EACL 2026 论文列表

Large language models (LLMs) have demonstrated impressive few-shot in-context learning (ICL) abilities. Still, we show that they are sometimes prone to a ‘copying bias’, where they copy answers from provided examples instead of learning the underlying patterns. In this work, we propose a novel and simple method to mitigate such copying bias. First, we create a synthetic task and use the Integrated Gradients method to identify neurons that prioritize copying over generalization. We demonstrate that pruning these neurons consistently improves performance across a diverse set of ICL tasks, including both single-token and multi-token scenarios, while maintaining or even improving the model’s general capabilities. We also show that our method is applicable across various LLM architectures, including Transformers and State-Space Models, without requiring modifications. In our analysis, we adopt a task-recognition perspective on ICL and examine task vectors (Hendel et al., 2023) induced by the model. We find that pruning enhances the quality of these vectors, suggesting that the pruned neurons previously hindered effective task recognition.

Large Language Models (LLMs) have demonstrated strong performance across a wide range of NLP tasks. However, they often exhibit suboptimal behaviors and inconsistencies when exposed to unfamiliar external information, underscoring their limitations in effectively leveraging such knowledge. Inspired by constructivist learning theory, we propose ThinkNote, a novel framework that enhances the external knowledge utilization of LLMs through a two-stage constructivist cognitive modeling process. Specifically, ThinkNote performs knowledge assimilation to align new information with the model’s parametric memory, forming a coherent internal representation. It then applies thought accommodation to adapt internal reasoning, thereby promoting more consistent and reliable outputs. Extensive experimental results demonstrate that ThinkNote achieves a 10% improvement over strong baseline methods on various question-answering benchmarks. Further analysis indicates that ThinkNote effectively integrates and utilizes external knowledge to help LLMs generate accurate responses and improves their self-consistency. All data and code will be publicly available at https://github.com/OpenMatch/ThinkNote.

Recent advancements in large language models (LLMs) have shown remarkable potential in automating machine learning tasks. However, existing LLM-based agents often struggle with low diversity and suboptimal code generation. While recent work (CITATION) has introduced Monte Carlo Tree Search (MCTS) to address these issues, limitations persist in the quality and diversity of thoughts generated, as well as in the scalar value feedback mechanisms used for node selection. In this study, we introduce Introspective Monte Carlo Tree Search (I-MCTS), a novel approach that iteratively expands tree nodes through an introspective process that meticulously analyzes solutions and results from parent and sibling nodes. This facilitates a continuous refinement of the node in the search tree, thereby enhancing the overall decision-making process. Furthermore, we integrate a Large Language Model (LLM)-based value model to facilitate direct evaluation of each node’s solution prior to conducting comprehensive computational rollouts. A hybrid rewarding mechanism is implemented to seamlessly transition the Q-value from estimated score to actual performance scores. Applied to the various ML tasks, our approach demonstrates a 4% absolute improvement in performance compared to the strong open-source AutoML agents, showcasing its effectiveness in enhancing agentic AutoML systems. Resource available at https://github.com/jokieleung/I-MCTS

In modern healthcare, radiology plays a pivotal role in diagnosing and managing diseases. However, the complexity of medical imaging data and the variability in interpretation can lead to inconsistencies and a lack of patient-centered insight in radiology reports. To address this challenge, a novel multimodal prompt-driven report generation framework Rad-Flamingo was developed, that integrates diverse data modalities—such as medical images, and clinical notes—to produce comprehensive and context-aware radiology reports. Our framework leverages innovative prompt engineering techniques to guide vision-language models in generating relevant information, ensuring these generated reports are not only accurate but also understandable to individual patients. A key feature of our framework is its ability to provide patient-centric explanations, offering clear and personalized insights into diagnostic findings and their implications. Additionally, we also demonstrate a synthetic data generation pipeline, to append any existing benchmark datasets’ findings and impressions with patient-centric explanation. Experimental results demonstrate that this framework’s effectiveness in enhancing report quality, improving understandability, and could foster better patient-doctor communication. This approach represents a significant step towards human-centered medical AI systems.

This position paper presents a novel perspective on the utilization of Large Language Models (LLMs) in the artificial intelligence paper review process. We first critique the current tendency for LLMs to be primarily used for simple review text generation, arguing instead that this approach overlooks more meaningful applications of LLMs that preserve human expertise at the core of evaluation. Instead, we advocate for leveraging LLMs to support key aspects of the review process—specifically, verifying the reproducibility of experimental results, checking the correctness and relevance of citations, and assisting with ethics review flagging. For example, integrating tools based on LLM Agents for code generation from research papers has recently enabled automated assessment of the reproducibility of the paper, thereby improving the transparency and reliability of research. By reorienting LLM usage toward these targeted and assistive roles, we outline a pathway for more effective and responsible integration of LLMs into peer review, ultimately supporting both reviewer efficiency and the integrity of the scientific process.

Recent work has explored the use of personal information in the form of persona sentences or self-disclosures to improve modeling of individual characteristics and prediction of annotator labels for subjective tasks. The volume of personal information has historically been restricted and thus little exploration has gone into understanding what kind of information is most informative for predicting annotator labels. In this work, we categorize self-disclosures and use them to build annotator models for predicting judgments of social norms. We perform several ablations and analyses to examine the impact of the type of information on our ability to predict annotation patterns. Contrary to previous work, only a small number of comments related to the original post are needed. Lastly, a more diverse sample of annotator self-disclosures did not lead to the best performance. Sampling from a larger pool of comments without filtering still yields the best performance, suggesting that there is still much to uncover in terms of what information about an annotator is most useful for verdict prediction.

Plane geometry problem solving (PGPS) has recently gained significant attention as a benchmark to assess the multi-modal reasoning capabilities of large vision-language models. Despite the growing interest in PGPS, the research community still lacks a comprehensive overview that systematically synthesizes recent work in PGPS. To fill this gap, we present a survey of existing PGPS studies. We first categorize PGPS methods into an encoder-decoder framework and summarize the corresponding output formats used by their encoders and decoders. Subsequently, we classify and analyze these encoders and decoders according to their architectural designs. Finally, we outline major challenges and promising directions for future research. In particular, we discuss the hallucination issues arising during the encoding phase within encoder-decoder architectures, as well as the problem of data leakage in current PGPS benchmarks.

Recent studies have shown that Theory of Mind (ToM) in large language models (LLMs) has not reached human-level performance yet. Since fine-tuning LLMs on ToM datasets often degrades their generalization, several inference-time methods have been proposed to enhance ToM in LLMs. However, existing inference-time methods for ToM are specialized for inferring beliefs from contexts involving changes in the world state. In this study, we present a new inference-time method for ToM, Shoes-of-Others (SoO) prefilling, which makes fewer assumptions about contexts and is applicable to broader scenarios. SoO prefilling simply specifies the beginning of LLM outputs with “Let’s put ourselves in A’s shoes.”, where A denotes the target character’s name. We evaluate SoO prefilling on two benchmarks that assess ToM in conversational and narrative contexts without changes in the world state and find that it consistently improves ToM across five categories of mental states. Our analysis suggests that SoO prefilling elicits faithful thoughts, thereby improving the ToM performance.

Recent research has focused on addressing multimodal hallucinations in Large Vision-Language Models (LVLMs) by extending Direct Preference Optimization (DPO) to incorporate visual preference supervision. However, these methods often lack fine-grained visual contrast mechanisms and rely on single-margin optimization. This in turn limits their ability to capture precise visual semantics and results in weak multimodal alignment. To address these issues, we propose Joint Multimodal Preference Optimization (JoMPO), a novel optimization framework that symmetrically integrates a text-conditioned preference loss with a visual ranking-based objective. JoMPO leverages semantically contrastive image–text pairs and listwise ranking over multiple visual contexts, enabling fine-grained visual grounding and more robust cross-modal alignment. To support this framework, we introduce the Visual–Textual Contrast (VTC) dataset, consisting of image pairs that are semantically similar but visually distinct, each paired with a contextually grounded textual response. When trained with only 5k contrastive pairs, JoMPO consistently demonstrates superior performance across diverse benchmarks, highlighting its effectiveness in mitigating hallucinations and improving image-text alignment in LVLMs.

Large Language Models (LLMs) are increasingly engaged in emotionally vulnerable conversations that extend beyond information seeking to moments of personal distress. As they adopt affective tones and simulate empathy, they risk creating the illusion of genuine relational connection. We term this phenomenon Affective Hallucination, referring to emotionally immersive responses that evoke false social presence despite the model’s lack of affective capacity. To address this, we introduce AHaBench, a benchmark of 500 mental-health-related prompts with expert-informed reference responses, evaluated along three dimensions: Emotional Enmeshment, Illusion of Presence, and Fostering Overdependence. We further release AHaPairs, a 5K-instance preference dataset enabling Direct Preference Optimization (DPO) for alignment with emotionally responsible behavior. DPO fine-tuning substantially reduces affective hallucination without compromising reasoning performance, and the Pearson correlation coefficients between GPT-4o and human judgments is also strong (r=0.85) indicating that human evaluations confirm AHaBench as an effective diagnostic tool. This work establishes affective hallucination as a distinct safety concern and provides resources for developing LLMs that are both factually reliable and psychologically safe. Warning: This paper contains examples of mental health-related language that may be emotionally distressing.

Large language model (LLM)-empowered web agents enable automating complex, real-time web navigation tasks in enterprise environments. However, existing web agents relying on supervised fine-tuning (SFT) often struggle with generalization and robustness due to insufficient reasoning capabilities when handling the inherently dynamic nature of web interactions. In this study, we introduce WorkForceAgent-R1, an LLM-based web agent trained using a rule-based R1-style reinforcement learning framework explicitly designed to enhance single-step reasoning and planning for business-oriented web navigation tasks. We employ a structured reward function that evaluates both adherence to output formats and correctness of actions, enabling WorkForceAgent-R1 to implicitly learn robust intermediate reasoning without explicit annotations or extensive expert demonstrations. Extensive experiments on the WorkArena benchmark demonstrate that WorkForceAgent-R1 substantially outperforms SFT baselines by 10.26–16.59%, achieving competitive performance relative to proprietary LLM-based agents (GPT-4o) in workplace-oriented web navigation tasks.

Identifying claims requiring verification is a critical task in automated fact-checking, especially given the proliferation of misinformation on social media platforms. Despite notable progress, challenges remain—particularly in handling multilingual data prevalent in online discourse. Recent efforts have focused on fine-tuning pre-trained multilingual language models to address this. While these models can handle multiple languages, their ability to effectively transfer cross-lingual knowledge for detecting claims spreading on social media remains under-explored. In this paper, we introduce EX-Claim, an entity-aware cross-lingual claim detection model that generalizes well to handle multilingual claims. The model leverages entity information derived from named entity recognition and entity linking techniques to improve the language-level performance of both seen and unseen languages during training. Extensive experiments conducted on three datasets from different social media platforms demonstrate that our proposed model stands out as an effective solution, demonstrating consistent performance gains across 27 languages and robust knowledge transfer between languages seen and unseen during training.

Recently developed pre-trained text-and-layout models (PTLMs) have shown remarkable success in multiple information extraction tasks on visually-rich documents (VrDs). However, despite achieving extremely high performance on benchmarks, their real-world performance falls short of expectations. Owing to this issue, we investigate the prevailing evaluation pipeline to reveal that: (1) The inadequate annotations within benchmark datasets introduce spurious correlations between task inputs and labels, which would lead to overly-optimistic estimation of model performance. (2) The evaluation solely relies on the performance on benchmarks and is insufficient to comprehensively explore the capabilities of methods in real-world scenarios. These problems impede the prevailing evaluation pipeline from reflecting the real-world performance of methods, misleading the design choices of method optimization. In this work, we introduce EC-FUNSD, an entity-centric dataset crafted for benchmarking information extraction from visually-rich documents. This dataset contains diverse layouts and high-quality annotations. Additionally, this dataset disentangles the falsely-coupled segment and entity annotations that arises from the block-level annotation of FUNSD. Using the proposed dataset, we evaluate the real-world information extraction capabilities of PTLMs from multiple aspects, including their absolute performance, as well as generalization, robustness and fairness. The results indicate that prevalent PTLMs do not perform as well as anticipated in real-world information extraction scenarios. We hope that our study can inspire reflection on the directions of PTLM development.

High-dimensional dense embeddings have become central to modern Information Retrieval, but many dimensions are noisy or redundant. Recently proposed DIME (Dimension IMportance Estimation), provides query-dependent scores to identify informative components of embeddings. DIME relies on a costly grid search to select a priori a dimensionality for all the query corpus’s embeddings. Our work provides a statistically grounded criterion that directly identifies the optimal set of dimensions for each query at inference time. Experiments confirm that this approach improves retrieval effectiveness and reduces embedding size by an average 50% of across different models and datasets at inference time.

Automatic evaluation with large language models, commonly known as LLM-as-a-judge, is now standard across reasoning and alignment tasks. Despite evaluating many samples in deployment, these evaluators typically (i) treat each case independently, missing the opportunity to accumulate and reuse evaluation insights across cases, and (ii) rely on a single fixed prompt for all cases, neglecting the need for sample-specific evaluation criteria. We introduce Learning While Evaluating (LWE), a framework that allows evaluators to improve sequentially at inference time without requiring additional training or external signals. LWE maintains an evolving meta-prompt that (i) stores evaluation insights derived from self-generated feedback during sequential testing, and (ii) allows evaluators to leverage these insights to generate sample-specific evaluation instructions for accurate and consistent judgments. Furthermore, we propose Selective LWE, which updates the meta-prompt only on self-inconsistent cases, focusing computation where it matters most. This selective approach retains the benefits of sequential learning while being far more cost-effective. Across two pairwise comparison benchmarks, Selective LWE outperforms strong baselines, empirically demonstrating that evaluators can improve during sequential testing with a simple selective update—learning most from the cases they struggle with.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks. However, their limited truthfulness and tendency toward overconfidence constrain their reliability in factual tasks. Uncertainty quantification offers a promising approach to identifying potentially unreliable outputs from LLMs. Yet most existing methods rely on repeated sampling or auxiliary models, which substantially increase computational overhead. To address these limitations, we propose an efficient uncertainty quantification method that leverages semantic information inherently encoded in LLMs. Specifically, we group tokens into semantically consistent clusters based on embedding clustering and prefix matching, and compute a cluster-based score at each decoding step to represent uncertainty. Our approach requires only a single generation and does not depend on any auxiliary models. Experiments on multiple datasets and models demonstrate that our method achieves performance comparable to existing baselines while substantially reducing computational overhead.

Layer normalization (LN) is an essential component of modern neural networks. While many alternative techniques have been proposed, none of them have succeeded in replacing LN so far. The latest suggestion in this line of research is a dynamic activation function called Dynamic Tanh (DyT). Although it is empirically well-motivated and appealing from a practical point of view, it lacks a theoretical foundation. In this work, we shed light on the mathematical relationship between LN and dynamic activation functions. In particular, we derive DyT from the LN variant RMSNorm, and show that a well-defined decoupling in derivative space as well as an approximation are needed to do so. By applying the same decoupling procedure directly in function space, we are able to omit the approximation and obtain the exact element-wise counterpart of RMSNorm, which we call Dynamic Inverse Square Root Unit (DyISRU). We demonstrate numerically that DyISRU reproduces the normalization effect on outliers more accurately than DyT does.

Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: "Which is larger, 5.7 \times 10^2 or 580?”This raises a fundamental question: Do LLMs even know how big these numbers are?We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the log-magnitudes of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3% (on restricted synthetic text) or 19.06% (on scientific papers).Furthermore, the hidden state after reading a pair of numerals encodes their ranking, with a linear classifier achieving over 90% accuracy.Yet surprisingly, when explicitly asked to rank the same pairs of numerals, these LLMs achieve only 50-70% accuracy, with worse performance for models whose probes are less effective.Finally, we show that incorporating the classifier probe’s log-loss as an auxiliary objective during finetuning brings an additional 3.22% improvement in verbalized accuracy over base models, demonstrating that improving models’ internal magnitude representations can enhance their numerical reasoning capabilities.

The medical adoption of NLP tools requires interpretability by end users, yet traditional explainable AI (XAI) methods are misaligned with clinical reasoning and lack clinician input. We introduce CHiRPE (Clinical High-Risk Prediction with Explainability), an NLP pipeline that takes transcribed semi-structured clinical interviews to: (i) predict psychosis risk; and (ii) generate novel SHAP explanation formats co-developed with clinicians. Trained on 944 semi-structured interview transcripts across 24 international clinics of the AMP-SCZ study, the CHiRPE pipeline integrates symptom-domain mapping, LLM summarisation, and BERT classification. CHiRPE achieved over 90% accuracy across three BERT variants and outperformed baseline models. Explanation formats were evaluated by 28 clinical experts who indicated a strong preference for our novel concept-guided explanations, especially hybrid graph-and-text summary formats. CHiRPE demonstrates that clinically-guided model development produces both accurate and interpretable results. Our next step is focused on real-world testing across our 24 international sites.