AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

ACL 2026

DualFact+: A Multimodal Fact Verification Framework for Procedural Video Captioning

Cennet Oguz, Yasser HAMIDULLAH, Josef van Genabith, Simon Ostermann

Evaluating factual correctness in procedural video captions is challenging because captions must reflect both the abstract procedural roles (e.g., actions, ingredients, tools, locations) and their visual execution. Existing evaluation metrics, which rely on lexical overlap or holistic semantic similarity, often miss role-specific omissions and misclassify visually present but task-irrelevant content as hallucinations. We introduce DualFact+, a role-aware, fact-level evaluation framework that distinguishes conceptual facts, encoding ontology-based role typing of procedural steps (Action, Object or Ingredient, Tool, Location), from contextual facts, encoding video-grounded predicate–argument relations that specify how these roles are instantiated during execution. To enable complete and role-consistent evaluation, DualFact+ incorporates visually grounded implicit arguments and contrastive fact sets, and operates in two complementary modes: DualFact-C for text-based verification and DualFact-V for video-grounded verification. Experiments on YouCook3-Fact and CraftBench-Fact show that state-of-the-art captioning models produce fluent but often incomplete descriptions with systematic role-level errors. DualFact+ achieves stronger correlation with human factuality judgments than standard lexical and embedding-based metrics, highlighting the importance of role-aware evaluation for procedural video understanding.

ACL 2026

Not all ANIMALs are equal: metaphorical framing through source domains and semantic frames

Yulia Otmakhova, Matteo Guida, Lea Frermann

Metaphors are powerful framing devices, yet their source domains alone do not fully explain the specific associations they evoke. We argue that the interplay between source domains and semantic frames determines how metaphors shape understanding of complex issues, and present a computational framework that allows to derive salient discourse metaphors through their source domains and semantic frames. Applying this framework to climate change news, we uncover not only well-known source domains but also reveal nuanced frame-level associations that distinguish how the issue is portrayed. In analyzing immigration discourse across political ideologies, we demonstrate that liberals and conservatives systematically employ different semantic frames within the same source domains, with conservatives favoring frames emphasizing uncontrollability and liberals choosing neutral or more “victimizing” semantic frames. Our work bridges conceptual metaphor theory and linguistics, providing the first NLP approach for discovery of discourse metaphors and fine-grained analysis of differences in metaphorical framing.

ACL 2026

Relaxing the Constraints: A Dual-Importance Projection Mechanism for Lifelong Model Editing

Zhenghai Chen, Senbin Xu, Jiaxi Tan, Xinhua Wu, Yan Zhang, Xiawu Zheng, Shengchuan Zhang, Ke Li, Sicheng Zhao, Liujuan Cao, Rongrong Ji

Factual knowledge stored in Large Language Models (LLMs) inevitably becomes outdated or erroneous over time, making it critical to update these models without incurring the high cost of retraining. Existing sequential knowledge editing methods predominantly rely on strict orthogonal projection to preserve previously edited knowledge. However, this excessive constraint limits gradient expressiveness, resulting in a significant degradation of model generalization and overall performance as the number of edits increases. To address this challenge, we propose Dual-Importance Projection Editing (DipEdit). This method leverages Singular Value Decomposition (SVD) to identify critical gradient subspaces and introduces a dual mechanism comprising “accumulated importance” and “projection importance.” Unlike traditional approaches that enforce strict orthogonality, DipEdit dynamically scales gradient components parallel to key subspaces based on their projection importance rather than discarding them directly. This approach enhances the model’s adaptability to new knowledge while maximally preserving historical knowledge. Extensive experiments conducted on five mainstream LLMs using the ZsRE and Counterfact datasets demonstrate that DipEdit effectively handles thousands of sequential edits. The proposed method achieves an average comprehensive performance improvement of 10.36% and effectively maintains the model’s general capabilities on downstream tasks. Code is available at: https://github.com/czhhhla/DipEdit.

ACL 2026

One Script Instead of Hundreds? On Pretraining Romanized Encoder Language Models

Benedikt Ebing, Lennart Keller, Goran Glavaš

Exposing latent lexical overlap, script romanization has emerged as an effective strategy for improving cross-lingual transfer (XLT) in multilingual language models (mLMs). Most prior work, however, focused on setups that favor romanization the most: (1) transfer from high-resource Latin-script to low-resource non-Latin-script languages and/or (2) between genealogically closely related languages with different scripts. It thus remains unclear whether romanization is a good representation choice for pretraining general-purpose mLMs, or, more precisely, if information loss associated with romanization harms performance for high-resource languages. We address this gap by pretraining encoder LMs from scratch on both romanized and original texts for six typologically diverse high-resource languages, investigating two potential sources of degradation: (i) loss of script-specific information and (ii) dilution of language-specific representations from increased subword overlap. Using two romanizers with different fidelity profiles, we observe negligible performance loss for languages with segmental scripts, whereas languages with morphosyllabic scripts (Chinese and Japanese) suffer degradation that higher-fidelity romanization mitigates but cannot fully recover. Importantly, comparing monolingual LMs with their mLM counterpart, we find no evidence that increased subword overlap dilutes language-specific representations. We further show that romanization improves encoding efficiency (i.e., fertility) for segmental scripts at a negligible performance cost.

ACL 2026

When Reviews Disagree: Fine-Grained Contradiction Analysis in Scientific Peer Reviews

Sandeep Kumar, Yash Kamdar, Abid Hossain, Bharti Kumari, Tanik Saikh, Asif Ekbal

Scientific peer reviews frequently contain conflicting expert judgments, and the increasing scale of conference submissions makes it challenging for Area Chairs and editors to reliably identify and interpret such disagreements. Existing approaches typically frame reviewer disagreement as binary contradiction detection over isolated sentence pairs, abstracting away the review-level context and obscuring differences in the severity of evaluative conflict. In this work, we introduce a fine-grained formulation of reviewer contradiction analysis that operates over full peer reviews by explicitly identifying contradiction evidence spans and assigning graded disagreement intensity scores. To support this task, we present RevCI, an expert-annotated benchmark of peer-review pairs with evidence-level contradiction annotations with graded intensity labels. We further propose IMPACT, a structured multi-agent framework that integrates aspect-conditioned evidence extraction, deliberative reasoning, and adjudication to model reviewer contradictions and their intensity. To support efficient deployment, we distill IMPACT into TIDE, a small language model that predicts contradiction evidence and intensity in a single forward pass. Experimental results show that IMPACT substantially outperforms strong single-agent and generic multi-agent baselines in both evidence identification and intensity agreement, while TIDE achieves competitive performance at significantly lower inference cost.

ACL 2026

Fine-Grained Detection of Context-Grounded Hallucinations Using LLMs

Yehonatan Peisakhovsky, Zorik Gekhman, Yosi Mass, Liat Ein-Dor, Roi Reichart

Context-grounded hallucinations are cases where model outputs contain information not verifiable against the source text. We study the applicability of LLMs for localizing such hallucinations, as a more practical alternative to existing complex evaluation pipelines. In the absence of established benchmarks for meta-evaluation of hallucinations localization, we construct one tailored to LLMs, involving a challenging human annotation of over 1,000 examples. We complement the benchmark with an LLM-based evaluation protocol, verifying its quality in a human evaluation. Since existing representations of hallucinations limit the types of errors that can be expressed, we propose a new representation based on free-form textual descriptions, capturing the full range of possible errors. We conduct a comprehensive study, evaluating four large-scale LLMs, which highlights the benchmark’s difficulty, as the best model achieves an F1 score of only 0.67. Through careful analysis, we offer insights into optimal prompting strategies for the task and identify the main factors that make it challenging for LLMs: (1) a tendency to incorrectly flag missing details as inconsistent, despite being instructed to check only facts in the output; and (2) difficulty with outputs containing factually correct information absent from the source - and thus not verifiable - due to alignment with the model’s parametric knowledge.

ACL 2026

Alignment Data Map for Efficient Preference Data Selection and Diagnosis

Seohyeong Lee, Eunwon Kim, Hwaran Lee, Buru Chang

Human preference data is essential for aligning large language models (LLMs) with human values, but collecting such data is often costly and inefficient-motivating the need for efficient data selection methods that reduce annotation costs while preserving alignment effectiveness. To address this issue, we propose Alignment Data Map, a data analysis tool for identifying and selecting effective preference data. We first evaluate alignment scores of the preference data by LLM-as-a-judge, explicit reward model, and reference-based approaches. The Alignment Data Map considers both response quality and inter-response variability based on the alignment scores. From our experimental findings, training on only 33% of samples that exhibit high-quality and low-variability, achieves comparable or superior alignment performance on MT-Bench, Evol-Instruct, and AlpacaEval, compared to training with the full dataset. In addition, Alignment Data Map detects potential label misannotations by analyzing correlations between annotated labels and alignment scores, improving annotation accuracy. The implementation is available at https://github.com/01choco/Alignment-Data-Map.

ACL 2026

Sentient Agent as a Judge: Evaluating Higher-Order Social Cognition in Large Language Models

Bang Zhang, Ruotian Ma, Qingxuan Jiang, Peisong Wang, Jiaqi Chen, Zheng Xie, Xingyu Chen, Yue Wang, Fanghua Ye, Jian Li, Yifan Yang, Zhaopeng Tu, Xiaolong Li

Assessing how well a large language model (LLM) understands human, rather than merely text, remains an open challenge.To bridge the gap, we introduce Sentient Agent as a Judge(SAGE), an automated evaluation framework that measures an LLM’s higher-order social cognition.SAGE instantiates a “Sentient Agent” – an LLM-powered agent that simulates human-like emotional changes and inner thoughts to provide a more realistic evaluation of the tested model in multi-turn conversations.At every turn, the agent reasons about (i) how its emotion changes, (ii) how it feels, and (iii) how it should reply, yielding a numerical emotion trajectory and interpretable inner thoughts.Experiments on 100 supportive-dialogue scenarios show that the final Sentient emotion score correlates strongly with Barrett-Lennard Relationship Inventory (BLRI) ratings and utterance-level empathy metrics, validating psychological fidelity. Human evaluation further demonstrates 85.3% consistency between the agent’s emotional reasoning and human judgments. We also build a public Sentient Leaderboard covering 18 commercial and open-source models that uncovers substantial gaps (up to 4\times) between frontier systems (GPT-4o-Latest, Gemini2.5-Pro) and earlier baselines, gaps not reflected in conventional leaderboards (e.g. Arena). SAGE thus provides a principled, scalable, and interpretable tool for tracking progress toward genuinely empathetic and socially adept language agents.

ACL 2026

Feedback Is The Key for Automated Survey Generation

Tianyi Xu, Zhe Zhao, Tianshuo Wei, Yiqun Kou, Liuliu Han, Ye Wei

The escalating demand for comprehensive literature surveys in rapidly evolving research areas makes manual writing increasingly impractical, underscoring the necessity of automation. Large Language Models (LLMs) provide a promising foundation for this task, yet guiding them to generate accurate, reliable content remains a fundamental challenge, as issues such as hallucinations and vague organization often persist. To address this, we propose FIKSurvey, a feedback-driven framework grounded in the idea that “Feedback is the key for automatic survey generation.” Specifically, FIKSurvey systematically incorporates feedback across three dimensions: outline feedback for structural clarity, citation feedback for evidence validation, and content feedback for readability and analytical depth. The framework also supports optional human-in-the-loop intervention for user-specific needs. Experiments confirm that FIKSurvey substantially improves both citation and content quality, demonstrating feedback as the critical mechanism for automatic survey generation.

ACL 2026

CoTEvol: Self-Evolving Chain-of-Thoughts for Data Synthesis in Mathematical Reasoning

Zhuo Wang, Zhuo Zhang, Yafu Li, Yu Cheng, Lizhen Qu, Zenglin Xu

Large Language Models (LLMs) exhibit strong mathematical reasoning when trained on high-quality Chain-of-Thought (CoT) that articulates intermediate steps, yet costly CoT curation hinders further progress. While existing remedies such as distillation from stronger LLMs and self-synthesis based on test-time search alleviate this issue, they often suffer from diminishing returns or high computing overhead. In this work, we propose CoTEvol, a genetic evolutionary framework that casts CoT generation as a population-based search over reasoning trajectories. Candidate trajectories are iteratively evolved through reflective global crossover at the trajectory level and local mutation guided by uncertainty at the step level, enabling holistic recombination and fine-grained refinement. Lightweight, task-aware fitness functions are designed to guide the evolutionary process toward accurate and diverse reasoning. Empirically, improves correct-CoT synthesis success by over 30% and enhances structural diversity, with markedly improved efficiency. LLMs trained on these evolutionary CoT data achieve an average gain of 6.6% across eight math benchmarks, outperforming previous distillation and self-synthesis approaches. These results underscore the promise of evolutionary CoT synthesis as a scalable and effective method for mathematical reasoning tasks.

ACL 2026

RAG over Tables: Hierarchical Memory Index, Multi-Stage Retrieval, and Benchmarking

Jiaru Zou, Dongqi Fu, Sirui Chen, Xinrui He, Zihao Li, Yada Zhu, Jiawei Han, Jingrui He

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by integrating them with an external knowledge base to improve the answer relevance and accuracy. In real-world scenarios, beyond pure text, a substantial amount of knowledge is stored in tables, and user questions often require retrieving answers that are distributed across multiple tables. Retrieving knowledge from a table corpora (i.e., various individual tables) for a question remains nascent, for (i) how to understand intra- and inter-table knowledge effectively, (ii) how to filter unnecessary tables and retrieve the most relevant tables efficiently, (iii) how to organize complex retrieved contexts for LLMs’ reasoning, and (iv) how to evaluate the corresponding performance in a realistic setting. Facing the above challenges, in this paper, we first propose a table-corpora-aware RAG framework, named T-RAG, which consists of the hierarchical memory index, multi-stage retrieval, and graph-aware context organization for effective and efficient table knowledge retrieval and inference. Then, we develop a multi-table question answering benchmark named MultiTableQA, which spans 3 different task types, 57,193 tables, and 23,758 questions in total, and the sources are all from real-world scenarios. Based on MultiTableQA, we perform a comprehensive comparison of table retrieval methods, RAG-based approaches, and table-to-graph representation learning methods. T-RAG consistently achieves state-of-the-art accuracy, recall, and runtime performance, with improvements of up to 9.4%. Moreover, T-RAG yields an average inference gain of 11.8% across different downstream backbone LLMs. Our code and data are available at https://github.com/jiaruzouu/T-RAG.

ACL 2026

MSCode: Advancing Human Motion-Language Understanding via Modality-Shared Codebook

Haoyu Shi, Huaiwen Zhang

Recently, human motion understanding has been a prominent area of research due to its critical importance in many fields. The key to advancing this understanding lies in the precise alignment between motion and linguistic modalities. Existing methods mainly follow two paradigms: global contrastive alignment and vocabulary space-based alignment. However, motion sequences exhibit sequential spatiotemporal dynamics while text conveys abstract semantics, leading to a fundamental mismatch in semantic levels and granularities. This undermines cross-modal alignment and results in suboptimal downstream performance. To alleviate this, we introduce a modality-shared codebook that enables unified representation learning and precise alignment of motion and linguistic modalities. Each codeword in the codebook is regularized to encode cross-modality shared semantics, and we leverage sparse activation and distribution consistency loss to enforce matched motion and text are represented by the same set of codewords. Additionally, we introduce a locality-aware Gaussian encoder to refine pose features and design a hard-negative guided loss to strengthen alignment discriminability. Extensive experiments across various language-motion evaluation, including text-motion retrieval, text-motion grounding, and motion caption, demonstrate that our model significantly surpasses current state-of-the-art methods.

ACL 2026

Autoregressive Semantic Visual Reconstruction Helps VLMs Understand Better

Dianyi Wang, Wei Song, Yikun Wang, Siyuan Wang, Kaicheng Yu, Zhongyu Wei, Jiaqi Wang

Typical large vision-language models (LVLMs) apply autoregressive supervision primarily to textual responses, without fully exploiting causal learning over rich visual inputs. As a result, these models often emphasize vision-to-language alignment while potentially overlooking fine-grained visual information. While prior work has explored autoregressive image generation, effectively leveraging autoregressive visual supervision to enhance image understanding remains an open challenge. In this paper, we introduce Autoregressive Semantic Visual Reconstruction (ASVR), which enables joint learning of visual and textual modalities within a unified autoregressive framework. ASVR trains models to autoregressively reconstruct the semantic content of input images, which consistently enhances multimodal comprehension. Notably, we show that even when provided with continuous image features as input, models can effectively reconstruct discrete semantic tokens, resulting in stable and consistent improvements across various multimodal understanding benchmarks. ASVR delivers significant performance gains and scalability across varying data scales, visual input, visual supervision and model architectures. In particular, ASVR generally improves baselines by 2-3% across 14 multimodal benchmarks.

ACL 2026

Why Can Distillation Work with Limited Resources? A Systematic Study

Xiao Hu, Xingyu Lu, Liyuan Mao, Yifan Zhang, Tianke Zhang, Bin Wen, Fan Yang, Tingting Gao, Guorui Zhou

Recently, large language models have made remarkable progress in reasoning, largely driven by scaling data and model size. In parallel, several studies argue that for smaller models, high-quality distillation can yield strong reasoning performance with minimal resources. However, a framework for understanding machine reasoning that explains why low-resource distillation can boost model performance is still missing. In this paper, we conduct a controlled case study: using less than 920 examples, a simple distillation based on the base model can actually achieve notable reasoning performance improvement, compared with the base model and even the zero-RL models. By analyzing the token frequency in model outputs, we find that the distilled model shows more flexible reasoning. It uses anthropomorphic tokens and logical connectors much more often than the base and zero-RL model. Further analysis reveals that distillation enhances the presence of two advanced cognitive behaviors: Multi-Perspective Thinking or Attempting and Metacognitive Awareness. Frequent occurrences of these two advanced cognitive behaviors give rise to flexible reasoning, which is essential for solving reasoning problems.

ACL 2026

AnchorAlign: A Novel Anchor Alignment-enhanced Generative Method for Joint Named Entity Recognition and Relation Extraction

Xiaolong Weng, Yuanyun Zhou, Boyu Qiu, Zehua Wang, Ying Xiong, Buzhou Tang

Named Entity Recognition (NER) and Relation Extraction (RE) are two fundamental and interdependent tasks in information extraction (IE), aiming to identify entities and relations from unstructured text. Recently, generative methods have become mainstream instead of discriminative methods for IE, especially joint multi-task IE, due to their promising performance and flexibility. For joint NER and RE, existing methods suffer from misalignment between entities and relations, as well as misalignment among relations. To address these issues, we propose AnchorAlign, a novel generative method enhanced by anchor alignment. Specifically, we first introduce an anchor entity selection mechanism to identify key entities in the text as anchor points, which serve as semantic pivots to bridge the two tasks. Then, we design a dual-level anchor alignment module: at the semantic level, we construct a cross-task semantic alignment space to align the semantic representations of anchor entities and their associated relations; at the generation level, we introduce an anchor-guided generation constraint to guide the model to generate entities and relations with strict alignment based on the anchor points. Extensive experiments on five benchmark datasets show that AnchorAlign outperforms state-of-the-art baselines, demonstrating its effectiveness. Our work provides a new perspective for optimizing the joint modeling of NER and RE, and has potential to be extended to more complex multi-task IE such as NER and Event Extraction (EE).

ACL 2026

Decoupled Reasoning with Implicit Fact Tokens (DRIFT): A Dual-Model Framework for Efficient Long-Context Inference

Wenxuan Xie, Xuhong Wang

The integration of extensive, dynamic knowledge into Large Language Models (LLMs) remains a significant challenge due to the inherent entanglement of factual data and reasoning patterns. Existing solutions, ranging from non-parametric Retrieval-Augmented Generation (RAG) to parametric knowledge editing, are often constrained in practice by finite context windows, retriever noise, or the risk of catastrophic forgetting. In this paper, we propose DRIFT, a novel dual-model architecture designed to explicitly decouple knowledge extraction from the reasoning process. Unlike static prompt compression, DRIFT employs a lightweight knowledge model to dynamically compress document chunks into implicit fact tokens conditioned on the query. These dense representations are projected into the reasoning model’s embedding space, replacing raw, redundant text while maintaining inference accuracy. Extensive experiments show that DRIFT significantly improves performance on long-context tasks, outperforming strong baselines among comparably sized models. Our approach provides a scalable and efficient paradigm for extending the effective context window and reasoning capabilities of LLMs. Our code and data will be made public upon publication.

ACL 2026

R-Search: Empowering LLM Reasoning with Search via Multi-Reward Reinforcement Learning

Qingfei Zhao, Ruobing Wang, Dingling Xu, Daren Zha, Ma Bowen, Zhichun Wang, Shijie Jia, Limin Liu, Xin Wang

Large language models (LLMs) have notably progressed in multi-step and long-chain reasoning. However, extending their reasoning capabilities to encompass deep interactions with search remains a non-trivial challenge, as models often fail to identify optimal reasoning–search interaction trajectories, resulting in suboptimal responses. We propose R-Search, a novel reinforcement learning framework for Reasoning–Search integration, designed to enable LLMs to autonomously execute multi-step reasoning with deep search interaction, and learn optimal reasoning–search interaction trajectories via multi-reward signals, improving response quality in complex logic- and knowledge-intensive tasks. R-Search guides the LLM to dynamically decide when to search or reason, while globally integrating key evidence to enhance deep knowledge interaction between reasoning and search. During RL training, R-Search provides multi-type rewards to jointly optimize the reasoning–search trajectory. Experiments on seven datasets show that R-Search significantly outperforms mainstream RAG baselines.

ACL 2026

Med-SRAF: A Multi-Agent Framework for Medical Reasoning via Semantic Routing and Agentic Fusion

Xiao Li, Zhuo Chen, Jun Xia, hongxin xiang, Chao Wang, Wenjie Du, Yang Wang

While Retrieval-Augmented Generation (RAG) has become a standard paradigm for mitigating hallucinations in Large Language Models (LLMs), its effectiveness in complex medical reasoning remains limited. Existing RAG methods suffer from two main challenges: First, Semantic Drift: without explicit domain constraints, LLM-driven query decomposition often deviates from the original clinical intent, introducing substantial noise that degrades retrieval relevance. Second, Concatenation Fallacy: retrieved evidence from different semantic aspects is aggregated in a naive, unstructured manner, without modeling their inter-dependencies and potential conflicts, which ultimately undermines downstream reasoning. To address these challenges, we propose Med-SRAF, a multi-agent retrieval augmentation framework guided by medical domain knowledge. This framework reconstructs the traditional RAG process through two core mechanisms: (1) Intent-driven Semantic Routing, where a UMLS-based NavigationAgent dynamically maps queries to medical dimensions for strategic search space pruning; and (2) Evidence-based Agentic Fusion, where a FusionAgent resolves conflicts among dimension-specific evidence to build logically consistent reasoning chains. Extensive experiments on five widely used medical benchmarks show that Med-SRAF consistently outperforms existing general RAG baselines, achieving an average accuracy improvement of over 4.9%, highlighting its effectiveness in robust and interpretable medical reasoning. Our code is at https://anonymous.4open.science/r/MultiAgent_RAG-F6DC.

ACL 2026

Single-Agent Generation Surpasses Multi-Agent Systems in Semantic Diversity

Encheng Cui, Shaowen Peng, Kazuhiro Ito, Jinsha Xu, Shohei Hisada, Shoko Wakamiya, Eiji Aramaki

Multi-Agent Systems (MAS) are commonly used to improve reasoning diversity and robustness by simulating interactions among agents with distinct roles. However, prior work often entangles the contribution of the multi-agent architecture with that of prompt conditioning, making the source of observed diversity gains unclear. We address this confound with a controlled study on divergent thinking tasks, using identical prompt conditioning for MAS and single agent baseline. Under these matched conditions, single agent setups consistently outperform multi-agent systems in semantic diversity. We attribute this gap to information visibility: parallel agents often converge on overlapping ideas, whereas a single agent model can condition on its own generation to avoid redundancy. We further find that a Multi-Output strategy, which prompts a single agent to produce multiple responses within a single inference pass, achieves the highest diversity without degrading logical validity. Together, these results point to a more efficient and effective way to expand diversity, with implications for the design of more efficient agentic frameworks.

ACL 2026

Debiasing LLMs by Masking Unfairness-Driving Attention Heads

Tingxu Han, Wei Song, Ziqi Ding, Ziming Li, Chunrong Fang, Yuekang Li, Dongfang Liu, Zhenyu Chen, Zhenting Wang

Large language models (LLMs) increasingly mediate decisions in domains where unfair treatment of demographic groups is unacceptable. Existing work probes when biased outputs appear, but gives little insight into the mechanisms that generate them, leaving existing mitigations largely fragile. In this paper, we conduct a systematic investigation of LLM unfairness and propose DiffHeads—a lightweight debiasing framework for LLMs. We first compare Direct-Answer (DA) prompting to Chain-of-Thought (CoT) prompting across eight representative open- and closed-source LLMs. DA will trigger the nature-bias component of the LLM and reduce measured unfairness by 391.9%- 534.5% in both one- and two-turn dialogues. Next, we define a token-to-head contribution score that traces each token’s influence back to individual attention heads. This reveals a small cluster of bias heads that activate under DA but stay largely dormant with CoT, providing the first causal link between prompting strategy and bias emergence. Finally, building on this insight, we propose DiffHeads, which identify bias heads through differential activation analysis between DA and CoT and selectively mask only those heads. DiffHeads reduces unfairness by 49.4%, and 40.3% under DA and CoT, respectively, without harming model utility.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.