EACL 2026 Papers

Large language models (LLMs) are considered valuable Intellectual Properties (IP) due to the enormous computational cost of training, making their protection against malicious stealing or unauthorized deployment crucial.Despite efforts in watermarking and fingerprinting, existing methods either affect text generation or rely on white-box access, limiting practicality.To address this, we propose DuFFin, a novel Dual-Level Fingerprinting framework for black-box ownership verification.DuFFin jointly extracts trigger patterns and knowledge-level fingerprints to identify the source of a suspect model.We conduct experiments on diverse open-source models, including four popular base LLMs and their fine-tuned, quantized, and safety-aligned variants released by large companies, start-ups, and individuals.Results show that DuFFin accurately verifies the copyright of protected LLMs on their variants, achieving an IP-ROC greater than 0.99.Our code is available at https://github.com/yuliangyan0807/llm-fingerprint.

Argument Mining (AM) aims to identify and interpret argumentative structures in unstructured text, with Argument Component Classification (ACC) as a core task. Despite significant advances, most ACC approaches rely on manually pre-segmented inputs, an assumption that rarely holds in practice due to the high cost and effort of expert human annotation, creating a major bottleneck for scalable AM systems. In this work, we focus on the foundation Argument Component Segmentation (ACS) task by proposing a fine-grained, paired-tag annotation schema that explicitly distinguishes between relevant and surrounding content, thus overcoming the limitations of previous single-separator approaches. Leveraging small and open Large Language Models (LLMs) fine-tuned on our paired-tag annotation schema, we can perform ACS with quality comparable to human expert annotators across multiple benchmark datasets. We further validate our approach on the downstream ACC task, showing that automated segmentation with fine-tuned LLMs yields ACC performances comparable to pipelines relying on human annotations. These findings suggest that reliable automated ACS via LLMs is both feasible and effective, paving the way for more scalable AM pipelines without human intervention.

Language models (LMs) may memorize personally identifiable information (PII) from training data, enabling adversaries to extract it during inference. Existing defense mechanisms such as differential privacy (DP) reduce this leakage, but incur large drops in utility. Based on a comprehensive study using circuit discovery to identify the computational circuits responsible PII leakage in LMs, we hypothesize that specific PII leakage circuits in LMs should be responsible for this behavior. Therefore, we propose PATCH: Privacy-Aware Targeted Circuit Patching, a novel approach that first identifies and subsequently directly edits PII circuits to reduce leakage. PATCH achieves better privacy-utility trade-off than existing defenses, e.g., reducing recall of PII leakage from LMs by up to 65%. Finally, PATCH can be combined with DP to reduce recall of residual leakage of an LM to as low as 0.01%. Our analysis shows that PII leakage circuits persist even after the application of existing defense mechanisms. In contrast, PATCH can effectively mitigate their impact.

Dense large language models (LLMs) face critical efficiency bottlenecks, as they rigidly activate all parameters regardless of input complexity. While existing sparsity methods (static pruning or dynamic activation) partially address this issue, they either lack adaptivity to contextual or model structural demands or incur prohibitive computational overhead. Inspired by the human brain’s dual-process mechanisms — predictive coding (N400) for backbone sparsity and structural reanalysis (P600) for complex contexts — we propose CLADA, a Cognitive-Load-Aware Dynamic Activation framework that synergizes statistical sparsity with semantic adaptability.Our key insight is that LLM activations exhibit two complementary patterns: 1. Global Statistical Sparsity driven by sequence-level prefix information, and 2. Local Semantic Adaptability modulated by cognitive load metrics (e.g., surprisal and entropy). CLADA employs a hierarchical thresholding strategy: a baseline derived from offline error-controlled optimization ensures over 40% sparsity, which is then dynamically adjusted using real-time cognitive signals. Evaluations across six mainstream LLMs and nine benchmarks demonstrate that CLADA achieves 20% average speedup with less than 2% accuracy degradation, outperforming Griffin (over 5% degradation) and TT (negligible speedup).Crucially, we establish the first formal connection between neurolinguistic event-related potential (ERP) components and LLM efficiency mechanisms through multi-level regression analysis (R^2 = 0.17), revealing a sparsity–adaptation synergy. Requiring no retraining or architectural changes, CLADA provides a deployable solution for resource-aware LLM inference while advancing biologically inspired AI design.

We present FLUKE (Framework for LingUistically-driven and tasK-agnostic robustness Evaluation), a framework for assessing model robustness through systematic minimal variations of test data. FLUKE introduces controlled variations across linguistic levels — from orthography to dialect and style — and leverages large language models (LLMs) with human validation to generate modifications. We demonstrate FLUKE’s utility by evaluating both fine-tuned models and LLMs across six diverse NLP tasks (four classification and two generation tasks), and reveal that (1) the impact of linguistic variations is highly task-dependent, with some tests being critical for certain tasks but irrelevant for others; (2) LLMs still exhibit significant brittleness to certain linguistic variations, with reasoning LLMs surprisingly showing less robustness on some tasks compared to base models, and scaling improving robustness only for surface-level modifications; (3) models are overall more brittle to natural, fluent modifications such as syntax or style changes (and especially to negation), compared to corruption-style tests such as letter flipping; (4) the ability of a model to use a linguistic feature in generation does not correlate to its robustness to this feature on downstream tasks. These findings highlight the importance of systematic robustness testing for understanding model behaviors.

Multi-agent debate – multiple instances of large language models discussing problems in turn-based interaction – has shown promise for solving knowledge and reasoning tasks. However, these methods show limitations when solving complex problems that require longer reasoning chains. We analyze how multi-agent debate drifts away from the initial problem over multiple turns, thus harming task performance. We define this phenomenon as problem drift and quantify its presence across ten tasks (i.e., three generative, three knowledge, three reasoning, and one instruction-following task). We find that generative tasks drift often due to the subjectivity of the answer space (76-89%), compared to high-complexity tasks (7-21%). To identify the reasons, eight human experts analyze 170 multi-agent debates suffering from problem drift. We find the most common issues related to this drift are the lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases). We propose DRIFTJudge, an LLM-as-a-judge method, as a first baseline to detect problem drift. We also propose DRIFTPolicy, which mitigates 31% of problem drift cases. Our study is a step toward understanding a key limitation of multi-agent debate, highlighting why longer debates can harm task performance and how problem drift could be addressed.

Large vision-language models (LVLMs) are gaining traction in clinical tasks such as diagnostic support, report generation, and medical question answering. Among post-training techniques, Direct Preference Optimization (DPO) has shown promise in aligning model outputs with human preferences, yet its effectiveness in high-stakes medical contexts remains underexplored. In this work, we present the first systematic evaluation of nine DPO variants applied to two leading medical LVLMs, LLaVA-Med and HuatuoGPT-Vision. We benchmark these models on five curated datasets covering diverse clinical tasks. Evaluations include both automated metrics and expert assessments. Our results show that while DPO improves alignment and reduces severe hallucinations, it yields inconsistent gains over supervised fine-tuning. We further introduce DPO variant that better handles visual misinterpretations and enhances clinical understanding. These findings reveal both the potential and limitations of DPO in medical AI. To support future research, we will release all DPO training data, model checkpoints, and expert annotations upon acceptance.

The growing use of large language models (LLMs) has increased the need for automatic evaluation systems, particularly to address the challenge of information hallucination. Although existing faithfulness evaluation approaches have shown promise, they are predominantly English-focused and often require expensive human-labeled training data for fine-tuning specialized models. As LLMs see increased adoption in multilingual contexts, there is a need for accurate faithfulness evaluators that can operate across languages without extensive labeled data. This paper presents STEMF (Self-Taught Evaluators for Multilingual Faithfulness), a framework that learns exclusively from synthetic multilingual data while leveraging cross-lingual transfer learning. Through experiments comparing language-specific and mixed-language fine-tuning approaches, we demonstrate a consistent relationship between an LLM’s general language capabilities and its performance in language-specific evaluation tasks. Our framework shows improvements over existing baselines, including state-of-the-art English evaluators and machine translation-based approaches.

Designing user-centered LLM systems requires understanding how people use them, but patterns of user behavior are often masked by the variability of queries. In this work, we introduce a new framework to describe request-making that segments user input into request content, roles assigned, query-specific context, and the remaining task-independent expressions. We apply the workflow to create and analyze a dataset of 211k real-world queries based on WildChat. Compared with similar human-human setups, we find significant differences in the language for request-making in the human-LLM scenario. Further, we introduce a novel and essential perspective of diachronic analyses with user expressions, which reveals fundamental and habitual user-LLM interaction patterns beyond individual task completion. We find that query patterns evolve from early ones emphasizing sole requests to combining more context later on, and individual users explore expression patterns but tend to converge with more experience. From there, we propose to understand communal trends of expressions underlying distinct tasks and discuss the preliminary findings. Finally, we discuss the key implications for user studies, computational pragmatics, and LLM alignment.

As large language models (LLMs) are increasingly deployed as multilingual services, keeping their factual knowledge accurate across languages has become both essential and challenging. However, most of the existing knowledge editing (KE) methods are static, in that they update parameters offline for given accumulated edits of knowledge, and are struggling to effectively propagate edits in one language to others, while avoiding side effects. To mitigate this issue, we propose CLICKER, a KE method with stepwise reasoning that dynamically retrieves only knowledge relevant to a given query and then edit, while maintaining cross-lingual consistency through: (1) relevance-aware knowledge retrieval, (2) on-demand in-context KE, and (3) language alignment of the outputs. To rigorously evaluate the locality of edits in cross-lingual KE, we develop Multi-CounterFact dataset that contain many semantically-similar but irrelevant prompts for the edit. Experiments on Multi-CounterFact and MzsRE with both open- and closed-source LLMs confirmed that CLICKER effectively localizes edits and resolves cross-lingual inconsistencies, outperforming dynamic KE baselines.

Recent reinforcement learning (RL)-trained language models have demonstrated strong performance on complex reasoning tasks by producing long and detailed reasoning traces. However, despite these advancements, they often struggle with finding the right balance in reasoning length: some terminate prematurely before reaching a correct answer (underthinking), while others continue reasoning beyond necessity, leading to inefficiency or even degraded accuracy (overthinking).To address these challenges, we propose a method for optimizing reasoning length via self-assessed confidence. By prompting the model to evaluate its own confidence at intermediate reasoning steps, we enable dynamic stopping once sufficient reasoning is achieved.Experiments across multiple reasoning benchmarks show that our approach improves computational efficiency without compromising answer quality. Furthermore, we find that confidence estimates from RL-trained reasoning models are more reliable than those from standard LLMs, making it a valuable internal signal for controlling reasoning depth.

Moderation layers are increasingly a core component of many products built on user- or model-generated content. However, drafting and maintaining domain-specific safety policies remains costly. We present Deep Policy Research (DPR), a minimal agentic system that drafts a full content moderation policy based on only human-written seed domain information. DPR uses a single web search tool and lightweight scaffolding to iteratively propose search queries, distill diverse web sources into policy rules, and organize rules into an indexed document. We evaluate DPR on (1) the OpenAI undesired content benchmark across five domains with two compact reader LLMs and (2) an in-house multimodal advertisement moderation benchmark. DPR consistently outperforms definition-only and in-context learning baselines, and in our end-to-end setting it is competitive with expert-written policy sections in several domains. Moreover, under the same seed specification and evaluation protocol, DPR outperforms a general-purpose deep research system, suggesting that a task-specific, structured research loop can be more effective than generic web research for policy drafting. We release our experiment code at https://github.com/xiaowu0162/deep-policy-research.

Length generalization is the ability of language models to maintain performance on inputs longer than those seen during pretraining. In this work, we introduce a simple yet powerful position encoding (PE) strategy, Random Float Sampling (RFS), that generalizes well to lengths unseen during pretraining or fine-tuning. In particular, instead of selecting position indices from a predefined discrete set, RFS uses randomly sampled continuous values, thereby avoiding out-of-distribution (OOD) issues on unseen lengths by exposing the model to diverse indices during training. Since assigning indices to tokens is a common and fundamental procedure in widely used PEs, the advantage of RFS can easily be incorporated into, for instance, the absolute sinusoidal encoding, RoPE, and ALiBi. Experiments corroborate its effectiveness by showing that RFS results in superior performance in length generalization tasks as well as zero-shot commonsense reasoning benchmarks.

Historically, LLMs have been trained using either autoregressive (AR) or masked language modeling (MLM) objectives, with AR models gaining dominance in recent years. However, AR models are inherently incapable of masked infilling, which is the ability to predict masked tokens between past and future context. In contrast, MLM models suffer from intrinsic computational inefficiencies during both training and inference that hinder their scalability. This work introduces MARIA (Masked and Autoregressive Infilling Architecture), a novel approach that leverages the strengths of both paradigms to achieve state-of-the-art masked infilling performance. MARIA combines a pre-trained MLM and AR model by training a linear decoder that takes their concatenated hidden states as input. This minimal modification enables the AR model to perform infilling while retaining its inherent advantages in terms of faster inference with KV caching. Our results demonstrate that MARIA significantly outperforms existing methods, namely discrete diffusion models, on masked infilling tasks.

Domain-specific Named Entity Recognition (NER) often requires data augmentation due to the scarcity of annotated corpora. Guidance Data Augmentation (GDA), a method utilizing Large Language Models (LLMs) to decompose sentences into abstract components, can lead to over-abstraction, resulting in undefined entity tags and sentences lacking domain-specific vocabulary. In this work, we propose Reflective GDA (R-GDA), a framework that introduces a multi-agent feedback loop to enhance augmentation quality. R-GDA incorporates two distinct agents: a Guidance Refiner (GR), which assesses the initial abstraction to prevent over-generalization, and an Augmentation Calibrator (AC), which validates the final generated sample for domain-fidelity and tag integrity. On the SciERC and NCBI-disease datasets, R-GDA improves F1-Score, validating its effectiveness. Concurrently, it achieves low BERTScore in most cases, indicating greater sentence diversity. For the FIN dataset, it achieves performance comparable to the GDA baseline. R-GDA consistently prevents errors regarding domain-specific tags, demonstrating that the reflective feedback mechanism enhances data fidelity by mitigating critical generation errors.

Safety-critical classification tasks face a persistent challenge: traditional models achieve high overall accuracy but inadequate performance on critical minority classes. We introduce a numbers to narratives framework that transforms tabular data into contextually rich descriptions, enabling language models to leverage pre-trained knowledge for minority class detection. Our approach integrates structured verbalization, linguistically-informed augmentation, and parameter-efficient fine-tuning to address the "minority class blind spot” in high-consequence domains. Using a significantly more efficient model architecture than existing approaches, our framework achieves superior minority class F1-scores: 78.76% for machine failures (+7.42 points over XGBoost), 65.87% for at-risk students (+12.12 points over MLP), and 32.00% for semiconductor failures (+1.01 points over XGBoost, despite 14:1 class imbalance). Our approach also improves overall accuracy by up to 22.43% in five of six datasets while maintaining computational feasibility. Ablation studies confirm that narrative-based verbalization enables effective reasoning about tabular data by contextualizing abstract numerical features. This work provides a practical, resource-efficient approach for enhancing minority class performance in safety-critical domains.

Large language models (LLMs) are often claimed to exhibit reasoning ability when supervised with chain-of-thought (CoT) traces. True reasoning, however, requires invariance: isomorphic problems should yield identical solutions regardless of superficial variation. We test this property by evaluating base and reasoning-optimized models—including LLaMA, Mistral, Qwen, GPT-OSS, and Deepseek—on isomorphic variants from GSM8K and MATH. All models exhibit substantial accuracy drops under perturbation. To assess whether training can induce invariance, we fine-tune models with Program-of-Thought (PoT) supervision under concrete and masked formulations. PoT fine-tuning increases behavioral cross-variant consistency but does not significantly reduce the accuracy gap, and these gains fail to transfer across prompting formats and domains. Our central finding is that models converge toward stable but systematically incorrect behaviors: consistency without correctness. This dissociation suggests that current reasoning supervision teaches models to reproduce solution templates rather than to abstract mathematical structure.

Ranking is a fundamental component in a wide range of AI applications. However, large language models (LLMs) remain unstable on long-context ranking. Sliding-window processing is costly and listwise prompting over full candidates still yields inconsistent orders. We show that sampling alone, even with selection-based methods, cannot stabilize ranking because LLM consistency decomposes into within-list order and cross-list preference, in which a single stochastic process cannot align. To address this, we introduce Self-Sorting (SS), which generates m candidate lists and performs n selection-time re-rankings over those lists. SS fuses explicit within-list positions with implicit cross-list preferences to score entities and return a top-k set. Experimental results on five widely used ranking benchmarks show significant improvements in nDCG@1,5,10, highlighting the critical role of implicit consistency.

Contrastive learning (CL) has achieved remarkable progress in natural language processing (NLP), primarily as a paradigm for pre-training and fine-tuning. However, its potential during the generation phase, particularly in in-context learning (ICL)-based retrieval-augmented summarization, remains largely unexplored. While previous studies have attempted to incorporate negative samples into ICL prompts, these methods do not enforce a true contrastive objective that encourages separation of positive and negative samples in the representation space. In this paper, we first demonstrate through preliminary experiments that small language models (SLMs) can interpret contrastive prompts and effectively distinguish between positive and negative samples during inference, without any parameter updates. Building on these findings, we propose ConRAS, a novel framework that injects contrastive objectives into ICL-based retrieval-augmented summarization. Extensive experiments and in-depth analysis on three summarization benchmarks using four SLMs show that ConRAS consistently outperforms state-of-the-art retrieval-augmented methods, achieving significant improvements in summary quality.