EACL 2026 Papers

Autoregressive (AR) models excel at generating temporally coherent audio by producing tokens sequentially, yet they often falter in faithfully following complex textual prompts—especially those describing complex sound events. We uncover a surprising capability in AR audio generators: their early prefix tokens implicitly encode global semantic attributes of the final output, such as event count and sound-object category, revealing a form of implicit planning. Building on this insight, we propose Plan-Critic, a lightweight auxiliary model trained with a Generalized Advantage Estimation (GAE)-inspired objective to predict final instruction-following quality from partial generations. At inference time, Plan-Critic enables guided exploration: it evaluates candidate prefixes early, prunes low-fidelity trajectories, and reallocates computation to high-potential planning seeds. Our Plan-Critic-guided sampling achieves up to a 10 points improvement in CLAP score over the AR baseline—establishing a new state of the art in AR text-to-audio generation—while maintaining computational parity with standard best-of-N decoding. This work bridges the gap between causal generation and global semantic alignment, demonstrating that even strictly autoregressive models can plan ahead.

Rhetorical Role Labeling (RRL) identifies the functional role of each sentence in a document, a key task for discourse understanding in domains such as law and medicine. While hierarchical models capture local dependencies effectively, they are limited in modeling global, corpus-level features. To address this limitation, we propose two prototype-based methods that integrate local context with global representations. Prototype-Based Regularization (PBR) learns soft prototypes through a distance-based auxiliary loss to structure the latent space, while Prototype-Conditioned Modulation (PCM) constructs corpus-level prototypes and injects them during training and inference.Given the scarcity of RRL resources, we introduce SCOTUS-Law, the first dataset of U.S. Supreme Court opinions annotated with rhetorical roles at three levels of granularity: category, rhetorical function, and step. Experiments on legal, medical, and scientific benchmarks show consistent improvements over strong baselines, with \sim4 Macro-F1 gains on low-frequency roles. We further analyze the implications in the era of Large Language Models and complement our findings with expert evaluation.

LLMs, while outperforming humans in a wide range of tasks, can still fail in unanticipated ways. We focus on two pervasive failure modes: (i) hallucinations, where models produce incorrect information about the world, and (ii) the low-resource effect, where the models show impressive performance in high-resource languages like English but the performance degrades significantly in low-resource languages like Bengali. We study the intersection of these issues and ask: do hallucination detectors suffer from the low-resource effect? We conduct experiments on five tasks across three domains (factual recall, STEM, and Humanities). Experiments with four LLMs and three hallucination detectors reveal a curious finding: As expected, the task accuracies in low-resource languages experience large drops (compared to English). However, the drop in detectors’ accuracy is often several times smaller than the drop in task accuracy. Our findings suggest that even in low-resource languages, the internal mechanisms of LLMs might encode signals about their uncertainty. Further, the detectors are robust within language (even for non-English) and in multilingual setups, but not in cross-lingual settings without in-language supervision.

Idioms present a unique challenge for language models due to their non-compositional figurative interpretations, which often strongly diverge from the idiom’s literal interpretation. In this paper, we employ causal tracing to systematically analyze how pretrained causal transformers deal with this ambiguity. We localize three mechanisms: (i) Early sublayers and specific attention heads retrieve figurative interpretation, while suppressing literal interpretation. (ii) When disambiguating context precedes the idiom, the model leverages it from the earliest layer and later layers refine the interpretation if the context conflicts with the retrieved interpretation. (iii) Then, selective, competing pathways carry both interpretations: an intermediate pathway that prioritizes the figurative interpretation and a parallel direct route that favors literal interpretation, ensuring that both readings remain available. Our findings provide mechanistic evidence for idiom comprehension in autoregressive transformers.

Most realistic task automation problems require large language models (LLMs) to call tools, which often return complex JSON responses. These responses must be further processed to derive the information necessary for task completion. The ability of LLMs to do so is under-studied. In this paper, we study the tool response processing task and LLMs’ abilities to process structured (JSON) responses. We created a dataset for this task, and evaluated 15 open and closed weight models using multiple prompting approaches. Our results show that JSON processing remains a difficult task even for frontier models across multiple prompting strategies. The optimal response processing strategy depends on both the nature and size of the tool outputs, as well as the complexity of the required reasoning. Variations in processing approaches can lead to performance differences ranging from 3% to 50%.

Test-time scaling (TTS) has enhanced the performance of Reasoning Models (RMs) on various tasks such as math and coding, yet its efficacy in machine translation (MT) remains underexplored. This paper investigates whether increased inference-time computation improves translation quality. We evaluate 12 RMs across a diverse suite of MT benchmarks spanning multiple domains, examining three scenarios: direct translation, forced-reasoning extrapolation, and post-editing. Our findings show that for general-purpose RMs, TTS provides limited and inconsistent benefits for direct translation, with performance quickly plateauing. However, the effectiveness of TTS is unlocked by domain-specific fine-tuning, which aligns a model’s reasoning process with task requirements, leading to consistent improvements up to an optimal, self-determined reasoning depth. We also find that forcing a model to reason beyond its natural stopping point consistently degrades translation quality. In contrast, TTS proves highly effective in a post-editing context, reliably turning self-correction into a beneficial process. These results indicate that the value of inference-time computation in MT lies not in enhancing single-pass translation with general models, but in targeted applications like multi-step, self-correction workflows and in conjunction with task-specialized models.

The increasing reliance on Large Language Models (LLMs) across various domains extends to education, where students progressively use generative AI as a tool for learning. While prior work has examined LLMs’ mathematical ability, their reliability in grading authentic student problem-solving processes and delivering effective feedback remains underexplored. This study introduces MathEDU, a dataset consisting of student problem-solving processes in mathematics and corresponding teacher-written feedback. We systematically evaluate the reliability of various models across three hierarchical tasks: answer correctness classification, error identification, and feedback generation. Experimental results show that fine-tuning strategies effectively improve performance in classifying correctness and locating erroneous steps. However, the generated feedback across models shows a considerable gap from teacher-written feedback. Critically, the generated feedback is often verbose and fails to provide targeted explanations for the student’s underlying misconceptions. This emphasizes the urgent need for trustworthy and pedagogy-aware AI feedback in education.

Current evaluation of German automatic text simplification (ATS) relies on general-purpose metrics such as SARI, BLEU, and BERTScore, which insufficiently capture simplification quality in terms of simplicity, meaning preservation, and fluency. While specialized metrics like LENS have been developed for English, corresponding efforts for German have lagged behind due to the absence of human-annotated corpora. To close this gap, we introduce DETECT, the first German-specific metric that holistically evaluates ATS quality across all three dimensions of simplicity, meaning preservation, and fluency, and is trained entirely on synthetic large language model (LLM) responses. Our approach adapts the LENS framework to German and extends it with (i) a pipeline for generating synthetic quality scores via LLMs, enabling dataset creation without human annotation, and (ii) an LLM-based refinement step for aligning grading criteria with simplification requirements. To the best of our knowledge, we also construct the largest German human evaluation dataset for text simplification to validate our metric directly. Experimental results show that DETECT achieves substantially higher correlations with human judgments than widely used ATS metrics, with particularly strong gains in meaning preservation and fluency. Beyond ATS, our findings highlight both the potential and the limitations of LLMs for automatic evaluation and provide transferable guidelines for general language accessibility tasks.

Current sentence embedding evaluations typically rely on static test beds like the Massive Text Embedding Benchmark (MTEB). While invaluable, repeated tuning on a fixed suite can inflate reported scores and obscure real-world robustness. We introduce the Paraphrasing Text Embedding Benchmark (PTEB), a dynamic protocol that stochastically generates meaning-preserving paraphrases at evaluation time and aggregates results across multiple runs. Using a cost-efficient LLM-based method grounded in gold ratings and human validation, we show that LLMs generate token-diverse but semantically preserving paraphrases. Across 7 MTEB tasks, we validate our hypothesis that the performance of sentence encoders is sensitive to changes in token space even when semantics remain fixed. We also observe that smaller models are not disproportionately affected relative to larger ones. Our results are statistically robust over multiple runs spanning 20 datasets and 25 languages. More generally, we aim to propose a new evaluation paradigm in NLP that relies less on static, pre-defined benchmarks but shifts towards dynamic, stochastic evaluation leveraging eval-time compute. We make the code to run PTEB publicly available.

Emotion classification plays a significant role in emotion prediction and harmful content detection. Recent advancements in NLP, particularly through large language models (LLMs), have greatly improved outcomes in this field. This study introduces ViGoEmotions - a Vietnamese emotion corpus comprising 20,664 social media comments in which each comment is classified into 27 fine-grained distinct emotions. To evaluate the quality of the dataset and its impact on emotion classification, eight pre-trained Transformer-based models were evaluated under three preprocessing strategies: preserving original emojis with rule-based normalization, converting emojis into textual descriptions, and applying ViSoLex, a model-based lexical normalization system. Results show that converting emojis into text often improves the performance of several BERT-based baselines, while preserving emojis yields the best results for ViSoBERT and CafeBERT. In contrast, removing emojis generally leads to lower performance. ViSoBERT achieved the highest Macro F1-score of 61.50% and Weighted F1-score of 63.26%. Strong performance was also observed from CafeBERT and PhoBERT. These findings highlight that while the proposed corpus can support diverse architectures effectively, preprocessing strategies and annotation quality remain key factors influencing downstream performance.

In this work, we focus on discovery of legal factors for a specific case type under consideration (e.g., vehicle insurance disputes). We refer to these legal factors more explicitly as “Argumentation and Judgement Factors” (AJFs). AJFs encode specific legal knowledge that is important for legal argumentation and judicial decision making. We propose a multi-step approach for discovering a list of AJFs for a given case type using a set of relevant legal documents (e.g., past judgements, relevant acts) and Symbolic Knowledge Distillation (SKD) from a Large Language Model (LLM). We propose a novel geneRatE-CRitic-reviEW (RECREW) prompting strategy for effective SKD. We construct and evaluate the discovered list of AJFs on two different types of cases (auto-insurance and life insurance) and show their utility in a dispute resolution application.

In this work, we propose an entity type-specific and knowledge-augmented token classification framework designed to improve encoder models’ performance on recipe texts. Our empirical analysis shows that this approach achieves state-of-the-art (SOTA) results on 5 out of 7 benchmark recipe datasets, significantly outperforming traditional token classification methods. We introduce a novel methodology leveraging curated domain-specific knowledge contexts to guide encoder models such as BERT and RoBERTa, which we refer to as RecipeBERT-KA and RecipeRoBERTa-KA. Additionally, we release a newly reprocessed entity type-specific and knowledge-enriched dataset that merges seven widely used food datasets, making it the largest annotated food-related dataset to date. Comparative analysis with SOTA large language models (GPT-4o, Mistral-7B, LLaMA 3-13B and LLaMA 3-70B) highlights the practical advantages of our smaller and specialised models. Finally, we analyse the impact of the different knowledge contexts, our models’ potential for transfer learning, the effect of combining the datasets and scenarios where traditional token classification may still perform competitively, offering nuanced insight into method selection.

Multimodal large language models (MLLMs) have shown impressive capabilities in vision-language tasks such as reasoning segmentation, where models generate segmentation masks based on textual queries. While prior work has primarily focused on perturbing image inputs, semantically equivalent textual paraphrases—crucial in real-world applications where users express the same intent in varied ways—remain underexplored. To address this gap, we introduce a novel adversarial paraphrasing task: generating grammatically correct paraphrases that preserve the original query meaning while degrading segmentation performance. To evaluate the quality of adversarial paraphrases, we develop a comprehensive automatic evaluation protocol validated with human studies. Furthermore, we introduce SPARTA—a black-box, sentence-level optimization method that operates in the low-dimensional semantic latent space of a text autoencoder, guided by reinforcement learning. SPARTA achieves significantly higher success rates, outperforming prior methods by up to 2x on both the ReasonSeg and LLMSeg-40k datasets. We use SPARTA and competitive baselines to assess the robustness of advanced reasoning segmentation models. We reveal that they remain vulnerable to adversarial paraphrasing—even under strict semantic and grammatical constraints. All code and data will be released publicly upon acceptance.

Large Language Models (LLMs) often struggle with formal logical reasoning, frequently conflating content plausibility with logical validity. This well-known content effect undermines their capacity to act as reliable deductive reasoners, particularly in multilingual contexts where both linguistic variability and world knowledge may deepen biases. Prior work shows that prompting and tuning interventions can alleviate these issues only partially, leaving models vulnerable to semantic interference.While previous studies have explored activation steering and other test-time interventions, this work has focused predominantly on English.To make reasoning more consistent, robust, and transferable across languages, we investigate the use of activation steering—an inference-time intervention that modulates internal representations towards a cross-lingual reasoning space. Our experiments demonstrate that steering techniques constructed for English-based syllogisms generalise effectively to multilingual datasets, yielding higher formal reasoning accuracy (up to +36%) while minimally affecting language modelling performance. Moreover, steering supports partial transfer to out-of-distribution tasks, highlighting its potential as a scalable mechanism for cross-lingual transferable reasoning. These findings advance the prospect of developing LLMs that can serve as reliable soft reasoners across language landscapes.

Medical multiple-choice question answering (MCQA) benchmarks show that models achieve near-human accuracy, with some benchmarks approaching saturation–leading to claims of clinical readiness. Yet a single accuracy score is a poor proxy for competence: models that change answers under minor input perturbations cannot be considered reliable. We argue that reliability underpins accuracy–only consistent predictions make correctness meaningful. We release ReMedQA, a new benchmark that augments three standard medical MCQA datasets with open-ended answers and systematically perturbed options. Building on this, we introduce ReAcc and ReCon, two reliability metrics: ReAcc measures the proportion of questions answered correctly across all variations, while ReCon measures the proportion answered consistently regardless of correctness. Our evaluation shows that high MCQA accuracy masks low reliability: models remain sensitive to format and perturbation changes, and domain specialization offers no robustness gain. MCQA underestimates smaller models while inflating large ones that exploit structural cues–with some exceeding 50% accuracy even when the original questions are hidden. This shows that, despite near-saturated accuracy, we are not yet done with medical MCQA benchmarks.

Historical language models play a crucial role in the study of languages, and can benefit tasks such as named-entity recognition (NER), part-of-speech (PoS) tagging, and post-OCR correction, among others. Despite their relevance, most efforts have been concentrated on English. To the best of our knowledge, no such model exists for historical Portuguese. In this work, we introduce PortOldBERT, the first historical Portuguese encoder language model. We demonstrate its usefulness by comparing PortOldBERT’s performance with Albertina, the encoder on which it is based, across multiple tasks—pseudo-perplexity, NER, PoS tagging, word error rate (WER) prediction, and OCR error detection—and for different historical periods. PortOldBERT consistently outperforms Albertina in historical data, demonstrating its ability to effectively integrate historical linguistic contexts while retaining the ability to process contemporary text.

Recent advances in multimodal and speech-native large language models (LLMs) have delivered impressive speech recognition, translation, understanding, and question-answering capabilities for high-resource languages. However, African languages and non-native French or English accents remain dramatically underrepresented in benchmarks limiting the understanding and applicability of leading LLMs for millions of francophone and anglophone users in low-resource settings. We presents AfriVox, an open-source benchmark (including novel domain-specific and unscripted datasets) across 20 African languages, African-accented French, Arabic, and 100+ African English accents, contrasting leading multimodal speech LLMs with traditional unimodal automatic speech transcription (ASR) and translation (AST) models. Our analysis reveals significant language coverage variation, surprising LLM translation performance gains (e.g. Gemini), robustness concerns with unscripted speech, and substantial performance disparities for “supported” African languages. We profile the strengths, limitations, and language support of each model, and conduct the first targeted fine-tuning of a modern speech LLM (Qwen2.5-Omni) for three Nigerian languages, exceeding SOTA, and achieving up to 54% relative WER reduction and significant BLEU gains, offering practical guidance for implementers seeking to serve local language users.

Novelty assessment is a central yet understudied aspect of peer review, particularly in high-volume fields like NLP where reviewer capacity is increasingly strained. We present a structured approach for automated novelty evaluation that models expert reviewer behavior through three stages: (i) content extraction from submissions, (ii) retrieval and synthesis of related work, and (iii) structured comparison for evidence-based assessment. Our method is informed by analysis of human-written novelty reviews and captures key patterns such as independent claim verification and contextual reasoning. Evaluated on 182 ICLR 2025 submissions with human-annotated reviewer novelty assessments, the approach achieves 86.5% alignment with human reasoning and 75.3% agreement on novelty conclusions, substantially outperforming existing LLM-based baselines. It produces detailed, literature-aware analysis and improves consistency over ad hoc reviewer judgments. These results highlight the potential for structured LLM-assisted approaches to support more rigorous and transparent peer review without displacing human expertise. The data and the code are available at https://ukplab.github.io/eacl2026-assessing-paper-novelty/

Supervised learning relies on data annotation which usually is time-consuming and therefore expensive. A longstanding strategy to reduce annotation costs is active learning, an iterative process, in which a human annotates only data instances deemed informative by a model. Research in active learning has made considerable progress, especially with the rise of large language models (LLMs). However, we still know little about how these remarkable advances have translated into real-world applications, or contributed to removing key barriers to active learning adoption. To fill in this gap, we conduct an online survey in the NLP community to collect previously intangible insights on current implementation practices, common obstacles in application, and future prospects in active learning. We also reassess the perceived relevance of data annotation and active learning as fundamental assumptions. Our findings show that data annotation is expected to remain important and active learning to stay highly relevant while benefiting from LLMs. Consistent with a community survey from over 15 years ago, three key challenges yet persist—setup complexity, uncertain cost reduction, and tooling—for which we propose alleviation strategies. We publish an anonymized version of the dataset.