EACL 2026 Papers

Benchmarks for language models have become essential tools for research. Yet, such benchmarks face a persistent contamination problem, with recent studies finding 25-50% of evaluation datasets appearing in training corpora. This is true even looking at the two-player zero-sum game setting, where most benchmarks are based on popular games, like chess, whose optimal strategies are all over the web. Such contamination hinders the possibility to differentiate memorization and reasoning skills. To rectify these problems, we introduce TCG-Bench, a benchmark based on a new two-player trading card game (TCG), similar in spirit to games like Magic: The Gathering. TCG-Bench offers three key innovations: (1) a contamination-resistant design by separating the publicly released game engine from hidden card implementations, (2) a continuous difficulty spectrum via Monte Carlo simulation that prevents benchmark saturation, and (3) a parallel implementation in English and Arabic, the first multilingual text-based game benchmark to do so. We also formalize a practical threat model and refresh protocol that preserves evaluation integrity even if specific cards leak.Our analysis across 17 models (50,000+ games) reveals that performance declines exponentially with difficulty, while model size correlates only weakly with strategic ability. We also observe cross-linguistic performance gaps between English and Arabic, with a gap of 47.4% at 32B, highlighting the need for multilingual game benchmarks that target reasoning capabilities in the target language. We host a leaderboard showcasing these results and welcome evaluation requests on our private cards.

Generative models are known to have reduced performance in different global cultural contexts and languages. While continual data updates have been known to be conducted to improve overall model performance, bolstering and evaluating this cross-cultural competence of generative AI models requires data resources to be intentionally expanded to include global contexts and languages. In this work, we construct a multi-pronged pipeline to collect and contribute culturally salient, multilingual data. We posit that such data can assess the state of the global applicability of our models and thus, in turn, help identify and improve upon cross-cultural gaps.

Large Language Models (LLMs) have been frequently used as automatic annotators for tasks such as Text Emotion Recognition (TER). We consider a scenario in which annotators assign at least one emotion label from a large set of options to a text snippet. For this emotion tagging task, we propose a novel zero-shot algorithm that leverages Best-Worst Scaling (BWS), prompting the LLM to choose the least and most suitable emotions for a given text from several label subsets. The LLM’s choices can be represented by a graph linking labels via worse-than relations. Random walks on this graph yield the final score for each label. We compare our algorithm with naive prompting approaches as well as an established BWS-based method. Extensive experiments demonstrate the suitability of the method. It proves to compare favorably to the benchmarks in terms of both accuracy and calibration with respect to human annotations. Moreover, our algorithm’s automatic annotations are shown to be suitable for finetuning lightweight emotion classification models. The proposed method consumes considerably fewer computational resources than the established BWS approach.

Effectively identifying threats and mitigating their potential damage during crisis situations, such as natural disasters or violent attacks, is paramount for safeguarding endangered individuals. To tackle these challenges, AI has been used to assist humans in emergency situations. Still, the use of NLP techniques remains limited and mostly focuses on classification tasks. The significant potential of timely warning message generation using NLG architectures, however, has been largely overlooked. In this paper, we present CrisiText, the first large-scale dataset for the generation of warning messages across 13 different types of crisis scenarios. The dataset contains more than 400,000 warning messages (spanning almost 18,000 crisis situations) aimed at assisting civilians during and after such events. To generate the dataset, we started from existing crisis descriptions and created chains of events related to the scenarios. Each event was then paired with a warning message. The generations follow expert’s written guidelines to ensure correct terminology and factuality of their suggestions. Additionally, each message is accompanied by three suboptimal variants to allow for the study of different NLG approaches. To this end, we conducted a series of experiments comparing supervised fine-tuning setups with preference alignment, zero-shot, and few-shot approaches. We further assessed model performance in out-of-distribution scenarios and evaluated the effectiveness of an automatic post-editor.

Labeling datasets for African languages poses substantial challenges due to the diverse settings in which annotations are collected, leading to highly variable labeling costs. These costs vary with task complexity, annotator expertise, and data availability. Yet, most active learning (AL) frameworks assume uniform annotation costs, limiting their applicability in real-world, resource-constrained scenarios. To address this, we introduce KnapsackBALD, a novel cost-aware active learning method that integrates the BatchBALD acquisition strategy with a 0-1 Knapsack optimization objective to select informative and budget-efficient samples. We evaluate KnapsackBALD on the MasakhaNEWS dataset, a multilingual news classification benchmark covering 11 African languages. Our method consistently outperforms seven strong active learning baselines, including BALD, BatchBALD, and stochastic sampling variants such as PowerBALD and Softmax-BALD, across all three cost scenarios. The performance gap widens as annotation cost imbalances become more extreme, demonstrating the robustness of KnapsackBALD in different cost settings. These findings show that when annotation costs are explicitly heterogeneous, cost-sensitive acquisition is critical for effective active learning, as demonstrated in African Languages NLP and similar settings. Our code base is open-sourced here.

As Large Language Models (LLMs) increasingly mediate global information access with the potential to shape public discourse, their alignment with universal human rights principles becomes important to ensure that these rights are abided by in high stakes AI-mediated interactions. In this paper, we evaluate how LLMs navigate trade-offs involving the Universal Declaration of Human Rights (UDHR), leveraging 1,152 synthetically generated scenarios across 24 rights articles and eight languages. Our analysis of eleven major LLMs reveals systematic biases where models: (1) accept limiting Economic, Social, and Cultural rights more often than Political and Civil rights, (2) demonstrate significant cross-linguistic variation with elevated endorsement rates of rights-limiting actions in Chinese and Hindi compared to English or Romanian, (3) show substantial susceptibility to prompt-based steering, and (4) exhibit noticeable differences between Likert and open-ended responses, highlighting critical challenges in LLM preference assessment.

Large language models (LLMs) have advanced code generation from single-function tasks to competitive-programming problems, but existing multi-agent solutions either rely on costly large-scale (> 30 B) models or collapse when downsized to small open-source models. We present MapCoder-Lite, a framework for distilling the complex reasoning of large, multi-agent coding systems into a single 7B model. Our contribution is a novel, three-pillar methodology that synergistically generates, refines, and encodes multi-agent knowledge: (i) pass-based trajectory distillation from strong LLMs fixes format fragility in retrieval and reduces failures in debugging, (ii) supervisor-guided correction with global feedback strengthens planning and coding agents, and (iii) agent-wise LoRA fine-tuning delivers memory-efficient specialisation.Comprehensive evaluation on xCodeEval, APPS, and CodeContests shows that MapCoder-Lite more than doubles xCodeEval accuracy (13.2% → 28.3%), eliminates all format failures, while reducing GPU memory and token-generation time by 4\times compared to a 32B model. It also achieves over 10% gains on simpler coding benchmarks, demonstrating broad improvements beyond competitive programming. These results demonstrate that careful agent-wise fine-tuning unleashes high-quality multi-agent coding on a small language model. Our code is publicly available at https://github.com/aiha-lab/MapCoder-Lite.

Automatic Speech Recognition (ASR) performance is heavily dependent on the availability of large-scale, high-quality datasets. For low-resource languages, existing open-source ASR datasets often suffer from insufficient quality and inconsistent annotation, hindering the development of robust models. To address these challenges, we propose a novel and generalizable data aggregation and preprocessing pipeline designed to construct high-quality ASR datasets from diverse, potentially noisy, open-source sources. Our pipeline incorporates rigorous processing steps to ensure data diversity, balance, and the inclusion of crucial features like word-level timestamps. We demonstrate the effectiveness of our methodology by applying it to Vietnamese, resulting in a unified, high-quality 500-hour dataset that provides a foundation for training and evaluating state-of-the-art Vietnamese ASR systems. Our project page is available at https://github.com/qualcomm-ai-research/PhoASR.

How is a factual claim made credible? We propose the novel task of Epistemic Appeal Identification, which identifies whether and how factual statements have been anchored by external sources or evidence. To advance research on this task, we present FactAppeal, a manually annotated dataset of 3,226 English-language news sentences. Unlike prior resources that focus solely on claim detection and verification, FactAppeal identifies the nuanced epistemic structures and evidentiary basis underlying these claims and used to support them. FactAppeal contains span-level annotations which identify factual statements and mentions of sources on which they rely. Moreover, the annotations include fine-grained characteristics of factual appeals such as the type of source (e.g. Active Participant, Witness, Expert, Direct Evidence), whether it is mentioned by name, mentions of the source’s role and epistemic credentials, attribution to the source via direct or indirect quotation, and other features. We model the task with a range of encoder models and generative decoder models in the 2B-9B parameter range. Our best performing model, based on Gemma 2 9B, achieves a macro-F1 score of 0.73.

Recent attempts to leverage large language models (LLMs) for reasoning and pre-trained knowledge in multi-modal reasoning focus on two main approaches: aligning image features with linguistic space, and converting images into textual cues to exploit the implicit reasoning capabilities of LLMs. Although they integrate visual information into the reasoning pipeline, they often treat visual perception and language reasoning as separate processes, limiting the potential for fully unified multi-modal reasoning. In this paper, we propose a novel method, Visual–Linguistic Abductive Reasoning (ViLA), inspired by human abductive reasoning processes. ViLA hypothesizes a plausible answer, generates the corresponding visual and textual premises, and employs fuzzy scoring to select the most coherent combination, thus deriving the final inference. This process integrates visual and linguistic modalities into interpretable abductive reasoning chains, enabling unified multi-modal reasoning. Without fine-tuning LLMs or retrieving external knowledge, ViLA improves performance by 2.31% on AOKVQA, 1.7% on OKVQA, and 1.7% on GQA over previous state-of-the-art models, while also improving interpretability and stability.

Despite significant progress in image captioning, generating accurate and descriptive captions remains a long-standing challenge. In this study, we propose Attention-Guided Image Captioning (AGIC), which amplifies salient visual regions directly in the feature space to guide caption generation. We further introduce a hybrid decoding strategy that combines deterministic and probabilistic sampling to balance fluency and diversity. To evaluate AGIC, we conduct extensive experiments on the Flickr8k, Flickr30k and MSCOCO datasets. The results show that AGIC matches or surpasses several state-of-the-art models while achieving faster inference. Moreover, AGIC demonstrates strong performance across multiple evaluation metrics, offering a scalable and interpretable solution for image captioning.

Tokenizer adaptation plays an important role in adapting pre-trained language models to new domains or languages. In this work, we address two complementary aspects of this process: vocabulary extension and pruning. The common approach to extension trains a new tokenizer on domain-specific text and appends the tokens that do not overlap with the existing vocabulary, which often results in many tokens that are unreachable or never used. We propose continued BPE training that extends a pre-trained tokenizer by continuing the BPE merge learning process on new data. Experiments across multiple languages and model families show that this approach improves tokenization efficiency and leads to better utilization of added vocabulary. We also introduce leaf-based vocabulary pruning, which removes redundant tokens while preserving model quality. Together, these methods provide practical tools for controlled vocabulary modification, which we release as an open-source toolkit.

Recent advancements in large vision-language models (VLMs) have primarily focused on English, with limited attention given to other languages. To address this gap, we introduce MEENA (also known as PersianMMMU), the first dataset designed to evaluate Persian VLMs across scientific, reasoning, and human-level understanding tasks. Our dataset comprises approximately 7,500 Persian and 3,000 English questions, covering a wide range of topics such as reasoning, mathematics, physics, diagrams, charts, and Persian art and literature. Key features of MEENA include: (1) diverse subject coverage spanning various educational levels, from primary to upper secondary school, (2) rich metadata, including difficulty levels and descriptive answers, (3) original Persian data that preserves cultural nuances, (4) a bilingual structure to assess cross-linguistic performance, and (5) a series of diverse experiments assessing various capabilities, including overall performance, the model’s ability to attend to images, and its tendency to generate hallucinations. We hope this benchmark contributes to enhancing VLM capabilities beyond English.

Modern language models (LMs) are trained in an autoregressive manner, conditioned only on the prefix. In contrast, sequence labeling (SL) tasks assign labels to each individual input token, naturally benefiting from bidirectional context. This discrepancy has historically led SL to rely on inherently bidirectional encoder-only models. However, the rapid development of decoder-only models has raised the question of whether they can be adapted to SL. While causal mask removal has emerged as a viable technique for adapting decoder-only models to leverage the full context for SL, it requires considerable changes to the base model functionality. In this work, we explore sequence repetition (SR) as a less invasive alternative for enabling bidirectionality in decoder-only models. Through fine-tuning experiments, we show that SR inherently makes decoders bidirectional, improving the quality of token-level embeddings and surpassing encoders and unmasked decoders. Contrary to earlier claims, we find that increasing the number of repetitions does not degrade SL performance. Finally, we demonstrate that embeddings from intermediate layers are highly effective for SR, comparable to those from final layers, while being significantly more efficient to compute. Our findings underscore that SR alleviates the structural limitations of decoders, enabling more efficient and adaptable LMs and broadening their applicability to other token-level tasks.

Large language models (LLMs) are increasingly applied in clinical decision support, yet current evaluations rarely reveal whether their outputs reflect genuine medical reasoning or superficial correlations. We introduce DeVisE (Demographics and Vital signs Evaluation), a behavioral testing framework that probes fine-grained clinical understanding through controlled counterfactuals. Using intensive care unit (ICU) discharge notes from MIMIC-IV, we construct both raw (real-world) and template-based (synthetic) variants with single-variable perturbations in demographic (age, gender, ethnicity) and vital sign attributes. We evaluate eight LLMs, spanning general-purpose and medical variants, under zero-shot setting. Model behavior is analyzed through (1) input-level sensitivity, capturing how counterfactuals alter perplexity, and (2) downstream reasoning, measuring their effect on predicted ICU length-of-stay and mortality. Overall, our results show that standard task metrics obscure clinically relevant differences in model behavior, with models differing substantially in how consistently and proportionally they adjust predictions to counterfactual perturbations

Recent advances in large language models have enabled mental health dialogue systems, yet existing approaches remain predominantly reactive, lacking systematic user state modeling for proactive therapeutic exploration. We introduce PsyProbe, a dialogue system designed for the exploration phase of counseling that systematically tracks user psychological states through the PPPPPI framework (Presenting, Predisposing, Precipitating, Perpetuating, Protective, Impact) augmented with cognitive error detection. PsyProbe combines State Builder for extracting structured psychological profiles, Memory Construction for tracking information gaps, Strategy Planner for Motivational Interviewing behavioral codes, and Response Generator with Question Ideation and Critic/Revision modules to generate contextually appropriate, proactive questions. We evaluate PsyProbe with 27 participants in real-world Korean counseling scenarios, including automatic evaluation across ablation modes, user evaluation, and expert evaluation by a certified counselor. The full PsyProbe model consistently outperforms baseline and ablation modes in automatic evaluation. User evaluation demonstrates significantly increased engagement intention and improved naturalness compared to baseline. Expert evaluation shows that PsyProbe substantially improves core issue understanding and achieves question rates comparable to professional counselors, validating the effectiveness of systematic state modeling and proactive questioning for therapeutic exploration.

Large language models (LLMs) are increasingly used as evaluators for natural language generation, applying human-defined rubrics to assess system outputs. However, human rubrics are often static and misaligned with how models internally represent language quality. We introduce GER-Eval (Generating Evaluation Rubrics for Evaluation) to investigate whether LLMs can design and use their own evaluation rubrics. We evaluate the semantic coherence and scoring reliability of LLM-defined criteria and their alignment with human criteria. LLMs reliably generate interpretable and task-aware evaluation dimensions and apply them within models, but their scoring reliability degrades in factual and knowledge-intensive settings. Closed-source models such as GPT-4o achieve higher agreement and cross-model generalization than open-weight models such as Llama. Our findings position evaluation as a learned linguistic capability of LLMs—consistent within models but fragmented across them—and call for new methods that jointly model human and LLM evaluative language to improve reliability and interpretability.