playing and work with the bands King Crimson and Bad Company. He died of a heart attack in Spain on 21 September 2006 aged 60.&quot;}}, {{{&quot;Text number&quot;: 1, &quot;Text content&quot;: &quot;Ariane Daniela Forster (17 January 1962 - 20 October 2010), known by her stage name Ari Up, was a German vocalist best known as a member of the English punk rock band The Slits.&quot;}}, {{{&quot;Text number&quot;: 2, &quot;Text content&quot;: &quot;Muzzle is an alternative rock band formed in 1994 by Ryan Maxwell, Wesley Nelson, Burke Thomas, and Greg Collinsworth. They have released two albums with Reprise Records: &quot;Betty Pickup&quot; in 1996 and &quot;Actual Size&quot; in 1999.&quot;}}, {{{&quot;Text number&quot;:&quot;3&quot;,&quot;Text content&quot;:&quot;Douglas Theodore &quot;Doug&quot; Pinnick (born September 3, 1950), sometimes stylized as dUg Pinnick or simply dUg, is an American musician best known as the bass guitarist, songwriter, and co-lead vocalist for the hard rock/progressive metal band King&#x27;s amplifiers.&quot;}}, {{{&quot;Text number&quot;:&quot;4&quot;,&quot;Text content&quot;:&quot;Leslie West (born Leslie Weinstein; October 22, 1945) is an American rock guitarist, vocalist, and songwriter. He is best known as a founding member of the hard rock band Mountain.&quot;}}}]]Output: 0#1
+Real query:
+YOUR QUERY
+textList:
+YOUR TEXTLIST</td></tr></table>
+
+Figure 7: The LLM prompts of rerank module.
+
+Figure 8: The LLM prompts of entity disambiguation.  
+```txt
+Your task is to output different names for the same entity based on the input text. One-Shot Demonstration: input text: Fred Gehrke. Clarence Fred Gehrke (April 24, 1918  $\backslash$ u2013 February 9, 2002) was an American football player and executive. He played in the National Football League (NFL) for the Cleveland / Los Angeles Rams, San Francisco 49ers and Chicago Cardinals from 1940 through 1950. To boost team morale, Gehrke designed and painted the Los Angeles Rams logo in 1948, which was the first painted on the helmets of an NFL team. He later served as the general manager of the Denver Broncos from 1977 through 1981. He is the great-grandfather of Miami Marlin Christian Yelich. output: ["Fred Gehrke","Clarence Fred Gehrke"] input text: YOUR TEXT
+```
+
+```json
+---Role--- You are an expert tasked with evaluating two answers to the same question based on three criteria: Correctness, Diversity, and Comprehension. --Goal--- You will evaluate two answers to the same question based on three criteria: Correctness, Diversity, and Comprehension. - Correctness: Does the answer correctly answer the question? - Diversity: Does the answer provide a variety of perspectives on the question? - Comprehension: Does the answer effectively help readers understand? For each criterion, choose the better answer (either Answer 1 or Answer 2) and explain why. Then, select an overall winner based on these three categories. Here is the question: {query} Here is the groud-truth answer: {groud-truth answer} Here are the two answers: Answer 1: {answer1} Answer 2: {answer2} Evaluate both answers using the three criteria listed above and provide detailed explanations for each criterion. Output your evaluation in the following JSON format: {{ "Correctness": {{ "Winner": ["Answer 1 or Answer 2"], "Explanation": ["Provide explanation here"]}}, "Diversity": {{ "Winner": ["Answer 1 or Answer 2"], "Explanation": ["Provide explanation here"]}}, "Comprehension": {{ "Winner": ["Answer 1 or Answer 2"], "Explanation": ["Provide explanation here"]}}, "Overall Winner": {{ "Winner": ["Answer 1 or Answer 2"], "Explanation": ["Summarize why this answer is the overall winner based on the three criteria"]}}}}}
+```
+
+Figure 9: The LLM prompts of question-answering quality evaluation.

2025/MS-RAG_ Simple and Effective Multi-Semantic Retrieval-Augmented Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19e00370f5cdf340be668a315f446837c15d2ab592dfaf970cdfd8f4158d9e25
+size 1061348

2025/MUCAR_ Benchmarking Multilingual Cross-Modal Ambiguity Resolution for Multimodal Large Language Models/aabf9fec-71eb-46aa-acbd-b7c977b28b1d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df15d45889eee70a7ecb1210cbb04ebd23b033febfb237c4b4943a60f89c8752
+size 30990146

2025/MUCAR_ Benchmarking Multilingual Cross-Modal Ambiguity Resolution for Multimodal Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d7fd34289e96892106f552c116a74b6cf51a5fd8ff76c89da86679a99b1138c
+size 2211910

2025/MULTIGUARD_ An Efficient Approach for AI Safety Moderation Across Languages and Modalities/6a831b8d-c698-467c-8423-29005465a436_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af776a2301b25cbf46b52d6bad38fb0658e81650cb6f5e815e0500739dc58c3c
+size 2722399

2025/MULTIGUARD_ An Efficient Approach for AI Safety Moderation Across Languages and Modalities/full.md

@@ -0,0 +1,354 @@
+# OMNIGUARD: An Efficient Approach for AI Safety Moderation Across Languages and Modalities
+
+Sahil Verma<sup>1</sup> Keegan Hines<sup>2</sup> Jeff Bilmes<sup>1</sup> Charlotte Siska<sup>2</sup>
+
+Luke Zettlemoyer<sup>1</sup> Hila Gonen<sup>1</sup> Chandan Singh<sup>2</sup>
+
+<sup>1</sup>University of Washington  <sup>2</sup>Microsoft
+
+# Abstract
+
+The emerging capabilities of large language models (LLMs) have sparked concerns about their immediate potential for harmful misuse. The core approach to mitigate these concerns is the detection of harmful queries to the model. Current detection approaches are fallible, and are particularly susceptible to attacks that exploit mismatched generalization of model capabilities (e.g., prompts in low-resource languages or prompts provided in non-text modalities such as image and audio). To tackle this challenge, we propose OMNI-GUARD, an approach for detecting harmful prompts across languages and modalities. Our approach (i) identifies internal representations of an LLM/MLLM that are aligned across languages or modalities and then (ii) uses them to build a language-agnostic or modality-agnostic classifier for detecting harmful prompts. OMNI-GUARD improves harmful prompt classification accuracy by  $11.57\%$  over the strongest baseline in a multilingual setting, by  $20.44\%$  for image-based prompts, and sets a new SOTA for audio-based prompts. By repurposing embeddings computed during generation, OMNI-GUARD is also very efficient ( $\approx 120\times$  faster than the next fastest baseline). Code and data are available at https://github.com/vsahil/Omniguard.
+
+# 1 Introduction
+
+The rapid rise of capabilities in large language models (LLMs) has created an urgent need for safeguards to prevent their immediate harmful misuse as they are deployed to human users en masse (Bommasani et al., 2022). Moreover, these safeguards are critical for defending against future potential harms from LLMs (Bengio et al., 2024). Standard safeguard approaches broadly include approaches such as safety training using reinforcement learning from human feedback (Ouyang et al., 2022a; Leike et al., 2018) or using pre-trained
+
+![](images/1581fd567b5693188a090a68b7799261de4bc124417370ba11376d967eab91a2.jpg)  
+Figure 1: OMNIGUARD builds a harmfulness classifier that operates on internal representations of an LLM (or MLLM). OMNIGUARD uses a custom metric (U-Score) to identify representations that generalize across languages and modalities. At inference time, OMNIGUARD re-uses the embeddings from the LLM/MLLM being used for generation, and thereby completely avoids the overhead of passing the inputs through a separate guard model for safety moderation.
+
+guard models that classify the safety of an input prompt (OpenAI, 2025; Inan et al., 2023; Han et al., 2024).
+
+With these safeguards in place, harmful prompts in high-resource languages, e.g., English, are successfully detected. However, harmful prompts in low-resource languages can often bypass these safeguards (Deng et al., 2024; Yong et al., 2024; Yang et al., 2024), i.e., jailbreaking the LLM. Modern LLMs are vulnerable to attacks not only from low-resource natural languages, but also from artificial cipher languages, e.g., base64 or caesar encoding of English prompts (Wei et al., 2023; Yuan et al., 2024a). This phenomenon also extends beyond text to jailbreaking multimodal LLMs (MLLMs) using modalities such as images (Gong et al., 2025; Liu et al., 2024b) or audio (Yang et al., 2025).
+
+Wei et al. (2023) argue that these attacks are successful due to mismatched generalization, a scenario in which the model's safety training does not generalize to other settings, but general performance does. This may happen because pretraining data often includes more diverse data than that available for safety finetuning (Ghosh et al., 2024b).
+
+In this work, we defend against attacks that exploit the mismatched generalization of the safety training of LLMs and MLLMs. Specifically, we defend against attacks that utilize low-resource languages, both natural and cipher languages, as well as attacks employing other modalities, such as images and audio.
+
+We introduce OMNIGUARD, an approach that builds a classifier using the internal representations of a model. These representations are extracted from specific layers that produce representations that are universally similar across multiple languages and across multiple modalities. OMNIGUARD's classifier trained on such representations, is able to accurately detect harmful inputs across 73 languages, with an average of  $86.22\%$  accuracy across 53 natural languages and an average of  $73.06\%$  accuracy across 20 cipher languages. OMNIGUARD can also detect harmful inputs provided as images with  $88.31\%$  and as audio with  $93.09\%$  accuracy respectively.
+
+In contrast to popular guard models such as LlamaGuard (Inan et al., 2023), AegisGuard (Ghosh et al., 2024a), or WildGuard (Han et al., 2024), OMNIGUARD does not require training a separate LLM specifically to detect harmfulness. By building a classifier that uses the internal representations of the main LLM or MLLM, OMNIGUARD avoids the overhead of passing the prompt through a separate guard model, making it very efficient.
+
+In summary, our contributions are the following: (1) We propose OMNIGUARD, an approach for detecting harmful prompts, (2) we show that OMNIGUARD accurately detects harmfulness across multiple languages and multiple modalities, (3) we show that OMNIGUARD is very sample-efficient during training, and (4) we show that OMNIGUARD is highly efficient at inference time.
+
+# 2 Methodology
+
+OMNIGUARD seeks to robustly detect harmful prompts, regardless of their language or modality. We first leverage the tendency of LLMs and MLLMs to create universal representations that are similar across languages (Wendler et al., 2024; Zhao et al., 2024) and across modalities (Wu et al., 2024; Zhuang et al., 2025) in Section 2.1, and then use them to train harmfulness classifiers that robustly detect harmful inputs in Section 2.2.
+
+# 2.1 Finding language-agnostic representations in an LLM
+
+The first step of OMNIGUID searches for internal representations of an LLM that are universally shared across languages. We prompt an LLM with English sentences and their translations to other languages, and extract their representations at different layers. For language-agnostic representations, we expect the similarity between the representations of English sentences and the representations of their translations to be similar, and we expect this similarity to be higher than the similarity between representations of two sentences that are not translations of each other (a random pair of sentences). We concretize this notion by defining the Universality Score (U-Score, Eq. 1), which is the difference between the average cosine similarities of pairs of sentences that are translations of each other and pairs of sentences that are not.
+
+$$
+U - S c o r e :=
+$$
+
+$$
+\frac {1}{N} \sum_ {i \in [ N ]} \operatorname {C o s S i m} \left(\operatorname {E m b} \left(e _ {i}\right), \operatorname {E m b} \left(l _ {i}\right)\right)
+$$
+
+$$
+- \frac {1}{N (N - 1)} \sum_ {\substack {i, j \in [ N ] \\ i \neq j}} \operatorname {CosSim} \left(\operatorname {Emb}\left(e _ {i}\right), \operatorname {Emb}\left(l _ {j}\right)\right) \tag{1}
+$$
+
+where  $e_i$  and  $l_i$  are sentences in English and their translations to another language.
+
+This procedure can be generalized to new different modalities rather than different languages by changing which embeddings are being used. For example, to determine if internal representations of an MLLM are aligned across modalities, we replace embeddings for a translated piece of text with embeddings from a different modality (e.g. a text caption and its corresponding image, or a text transcription and its corresponding audio clip). See experimental details in Section 3.
+
+# 2.2 Fitting a harmfulness classifier
+
+After selecting the layer that maximizes the U-Score, we extract embeddings from that layer and use them as inputs to fit a lightweight, supervised classifier that predicts harmfulness. In our experiments, the classifier is a multilayer perceptron with 2 hidden layers (with hidden sizes 512 and 256). At inference time, when a prompt is passed to a model for generation, OMNIGUARD applies
+
+![](images/9fbe92e3ad8a6fb8068dfbe1ace5798c7d1a8f3c1e12907b0aec05f3b25ff74f.jpg)  
+Figure 2: The  $U$ -Score across different layers for different modalities. (A) Different layers of the model Llama3.3-70B-Instruct for different languages. (B) The Cross-Model Alignment Score at different layers of the model (Molmo-7B) for similarity between images and captions. The highest values are obtained with at layers 21-25, indicating better alignment between images and their text captions at these layers. (C) The Cross-Model Alignment Score at different layers of the model (Llama-Omni 8B) for similarity between audios and transcriptions. The highest values are obtained with at layers 20-23, indicating better alignment between audios and their text transcriptions at these layers.
+
+this classifier to the embeddings generated by the model, incurring minimal overhead at inference time for safety classification. Note, however, that this approach only applies to open-source models, for which OMNIGUARD can build a classifier by obtaining embeddings. During training, only the lightweight classifier's parameters are learned (the original model is never modified), making the training process data-efficient and inexpensive.
+
+# 3 Experimental Setup
+
+Table 1 and Table 2 give details on all the models and datasets for this section.
+
+# 3.1 Selecting universal layers via the U-Score
+
+Selecting language-agnostic layers To select language-agnostic layers, we use a dataset of translated sentence pairs spanning various languages. Specifically, we use sentences in 53 natural languages from the Flores200 dataset and additionally translate the sentences into 20 cipher languages (using encodings such as Caesar shifts, base64, hexadecimal); see a full list in Appendix A. We extract embeddings from each layer of Llama3.3-70B-Instruct for the sentences in all 73 languages and use them to compute the U-Score (averaged over languages). Fig. 2 shows the U-Score as a function of layer depth. For natural languages (blue curve), the U-Score peaks in the middle layers of
+
+the model, with the highest values in layer 57 (out of 81 layers). For cipher languages (red curve), the U-Score is much lower than for natural languages, suggesting the model fails to represent semantic similarity in these languages (see analysis in Section 5).
+
+Selecting modality-agnostic layers To select layers aligned between images and captions, we use the MM-Vet v2 dataset, a popular dataset for MLLM evaluation containing 517 examples, each consisting of a text question paired with one or more images. We generate captions for each image using a captioning model (Molmo-7B) and then extract embeddings for each image and its corresponding caption using an MLLM (also Molmo-7B) and use them to compute the U-Score, which peaks in layer 22 (out of 28 layers; see Fig. 2 green curve).
+
+To select layers aligned between text and audio, we use the audio version of the Alpacaeval dataset from VoiceBench, a dataset of 636 audio-transcript pairs. We extract embeddings from each layer of an MLLM (LLaMA-Omni 8B) and use them to compute the U-Score, which peaks at layer 21 (out of 32 layers; see Fig. 2 purple curve).
+
+Overall, we see that LLMs and MLLMs generate representations that are shared across languages and modalities.
+
+# 3.2 Training and evaluating the harmfulness classifier
+
+# 3.2.1 Setup for multilingual text attacks
+
+OMNIGUARD classifier. Following Section 3.1, we build a classifier that takes as input embeddings from layer 57 of Llama3.3-70B-Instruct. As training data, we randomly select 2,800 examples from the Aegis AI Content Safety dataset, balancing the benign and harmful classes. Notably, this dataset is about  $18 \times$  smaller than the training data used by our baseline methods. We translate these English examples to 52 other natural languages (via the Google Translate API) and 20 cipher languages (using fixed rules), totaling 73 languages. We train OMNIGUARD using only half the languages (see list in Appendix A).
+
+Baselines. We compare to many popular guard models (see Table 2) middle row. Notably, DuoGuard and PolyGuard were trained to detect harmful prompts across multiple languages. For a more direct comparison, we also compare to finetuned
+
+<table><tr><td></td><td>Dataset name</td><td>Citation</td><td>HuggingFace ID</td><td>Number of examples</td></tr><tr><td rowspan="3">General</td><td>Flores200</td><td>(Team et al., 2022)</td><td>Muennighoff/flores200</td><td>997</td></tr><tr><td>MM-Vet v2</td><td>(Yu et al., 2024b)</td><td>whyu/mm-vet-v2</td><td>517</td></tr><tr><td>SST-2</td><td>(Socher et al., 2013)</td><td>stanfordlp/sst2</td><td>1000</td></tr><tr><td rowspan="15">Text</td><td>Aegis AI Content Safety Dataset</td><td>(Ghosh et al., 2024b)</td><td>nvidia/Aegis-AI-Content-Safety-Dataset-1.0</td><td>10,800</td></tr><tr><td>MultiJail</td><td>(Deng et al., 2024)</td><td>DAMO-NLP-SG/MultiJail</td><td>315</td></tr><tr><td>Xsafety</td><td>(Wang et al., 2024a)</td><td>ToxicityPrompts/Xsafety</td><td>28,000</td></tr><tr><td>RTP-LX</td><td>(de Wynter et al., 2025)</td><td>ToxicityPrompts/RTP-LX</td><td>30,300</td></tr><tr><td>AyaRedTeaming</td><td>(Aakanksha et al., 2024)</td><td>CohereLabs/aya_redteaming</td><td>2662</td></tr><tr><td>Thai Toxicity tweets</td><td>(SiriHattasak et al., 2018)</td><td>tmu-nlp/thai_toxicity tweet</td><td>3,300</td></tr><tr><td>Ukr Toxicity</td><td>(Dementieva et al., 2024)</td><td>ukr-detect/ukr-toxicity-dataset</td><td>5,000</td></tr><tr><td>HarmBench (HB)</td><td>(Mazeika et al., 2024)</td><td>walledai/HarmBench</td><td>400</td></tr><tr><td>Forbidden Questions (FQ)</td><td>(Shen et al., 2024a)</td><td>TrustAIRLab/ForbiddenQuestion_set</td><td>390</td></tr><tr><td>Simple Safety Tests</td><td>(Vidgen et al., 2024)</td><td>walledai/SimpleSafetyTests</td><td>100</td></tr><tr><td>SaladBench (SaladB)</td><td>(Li et al., 2024a)</td><td>walledai/SaladBench</td><td>26,500</td></tr><tr><td>Toxicity Jigsaw (TJS)</td><td>(cjadams et al., 2017)</td><td>Arive/toxicity_classification_jigsaw</td><td>26,000</td></tr><tr><td>Toxic Text</td><td>(Correa, 2023)</td><td>nicholasKluge/toxic-text</td><td>41,800</td></tr><tr><td>AdvBench</td><td>(Zou et al., 2023a)</td><td>walledai/AdvBench</td><td>520</td></tr><tr><td>CodeAttack</td><td>(Ren et al., 2024)</td><td>https://github.com/AI45Lab/CodeAttack</td><td>3120</td></tr><tr><td rowspan="9">Vision</td><td>JailBreakV-28K</td><td>(Luo et al., 2024)</td><td>JailbreakV-28K/JailBreakV-28k</td><td>8,000</td></tr><tr><td>VLSafe</td><td>(Chen et al., 2024c)</td><td>YangyiYY/LVLM_NLF</td><td>1,110</td></tr><tr><td>FigStep</td><td>(Gong et al., 2025)</td><td>https://github.com/wangyu-ovo/MML</td><td>500</td></tr><tr><td>MML SafeBench</td><td>(Wang et al., 2024b)</td><td>https://github.com/wangyu-ovo/MML</td><td>2,510</td></tr><tr><td>Hades</td><td>(Li et al., 2024e)</td><td>Monosail/HADES</td><td>750</td></tr><tr><td>SafeBench</td><td>(Ying et al., 2024)</td><td>Zonghao2025/safebench</td><td>2,300</td></tr><tr><td>MM SafetyBench</td><td>(Liu et al., 2024b)</td><td>PKU-Alignment/MM-SafetyBench</td><td>1680</td></tr><tr><td>RedTeamVLM</td><td>(Li et al., 2024b)</td><td>MMInstruction/RedTeamingVLM</td><td>200</td></tr><tr><td>VLSBench</td><td>(Hu et al., 2025)</td><td>Foreshhh/vlsbench</td><td>2,240</td></tr><tr><td rowspan="2">Audio</td><td>VoiceBench (Alpacaeval)</td><td>(Chen et al., 2024d)</td><td>hl&#x27;t-lab/voicebench</td><td>636</td></tr><tr><td>AIAH</td><td>(Yang et al., 2025)</td><td>https://github.com/YangHao97/RedteamAudioLMMs</td><td>350</td></tr></table>
+
+Table 1: Details of datasets used for training and evaluation. Some of the text datasets are inherently multilingual : MultiJail (10 languages), XSafety (10 languages), RTP-LX (28 languages), Aya RedTeaming (8 languages), Thai Toxicity tweets (prompts in Thai), and Ukr Toxicity (prompts in Ukrainian). The remaining text datasets are English-only, and were translated to 72 other languages (52 natural and 20 cipher): HarmBench (HB), Forbidden Questions (FQ), Simple Safety Tests, SaladBench (SaladB), Toxicity Jigsaw (TJS), Toxic Text, and AdvBench.
+
+versions of DuoGuard and PolyGuard using the same 37 languages we use to train OMNIGUARD; Following the original PolyGuard paper, we finetuned these models using LoRA (Hu et al., 2021) for all linear layers with rank 8 and alpha 16 for one epoch with a learning rate of  $2e - 4$ .
+
+Datasets. We evaluate on several common text attack benchmarks (see Table 1). We additionally evaluate on three benchmarks from CodeAttacks that transform a harmful query as a list, a stack, or as a string in a Python program, obfuscating the harmfulness. For evaluation in this setup, we transform the harmful prompts from AdvBench and benign prompts from Toxicity Jigsaw datasets in the three code formats and subsample the Toxicity Jigsaw dataset to be of the same size as Advbench. Note that for this experiment, we only trained OMNIGUIDR on the English subset of the training dataset.
+
+# 3.2.2 Setup for vision attacks
+
+OMNIGUARD classifier. Following Section 3.1, we build a classifier that takes as input embeddings from layer 22 of Molmo-7B. As training data, we use 2000 image-query pairs randomly sampled from the JailBreakV-28K dataset and 1024 image-query pairs sampled from the VLSafe dataset as the harmful datapoints and 517 image-query from the MM-Vet v2 dataset as the benign datapoints.
+
+Baselines. We compare to guard models that take an image or image-text pair and output a binary harmfulness classification (see Table 2 bottom row). We train VLMGuard on the same training data as OMNIGUARD.
+
+Datasets. We evaluate detecting image/text attacks using several datasets (see Table 1). Fig-Step and MML Safebench are typographic attacks that embed a harmful prompt in an image. MML Safebench further encrypts a harmful prompt in several variants, such as rotation, mirror images, word replacement, and with base64 encoding. Hades and Safebench consist of images and text queries where the text itself is harmful. MM-safetybench, RTVLM, and VLSBench consist of an image and a query where the text query is seemingly benign, but when combined with the respective image, it is harmful (e.g. see Figure 1).
+
+# 3.2.3 Setup for audio attacks
+
+OMNIGUARD classifier. Following Section 3.1, we build a classifier that takes as input embeddings from layer 21 of Llama-Omni-8B. We train the classifier on the English portion of the training data we use for the text setting, by using a text-to-speech model to convert the text into audio. We use the open-source Kokoro model as the text-to-speech model.
+
+<table><tr><td></td><td>Model name</td><td>Citation</td><td>HuggingFace ID</td><td>Rough Parameter Count</td></tr><tr><td rowspan="4">General</td><td>Llama3.3-70B-Instruct</td><td>(Grattafiori et al., 2024)</td><td>meta-llama/Llama-3.3-70B-Instruct</td><td>70B</td></tr><tr><td>Molmo-7B</td><td>(Deitke et al., 2024)</td><td>allenai/Molmo-7B-D-0924</td><td>7B</td></tr><tr><td>LLama-Omni 8B</td><td>(Fang et al., 2025)</td><td>ICTNLP/Llama-3.1-8B-Omni</td><td>8B</td></tr><tr><td>Kokoro</td><td>(Hexgrad, 2025)</td><td>hexgrad/Kokoro-82M</td><td>82M</td></tr><tr><td rowspan="10">Text</td><td>LlamaGuard 1</td><td>(Inan et al., 2023)</td><td>meta-llama/LlamaGuard-7b</td><td>7B</td></tr><tr><td>LlamaGuard 2</td><td>(Inan et al., 2023)</td><td>meta-llama/Meta-Llama-Guard-2-8B</td><td>8B</td></tr><tr><td>LlamaGuard 3</td><td>(Inan et al., 2023)</td><td>meta-llama/Llama-Guard-3-8B</td><td>8B</td></tr><tr><td>AegisGuard Permissive</td><td>(Ghosh et al., 2024a)</td><td>nvidia/Aegis-AI-Content-Safety-LlamaGuard-Permissive-1.0</td><td>7B</td></tr><tr><td>AegisGuard Defensive</td><td>(Ghosh et al., 2024a)</td><td>nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0</td><td>7B</td></tr><tr><td>WildGuard</td><td>(Han et al., 2024)</td><td>allenai/wildguard</td><td>7B</td></tr><tr><td>HarmBench (mistral)</td><td>(Mazeika et al., 2024)</td><td>cais/HarmBench-Mistral-7b-val-cls</td><td>7B</td></tr><tr><td>HarmBench (llama)</td><td>(Mazeika et al., 2024)</td><td>cais/HarmBench-Llama-2-13b-cls</td><td>13B</td></tr><tr><td>DuoGuard</td><td>(Deng et al., 2025)</td><td>DuoGuard/DuoGuard-1B-Llama-3.2-transfer</td><td>1B</td></tr><tr><td>PolyGuard</td><td>(Kumar et al., 2025)</td><td>ToxicityPrompts/PolyGuard-Qwen</td><td>7B</td></tr><tr><td rowspan="3">Vision</td><td>Llama Guard 3 Vision</td><td>(Chi et al., 2024)</td><td>meta-llama/Llama-Guard-3-11B-Vision</td><td>11B</td></tr><tr><td>VLMGuard</td><td>(Du et al., 2024)</td><td>-</td><td>2.2M</td></tr><tr><td>LLavaGuard</td><td>(Helff et al., 2025)</td><td>AIML-TUDA/LlavaGuard-7B-hf</td><td>7B</td></tr></table>
+
+Table 2: Model and baseline details.
+
+Baselines. We are unaware of any existing models for detecting harmful audio input. The most relevant approach, SpeechGuard (Peri et al., 2024) adds noise as a defense against potentially harmful audio inputs but does not directly classify harmfulness. To contextualize our results for audio benchmarks, we compare performance to guard models that directly classify the raw text present in the audio (OMNIGUARD and LlamaGuard3).
+
+Datasets. We use the two audio benchmarks (see Table 1 bottom row). We also evaluate on several text jailbreak benchmarks using Kokoro to convert them from text to speech: HB, FQ, Simple Safety Tests, SaladB, and TJS. We use Kokoro for generating text-to-speech versions.
+
+# 4 Results
+
+Defending against multilingual text attacks Table 3 compares the accuracy of detecting harmful prompts for text benchmarks. Table 3(A) shows results for multilingual benchmarks, where OMNIGUIDArs achieves the highest accuracy  $(86.36\%)$  compared to the baselines, and achieves new stateof-the-art performance for 3 benchmarks: MultiJail, RTP-LX, and AyaRedTeaming. The strongest baseline is Polyguard, which yields an average accuracy of  $83.19\%$  despite being trained on a much larger dataset (1.91M examples for Polyguard versus 103K examples for OMNIGUIDArs). In benchmarks that were translated from English to various other languages, including cipher languages, we again see that OMNIGUIDArs achieves the highest accuracy (Table 3(B)). Finally, Table 3(C) shows that OMNIGUIDArs outperforms finetuned versions of DuoGuard and Polyguard on unseen languages, demonstrating that OMNIGUIDArs can outperform methods that were trained specifically for multilingual harmfulness classification.
+
+Defending against image-based attacks Table 4 shows the accuracy of detecting harmful image and text prompts for (A) pairs consisting of images and text queries, where either the image or both the image and query can be harmful and (B) typographic images with various encryptions. OMNIGUARD achieves the highest performance for both sets of benchmarks  $(95.44\%$  and  $79.76\%)$  while being trained using only about 3500 image-query pairs (compared to about 5500 datapoints used by LlavaGuard). The only benchmark where OMNIGUARD fails to detect harmful prompts is MML Base64, which consists of typographed images of prompts encrypted using base64 encoding.
+
+Defending against audio-based attacks Table 5 shows the accuracy of detecting harmful audio prompts. OMNIGUARD detects harmful audio input with high accuracy across all benchmarks. As we are not aware of any existing defenses for audio jailbreaks, we compare against OMNIGUARD and LlamaGuard3's accuracy in detecting harmful prompts when the same inputs are provided in English text. The accuracy OMNIGUARD achieves in detecting harmful audio inputs is similar to or higher than its performance for detecting harmful text inputs.
+
+Data-efficient adaptation We also evaluate the accuracy of OMNIGUARD and baselines in adapting to out-of-distribution code attacks given very few samples. In this setting, some prior work has speculated that guard models may be very data efficient, as they can make use of few-shot examples in-context (Inan et al., 2023). However, we find that baseline guard models generally struggle to rapidly adapt to this setting given few-shot examples (Figure 3). In contrast, OMNIGUARD is
+
+<table><tr><td colspan="2"></td><td>MultiJail</td><td>Xsafety</td><td>RTP-LX</td><td colspan="2">Aya RedTeaming</td><td>Thai Tox</td><td>Ukr Tox</td><td>Avg.</td></tr><tr><td rowspan="12">(A) Multilingual text benchmarks</td><td>LlamaGuard 1</td><td>39.27</td><td>57.01</td><td>48.66</td><td colspan="2">54.49</td><td>41.31</td><td>53.99</td><td>49.12</td></tr><tr><td>LlamaGuard 2</td><td>48.69</td><td>52.66</td><td>34.69</td><td colspan="2">58.58</td><td>42.86</td><td>51.79</td><td>48.21</td></tr><tr><td>LlamaGuard 3</td><td>66.87</td><td>64.34</td><td>45.57</td><td colspan="2">63.83</td><td>46.73</td><td>51.79</td><td>56.52</td></tr><tr><td>AegisGuard (P)</td><td>61.49</td><td>79.78</td><td>75.07</td><td colspan="2">78.88</td><td>56.09</td><td>65.75</td><td>69.51</td></tr><tr><td>AegisGuard (D)</td><td>79.71</td><td>90.77</td><td>92.17</td><td colspan="2">89.78</td><td>63.34</td><td>67.95</td><td>80.62</td></tr><tr><td>WildGuard</td><td>42.55</td><td>71.23</td><td>71.94</td><td colspan="2">61.45</td><td>40.42</td><td>55.03</td><td>57.10</td></tr><tr><td>HarmBench (llama)</td><td>0.22</td><td>0.14</td><td>0.0</td><td colspan="2">0.03</td><td>39.04</td><td>50.1</td><td>14.92</td></tr><tr><td>HarmBench (mistral)</td><td>2.4</td><td>5.65</td><td>5.14</td><td colspan="2">7.39</td><td>40.42</td><td>50.55</td><td>18.59</td></tr><tr><td>MD-Judge</td><td>25.78</td><td>53.58</td><td>66.46</td><td colspan="2">46.20</td><td>39.48</td><td>53.89</td><td>47.56</td></tr><tr><td>DuoGuard</td><td>39.20</td><td>63.42</td><td>66.57</td><td colspan="2">61.80</td><td>45.63</td><td>50.75</td><td>54.56</td></tr><tr><td>PolyGuard</td><td>82.00</td><td>96.41</td><td>83.86</td><td colspan="2">90.34</td><td>70.43</td><td>76.07</td><td>83.19</td></tr><tr><td>OMNIGUARD</td><td>93.83</td><td>93.64</td><td>94.55</td><td colspan="2">94.31</td><td>68.7</td><td>73.1</td><td>86.36</td></tr><tr><td colspan="2"></td><td>HarmBench</td><td>FQ</td><td>SimpleST</td><td>SaladB</td><td>TJS</td><td>ToxBench</td><td>Avg.</td><td></td></tr><tr><td rowspan="12">(B) Translated text benchmarks</td><td>LlamaGuard 1</td><td>32.47</td><td>23.75</td><td>34.32</td><td>23.27</td><td>62.49</td><td>65.55</td><td>34.39</td><td>39.46</td></tr><tr><td>LlamaGuard 2</td><td>57.19</td><td>43.72</td><td>50.71</td><td>34.54</td><td>58.21</td><td>62.17</td><td>56.95</td><td>51.93</td></tr><tr><td>LlamaGuard 3</td><td>70.02</td><td>53.25</td><td>67.81</td><td>46.30</td><td>62.33</td><td>70.87</td><td>70.26</td><td>62.98</td></tr><tr><td>AegisGuard (P)</td><td>62.16</td><td>43.01</td><td>56.55</td><td>44.92</td><td>73.69</td><td>72.80</td><td>62.12</td><td>59.32</td></tr><tr><td>AegisGuard (D)</td><td>88.53</td><td>76.67</td><td>87.64</td><td>78.27</td><td>71.38</td><td>68.72</td><td>90.77</td><td>80.28</td></tr><tr><td>WildGuard</td><td>33.64</td><td>31.20</td><td>33.90</td><td>27.37</td><td>66.61</td><td>67.27</td><td>39.98</td><td>42.85</td></tr><tr><td>HarmBench (llama)</td><td>0.03</td><td>0.11</td><td>0.01</td><td>0.07</td><td>48.62</td><td>49.97</td><td>0.01</td><td>14.12</td></tr><tr><td>HarmBench (mistral)</td><td>2.32</td><td>1.75</td><td>2.04</td><td>1.66</td><td>50.53</td><td>50.69</td><td>1.7</td><td>15.81</td></tr><tr><td>MD-Judge</td><td>16.19</td><td>12.11</td><td>22.29</td><td>13.81</td><td>65.34</td><td>64.26</td><td>25.67</td><td>31.38</td></tr><tr><td>DuoGuard</td><td>20.44</td><td>44.36</td><td>28.79</td><td>36.88</td><td>68.57</td><td>69.07</td><td>28.58</td><td>42.38</td></tr><tr><td>PolyGuard</td><td>66.22</td><td>56.05</td><td>62.53</td><td>54.88</td><td>78.34</td><td>76.52</td><td>67.96</td><td>66.07</td></tr><tr><td>OMNIGUARD</td><td>89.13</td><td>89.57</td><td>89.62</td><td>87.30</td><td>76.68</td><td>75.07</td><td>86.59</td><td>84.85</td></tr><tr><td colspan="2"></td><td>HarmBench</td><td>FQ</td><td>SimpleST</td><td>SaladB</td><td>TJS</td><td>ToxBench</td><td>Avg.</td><td></td></tr><tr><td rowspan="3">(C) Unseen langs.</td><td>FT DuoGuard</td><td>23.59</td><td>39.08</td><td>28.14</td><td>33.29</td><td>54.1</td><td>53.23</td><td>28.29</td><td>37.1</td></tr><tr><td>FT PolyGuard</td><td>72.45</td><td>79.84</td><td>76.81</td><td>76.85</td><td>74.07</td><td>72.33</td><td>73.55</td><td>75.13</td></tr><tr><td>OMNIGUARD</td><td>86.51</td><td>86.65</td><td>86.42</td><td>85.01</td><td>72.82</td><td>71.44</td><td>84.29</td><td>81.88</td></tr></table>
+
+Table 3: Accuracy of detecting harmful prompts for text attack benchmarks that are (A) multilingual benchmarks, (B) English translated to 73 languages, and (C) English translated to languages not seen at training time. In all settings, OMNIGUARD achieves the highest performance. Table B1 further stratifies these results by high-resource, low-resource, and cipher languages.  
+
+<table><tr><td colspan="2"></td><td>Hades</td><td>VLSBench</td><td>MM-SafetyBench</td><td>SafeBench</td><td>RTVLM</td><td>FigStep</td><td>Avg.</td></tr><tr><td rowspan="4">(A) +Query</td><td>Llama3 Vision GRD</td><td>76.00</td><td>3.97</td><td>31.90</td><td>68.40</td><td>56.50</td><td>47.40</td><td>47.36</td></tr><tr><td>VLMGuard</td><td>98.00</td><td>74.56</td><td>92.20</td><td>73.90</td><td>94.00</td><td>99.80</td><td>88.74</td></tr><tr><td>LLavaGuard</td><td>23.73</td><td>42.08</td><td>10.95</td><td>12.10</td><td>18.50</td><td>3.40</td><td>18.46</td></tr><tr><td>OMNIGUARD</td><td>100.00</td><td>92.24</td><td>99.82</td><td>91.60</td><td>89.00</td><td>100.00</td><td>95.44</td></tr><tr><td colspan="2"></td><td>MML Rotate</td><td>MML Mirror</td><td>MML W.R.</td><td>MML Q.R.</td><td colspan="2">MML Base64</td><td>Avg.</td></tr><tr><td rowspan="4">(B) Typographical image</td><td>Llama3 Vision GRD</td><td>83.20</td><td>68.00</td><td>96.40</td><td>25.40</td><td colspan="2">98.80</td><td>74.36</td></tr><tr><td>VLMGuard</td><td>6.80</td><td>21.00</td><td>100.0</td><td>86.20</td><td colspan="2">0.20</td><td>42.84</td></tr><tr><td>LLavaGuard</td><td>0.00</td><td>0.00</td><td>0.00</td><td>11.40</td><td colspan="2">0.00</td><td>2.28</td></tr><tr><td>OMNIGUARD</td><td>100.0</td><td>100.0</td><td>99.60</td><td>98.80</td><td colspan="2">0.40</td><td>79.76</td></tr></table>
+
+Table 4: Accuracy of detecting harmful queries in multimodal benchmarks for (A) image-query pairs and (B) typographed images with encrypted text. OMNIGUARD achieves the highest performance for both kinds of benchmarks.  
+
+<table><tr><td></td><td>AIAH</td><td>SafeBench (M)</td><td>SafeBench (F)</td><td>HB</td><td>FQ</td><td>SimpleST</td><td>SaladB</td><td>TJS</td><td>AdvBench</td></tr><tr><td>OMNIGUARD (Audio)</td><td>91.14</td><td>94.4</td><td>93.8</td><td>95.98</td><td>90.42</td><td>97.0</td><td>94.21</td><td>82.03</td><td>98.85</td></tr><tr><td>OMNIGUARD (text-en)</td><td>-</td><td>-</td><td>-</td><td>92.0</td><td>93.3</td><td>93.0</td><td>90.2</td><td>93.2</td><td>90.0</td></tr><tr><td>LlamaGuard3 (text-en)</td><td>-</td><td>-</td><td>-</td><td>97.32</td><td>78.75</td><td>99.0</td><td>67.03</td><td>72.16</td><td>98.07</td></tr></table>
+
+Table 5: Accuracy of detecting harmful queries in audio. OMNIGUARD is able to detect harmful audio inputs with high accuracy across all benchmarks. Since there are no baselines for detecting harmful prompts in audio, we compare the performance against OMNIGUARD's and LlamaGuard3 when the same benchmarks are provided as text in English.
+
+![](images/bf809896c714d38bcd4b77a3ad3b42c8e2082a0adf331cc241c0960b4191aa58.jpg)
+
+![](images/de5890ec734eb22e478ab65e903ebb87fbf48bb43b23d99911205b3a893d1db7.jpg)
+
+![](images/73a72d99afb86431a7f8ee9dbfff96653f7aa9d0cc0f1e185263bb5e665357d6.jpg)  
+Figure 3: Accuracy of detecting harmful prompts in a few-shot setting. As few-shot examples are provided, OMNIGUARD quickly achieves near-perfect accuracy, despite the attacks being quite different from its training data (e.g. without any few-shot examples, OMNIGUARD's accuracy is close to  $50\%$ ). In contrast, the guard model baselines improve their accuracy slowly in a few-shot setting, despite sometimes having seen similar code attacks in their training data. Accuracies are averaged over 50 random sets of few-shot examples; error bars show the standard error of the mean.
+
+able to rapidly achieve close to  $100\%$  accuracy for all three benchmarks by updating its lightweight parameters using less than five examples.
+
+# 5 Analysis
+
+Effect of U-Score-based layer selection. We perform ablation experiments to determine the effect of selecting the appropriate layer for training the OMNIGUARD classifier. For the text-only model, we compare the U-Score-selected layer (57) to 3
+
+their training data likely explicitly includes code attacks.
+
+<table><tr><td></td><td>Thai Tox</td><td>Ukr Tox</td><td>TJS</td><td>Toxtext</td><td>Avg.</td></tr><tr><td>Layer 10</td><td>62.1</td><td>65.5</td><td>66.95</td><td>61.89</td><td>64.42</td></tr><tr><td>Layer 75</td><td>65.2</td><td>66.4</td><td>70.72</td><td>65.79</td><td>68.26</td></tr><tr><td>Last Layer</td><td>63.1</td><td>51.2</td><td>61.33</td><td>56.76</td><td>59.05</td></tr><tr><td>U-Score selected layer (57)</td><td>68.7</td><td>73.1</td><td>76.8</td><td>75.07</td><td>73.4</td></tr></table>
+
+Table 6: OMNIGUARD's accuracy of detecting harmful prompts when trained using representations from different model layers.  
+
+<table><tr><td>Guard Method</td><td>Inference Time (s) ↓</td></tr><tr><td>LlamaGuard 3</td><td>87.25</td></tr><tr><td>AegisGuard (D)</td><td>152.26</td></tr><tr><td>WildGuard</td><td>306.14</td></tr><tr><td>MD-Judge</td><td>128.26</td></tr><tr><td>DuoGuard</td><td>4.85</td></tr><tr><td>PolyGuard</td><td>409.90</td></tr><tr><td>OMNIGUARD</td><td>0.04</td></tr></table>
+
+Table 7: Average inference time required for harmfulness prediction on the AdvBench dataset (averaged over 5 languages). OMNIGUIDR is about  $120\times$  faster than the fastest baseline (DuoGuard).
+
+other layers (layer 10, layer 75, and the last layer) when used for a set of toxicity prediction tasks. Table 6 shows that the representations from the layer with the highest  $U$ -Score result in significantly better harmfulness classification accuracy, improving between  $5\%$  and  $14\%$  compared to the other layers. We show ablation over more layers in Table B2.
+
+OMNIGUARD's efficiency OMNIGUARD is highly efficient at inference time because it reuses the internal representations of the main LLM that is already processing the user query for generation. Therefore, its compute time is only that of a lightweight multilayer perceptron, making it much faster than baseline guard models (note that this does limit OMNIGUARD to only work when the generation model is open-source, so embeddings can be extracted). Table 7 shows the inference time required by various guard models to predict the harmfulness of prompts in the AdvBench dataset in English, translated to Spanish, French, Telugu, and base64 encoding. OMNIGUARD is the fastest and is about  $120 \times$  faster than the fastest baseline (DuoGuard). Inference time as measured on a machine with 1 L40 GPU, 4 CPUs, and 50 GB RAM.
+
+Performance comparison across base LLMs We compare OMNIGUARD's accuracy when using different base LLMs in Table B3. We trained the classifiers on the layers with the best U-scores for each model. We find that the average accuracy for the moderator model trained using smaller LLMs is lower than the moderator model trained using
+
+![](images/8bde17c9b1ca6dad38eab253337c769dad522e76fbadc88b924ab651c5881466.jpg)  
+Figure 4: Comparison of accuracy of classifying sentiments in various languages compared to detecting harmful prompts in those languages using OMNIGUARD. In both cases the LLM is Llama3.3-70B-Instruct.
+
+the larger Llama3.3-70B-Instruct model.
+
+# Performance comparison across languages.
+
+We now analyze the harmfulness classification accuracy of OMNIGUARD by language, and compare it to the underlying LLM's sentiment classification accuracy for the same language (Fig. 4). We measure harmfulness classification accuracy using OMNIGUARD on all the datasets in Table 3 and sentiment classification accuracy using Llama3.3-70B-Instruct with zero-shot prompting on 72 translated versions of the SST-2 dataset (translated to all the languages we consider).
+
+We observe that the accuracies are generally correlated, indicating that OMNIGUIDR is able to defend well in languages for which the LLM is more coherent/susceptible to attack. Unsurprisingly, the accuracies for natural languages are higher than the accuracies for cipher languages. Nevertheless, harmfulness classification accuracy can be fairly high, even when sentiment classification accuracy is near chance (50%).
+
+# 6 Related Work
+
+Jailbreak Attacks in LLMs Several techniques have recently emerged to attack or jailbreak LLMs. Early techniques relied on manual effort and were very time-intensive (Shen et al., 2024b; Andriushchenko et al., 2024). Later techniques automated this process, e.g., Zou et al. (2023b); Jones et al. (2023); Zhu et al. (2023) proposed gradient
+
+based approaches to identify inputs to jailbreak LLMs with white-box access. Another set of techniques start from a set of human written prompts and modify them using approaches like genetic algorithms (Liu et al., 2024a; Lapid et al., 2024; Li et al., 2024d), fuzzing (Yu et al., 2024a), or reinforcement learning (Chen et al., 2024b) to automatically produce prompts for jailbreaking. Another set of techniques, use a helper LLM to generate prompts that attack a target LLM (Chao et al., 2024; Ding et al., 2024; Mehrotra et al., 2024). Finally, Wei et al. (2024); Wang et al. (2023); Anil et al. (2024); Pernisi et al. (2024) use simple in-context demonstrations to jailbreak the models by overcoming its safety training and Russinovich et al. (2024); Li et al. (2024c) propose using multi-turn dialogues to jailbreak models.
+
+Multilingual Jailbreak Attacks Most of the aforementioned jailbreak techniques focus on attacks in English, against which significant defense exists both at the model and system level. To tackle this, a novel set of techniques have emerged that attack models using inputs in various languages or obfuscations that are able to bypass the safety guardrails. Deng et al. (2024); Yong et al. (2024); Wang et al. (2024a); Yang et al. (2024); Yoo et al. (2024); Upadhayay and Behzadan (2024); Song et al. (2024) demonstrated that attacking models using mid and low resources languages led to higher attack success rates, compared to the case of attacking the model in high-resource languages like English.
+
+Going beyond natural languages, a newer set of works propose using cipher characters or languages to evade the safety filters, e.g., Jin et al. (2024) propose interspersing cipher characters in between text, Jiang et al. (2024) propose replacing the unsafe words with their ASCII art versions, and Yuan et al. (2024a) propose prompting models in cipher languages like Morse, Atbash, Caesar.
+
+Multimodal Jailbreak Attacks Using modalities apart from text aims to explore a completely new attack surface, like images or audios. Several recent works have shown that MLLMs remain vulnerable to being jailbroken when prompted with images or audios that have a harmful query (the same harmful query in text would be easily detected as harmful). Liu et al. (2024b) show that using a prompt with a correlated image, e.g., using an image of a bomb when asking the model to answer the question: How to make a bomb? is
+
+more likely to jailbreak a model than when using an uncorrelated image. Hu et al. (2025) argue that providing a harmful image with a benign query (see Figure 1) further increases the potential of jailbreaking the model. Gong et al. (2025) and Wang et al. (2024b) demonstrate jailbreaking models by simply typographically embedding harmful queries in an image.
+
+Safety moderation in LLMs Safety moderation in LLMs broadly fits into two categories: intrinsic and extrinsic. Intrinsic mechanisms include finetuning or RLHF training on an LLM (Bianchi et al., 2024; Chen et al., 2024a; Yuan et al., 2024b; Ouyang et al., 2022b; Bai et al., 2022; Dai et al., 2023). Extrinsic safety mechanisms utilize external models to detect harmful inputs and responses; these models can either be simple filters or use guard models. Jain et al. (2023); Alon and Kamfonas (2023); Hu et al. (2024) propose using perplexity filtering for detecting harmful prompts. Guard models defend LLMs by training separate LLMs to detect harmful text (see Table 2). Separately, a line of work has introduced interpretability methods to transparently expose safety concerns in LLMs (Bereska and Gavves, 2024; Singh et al., 2023; Arditi et al., 2024; Benara et al., 2024).
+
+A few other works also use internal model representations to defend against harmful inputs. However, they defer from our work: OMNIGUARD is a standalone content safety classification model while the other approaches like Jailbreak Antidote (Shen et al., 2025) and AdaSteer (Zhao et al., 2025) directly change the internal representations to defend against harmful inputs. Therefore, OMNIGUARD is similar to other safety classification models like LlamaGuard and WildGuard. Additionally, OMNIGUARD can detect harmful prompts in multiple languages (both natural and cipher) and multiple modalities using the same method while the other approaches only defend against attacks in English text. OMNIGUARD is also extremely efficient in producing safety predictions while the other approaches take as much time as the inference of the underlying LLM, which can typically take several seconds to minutes.
+
+Safety against multilingual attacks DuoGuard (Deng et al., 2025) and PolyGuard (Kumar et al., 2025) are the two previous guard models that were specifically trained to defend against multilingual attacks. DuoGuard uses a two-player RL-driven mechanism to generate harmful data
+
+in multiple languages and uses that to finetune a Llama3.2-1B model. PolyGuard collects an extensive dataset of 1.91M samples of harmful and benign datapoints in 17 languages and uses that to finetune a Qwen-2.5-7B-Instruct model.
+
+Safety moderation in MLLMs Relatively few works have tackled detecting harmful prompts in multimodal settings (see Table 1 and Table 2). Chi et al. (2024) propose LlamaGuard3-11B-Vision (a finetuned version of Llama-3-11B-Vision) for detecting unsafe inputs in images and texts. Du et al. (2024) and Helff et al. (2025) propose other approaches for the same task. OMNIGUARD achieves higher accuracy in detecting harmful images and prompts compared to these approaches, and to the best of our knowledge is the first guard model for harmful audio inputs.
+
+# 7 Conclusions
+
+We propose OMNIGUARD, an approach for training a safety moderation classifier using the internal representations of an LLM or MLLM that are universally similar across languages and modalities. Our approach consists of two steps: first, we identify these universally similar representations and then we use them to train a harmfulness classifier. We find that OMNIGUARD accurately detects harmful prompts across languages, including low-resource languages as well as cipher languages, and also across modalities - images and audios. We show that OMNIGUARD allows to train more efficient safety moderation classifiers (both in training time and in inference time) compared to standard guard models, and conclude that our approach is superior in both accuracy and efficiency across languages and modalities.
+
+# Limitations
+
+While OMNIGUARD achieves state-of-the-art performance for detecting harmful prompts across languages and modalities, its performance depends on the underlying model. If the underlying model does not understand the language or an image or audio input, OMNIGUARD might not be able to detect if the input is harmful. However, this limitation is not unique to OMNIGUARD, and existing approaches suffer from the same limitation.
+
+Our approach also relies on the existence of universally similar representations, which we empirically found to exist across models and modalities. However, we did not exhaustively check all models
+
+and this assumption might not hold for models that we have not used in this work. Moreover, OMNI-GUARD requires access to internal representations of a model, making it inapplicable to closed-source models.
+
+Lastly, the results we report are based on a fixed set of evaluation datasets that are standard benchmarks used in the research area of AI safety moderation. While OMNIGUARD performs well across the datasets we experiment with, its performance in real-world settings might differ.
+
+Ethics. While this work seeks to mitigate the risks of LLM deployment in high-risk scenarios, OMNIGUARD is not a perfect classifier and unexpected failures may allow for the harmful misuse of LLMs.
+
+# References
+
+Aakanksha, Arash Ahmadian, Beyza Ermis, Seraphina Goldfarb-Tarrant, Julia Kreutzer, Marzieh Fadaee, and Sara Hooker. 2024. The multilingual alignment prism: Aligning global and local preferences to reduce harm. Preprint, arXiv:2406.18682.  
+Gabriel Alon and Michael Kamfonas. 2023. Detecting language model attacks with perplexity. Preprint, arXiv:2308.14132.  
+Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. 2024. Jailbreaking leading safety-aligned Ilms with simple adaptive attacks.  
+Cem Anil, Esin Durmus, ..., and David Duvenaud. 2024. Many-shot jailbreaking.  
+Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. Advances in Neural Information Processing Systems, 37:136037-136083.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, and 12 others. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback.  
+Vinamra Benara, Chandan Singh, John X Morris, Richard Antonello, Ion Stoica, Alexander G Huth, and Jianfeng Gao. 2024. Crafting interpretable embeddings by asking llms questions. arXiv preprint arXiv:2405.16714.  
+Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, and 1 others. 2024. Managing extreme ai risks amid rapid progress. Science, 384(6698):842-845.  
+Leonard Bereska and Efstratios Gavves. 2024. Mechanistic interpretability for ai safety-a review. arXiv preprint arXiv:2404.14082.
+
+Federico Bianchi, Mirac Suzgun, Giuseppe Attanasio, Paul Rottger, Dan Jurafsky, Tatsunori Hashimoto, and James Zou. 2024. Safety-tuned LLaMAs: Lessons from improving the safety of large language models that follow instructions. In The Twelfth International Conference on Learning Representations.  
+Rishi Bommasani, Drew A. Hudson, ..., and Percy Liang. 2022. On the opportunities and risks of foundation models. Preprint, arXiv:2108.07258.  
+Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J. Pappas, and Eric Wong. 2024. Jailbreaking black box large language models in twenty queries. Preprint, arXiv:2310.08419.  
+Kai Chen, Chunwei Wang, Kuo Yang, Jianhua Han, Lanqing Hong, Fei Mi, Hang Xu, Zhengying Liu, Wenyong Huang, Zhenguo Li, Dit-Yan Yeung, Lifeng Shang, Xin Jiang, and Qun Liu. 2024a. Gaining wisdom from setbacks: Aligning large language models via mistake analysis.  
+Xuan Chen, Yuzhou Nie, Wenbo Guo, and Xiangyu Zhang. 2024b. When llm meets drl: Advancing jailbreaking efficiency via drl-guided search. Preprint, arXiv:2406.08705.  
+Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. 2024c. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. Preprint, arXiv:2311.10081.  
+Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T. Tan, and Haizhou Li. 2024d. Voicebench: Benchmarking llm-based voice assistants. Preprint, arXiv:2410.17196.  
+Jianfeng Chi, Ujjwal Karn, Hongyuan Zhan, Eric Smith, Javier Rando, Yiming Zhang, Kate Plawiak, Zacharie Delpierre Coudert, Kartikeya Upasani, and Mahesh Pasupuleti. 2024. Llama guard 3 vision: Safeguarding human-ai image understanding conversations. Preprint, arXiv:2411.10414.  
+cjadams, Jeffrey Sorensen, Julia Elliott, Lucas Dixon, Mark McDonald, nithum, and Will Cukierski. 2017. Toxic comment classification challenge.  
+Nicholas Kluge Correa. 2023. Aira.  
+Josef Dai, Xuehai Pan, Ruiyang Sun, Jiaming Ji, Xinbo Xu, Mickel Liu, Yizhou Wang, and Yaodong Yang. 2023. Safe rlhf: Safe reinforcement learning from human feedback. Preprint, arXiv:2310.12773.  
+Adrian de Wynter, Ishaan Watts, Tua Wongsangaroonsri, Minghui Zhang, Noura Farra, Nektar Ege Altintoprak, Lena Baur, Samantha Claudet, Pavel Gajdušek, Qilong Gu, Anna Kaminska, Tomasz Kaminski, Ruby Kuo, Akiko Kyuba, Jongho Lee, Kartik Mathur, Petter Merok, Ivana Milovanović, Nani Paananen, and 13 others. 2025. Rtp-lx: Can llms evaluate toxicity in multilingual scenarios? Proceedings of the AAAI Conference on Artificial Intelligence, 39(27):27940-27950.  
+Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, Jiasen Lu, Taira Anderson, Erin Bransom, Kiana Ehsani, Huong Ngo, YenSung Chen, Ajay Patel, Mark Yatskar, Chris Callison-Burch, and 31 others. 2024. Molmo and pixmo: Open weights and open data for state-of-the-art vision-language models. Preprint, arXiv:2409.17146.
+
+Daryna Dementieva, Valeriia Khylenko, Nikolay Babakov, and Georg Groh. 2024. Toxicity classification in Ukrainian. In Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024), pages 244-255, Mexico City, Mexico. Association for Computational Linguistics.  
+Yihe Deng, Yu Yang, Junkai Zhang, Wei Wang, and Bo Li. 2025. Duoguard: A two-player rl-driven framework for multilingual llm guardrails. Preprint, arXiv:2502.05163.  
+Yue Deng, Wenxuan Zhang, Sinno Jialin Pan, and Lidong Bing. 2024. Multilingual jailbreak challenges in large language models. In The Twelfth International Conference on Learning Representations.  
+Peng Ding, Jun Kuang, Dan Ma, Xuezhi Cao, Yunsen Xian, Jiajun Chen, and Shujian Huang. 2024. A wolf in sheep's clothing: Generalized nested jailbreak prompts can fool large language models easily. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 2136-2153, Mexico City, Mexico. Association for Computational Linguistics.  
+Xuefeng Du, Reshmi Ghosh, Robert Sim, Ahmed Salem, Vitor Carvalho, Emily Lawton, Yixuan Li, and Jack W. Stokes. 2024. Vlmguard: Defending vlms against malicious prompts via unlabeled data. Preprint, arXiv:2410.00296.  
+Qingkai Fang, Shoutao Guo, Yan Zhou, Zhengrui Ma, Shaolei Zhang, and Yang Feng. 2025. Llama-omni: Seamless speech interaction with large language models. Preprint, arXiv:2409.06666.  
+Shaona Ghosh, Prasoon Varshney, Erick Galinkin, and Christopher Parisien. 2024a. Aegis: Online adaptive ai content safety moderation with ensemble of llm experts.  
+Shaona Ghosh, Prasoon Varshney, Makes Narsimhan Sreedhar, Aishwarya Padmakumar, Traian Rebedea, Jibin Rajan Varghese, and Christopher Parisien. 2024b. Aegis2.0: A diverse ai safety dataset and risks taxonomy for alignment of llm guardrails. In Neurips Safe Generative AI Workshop.  
+Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. 2025. Figstep: Jailbreaking large vision-language models via typographic visual prompts. Preprint, arXiv:2311.05608.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. 2024. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms. Preprint, arXiv:2406.18495.  
+Lukas Helff, Felix Friedrich, Manuel Brack, Kristian Kersting, and Patrick Schramowski. 2025. Llavaguard: An open vlm-based framework for safeguarding vision datasets and models. Preprint, arXiv:2406.05113.  
+Hexgrad. 2025. Kokoro-82m.
+
+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. Preprint, arXiv:2106.09685.  
+Xuhao Hu, Dongrui Liu, Hao Li, Xuanjing Huang, and Jing Shao. 2025. Vlsbench: Unveiling visual leakage in multimodal safety. Preprint, arXiv:2411.19939.  
+Zhengmian Hu, Gang Wu, Saayan Mitra, Ruiyi Zhang, Tong Sun, Heng Huang, and Viswanathan Swaminathan. 2024. Token-level adversarial prompt detection based on perplexity measures and contextual information. Preprint, arXiv:2311.11509.  
+Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, and Madian Khabsa. 2023. Llama guard: Llm-based input-output safeguard for humanai conversations. Preprint, arXiv:2312.06674.  
+Neel Jain, Avi Schwarzschild, Yuxin Wen, Gowthami Somepalli, John Kirchenbauer, Ping yeh Chiang, Micah Goldblum, Aniruddha Saha, Jonas Geiping, and Tom Goldstein. 2023. Baseline defenses for adversarial attacks against aligned language models. Preprint, arXiv:2309.00614.  
+Fengqing Jiang, Zhangchen Xu, Luyao Niu, Zhen Xiang, Bhaskar Ramasubramanian, Bo Li, and Radha Poovendran. 2024. Artprompt: Ascii art-based jailbreak attacks against aligned llms.  
+Haibo Jin, Andy Zhou, Joe D. Menke, and Haohan Wang. 2024. Jailbreaking large language models against moderation guardrails via cipher characters. Preprint, arXiv:2405.20413.  
+Erik Jones, Anca Dragan, Aditi Raghunathan, and Jacob Steinhardt. 2023. Automatically auditing large language models via discrete optimization. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 15307-15329. PMLR.  
+Priyanshu Kumar, Devansh Jain, Akhila Yerukola, Liwei Jiang, Himanshu Beniwal, Thomas Hartvigsen, and Maarten Sap. 2025. Polyguard: A multilingual safety moderation tool for 17 languages. Preprint, arXiv:2504.04377.  
+Raz Lapid, Ron Langberg, and Moshe Sipper. 2024. Open sesame! universal black box jailbreaking of large language models. Preprint, arXiv:2309.01446.  
+Jan Leike, David Krueger, Tom Everitt, Miljan Martic, Vishal Maini, and Shane Legg. 2018. Scalable agent alignment via reward modeling: a research direction. *Preprint*, arXiv:1811.07871.  
+Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. 2024a. SALAD-bench: A hierarchical and comprehensive safety benchmark for large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 3923-3954, Bangkok, Thailand. Association for Computational Linguistics.  
+Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. 2024b. Red teaming visual language models. Preprint, arXiv:2401.12915.
+
+Nathaniel Li, Ziwen Han, Ian Steneker, Willow Primack, Riley Goodside, Hugh Zhang, Zifan Wang, Cristina Menghini, and Summer Yue. 2024c. Llm defenses are not robust to multi-turn human jailbreaks yet.  
+Xiaoxia Li, Siyuan Liang, Jiyi Zhang, Han Fang, Aishan Liu, and Ee-Chien Chang. 2024d. Semantic mirror jailbreak: Genetic algorithm based jailbreak prompts against open-source llms. Preprint, arXiv:2402.14872.  
+Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and Ji-Rong Wen. 2024e. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In ECCV 2024, page 174-189, Berlin, Heidelberg. Springer-Verlag.  
+Xiaogeng Liu, Nan Xu, Muhao Chen, and Chaowei Xiao. 2024a. AutoDAN: Generating stealthy jailbreak prompts on aligned large language models. In The Twelfth International Conference on Learning Representations.  
+Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. 2024b. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In ECCV 2024, page 386-403, Berlin, Heidelberg. Springer-Verlag.  
+Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. 2024. Jailbreakv-28k: A benchmark for assessing the robustness of multimodal large language models against jailbreak attacks. Preprint, arXiv:2404.03027.  
+Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, David Forsyth, and Dan Hendrycks. 2024. Harmbench: A standardized evaluation framework for automated red teaming and robust refusal.  
+Anay Mehrotra, Manolis Zampetakis, Paul Kassianik, Blaine Nelson, Hyrum Anderson, Yaron Singer, and Amin Karbasi. 2024. Tree of attacks: Jailbreaking black-box llms automatically. Preprint, arXiv:2312.02119.  
+OpenAI. 2025. Openai moderation endpoint guide. Accessed: 2025-05-18.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022a. Training language models to follow instructions with human feedback. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA. Curran Associates Inc.  
+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. 2022b. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, volume 35, pages 27730-27744. Curran Associates, Inc.  
+Raghuveer Peri, Sai Muralidhar Jayanthi, Srikanth Ronanki, Anshu Bhatia, Karel Mundnich, Saket Dingliwal, Nilaksh Das, Zejiang Hou, Goeric Huybrechts, Srikanth Vishnubhotla, Daniel Garcia-Romero, Sundararajan Srinivasan,
+
+Kyu Han, and Katrin Kirchhoff. 2024. Speechguard: Exploring the adversarial robustness of multi-modal large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 10018-10035, Bangkok, Thailand. Association for Computational Linguistics.  
+Fabio Pernisi, Dirk Hovy, and Paul Röttger. 2024. Compro-messo! Italian many-shot jailbreaks undermine the safety of large language models.  
+Qibing Ren, Chang Gao, Jing Shao, Junchi Yan, Xin Tan, Wai Lam, and Lizhuang Ma. 2024. CodeAttack: Revealing safety generalization challenges of large language models via code completion. In *Findings of the Association for Computational Linguistics* ACL 2024, pages 11437-11452. Association for Computational Linguistics.  
+Mark Russinovich, Ahmed Salem, and Ronen Eldan. 2024. Great, now write an article about that: The crescendo multiturn llm jailbreak attack. Preprint, arXiv:2404.01833.  
+Guobin Shen, Dongcheng Zhao, Yiting Dong, Xiang He, and Yi Zeng. 2025. Jailbreak antidote: Runtime safety-utility balance via sparse representation adjustment in large language models. In The Thirteenth International Conference on Learning Representations.  
+Xinyue Shen, Zeyuan Chen, Michael Backes, Yun Shen, and Yang Zhang. 2024a. "do anything now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, CCS '24, page 1671-1685, New York, NY, USA. Association for Computing Machinery.  
+Xinyue Shen, Zeyuan Chen, Michael Backes, Yun Shen, and Yang Zhang. 2024b. "do anything now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models. Preprint, arXiv:2308.03825.  
+Chandan Singh, Armin Askari, Rich Caruana, and Jianfeng Gao. 2023. Augmenting interpretable models with large language models during training. Nature Communications, 14(1):7913.  
+Sugan Sirihattasak, Mamoru Komachi, and Hiroshi Ishikawa. 2018. Annotation and classification of toxicity for thai twitter. In Proceedings of LREC 2018 Workshop and the 2nd Workshop on Text Analytics for Cybersecurity and Online Safety (TA-COS'18), Miyazaki, Japan.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics.  
+Jiayang Song, Yuheng Huang, Zehhua Zhou, and Lei Ma. 2024. Multilingual blending: Llm safety alignment evaluation with language mixture. Preprint, arXiv:2407.07342.  
+NLLB Team, Marta R. Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2022. No language left behind: Scaling human-centered machine translation. Preprint, arXiv:2207.04672.
+
+Bibek Upadhayay and Vahid Behzadan. 2024. Sandwich attack: Multi-language mixture adaptive attack on llms. Preprint, arXiv:2404.07242.  
+Bertie Vidgen, Nino Scherrer, Hannah Rose Kirk, Rebecca Qian, Anand Kannappan, Scott A. Hale, and Paul Röttger. 2024. Simplesafetytests: a test suite for identifying critical safety risks in large language models. Preprint, arXiv:2311.08370.  
+Jiongxiao Wang, Zichen Liu, Keun Hee Park, Zhuojun Jiang, Zhaoheng Zheng, Zhuofeng Wu, Muhao Chen, and Chaowei Xiao. 2023. Adversarial demonstration attacks on large language models. Preprint, arXiv:2305.14950.  
+Wenxuan Wang, Zhaopeng Tu, Chang Chen, Youliang Yuan, Jen tse Huang, Wenxiang Jiao, and Michael R. Lyu. 2024a. All languages matter: On the multilingual safety of large language models.  
+Yu Wang, Xiaofei Zhou, Yichen Wang, Geyuan Zhang, and Tianxing He. 2024b. Jailbreak large vision-language models through multi-modal linkage. Preprint, arXiv:2412.00473.  
+Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. 2023. Jailbroken: How does LLM safety training fail? In Thirty-seventh Conference on Neural Information Processing Systems.  
+Zeming Wei, Yifei Wang, Ang Li, Yichuan Mo, and Yisen Wang. 2024. Jailbreak and guard aligned language models with only few in-context demonstrations. Preprint, arXiv:2310.06387.  
+Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. 2024. Do llamas work in english? on the latent language of multilingual transformers.  
+Zhaofeng Wu, Xinyan Velocity Yu, Dani Yogatama, Jiasen Lu, and Yoon Kim. 2024. The semantic hub hypothesis: Language models share semantic representations across languages and modalities. Preprint, arXiv:2411.04986.  
+Hao Yang, Lizhen Qu, Ehsan Shareghi, and Gholamreza Haffari. 2025. Audio is the achilles' heel: Red teaming audio large multimodal models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, pages 9292-9306, Albuquerque, New Mexico.  
+Yahan Yang, Soham Dan, Dan Roth, and Insup Lee. 2024. Benchmarking llm guardrails in handling multilingual toxicity. Preprint, arXiv:2410.22153.  
+Zonghao Ying, Aishan Liu, Siyuan Liang, Lei Huang, Jinyang Guo, Wenbo Zhou, Xianglong Liu, and Dacheng Tao. 2024. Safebench: A safety evaluation framework for multimodal large language models. Preprint, arXiv:2410.18927.  
+Zheng-Xin Yong, Cristina Menghini, and Stephen H. Bach. 2024. Low-resource languages jailbreak gpt-4. Preprint, arXiv:2310.02446.  
+Haneul Yoo, Yongjin Yang, and Hwaran Lee. 2024. Csrt: Evaluation and analysis of llms using code-switching red-teaming dataset. Preprint, arXiv:2406.15481.  
+Jiahao Yu, Xingwei Lin, Zheng Yu, and Xinyu Xing. 2024a. Gptfuzzer: Red teaming large language models with autogenerated jailbreak prompts. Preprint, arXiv:2309.10253.
+
+Weihao Yu, Zhengyuan Yang, Lingfeng Ren, Linjie Li, Jianfeng Wang, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Lijuan Wang, and Xinchao Wang. 2024b. Mm-vet v2: A challenging benchmark to evaluate large multimodal models for integrated capabilities. Preprint, arXiv:2408.00765.  
+Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Pinjia He, Shuming Shi, and Zhaopeng Tu. 2024a. GPT-4 is too smart to be safe: Stealthy chat with LLMs via cipher. In The Twelfth International Conference on Learning Representations.  
+Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Jiahao Xu, Tian Liang, Pinjia He, and Zhaopeng Tu. 2024b. Refuse whenever you feel unsafe: Improving safety in llms via decoupled refusal training.  
+Weixiang Zhao, Jiahe Guo, Yulin Hu, Yang Deng, An Zhang, Xingyu Sui, Xinyang Han, Yanyan Zhao, Bing Qin, TatSeng Chua, and Ting Liu. 2025. Adastear: Your aligned llm is inherently an adaptive jailbreak defender.  
+Yiran Zhao, Wenxuan Zhang, Guizhen Chen, Kenji Kawaguchi, and Lidong Bing. 2024. How do large language models handle multilingualism?  
+Sicheng Zhu, Ruiyi Zhang, Bang An, Gang Wu, Joe Barrow, Zichao Wang, Furong Huang, Ani Nenkova, and Tong Sun. 2023. Autodan: Interpretable gradient-based adversarial attacks on large language models. Preprint, arXiv:2310.15140.  
+Yufan Zhuang, Chandan Singh, Liyuan Liu, Jingbo Shang, and Jianfeng Gao. 2025. Vector-icl: In-context learning with continuous vector representations. In The Thirteenth International Conference on Learning Representations.  
+Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, and Matt Fredrikson. 2023a. Universal and transferable adversarial attacks on aligned language models. Preprint, arXiv:2307.15043.  
+Andy Zou, Zifan Wang, J. Zico Kolter, and Matt Fredrikson. 2023b. Universal and transferable adversarial attacks on aligned language models. Preprint, arXiv:2307.15043.
+
+# A Languages Used in Our Approach
+
+We use the following languages in our experiments:
+
+1. Natural Languages: English, French, German, Spanish, Persian, Arabic, Croatian, Japanese, Polish, Russian, Swedish, Thai, Hindi, Italian, Korean, Bengali, Portuguese, Chinese, Hebrew, Serbian, Danish, Turkish, Greek, Indonesian, Zulu, Hungarian, Basque, Swahili, Afrikaans, Bosnian, Lao, Romanian, Slovenian, Ukrainian, Finnish, Malay, Javanese, Welsh, Bulgarian, Armenian, Icelandic, Vietnamese, Sinhalese, Maori, Gujarati, Kannada, Marathi, Tamil, Telugu, Amharic, Norwegian, Czech, Dutch.  
+2. Cipher Languages: Caesar1, Caesar2, Caesar3, Caesar4, Caesar5, Caesar6, Caesar7, Caesarneg1, Caesarneg2, Caesarneg3, Caesarneg4, Caesarneg5, Caesarneg6, Caesarneg7, Ascii, Hexadecimal, Base64, Leet, Vowel, Alphanumeric. A number in front of Caesar cipher means that the English alphabets were shifted by that much forward and a number in front of Caesarneg cipher means that the English alphabets were shifted by that much backward.
+
+Out of these languages, we use the following for training our classifier: Arabic, Chinese, Czech, Dutch, English, French, German, Hindi, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Bosnian, Turkish, Finnish, Indonesian, Bengali, Swahili, Vietnamese, Tamil, Telugu, Greek, Maori, Javanese, Caesar1, Caesar2, Caesar4, Caesarneg2, Caesarneg4, Caesarneg6, Ascii, Hexadecimal
+
+And these for testing: Persian, Croatian, Hebrew, Serbian, Danish, Zulu, Hungarian, Basque, Afrikaans, Lao, Romanian, Slovenian, Ukrainian, Malay, Welsh, Bulgarian, Armenian, Icelandic, Sinhalese, Gujarati, Kannada, Marathi, Amharic, Norwegian, Caesar, Caesar5, Caesar7, Caesarneg3, Caesarneg1, Caesar6, Caesarneg7, Caesarneg5, Base64, Alphanumeric, Vowel, LeetSpeak.
+
+# B Datasets and models
+
+# C Experimental Details of Filtering of Wikitext and its Translation
+
+# D OMNIGUARD's Performance With Different Base LLMs
+
+<table><tr><td></td><td>High-Res</td><td>Low-Res</td><td>Cipher</td></tr><tr><td>LlamaGuard 1</td><td>69.92</td><td>41.25</td><td>16.07</td></tr><tr><td>LlamaGuard 2</td><td>75.28</td><td>62.2</td><td>16.2</td></tr><tr><td>LlamaGuard 3</td><td>82.23</td><td>75.84</td><td>24.74</td></tr><tr><td>AegisGuard (P)</td><td>83.36</td><td>59.06</td><td>44.22</td></tr><tr><td>AegisGuard (D)</td><td>88.14</td><td>76.26</td><td>83.21</td></tr><tr><td>WildGuard</td><td>81.35</td><td>43.51</td><td>16.53</td></tr><tr><td>HarmBench (llama)</td><td>14.25</td><td>14.08</td><td>14.11</td></tr><tr><td>HarmBench (mistral)</td><td>17.6</td><td>15.9</td><td>14.46</td></tr><tr><td>MD-Judge</td><td>59.51</td><td>30.21</td><td>15.44</td></tr><tr><td>DuoGuard</td><td>71.4</td><td>46.26</td><td>15.77</td></tr><tr><td>PolyGuard</td><td>94.47</td><td>79.22</td><td>21.28</td></tr><tr><td>OMNIGUARD</td><td>88.25</td><td>85.56</td><td>73.06</td></tr></table>
+
+Table B1: Accuracy of detecting harmful prompts stratified by high-resource natural, low-resource natural, and cipher languages.
+
+<table><tr><td></td><td>Thai Tox</td><td>Ukr Tox</td><td>TJS</td><td>Toxtext</td><td>Avg.</td></tr><tr><td>Layer 10</td><td>62.1</td><td>65.5</td><td>66.95</td><td>61.89</td><td>64.42</td></tr><tr><td>Layer 55</td><td>67.4</td><td>73.0</td><td>76.91</td><td>74.96</td><td>73.06</td></tr><tr><td>Layer 56</td><td>66.8</td><td>71.5</td><td>74.54</td><td>71.82</td><td>71.17</td></tr><tr><td>Selected layer 57</td><td>68.7</td><td>73.1</td><td>76.8</td><td>75.07</td><td>73.40</td></tr><tr><td>Layer 58</td><td>66.6</td><td>73.2</td><td>74.92</td><td>72.63</td><td>71.84</td></tr><tr><td>Layer 59</td><td>67.3</td><td>73.3</td><td>76.44</td><td>74.22</td><td>72.82</td></tr><tr><td>Layer 60</td><td>67.5</td><td>72.6</td><td>76.43</td><td>74.46</td><td>72.75</td></tr><tr><td>Layer 61</td><td>67.8</td><td>70.9</td><td>74.77</td><td>72.76</td><td>71.56</td></tr><tr><td>Layer 62</td><td>66.1</td><td>72.3</td><td>74.78</td><td>72.83</td><td>71.50</td></tr><tr><td>Layer 75</td><td>65.2</td><td>66.4</td><td>70.72</td><td>65.79</td><td>68.26</td></tr><tr><td>Last Layer</td><td>63.1</td><td>51.2</td><td>61.33</td><td>56.76</td><td>59.05</td></tr></table>
+
+Table B2: OMNIGUARD's accuracy of detecting harmful prompts when trained using representations from different model layers.
+
+<table><tr><td>Models</td><td>HarmBench</td><td>FQ</td><td>SimpleST</td><td>SaladB</td><td>TJS</td><td>ToxText</td><td>AdvBench</td><td>Avg.</td></tr><tr><td>Llama3.1-8B-Instruct</td><td>81.16</td><td>88.72</td><td>88.78</td><td>87.35</td><td>71.32</td><td>69.20</td><td>88.62</td><td>82.16</td></tr><tr><td>Gemma-3-4B-Instruct</td><td>78.03</td><td>88.23</td><td>93.86</td><td>82.09</td><td>53.50</td><td>51.45</td><td>89.07</td><td>76.60</td></tr><tr><td>Qwen-3-4B-Instruct</td><td>82.06</td><td>86.25</td><td>80.99</td><td>82.15</td><td>58.83</td><td>57.70</td><td>84.80</td><td>76.11</td></tr><tr><td>Olmo-2-7B-Instruct</td><td>80.57</td><td>88.86</td><td>92.14</td><td>88.68</td><td>70.65</td><td>64.79</td><td>88.24</td><td>81.99</td></tr><tr><td>Mistral-8B-Instruct</td><td>84.27</td><td>87.23</td><td>89.82</td><td>88.08</td><td>72.49</td><td>68.71</td><td>89.07</td><td>82.81</td></tr><tr><td>Llama3.3-70B-Instruct</td><td>89.13</td><td>89.57</td><td>89.62</td><td>87.30</td><td>76.68</td><td>75.07</td><td>86.59</td><td>84.85</td></tr></table>
+
+Table B3: OMNIGUARD's accuracy of detecting harmful prompts when paired with different underlying LLMs across multiple benchmarks.

2025/MULTIGUARD_ An Efficient Approach for AI Safety Moderation Across Languages and Modalities/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c5e81ecaeb4bcfad4f52051bfd0998e398b4edabde4925f82c2177effdc8b1f
+size 880124

2025/MULTIVOX_ A Benchmark for Evaluating Voice Assistants for Multimodal Interactions/fd3c0a4a-5db1-4e4e-9097-8d01504aa0be_content_list.json

@@ -0,0 +1,2243 @@
+[
+    {
+        "type": "text",
+        "text": "MULTIVOX: A Benchmark for Evaluating Voice Assistants for Multimodal Interactions",
+        "text_level": 1,
+        "bbox": [
+            194,
+            90,
+            805,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ramaneswaran Selvakumar*, Ashish Seth*, Nishit Anand, Utkarsh Tyagi, Sonal Kumar, Sreyan Ghosh, Dinesh Manocha",
+        "bbox": [
+            228,
+            152,
+            766,
+            185
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Maryland, College Park",
+        "bbox": [
+            344,
+            186,
+            652,
+            202
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{ramans, aseth125, dmanocha}@umd.edu",
+        "bbox": [
+            317,
+            203,
+            680,
+            218
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The rapid progress of Large Language Models (LLMs) has empowered omni models to act as voice assistants capable of understanding spoken dialogues. These models can process multimodal inputs beyond text, such as speech and visual data, enabling more context-aware interactions. However, current benchmarks fall short in comprehensively evaluating how well these models generate context-aware responses, particularly when it comes to implicitly understanding fine-grained speech characteristics, such as pitch, emotion, timbre, and volume or the environmental acoustic context such as background sounds. Additionally, they inadequately assess the ability of models to align paralinguistic cues with complementary visual signals to inform their responses. To address these gaps, we introduce MULTIVox, the first omni voice assistant benchmark designed to evaluate the ability of voice assistants to integrate spoken and visual cues including paralinguistic speech features for truly multimodal understanding. Specifically, MULTIVox includes 1000 human-annotated and recorded speech dialogues that encompass diverse paralinguistic features and a range of visual cues such as images and videos. Our evaluation on 10 state-of-the-art models reveals that, although humans excel at these tasks, current open-source models consistently struggle to produce contextually grounded responses. $^{1}$",
+        "bbox": [
+            141,
+            282,
+            460,
+            724
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            731,
+            258,
+            745
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "With recent advancements in Multimodal Large Language Models (MLLMs) (Xu et al., 2025; Microsoft et al., 2025), there is a growing interest in developing models that can understand and generate information across multiple modalities, such as images, video, and audio-simultaneously. This evolution is paving the way for the development of Omni Language Models (OLMs), which are crucial for building efficient and versatile Artificial",
+        "bbox": [
+            112,
+            755,
+            489,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1https://github.com/ramaneswaran/multivox",
+        "bbox": [
+            134,
+            906,
+            447,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/bec8965f0dfe9fb92a090577c3bd84669b05cf836b7157cca4f67f1871ce6796.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            261,
+            873,
+            362
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/9d62ab35f7747b122502b6feb2dab4dc9a909ff13e3b722b551c79ccb87e9912.jpg",
+        "image_caption": [
+            "Figure 1: Comparison of existing benchmark with MULTIVOX. Existing benchmark for omni-modal voice assistant evaluation are derived from vision-centric text VQA benchmarks. In MULTIVOX, models need leverage not only visual cues but also non-verbal speech signals."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            363,
+            873,
+            558
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "General Intelligence (AGI) (Bubeck et al., 2023; Morris et al., 2024). While OLMs provide a wide range of applications (Xu et al., 2025), one of their primary use cases is developing omni-modal voice assistants (OVA) (Huang et al., 2024). Unlike traditional speech voice assistants that rely solely on speech instruction, OVAs powered by OLMs such as GPT-4o (OpenAI, 2024) and Qwen2.5 Omni (Xu et al., 2025), can understand speech dialogues and reason over multimodal inputs, including images and videos.",
+        "bbox": [
+            507,
+            658,
+            884,
+            834
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Advancing the application of OLMs in voice assistants poses challenges not only in model development but also in constructing effective evaluation benchmarks. While existing OLM benchmarks like OmniBench (Li et al., 2024) incorpo",
+        "bbox": [
+            507,
+            841,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "28470",
+        "bbox": [
+            473,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 28470-28482 November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            152,
+            945,
+            843,
+            973
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "rate multimodal inputs such as images and video, they lack spoken dialogues—an essential modality for assessing the conversational and auditory capabilities required of voice assistants. On the other hand, current voice assistant benchmarks such as VoXDialogue (Cheng et al., 2025), SD-Eval (Ao et al., 2025), and S2S-Arena (Jiang et al., 2025) focus primarily on evaluating a model's ability to generate contextually appropriate responses based on auditory cues like content, emotion, or speaker demographics embedded in a speech instruction. However, these benchmarks fall short of capturing the full multimodal reasoning abilities expected of OVAs, particularly in integrating visual cues alongside speech instructions.",
+        "bbox": [
+            112,
+            84,
+            492,
+            324
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address this gap, we introduce MULTIVOX, a novel benchmark designed to evaluate an OLMs ability to incorporate multimodal cues to provide accurate and contextual responses. MULTIVOX includes 1000 questions consisting of human spoken questions paired with either a video or an image. Unlike existing benchmarks which primarily test visual grounding and use speech to deliver the content of a straightforward instruction, MULTIVOX consists of questions which require a model to combine visions skills such as object recognition, scene understanding, scene text understanding with speech skills such as acoustic scene understanding, paralanguage understanding and speaker profiling (See fig. 1). The spoken questions in MULTIVOX are recorded by professional voice actors to cover a diverse range of paralinguistic and emotional features that are not possible with current text-to-speech systems. A key problem in benchmarks that evaluate multi-modal reasoning capability is that models can take shortcuts by exploiting priors from other modalities, to mitigate this we introduce confounding samples in MULTIVOX. Specifically, each question in our benchmark has another associated question which has the same textual and visual content but their speech property is flipped such that expected answers should be different. Our key contributions are:",
+        "bbox": [
+            115,
+            326,
+            490,
+            774
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We present MULTIVOX, the first benchmark designed to evaluate omni-modal language models (OLMs) using human-spoken queries paired with visual inputs. The 1000 examples require models to jointly ground visual and paralinguistic speech cues to produce accurate, context-aware responses.",
+            "2. We evaluate 10 omni-modal models on MUL-"
+        ],
+        "bbox": [
+            127,
+            784,
+            489,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "TIVOX and find that, while humans excel with ease, current OLMs consistently struggle, particularly with grounding speech signals, revealing a critical bottleneck in their capabilities.",
+        "bbox": [
+            544,
+            84,
+            884,
+            162
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. We perform extensive qualitative and quantitative analysis on model's responses and uncover key insights: Models exhibit strong visual grounding but rely heavily on textual cues for speech-related tasks; they often ignore non-verbal audio signals like tone or background sounds.",
+        "bbox": [
+            524,
+            177,
+            885,
+            288
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            315,
+            665,
+            330
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Omni Voice Assistants The recent development of Omni Language Models (OLMs) has enabled development of omni-modal voice assistants that can simultaneously infer across both visual and speech inputs. Recent iterations of previously mentioned voice assistants now support visual inputs in form of either images like in the case of MiniOmni2 (Xie and Wu, 2024) and MoshiVis (Royer et al., 2025) or video inputs such as Qwen2.5-Omni (Xu et al., 2025). While these models demonstrate impressive instruction-following capabilities as voice assistants, when extensively evaluated on MULTIVOX, we find that they often overlook crucial paralinguistic cues—such as tone, emotion, and pitch—in speech input, which are essential for generating context-aware responses.",
+        "bbox": [
+            507,
+            341,
+            885,
+            598
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Benchmarks For Voice Assistants While there are works such as OmniBench (Li et al., 2024) and OmniXR (Chen et al., 2024a) that evaluate OLMs, there have been few efforts to standardize the evaluation of omni voice assistants. Recent work such as Lyra (Zhong et al., 2024) repurpose existing visual question answering (VQA) benchmarks by converting the textual questions to speech. However, such approach overlooks crucial non-verbal information typically present in spoken conversations. While some progress has been made in evaluating speech VAs, many existing benchmarks still fall short. VoiceBench (Chen et al., 2024b) assesses capabilities like world knowledge and instruction following by converting textual benchmarks like MMLU (Hendrycks et al., 2021) and AlpacaEval (Dubois et al., 2024) to speech, but overlooks the non-verbal speech information. SD-Eval (Ao et al., 2025) is a pioneering work in evaluating paralinguistic features but is limited to only four",
+        "bbox": [
+            507,
+            599,
+            885,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "28471",
+        "bbox": [
+            475,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4996194c6be72c13c7638fe9c5513567b0af8eff14005b9766c931fcd72a536d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            83,
+            884,
+            262
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a4eb9aa53f7d6c5c69283de0604f4e1e930ba105d686a54524de1183774fad55.jpg",
+        "image_caption": [
+            "Figure 2: Illustration of various types of questions in MULTIVox. We broadly define three categories of speech-understanding skills that a voice assistant needs in order to provide an accurate and contextual response. Each question in MULTIVox has a speech confounder, where the textual question remains same but the speech property is flipped to counter the possibility of models exploiting unimodal priors"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            268,
+            878,
+            448
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "categories with a narrow range of labels in each category. While VoxDialogue (Cheng et al., 2025) covers more diverse speech attributes, it relies heavily on synthetic speech generated through TTS systems that struggle to accurately convey emotional nuances and prosodic variations present in natural human speech.",
+        "bbox": [
+            112,
+            536,
+            489,
+            650
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 MULTIVox",
+        "text_level": 1,
+        "bbox": [
+            112,
+            663,
+            253,
+            680
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We introduce MULTIVOX, a novel benchmark for evaluating omni-modal language models on their ability to jointly interpret speech and visual inputs, and integrate them with world knowledge and reasoning to produce contextually appropriate responses. This section outlines the benchmark's design goals, construction process, and key dataset statistics.",
+        "bbox": [
+            112,
+            690,
+            489,
+            819
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Benchmark Design",
+        "text_level": 1,
+        "bbox": [
+            112,
+            834,
+            312,
+            850
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Motivation Real-world communication is inherently multimodal, combining what is said with how, by whom, and in what environment. To achieve human-like understanding, an omni-modal voice",
+        "bbox": [
+            112,
+            856,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "assistant\nassistant\nassistant\nassistant\nassistant\nassistant\nassistant\nassistant\nassistant\nassistant\nassistant\nassistant\nassistant",
+        "bbox": [
+            507,
+            536,
+            884,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "While recent progress has advanced visual understanding, speech remains a key bottleneck. Most benchmarks treat spoken input as text, overlooking critical non-verbal signals. Yet in real interactions, these cues determine how instructions are interpreted. For example, a user asking \"Am I being too loud?\" in a library depends on speech volume, not words alone.",
+        "bbox": [
+            507,
+            604,
+            884,
+            732
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To address the limitations of current benchmarks, MULTIVOX evaluates OLMs across three core skill domains—with a particular emphasis on speech grounding, which is an underdeveloped area in existing omni-modal benchmarks.",
+        "bbox": [
+            507,
+            736,
+            884,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Speech Skills The ability to infer non-verbal attributes from speech, such as emotion, background sounds, speaker age, or tone—beyond what is conveyed in textual content.",
+        "bbox": [
+            507,
+            820,
+            884,
+            884
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Vision Skills The ability to recognize and interpret visual elements in images or videos, including ob",
+        "bbox": [
+            507,
+            889,
+            884,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "28472",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/47fd63395574134993f4596b1b7dd43285716b670474ccd01068f7308553c2d3.jpg",
+        "image_caption": [
+            "Figure 3: Data Curation And Annotation Pipeline Of MULTIVOX. The pipeline begins with task design and a taxonomy of speech traits, followed by expert annotation using multimodal sources. Professionally recorded prompts undergo quality assessment, with final review guiding benchmark construction."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            82,
+            867,
+            206
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ject identity, scene understanding, reading scenetext etc.",
+        "bbox": [
+            112,
+            284,
+            485,
+            313
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "General Skills The ability to integrate visual and auditory inputs with world knowledge and common sense to make contextually appropriate decisions.",
+        "bbox": [
+            112,
+            317,
+            485,
+            366
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Confounder Pairs To test whether models genuinely ground non-verbal speech attributes—rather than relying on linguistic or visual shortcuts—we introduce confounder pairs. Each pair consists of two instances that share identical textual content and visual input, but differ in a targeted speech property (e.g., tone, emotion, background sound). Crucially, this difference is sufficient to invert the expected answer.",
+        "bbox": [
+            112,
+            367,
+            485,
+            511
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For example, a user asking \"Am I being too loud?\" in a library setting may expect reassurance when spoken softly, but concern or correction when spoken in a normal tone. By controlling for all other modalities, confounder pairs isolate the model's ability to interpret non-verbal speech signals.",
+        "bbox": [
+            112,
+            513,
+            487,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Benchmark Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            640,
+            361,
+            655
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Task Definition Our benchmark is organized around a three-level taxonomy of speech-related categories. The top level defines three domains: acoustic scene, speaker profile, and paralanguage. We then worked with eight expert annotators (graduate students in speech processing) to expand each domain into:",
+        "bbox": [
+            112,
+            663,
+            487,
+            774
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Categories: High-level categories of non-verbal speech attributes (e.g., emotion, speaker age, ambient environment)",
+            "- Sub-categories: Fine-grained, measurable skills within each group which can be specified as a measurable task (e.g., detect emotion, estimate age as elderly or young adult)."
+        ],
+        "bbox": [
+            136,
+            793,
+            487,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each task, experts authored a specification card detailing the speech attribute being evaluated. These cards included a clear definition of the target attribute, illustrative scenarios combining speech and vision, and structured guidelines for how annotators should construct samples and determine correct answers.",
+        "bbox": [
+            507,
+            284,
+            884,
+            395
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Expert Annotation Each expert was assigned a subset of tasks and followed a standardized pipeline to create benchmark samples. For each task, annotators first constructed a realistic scenario in which an OVA must correctly interpret the target speech attribute. They also created a corresponding confounder scenario, identical in text and visual input but differing in the relevant speech property, to ensure robust grounding.",
+        "bbox": [
+            507,
+            399,
+            882,
+            543
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To pair each question with appropriate visual content, annotators were given access to stock media libraries, prioritizing real videos. If no suitable match was found, AI-generated videos were used, though these were often insufficient for scenes involving text or fine-grained detail. In such cases, high-quality images—either retrieved or generated—were used instead. For tasks involving ambient audio (e.g., acoustic scenes), annotators selected relevant background sounds or music from curated sound libraries. In the end we collect 206 unique images and 287 unique videos.",
+        "bbox": [
+            507,
+            546,
+            882,
+            739
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Each finalized sample includes a textual query, accompanying visual input (image/video), and any required background audio. For paralanguage-based tasks, annotators also specified detailed voice delivery guidelines to guide later voice recording.",
+        "bbox": [
+            507,
+            741,
+            882,
+            822
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Reference Answers Annotators also authored reference answers and rationales for each sample, describing the expected model behavior and explaining how the relevant speech (and visual) cues should inform the response. These were later used for evaluation.",
+        "bbox": [
+            507,
+            825,
+            882,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "28473",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Voice Recording We employed professional voice actors to record the spoken queries, providing delivery guidelines based on the target speech property. To preserve authenticity, actors were given creative freedom in expression as long as the intended cue was conveyed. Due to ethical concerns we use TTS systems for children voice. A professional audio engineer handled background sound overlays where applicable.",
+        "bbox": [
+            110,
+            84,
+            487,
+            227
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Quality Control Two annotators independently verified whether the intended speech attribute was clearly conveyed in each recording. Recordings without unanimous approval were revised based on feedback—either through re-recording or by adjusting the script to better support the desired delivery.",
+        "bbox": [
+            110,
+            229,
+            485,
+            340
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Expert Analysis In the final review stage, annotators assessed each completed sample for overall quality. They were instructed to verify that (1) the scenario was realistic, (2) the sample minimized reliance on language or vision priors, and (3) the recorded speech clearly conveyed the intended attribute.",
+        "bbox": [
+            112,
+            342,
+            487,
+            454
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Evaluation Criteria",
+        "text_level": 1,
+        "bbox": [
+            112,
+            464,
+            314,
+            479
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The goal of MULTIVOX is to assess how well omnimodal language models integrate speech and vision to produce contextually grounded responses. To enable fine-grained diagnosis, we adopt a modular evaluation framework that tests both multimodal integration and unimodal grounding. Each benchmark sample is designed to probe one or more of the following components:",
+        "bbox": [
+            110,
+            486,
+            487,
+            615
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Speech Grounding Evaluates whether the OLM correctly interprets paralinguistic cues such as emotion or ambient sound. Each question contains a speech hook, a non-verbal attribute that is critical for answering correctly. These tasks help isolate model sensitivity to speech signals beyond text.",
+        "bbox": [
+            112,
+            615,
+            487,
+            711
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Visual Grounding Evaluates whether the OLM can interpret and incorporate key visual cues. Each sample includes a visual hook—a necessary visual detail (e.g., object, background element) that the model must recognize to respond appropriately.",
+        "bbox": [
+            112,
+            711,
+            487,
+            791
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Contextual Appropriateness We adopt appropriateness (Chen et al., 2023) as our core evaluation metric which measures how well an OLM produces a response that aligns with the intent, context, and modality of the input.",
+        "bbox": [
+            112,
+            791,
+            487,
+            873
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate contextual appropriateness using sample-specific rubrics that guide judgment based on three elements: (1) a reference answer authored",
+        "bbox": [
+            112,
+            873,
+            487,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/79a09e1625eeca97edd7d652bd62af38e52bb8bac9daadbbbe17baa786370235.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>Name</td><td>Vis.</td><td>Para.</td><td>Src.</td><td>Conf.</td></tr><tr><td>Foundation</td><td>AudioBench</td><td>X</td><td>✓</td><td>Human</td><td>X</td></tr><tr><td>Foundation</td><td>MMAU</td><td>X</td><td>✓</td><td>Mixed</td><td>X</td></tr><tr><td>Foundation</td><td>OmniBench</td><td>✓</td><td>✓</td><td>Human</td><td>X</td></tr><tr><td>Chat</td><td>SD-Eval</td><td>X</td><td>✓</td><td>Human</td><td>X</td></tr><tr><td>Chat</td><td>VoxDialog</td><td>X</td><td>✓</td><td>Synthetic</td><td>X</td></tr><tr><td>Chat</td><td>S2S-Arena</td><td>X</td><td>✓</td><td>Mixed</td><td>X</td></tr><tr><td>Chat</td><td>Lyra SVQA</td><td>✓</td><td>X</td><td>Synthetic</td><td>X</td></tr><tr><td>Chat</td><td>Ours</td><td>✓</td><td>✓</td><td>Human</td><td>✓</td></tr></table>",
+        "bbox": [
+            510,
+            82,
+            880,
+            200
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Comparison of MULTIVOX with related benchmarks.",
+        "bbox": [
+            507,
+            209,
+            882,
+            237
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "by the expert annotator, (2) a short rationale explaining what cues are necessary to arrive at the correct answer, and (3) task-level metadata specifying which modality is critical (e.g., speech hook, visual hook). These are provided to a GPT-4 judge, which scores model responses on a 1–5 scale, reflecting increasing levels of multimodal integration and contextual fidelity. To evaluate speech and vision grounding, we again utilize a LLM judge. To prevent score hacking using modality shortcuts we penalize OLMs that explicitly use text or visual content to respond to speech grounding",
+        "bbox": [
+            507,
+            266,
+            884,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Comparison With Other Benchmark",
+        "text_level": 1,
+        "bbox": [
+            507,
+            475,
+            843,
+            489
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section we highlight how MULTIVOX is different in terms of question types, modality coverage, speech source, and diagnostic power. Table 1 summarizes these differences.",
+        "bbox": [
+            507,
+            498,
+            882,
+            561
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Chat-based Questions Benchmarks such as AudioBench (Wang et al., 2025), MMAU (Sakshi et al., 2024), and OmniBench (Li et al., 2024) primarily test foundational tasks, with only the latter supporting full omni-modality. In contrast, MULTI-Vox is grounded in chat-style interaction, reflecting how OVAs are deployed in real-world use. This setting demands deeper contextual understanding and flexible reasoning.",
+        "bbox": [
+            507,
+            564,
+            882,
+            709
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Multi-Modal Inputs Benchmarks like SD-Eval (Ao et al., 2025), VoxDialog (Cheng et al., 2025), and S2S-Arena focus on speech-only inputs, targeting paralanguage or acoustic scene understanding in isolation. Others like Lyra SVQA (Zhong et al., 2024) incorporate visual input but neglect paralinguistic cues. MULTIVox is unique in requiring models to jointly interpret speech, vision, and background context, aligning with the OVA use case.",
+        "bbox": [
+            507,
+            709,
+            882,
+            871
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Human Speech Most existing chat benchmarks rely on synthetic speech, which current TTS systems struggle to render with accurate emotion",
+        "bbox": [
+            507,
+            873,
+            882,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "28474",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/eb27e7e2e82f0d2be2c7fc991ae6674389bc8c99ee77001ecd938c7f0770269a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Name</td><td rowspan=\"2\">Size</td><td colspan=\"3\">Acoustic Scene</td><td colspan=\"3\">Paralanguage</td><td colspan=\"3\">Speaker Profile</td><td rowspan=\"2\">Avg. CA</td></tr><tr><td>VG</td><td>SG</td><td>CA</td><td>VG</td><td>SG</td><td>CA</td><td>VG</td><td>SG</td><td>CA</td></tr><tr><td>Human</td><td>-</td><td>95.30</td><td>82.50</td><td>4.37</td><td>96.00</td><td>92.5</td><td>4.33</td><td>96.50</td><td>95.10</td><td>4.36</td><td>4.35</td></tr><tr><td colspan=\"12\">Open Source Models</td></tr><tr><td>Mini Omni2</td><td>7.0B</td><td>79.24</td><td>16.14</td><td>1.53</td><td>79.20</td><td>23.12</td><td>1.79</td><td>84.50</td><td>09.00</td><td>2.01</td><td>1.74</td></tr><tr><td>VITA 1.5</td><td>1.6B</td><td>78.12</td><td>16.50</td><td>2.60</td><td>81.14</td><td>34.57</td><td>2.56</td><td>88.60</td><td>14.20</td><td>3.01</td><td>2.69</td></tr><tr><td>VideoLlama2</td><td>7.0B</td><td>68.12</td><td>28.75</td><td>1.52</td><td>73.42</td><td>2.71</td><td>1.59</td><td>79.20</td><td>14.79</td><td>1.38</td><td>1.50</td></tr><tr><td>Baichuan-Omni</td><td>7.0B</td><td>76.15</td><td>34.37</td><td>1.90</td><td>77.24</td><td>32.71</td><td>2.25</td><td>84.20</td><td>16.20</td><td>2.01</td><td>2.02</td></tr><tr><td>Mini CPM</td><td>8.0B</td><td>87.62</td><td>35.35</td><td>2.87</td><td>89.14</td><td>39.28</td><td>2.35</td><td>88.40</td><td>25.40</td><td>2.90</td><td>2.69</td></tr><tr><td>Intern Omni</td><td>8.7B</td><td>80.75</td><td>20.25</td><td>2.54</td><td>80.71</td><td>14.57</td><td>1.94</td><td>82.60</td><td>06.60</td><td>2.64</td><td>2.35</td></tr><tr><td>phi4 multimodal</td><td>5.6B</td><td>81.24</td><td>23.12</td><td>2.26</td><td>79.14</td><td>33.57</td><td>2.63</td><td>84.57</td><td>12.40</td><td>2.48</td><td>2.44</td></tr><tr><td>Qwen 2.5 Omni</td><td>7.0B</td><td>84.87</td><td>15.37</td><td>3.19</td><td>89.42</td><td>38.71</td><td>2.98</td><td>91.40</td><td>11.20</td><td>3.06</td><td>3.08</td></tr><tr><td>Qwen 2.5 Omni COT</td><td>7.0B</td><td>83.50</td><td>24.50</td><td>3.27</td><td>88.28</td><td>26.71</td><td>3.00</td><td>88.80</td><td>18.00</td><td>3.33</td><td>3.19</td></tr><tr><td colspan=\"12\">Proprietary Models</td></tr><tr><td>Gemini 2.5 Flash</td><td>-</td><td>89.50</td><td>59.00</td><td>3.55</td><td>91.14</td><td>75.42</td><td>3.19</td><td>92.20</td><td>65.60</td><td>3.64</td><td>3.44</td></tr><tr><td>Gemini 2.5 Pro</td><td>-</td><td>91.25</td><td>54.75</td><td>3.65</td><td>92.00</td><td>77.42</td><td>3.32</td><td>91.60</td><td>71.60</td><td>3.74</td><td>3.56</td></tr></table>",
+        "bbox": [
+            112,
+            80,
+            880,
+            366
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Performance breakdown of human and model responses on MULTIVOX across key skill domains. Visual Grounding (VG) and Speech Grounding (SG) evaluates the ability to perceive specific information in the modality needed for answering the question. Contextual Appropriateness (CA) evaluates the ability to produce contextually appropriate and accurate answers given the multimodal cues",
+        "bbox": [
+            112,
+            376,
+            882,
+            426
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "or prosody (Wu et al., 2024; Tang et al., 2023). In MULTIVOX, all spoken queries are recorded by professional voice actors to preserve natural paralinguistic signals. To validate this, we conduct a user study across 100 paralanguage-focused samples, comparing human-recorded speech to CosyVoice (Du et al., 2024) and ElevenLabs TTS (ElevenLabs Inc., 2025). Ten annotators rated speech on (1) attribute match and (2) naturalness. Human speech scored 4.6/4.5, compared to 2.4/2.1 (CosyVoice) and 3.1/3.3 (ElevenLabs), supporting the decision to use professional recordings.",
+        "bbox": [
+            110,
+            450,
+            487,
+            643
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Confounders Many existing benchmarks are susceptible to shortcut exploitation via textual or visual priors (Kiela et al., 2021; Goyal et al., 2017). MULTIVox introduces confounder pairs, where paralinguistic speech properties are inverted while keeping textual and visual input constant. This isolates the model's ability to process non-verbal speech information and provides a more diagnostic and fine-grained evaluation framework.",
+        "bbox": [
+            112,
+            644,
+            489,
+            788
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            799,
+            260,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluated Models We evaluate a wide range of OLMs. The proprietary model assessed is Gemini2.0-flash (Developers, 2024). For open-source OLMs we use Mini-Omni2 (Xie and Wu, 2024), VideoLLama (Zhang et al., 2023), MiniCPM-o2.6 (Hu et al., 2024), Phi4-MM (Microsoft",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "et al., 2025), VITA1.5 (Fu et al., 2025), Baichuan Omni (Li et al., 2025), Intern Omni (Chen et al., 2024c).",
+        "bbox": [
+            507,
+            450,
+            884,
+            497
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Main Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            508,
+            660,
+            524
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2 summarizes the performance of several proprietary and open-source OVAs on MULTIVox. We highlight three key findings:",
+        "bbox": [
+            507,
+            533,
+            882,
+            582
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- MULTIVox is challenging. Although the tasks are straightforward for humans (average CA score: 4.33), the best-performing model (Gemini) only achieves 3.56—indicating that current OLMs struggle to integrate multimodal cues, particularly non-verbal speech signals, even in seemingly simple scenarios.",
+            "- Gap between proprietary and open-source models. Gemini 2.5 Flash and Pro models outperform open-source models in their ability to ground in multi-modal inputs and provide contextual responses.",
+            "- Speech grounding remains the bottleneck. All models show relatively strong visual grounding, but consistently struggle to interpret non-verbal speech cues such as tone, emotion, and background sounds."
+        ],
+        "bbox": [
+            531,
+            589,
+            884,
+            881
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To better understand the sources of these limitations, we conduct a detailed analysis of Gemini",
+        "bbox": [
+            507,
+            889,
+            882,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "28475",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2.5 Pro, the best-performing model on MULTIVOX. We examine its behavior across core skill domains, focusing on how well it grounds responses in visual and speech cues, and identifying the types of errors that arise.",
+        "bbox": [
+            112,
+            84,
+            487,
+            164
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Where do models fall short?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            177,
+            381,
+            192
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We analyze Gemini 2.5 Pro and Qwen 2.5 Omni performance across three skill domains: visual grounding, speech grounding, and multimodal reasoning. This breakdown highlights where current OLMs are reliable—and where they still fall short.",
+        "bbox": [
+            112,
+            199,
+            489,
+            279
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1.1 Visual Grounding: A Strength",
+        "text_level": 1,
+        "bbox": [
+            112,
+            290,
+            413,
+            306
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2 shows the vision grounding scores for Gemini and Qwen 2.5 Omni. We find that these models demonstrate consistently strong visual grounding capabilities across tasks like object detection, scene understanding, and scene text recognition. These results indicate robust understanding of both fine-grained visual elements and broader scene context. While Qwen performs competitively with Gemini, Gemini performs better in acoustic scene category, where there are several tasks that require scene text understanding. Apart from this, most errors arise from under-specificity in open-ended scenarios - for instance, recognizing a \"person holding game controllers\" without identifying attributes like \"retro game controllers\". This level of ambiguity is expected in real-world scenes and the high performance even in such conditions represents strong visual grounding capabilities.",
+        "bbox": [
+            112,
+            310,
+            489,
+            600
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1.2 Speech Grounding: A Bottleneck",
+        "text_level": 1,
+        "bbox": [
+            112,
+            611,
+            433,
+            626
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In contrast to visual grounding strengths, our analysis reveals weaknesses in speech grounding capabilities (Table 2). We analyze the final responses generated by the model to evaluate if it is able to integrate audio characteristics in its responses.",
+        "bbox": [
+            112,
+            631,
+            489,
+            709
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Models hear the voice, but doesn't \"recognizes\" the speaker profile We check the model's ability to infer demographic characteristics (age and gender) from speech cues. Analysis of Qwen's responses show that the model relies on textual rather than acoustic content to identify speaker attributes in  $30\\%$  of cases. Additionally, the model rarely commits to definitive answers  $(16.0\\%)$  of cases), instead offering ambiguous answers  $(14.6\\%)$ , expressing uncertainty  $(14.4\\%)$ , or refusing to respond  $(65.6\\%)$ . In contrast to open-source models, we observe that Gemini consistently performs much better at perceiving speaker attributes such",
+        "bbox": [
+            112,
+            711,
+            489,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "as age and gender.",
+        "bbox": [
+            509,
+            85,
+            648,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We further analyze the final responses to evaluate how well OLMs utilize this speech information in their final response. In case of Qwen, for questions requiring gender inference, the model provides neutral responses in  $60\\%$  of cases, with the remaining responses showing no significant bias toward either male or female speakers. Notably, when the model does make a gender inference, it appears to be influenced primarily by visual context and the question's content rather than speech characteristics. Gemini surprisingly shows similar trends, indicating that OLMs prefer a neutral response even if it can accurately infer their gender. Age-related inferences on the other hand reveal a stronger bias pattern, with both OLMs overwhelmingly favoring young-to-middle-aged adult in their responses regardless of the speaker's actual age. While Gemini is able to make accurate inferences regarding age, when generating final responses, this aspect is not integrated in its final responses. These errors have functional implications. For instance, these OLMs suggested potentially unsafe activities to speakers with children's voices in  $20\\%$  of the cases.",
+        "bbox": [
+            507,
+            101,
+            884,
+            469
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Models struggle to utilize background sounds In this category we test OLMs understanding of background music, sound and ambient noise. The evaluation of acoustic scene understanding reveals significant limitations, with Qwen achieving  $15\\%$  grounding score. Among these correct inferences, there is no significant difference in performance among music recognition and environmental sound recognition. Gemini performs much better at understanding background sound and music, achieving  $54\\%$  accuracy and we observe that it is also notably better at music understanding tasks. In questions with ambient noise, we observe that both OLMs demonstrate resilience to noise when processing queries. However, they still are limited in their capability to perceive noise. For instance, Qwen appears to rely predominantly on visual cues rather than acoustic features, as evidenced in approximately  $39.5\\%$  of cases where noisy environments were identified primarily through visual context (such as crowded airports).",
+        "bbox": [
+            507,
+            470,
+            884,
+            808
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Models rely on textual cues for emotion understanding While OLMs perform comparatively better in emotion understanding, we observe that they tend to rely on textual content rather than acoustic features. For instance Qwen explicitly relies on text in  $71\\%$  of cases and  $26\\%$  of the cases with Gemini. While emotions typically have a strong",
+        "bbox": [
+            507,
+            809,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "28476",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/17e1ba11acf4aeeb5dda0d446ccfb79158f0781574a2c8530104f6fd8cb1bac2.jpg",
+        "image_caption": [
+            "Figure 4: Distribution of error types in Gemini's responses on the MULTIVOX benchmark. The majority of errors (75.3%) stem from speech grounding, indicating difficulty in interpreting non-textual auditory cues."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            120,
+            455,
+            329
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "correlation with text, we deliberately introduced adversarial samples (Fig 2 where textual and acoustic emotional cues conflict. Moreover, the confounder pairs have different emotions and we observe that Qwen only gets both emotions right in  $27\\%$  cases and Gemini in  $50\\%$  of the cases. We observe similar performance trends in other paralanguage categories where there is an over-reliance on textual cues and the overall visual context.",
+        "bbox": [
+            112,
+            449,
+            489,
+            593
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitation in spoken instruction following We observe that open-source OLMs are limited in their spoken instruction following (Chen et al., 2024a), especially for instructions used in speech grounding where a grounding questions precedes the actual sample in our benchmark. Moreover, detailed explanations for answer could help in further detecting and penalizing modality shortcuts. For future work, we could consider using text modality as input to speech grounding questions for deeper analysis with instructions to explain its response.",
+        "bbox": [
+            112,
+            595,
+            489,
+            772
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 What causes these errors?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            785,
+            366,
+            800
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conducted a manual error analysis of Gemini's outputs on MULTIVOX to identify underlying failure patterns. Fig. 4 shows that perception errors dominate, accounting for  $75.3\\%$  of all failures, primarily reflecting the model's inability to ground to speech cues. Reasoning failures constitute  $7.0\\%$  of errors, indicating that even when the model suc",
+        "bbox": [
+            112,
+            808,
+            490,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/69e906fa98ee8f5cfbaef8dca9cbf5a6180581a6be71767055770a37521369f2.jpg",
+        "image_caption": [
+            "Figure 5: Distribution of model responses across confounder pairs. Bars indicate whether model answers were the same or different when speech cues were flipped. Ideally, answers should differ (left bar), but models often give identical responses (right bar), showing insensitivity to non-verbal speech cues."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            162,
+            884,
+            340
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "cessfully perceives multimodal inputs, it struggles to effectively integrate this information to generate appropriate responses. Instruction understanding failures represent a similarly significant error category, where the model defaults to describing visual content rather than addressing the intended query. The overwhelming majority of speech perception errors indicates a clear bottleneck: improving speech perception capabilities is essential for building effective OVAs in multimodal contexts.",
+        "bbox": [
+            505,
+            449,
+            884,
+            609
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3 Do models really listen to speech cues?",
+        "text_level": 1,
+        "bbox": [
+            507,
+            625,
+            858,
+            640
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To assess whether models are truly leveraging speech cues, we analyze their responses across confounder pairs, focusing on the top two performers: Gemini 2.5 Pro and Qwen2.5-Omni (Fig 5). Without considering confounders, both models appear moderately accurate, getting around  $50\\%$  of responses correct across these pairs. However, this aggregate accuracy can be misleading. When we examine whether models actually change their answers in response to flipped speech cues, we find that in a majority of the confounder pairs (57% for Gemini, 51% for Qwen), the model outputs are paraphrases, i.e., they show NO FLIP in answer. Crucially, within those NO FLIP cases, most instances with one correct answer suggest that the correctness arises from chance or visual/textual bias, not from grounding in speech. This indicates that,",
+        "bbox": [
+            505,
+            646,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "28477",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "despite non-trivial accuracy, models are largely ignoring non-verbal speech cues when answering.",
+        "bbox": [
+            112,
+            84,
+            489,
+            116
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            128,
+            247,
+            142
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We introduce MULTIVOX, the first benchmark designed to evaluate Omni Language Models (OLMs) as omni-modal voice assistants (OVAs) that integrate speech and vision for context-aware reasoning. Unlike prior benchmarks that rely on synthetic speech or focus only on unimodal cues, MULTIVOX includes 1000 professionally recorded, human-spoken questions paired with images or videos, emphasizing paralinguistic signals like tone, emotion, and background noise. A key innovation is the use of confounder pairs—speech variants with identical text and visuals—to ensure models attend to speech beyond surface cues. MULTIVOX enables fine-grained diagnosis across speech, vision, and general reasoning skills. Evaluation of 10 state-of-the-art models shows that, while visual grounding is robust, speech grounding remains a significant bottleneck. Our benchmark will be open-sourced to support the development of truly multimodal voice assistants.",
+        "bbox": [
+            115,
+            152,
+            489,
+            473
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            485,
+            220,
+            499
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We limit ourselves to questions to the english language, extending to multilingual settings is an important future direction to assess OLM generalization across languages.",
+            "- In this work, we limit our evaluation to the content of the speech outputs and not the speech quality of the output, such as naturalness and appropriateness. Evaluating speech synthesis and conversational prosody is an important but orthogonal direction left for future benchmarks."
+        ],
+        "bbox": [
+            136,
+            510,
+            487,
+            695
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            731,
+            213,
+            746
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Junyi Ao, Yuancheng Wang, Xiaohai Tian, Dekun Chen, Jun Zhang, Lu Lu, Yuxuan Wang, Haizhou Li, and Zhizheng Wu. 2025. Sd-eval: A benchmark dataset for spoken dialogue understanding beyond words. Preprint, arXiv:2406.13340.",
+            "Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4. Preprint, arXiv:2303.12712."
+        ],
+        "bbox": [
+            115,
+            753,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Bao Chen, Yuanjie Wang, Zeming Liu, and Yuhang Guo. 2023. Automatic evaluate dialogue appropriateness by using dialogue act. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 7361-7372, Singapore. Association for Computational Linguistics.",
+            "Lichang Chen, Hexiang Hu, Mingda Zhang, Yiwen Chen, Zifeng Wang, Yandong Li, Pranav Shyam, Tianyi Zhou, Heng Huang, Ming-Hsuan Yang, and Boqing Gong. 2024a. Omnixr: Evaluating omnimodality language models on reasoning across modalities. Preprint, arXiv:2410.12219.",
+            "Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T. Tan, and Haizhou Li. 2024b. VoiceBench: Benchmarking LLM-Based Voice Assistants. arXiv preprint. ArXiv:2410.17196.",
+            "Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2024c. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. Preprint, arXiv:2312.14238.",
+            "Xize Cheng, Ruofan Hu, Xiaoda Yang, Jingyu Lu, Dongjie Fu, Zehan Wang, Shengpeng Ji, Rongjie Huang, Boyang Zhang, Tao Jin, and Zhou Zhao. 2025. Voxdialogue: Can spoken dialogue systems understand information beyond words? In The Thirteenth International Conference on Learning Representations.",
+            "Google Developers. 2024. Gemini 2.0: Level up your apps with real-time multimodal interactions. Accessed: 2024-02-12.",
+            "Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, Changfeng Gao, Hui Wang, Fan Yu, Huadai Liu, Zhengyan Sheng, Yue Gu, Chong Deng, Wen Wang, Shiliang Zhang, Zhijie Yan, and Jingren Zhou. 2024. Cosyvoice 2: Scalable streaming speech synthesis with large language models. Preprint, arXiv:2412.10117.",
+            "Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B. Hashimoto. 2024. Length-controlled alpacaeval: A simple way to debias automatic evaluators. Preprint, arXiv:2404.04475.",
+            "ElevenLabs Inc. 2025. Elevenlabs: Ai voice generator and text-to-speech platform. https://elevenlabs.io/. Accessed: 2025-05-20.",
+            "Chaoyou Fu, Haojia Lin, Xiong Wang, Yi-Fan Zhang, Yunhang Shen, Xiaoyu Liu, Haoyu Cao, Zuwei Long, Heting Gao, Ke Li, Long Ma, Xiawu Zheng, Rongrong Ji, Xing Sun, Caifeng Shan, and Ran He. 2025. Vita-1.5: Towards gpt-4o level real-time vision and speech interaction. Preprint, arXiv:2501.01957.",
+            "Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in"
+        ],
+        "bbox": [
+            509,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "28478",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "vqa matter: Elevating the role of image understanding in visual question answering. Preprint, arXiv:1612.00837.",
+            "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Preprint, arXiv:2009.03300.",
+            "Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zheng Leng Thai, Kaihuo Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, and 6 others. 2024. Minicpm: Unveiling the potential of small language models with scalable training strategies. Preprint, arXiv:2404.06395.",
+            "Shaoshuai Huang, Xuandong Zhao, Dapeng Wei, Xinheng Song, and Yuanbo Sun. 2024. Chatbot and fatigued driver: Exploring the use of llm-based voice assistants for driving fatigue. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, CHI EA '24, New York, NY, USA. Association for Computing Machinery.",
+            "Feng Jiang, Zhiyu Lin, Fan Bu, Yuhao Du, Benyou Wang, and Haizhou Li. 2025. S2s-arena, evaluating speech2speech protocols on instruction following with paralinguistic information. Preprint, arXiv:2503.05085.",
+            "Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. 2021. The hateful memes challenge: Detecting hate speech in multimodal memes. Preprint, arXiv:2005.04790.",
+            "Yadong Li, Jun Liu, Tao Zhang, Tao Zhang, Song Chen, Tianpeng Li, Zehuan Li, Lijun Liu, Lingfeng Ming, Guosheng Dong, Da Pan, Chong Li, Yuanbo Fang, Dongdong Kuang, Mingrui Wang, Chenglin Zhu, Youwei Zhang, Hongyu Guo, Fengyu Zhang, and 74 others. 2025. Baichuan-omni-1.5 technical report. Preprint, arXiv:2501.15368.",
+            "Yizhi Li, Ge Zhang, Yinghao Ma, Ruibin Yuan, Kang Zhu, Hangyu Guo, Yiming Liang, Jiaheng Liu, Zekun Wang, Jian Yang, Siwei Wu, Xingwei Qu, Jinjie Shi, Xinyue Zhang, Zhenzhu Yang, Xiangzhou Wang, Zhaoxiang Zhang, Zachary Liu, Emmanouil Benetos, and 2 others. 2024. OmniBench: Towards The Future of Universal Omni-Language Models. arXiv preprint. ArXiv:2409.15272 [cs].",
+            "Microsoft, :: Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dong-dong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Dai, and 57 others. 2025. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras. Preprint, arXiv:2503.01743."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. 2024. Position: Levels of AGI for operationalizing progress on the path to AGI. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 36308-36321. PMLR.",
+            "Eliya Nachmani, Alon Levkovitch, Roy Hirsch, Julian Salazar, Chulayuth Asawaroengchai, Soroosh Mariooryad, Ehud Rivlin, RJ Skerry-Ryan, and Michelle Tadmor Ramanovich. 2024. Spoken question answering and speech continuation using spectrogram-powered llm. Preprint, arXiv:2305.15255.",
+            "OpenAI. 2024. Gpt-4o system card. Preprint, arXiv:2410.21276.",
+            "Amélie Royer, Moritz Böhle, Gabriel de Marmiesse, Laurent Mazaré, Alexandre Défossez, Neil Zeghidour, and Patrick Pérez. 2025. Vision-speech models: Teaching speech models to converse about images. ArXiv.",
+            "S Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran Selvakumar, Oriol Nieto, Ramani Duraiswami, Sreyan Ghosh, and Dinesh Manocha. 2024. Mmau: A massive multi-task audio understanding and reasoning benchmark. Preprint, arXiv:2410.19168.",
+            "Haobin Tang, Xulong Zhang, Jianzong Wang, Ning Cheng, and Jing Xiao. 2023. Emomix: Emotion mixing via diffusion models for emotional speech synthesis. In *Interspeech*.",
+            "Bin Wang, Xunlong Zou, Geyu Lin, Shuo Sun, Zhuohan Liu, Wenyu Zhang, Zhengyuan Liu, AiTi Aw, and Nancy F. Chen. 2025. Audiobench: A universal benchmark for audio large language models. Preprint, arXiv:2406.16020.",
+            "Haibin Wu, Xiaofei Wang, Sefik Emre Eskimez, Manthan Thakker, Daniel Tompkins, Chung-Hsien Tsai, Canrun Li, Zhen Xiao, Sheng Zhao, Jinyu Li, and Naoyuki Kanda. 2024. Laugh now cry later: Controlling time-varying emotional states of flow-matching-based zero-shot text-to-speech. 2024 IEEE Spoken Language Technology Workshop (SLT), pages 690-697.",
+            "Zhifei Xie and Changqiao Wu. 2024. MiniOmni2: Towards Open-source GPT-4o with Vision, Speech and Duplex Capabilities. arXiv preprint. ArXiv:2410.11190.",
+            "Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. 2025. Qwen2.5-omni technical report. Preprint, arXiv:2503.20215."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "28479",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Hang Zhang, Xin Li, and Lidong Bing. 2023. Video-llama: An instruction-tuned audio-visual language model for video understanding. Preprint, arXiv:2306.02858.",
+        "bbox": [
+            114,
+            85,
+            489,
+            137
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Zhisheng Zhong, Chengyao Wang, Yuqi Liu, Senqiao Yang, Longxiang Tang, Yuechen Zhang, Jingyao Li, Tianyuan Qu, Yanwei Li, Yukang Chen, Shaozuo Yu, Sitong Wu, Eric Lo, Shu Liu, and Jiaya Jia. 2024. Lyra: An efficient and speech-centric framework for omni-cognition. Preprint, arXiv:2412.09501.",
+        "bbox": [
+            114,
+            148,
+            487,
+            227
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            240,
+            236,
+            255
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In the Appendix, we provide:",
+        "bbox": [
+            114,
+            265,
+            332,
+            280
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Section B: Other Dataset Details",
+            "2. Section C: Annotator Details",
+            "3. Section D: LLM-as-a-judge Details",
+            "4. Section E: Voice Quality Assessment"
+        ],
+        "bbox": [
+            127,
+            291,
+            428,
+            386
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Other Dataset Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            398,
+            337,
+            413
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Here we detail the categories and sub-categories present in our benchmark.",
+        "bbox": [
+            112,
+            424,
+            485,
+            455
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B.1 Acoustic Scene Understanding",
+        "text_level": 1,
+        "bbox": [
+            114,
+            467,
+            403,
+            482
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Background Music Understanding: These tasks require the model to detect and interpret the presence and nature of background music in spoken queries. This includes genre classification, mood inference from music, and distinguishing music from speech or noise.",
+        "bbox": [
+            112,
+            487,
+            485,
+            582
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Sound Event Recognition: This task evaluates the model's ability to detect and categorize discrete, identifiable audio events (e.g., dog barking, glass breaking, door closing) that occur within the auditory scene alongside spoken content.",
+        "bbox": [
+            112,
+            585,
+            485,
+            663
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Ambient Environment Sound: Models are tested on their capacity to recognize broader acoustic environments (e.g., airport, cafe, subway) based on background audio cues. We construct these scenes using the MS-Noise (Nachmani et al., 2024) dataset, overlaying clean speech with environmental recordings at a signal-to-noise ratio (SNR) of -2 dB to simulate challenging real-world conditions.",
+        "bbox": [
+            112,
+            665,
+            485,
+            791
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B.2 Paralanguage Understanding",
+        "text_level": 1,
+        "bbox": [
+            114,
+            804,
+            394,
+            819
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Detailed distribution of the categories here are listed in Fig 7",
+        "bbox": [
+            112,
+            825,
+            485,
+            854
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "**Emotion:** The task focuses on identifying the emotional state of the speaker as conveyed through prosodic features (e.g., pitch, energy), independent of lexical content.",
+        "bbox": [
+            112,
+            857,
+            485,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Voice Modulation: Evaluates the model's sensitivity to dynamic vocal variations such as emphasis, intonation, and expressiveness, which can affect intent or meaning.",
+        "bbox": [
+            507,
+            84,
+            882,
+            148
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Pronunciation: Tasks involve recognizing deviations from standard pronunciation, which may signal emotion, emphasis, or speaker background.",
+        "bbox": [
+            507,
+            149,
+            882,
+            196
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Volume: Assesses the ability to perceive and reason about loudness cues, which can convey urgency, emotion, or social context.",
+        "bbox": [
+            507,
+            197,
+            882,
+            243
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Pace: Evaluates how well the model understands speech rate—e.g., rushed versus slow delivery—as a cue to emotional or cognitive states.",
+        "bbox": [
+            507,
+            244,
+            880,
+            292
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Stuttering: Measures recognition and interpretation of speech disfluencies, including repeated sounds or syllables, as part of speaker modeling or intent understanding.",
+        "bbox": [
+            507,
+            294,
+            882,
+            357
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Breathiness: Focuses on detecting breathy voice quality, which may indicate fatigue, emotion, or affective state.",
+        "bbox": [
+            507,
+            357,
+            880,
+            404
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B.3 Speaker Profiling",
+        "text_level": 1,
+        "bbox": [
+            509,
+            416,
+            695,
+            432
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Biological Gender Estimation: The task is to estimate the speaker's biological gender based solely on voice characteristics, controlling for content and visual input.",
+        "bbox": [
+            507,
+            437,
+            882,
+            500
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Age Group Estimation: Models must infer the age group (e.g., child, adult, elderly) of the speaker using acoustic cues.",
+        "bbox": [
+            507,
+            502,
+            880,
+            548
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Annotator Details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            561,
+            702,
+            576
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "1. Annotator Composition",
+        "text_level": 1,
+        "bbox": [
+            509,
+            586,
+            719,
+            602
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We formed a panel of six domain experts for our dataset creation and filtering process and our dataset review process. The panel consisted of four Ph.D students pursuing speech and audio-visual research and two MS students having research in speech and audio processing. The expertise of all domain experts is evidenced by their research publications and contributions to the field.",
+        "bbox": [
+            507,
+            606,
+            882,
+            734
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "2. Meetings to decide question creation process",
+        "text_level": 1,
+        "bbox": [
+            507,
+            746,
+            880,
+            762
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "All six annotators had three 2-hour online meetings to discuss the question and corresponding answers creation process to reach a consensus of the pipeline to be followed for dataset creation. The online meetings covered these details:",
+        "bbox": [
+            507,
+            766,
+            882,
+            845
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "- Multimodal Question-Answering Foundations: Aligning nonverbal audio cues (e.g., tone, background sounds) with visual context (e.g., scene imagery).",
+        "bbox": [
+            531,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "28480",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Confounder Pair Design: Generating minimal audio variants that invert answers (e.g., adding subtle noise to flip “quiet” vs. “noisy”).",
+            "- Annotation Platform & Guidelines: Hands-on use of our custom interface and details about input/output formats, edge cases, and scoring rubrics."
+        ],
+        "bbox": [
+            136,
+            84,
+            487,
+            227
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Following the online meetings, annotators jointly labeled 50 pilot samples and attained  $\\geq 90\\%$  inter-annotator agreement before proceeding to full-scale work.",
+        "bbox": [
+            112,
+            243,
+            487,
+            306
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "3. Question Creation Process",
+        "text_level": 1,
+        "bbox": [
+            112,
+            321,
+            346,
+            336
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Each annotator followed a three-step pipeline:",
+        "bbox": [
+            112,
+            343,
+            457,
+            360
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Scenario Drafting: Write a conversational prompt targeting modality cues (e.g., \"Can you tell if this room is too loud for a conference call?\").",
+            "- Confounder Generation: Create a paired version of the prompt with the same text and visuals but alter one audio attribute (e.g., add background machinery noise).",
+            "- Media Pairing: Select or synthesize matching audio and visual assets from our curated libraries to illustrate both the original and confounded conditions."
+        ],
+        "bbox": [
+            136,
+            376,
+            487,
+            596
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "This ensured every question/confounder pair isolated the intended cue and prevented shortcut learning by downstream models.",
+        "bbox": [
+            112,
+            613,
+            489,
+            661
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "4. Answer Creation Process",
+        "text_level": 1,
+        "bbox": [
+            112,
+            676,
+            336,
+            690
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "For each question instance, annotators produced:",
+        "bbox": [
+            112,
+            697,
+            477,
+            714
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Reference Answer: A concise response directly addressing the prompt (e.g., \"No, it's quiet enough for clear conversation\").",
+            "- Rationale Statement: A brief explanation linking the critical cue to the answer (e.g., \"The low ambient noise level confirms a silent office setting\")."
+        ],
+        "bbox": [
+            136,
+            730,
+            487,
+            856
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "These reference answers and rationales formed the ground truth for our GPT-4-based judgment of model responses.",
+        "bbox": [
+            112,
+            873,
+            487,
+            921
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "5. Annotation Criteria & Quality Control",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            838,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "All questions and answers were crafted according to these overarching principles:",
+        "bbox": [
+            507,
+            105,
+            880,
+            137
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Clarity & Brevity: Simple, conversational language devoid of syntactic complexity.",
+            "- Modality Isolation: Exactly one audio or visual \"hook\" per item, ensuring focused evaluation.",
+            "- Balanced Distribution of Skills: Even distribution of questions across different skills."
+        ],
+        "bbox": [
+            531,
+            149,
+            882,
+            282
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D LLM-as-a-Judge Details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            296,
+            757,
+            313
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "1. Human vs. LLM Evaluation Experiment:",
+        "text_level": 1,
+        "bbox": [
+            524,
+            322,
+            880,
+            337
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To validate whether a large language model (LLM) could reliably substitute for our expert reviewers, we conducted a blind evaluation on a sample of 300 question-answer pairs drawn evenly from our benchmark. Each pair was independently graded on a 1-5 scale by:",
+        "bbox": [
+            544,
+            338,
+            880,
+            434
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Three domain experts from our panel following a predecided grading criteria",
+            "- An LLM-as-a-Judge, implemented via a single GPT-4.1-mini call per sample."
+        ],
+        "bbox": [
+            566,
+            444,
+            880,
+            510
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We computed average scores for each item under both conditions and measured interrater agreement using Cohen's  $\\kappa$ . Across all 300 items, human-human agreement averaged  $\\kappa = 0.82$ , while human-LLM agreement reached  $\\kappa = 0.78$ , indicating the LLM's judgments were almost as consistent with the experts' as the experts were with one another. Overall the LLM's ratings fell within one point of the experts' in  $92\\%$  of cases.",
+        "bbox": [
+            544,
+            520,
+            882,
+            680
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Given these results, we adopted the LLM-as-a-Judge for large-scale scoring in subsequent experiments, leveraging its efficiency without materially sacrificing quality.",
+        "bbox": [
+            544,
+            686,
+            882,
+            750
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "2. LLM-as-a-Judge Criteria:",
+        "text_level": 1,
+        "bbox": [
+            522,
+            762,
+            754,
+            777
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We present each question to the LLM-as-a-Judge by supplying both the corresponding speech and visual inputs. The model is then asked to rate the provided answer on a five-point scale according to the following criteria:",
+        "bbox": [
+            544,
+            783,
+            882,
+            863
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "- Score 1: The response is often off-topic or incorrect and fails to recognize or use the specified speaker characteristic.",
+        "bbox": [
+            566,
+            873,
+            880,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "28481",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a4efbfdd60f6b1e089dc4b15be7ba584e7d02385f3479618dfeed3920472dec.jpg",
+        "image_caption": [
+            "Figure 6: Pie charts showing different Speaker Profile categories in terms of Gender and Age"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            97,
+            295,
+            222
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/571792c3fc2d397859449510fde26698dd326e5c98502e0b0f7a33d1bcb69846.jpg",
+        "image_caption": [
+            "Figure 7: Pie charts showing different Paralanguage Understanding Categories and different Emotion Categories in the benchmark"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            332,
+            89,
+            475,
+            225
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Score 2: The response occasionally addresses the prompt but handles the speaker characteristic inconsistently or superficially.",
+            "- Score 3: The response shows a basic understanding of intent, with partial integration of the speaker characteristic but lacks depth or precision.",
+            "- Score 4: The response delivers relevant and mostly accurate content that usually incorporates the speaker characteristic, with only minor lapses.",
+            "- Score 5: The response consistently produces accurate, context-rich answers that fully and effectively integrate the speaker characteristic."
+        ],
+        "bbox": [
+            171,
+            305,
+            485,
+            565
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "E Voice Quality Assessment",
+        "text_level": 1,
+        "bbox": [
+            114,
+            577,
+            371,
+            595
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We generated synthetic audio for 100 benchmark questions using two text-to-speech systems: CosyVoice (open source) (Du et al., 2024) and ElevenLabs (ElevenLabs Inc., 2025) (commercial). We also recorded these questions by professional voice actors.",
+        "bbox": [
+            112,
+            602,
+            487,
+            697
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "To compare these renderings, we conducted a Mean Opinion Score (MOS) Test on a random subset of 100 paralanguage-focused queries (as in Sec. 3.4). Ten expert annotators rated each audio on two dimensions, using a 5-point scale:",
+        "bbox": [
+            112,
+            700,
+            487,
+            781
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Attribute match: How accurately the intended paralinguistic cue (e.g., emotion, background noise) was conveyed.",
+            "- Naturalness: The overall human-likeness of the voice."
+        ],
+        "bbox": [
+            134,
+            789,
+            485,
+            876
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The MOS results were as follows: Professional human recordings: 4.6 (attribute match) / 4.5 (nat",
+        "bbox": [
+            112,
+            889,
+            487,
+            920
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/665f4ced0d90cc32e95279d646c5c00351a732ae6161c81d0f2480ce076452a0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            87,
+            662,
+            225
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/cd054fa14779b207ee832796054ab8f0b0420fc4014193464ebd001e6454f79a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            92,
+            873,
+            221
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "uralness), CosyVoice TTS: 2.4 / 2.1, ElevenLabs TTS: 3.1 / 3.3.",
+        "bbox": [
+            507,
+            305,
+            880,
+            335
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "These findings confirm that, while modern TTS can approximate certain prosodic features, professional voice actors remain far superior in both fidelity and naturalness, which led us to use human-recorded queries by preprofessional voice actors throughout the benchmark.",
+        "bbox": [
+            507,
+            338,
+            880,
+            432
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "F Voice Data Collection",
+        "text_level": 1,
+        "bbox": [
+            509,
+            445,
+            731,
+            460
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "To ensure natural and expressive speech, we employed four professional voice actors, four male and two female, contracted via Fiverr. Each actor was compensated according to the standard rates listed on the platform. This approach allowed us to capture high-quality, emotionally varied recordings with realistic prosody and delivery across all tasks. Our institution's Institutional Review Board (IRB) has granted approval for this data collection.",
+        "bbox": [
+            507,
+            470,
+            882,
+            615
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "G Additional Details: Auxiliary",
+        "text_level": 1,
+        "bbox": [
+            507,
+            627,
+            800,
+            643
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Compute Infrastructure: All our experiments are conducted on a single NVIDIA A6000 GPU. No training is required, and depending on the downstream task, a single inference run on a benchmark requires anywhere between 30 minutes to 2 hours.",
+        "bbox": [
+            507,
+            652,
+            882,
+            732
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Implementation Software and Packages: We use the official implementation of the OLMs we benchmark. For LLM-as-judge implementation, we utilize the official OpenAI APIs.",
+        "bbox": [
+            507,
+            733,
+            882,
+            796
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Potential Risks: We manually curate all our questions to avoid any potential harmful or biased samples.",
+        "bbox": [
+            507,
+            797,
+            882,
+            845
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "28482",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]