EACL 2026 Papers

Transformers have become the standard in Natural Language Processing (NLP) and Computer Vision (CV) due to their strong performance, yet they remain highly sensitive to small input changes, often referred to as adversarial attacks, such as synonym swaps in text or pixel-level perturbations in images. These adversarial attacks can mislead predictions, while existing defenses are often domain-specific or lack formal robustness guarantees. We propose the Attention-Regularized Transformer (ART), a framework that enhances robustness across modalities. ART builds on the Attention Sensitivity Tensor (AST), which quantifies the effect of input perturbations on attention outputs. By incorporating an AST-based regularizer into training, ART encourages stable attention maps under adversarial perturbations in both text and image tasks. We evaluate ART on IMDB, QNLI, CIFAR-10, CIFAR-100, and Imagenette. Results show consistent robustness gains over strong baselines such as FreeLB and DSRM: up to +36.9\% robust accuracy on IMDB and QNLI, and +5–25% on image benchmarks across multiple Vision Transformer (ViT) architectures, while maintaining or improving clean accuracy. ART is also highly efficient, training over 10\times faster than adversarial methods on text and requiring only 1.25\times the cost of standard training on images, compared to 1.5–5.5\times for recent robust ViTs. Codes are available at https://github.com/cliclab-um6p/ART

Do large language models (LLMs) model linguistic variation? We investigate this question through Hindi-English (Hinglish) verb code-mixing, where speakers can use either a Hindi verb or an English verb with the light verb karna (’do’). Both forms are grammatical, but speakers show unexplained variation in language choice for the verb. We compare human preferences on controlled code-mixed minimal pairs to LLM perplexities spanning families, sizes, and training language compositions. We find that current LLMs do not reliably classify verb language preferences to match native speaker judgments. We also see that with specific supervision, some models do predict human preference to an extent. We release native speaker acceptability judgments on 30 verb pairs, perplexity ratios for 4,279 verb pairs across 7 models, and experimental materials.

Coding remains one of the most fundamental modes of interaction between humans and machines. With the rapid advancement of Large Language Models (LLMs), code generation capabilities have begun to significantly reshape programming practices. This development prompts a central question: Have LLMs transformed code style, and how can such transformation be characterized? In this paper, we present a pioneering study that investigates the impact of LLMs on code style, with a focus on naming conventions, complexity, maintainability, and similarity. By analyzing code from over 20,000 GitHub repositories linked to arXiv papers published between 2020 and 2025, we identify measurable trends in the evolution of coding style that align with characteristics of LLM-generated code. For instance, the proportion of snake_case function names in Python code increased from 40.7% in Q1 2023 to 49.8% in Q3 2025. Furthermore, we investigate how LLMs approach algorithmic problems by examining their reasoning processes. Our experimental results may provide the first large-scale empirical evidence that LLMs affect real-world programming style.

Text embedding models are widely used in natural language processing applications. However, their capability is often benchmarked on tasks that do not require understanding nuanced numerical information in text. As a result, it remains unclear whether current embedding models can precisely encode numerical content, such as numbers, into embeddings. This question is critical because embedding models are increasingly applied in domains where numbers matter, such as finance and healthcare. For example, ”Company X’s market share grew by 2%” should be interpreted very differently from ”Company X’s market share grew by 20%” , even though both indicate growth in market share. This study aims to examine whether text embedding models can capture such nuances. Using synthetic data in a financial context, we evaluate 13 widely used text embedding models and find that they generally struggle to capture numerical details accurately. Our further analyses provide deeper insights into embedding numeracy, informing future research to strengthen embedding model-based NLP systems with improved capacity for handling numerical content.

Existing Chinese preference datasets suffer from limited scale, restricted domain coverage, and insufficiently rigorous data validation. Human annotation significantly limits the scalability of human preference datasets. As a result, Chinese Alignment and Chinese Reward Models (CRM) have not yet been thoroughly explored. To address these challenges, we design an LLM-based data annotation pipeline with no human intervention. Based on this pipeline, we curate COIG-P (Chinese Open Instruction Generalist - Preference), a high-quality, large-scale Chinese preference dataset consisting of 1M Chinese preference pairs and 92k carefully curated Chinese queries across diverse domains, including Chat, Coding, Maths, and others. We conduct experiments to verify the quality of COIG-P from two perspectives. (1) COIG-P brings significant performance improvements for the Qwen2/2.5 and Infinity-Instruct model series on AlignBench through DPO, with gains ranging from 2% to 12%. Furthermore, it significantly outperforms other existing Chinese preference datasets. (2) We train an 8B-sized CRM and manually annotate a Chinese Reward Benchmark (CRBench). Our CRM demonstrates robust scoring ability on CRBench. In addition, in practical data construction experiments, the quality of the data constructed by our CRM is comparable to that produced by GPT-4o.

Current multi-modal benchmarks primarily focus on facts within individual images. However, they overlook the associative relations among multiple images, which necessitate conducting commonsense reasoning grounded in associated knowledge at different granularities (i.e., image-level and entity-level) as well as the ability to perceive the order of images. Therefore, we propose a multi-image relational association task and a meticulously curated Multi-granularity Multi-image Relational Association (MMRA) benchmark, comprising 1,024 samples. To systematically evaluate current LVLMs, we establish a system of associative relations among images that contains 11 subtasks (e.g., UsageSimilarity, SubEvent, etc.) at two granularity levels (i.e., image-level and entity-level), based on relations in ConceptNet. Our experiments reveal that entity-level multi-image perception tasks pose greater challenges for LVLMs than image-level tasks. Moreover, LVLMs perform poorly on spatial-related tasks, indicating limited spatial awareness. Furthermore, we find that LVLMs exhibit weak image order perception capabilities, and we design a method to significantly improve this ability, demonstrating that most current LVLMs do not adequately consider image order perception during pre-training.

We investigate the robustness of fine-tuned Large Language Models (LLMs) for the task of Natural Language Inference (NLI), finding that the in-distribution gains from fine-tuning correspond to a large drop in out-of-distribution (OOD) performance. Despite the widespread use of closed-source LLMs, there are no robustness mitigation methods that work under their API fine-tuning constraints. Existing methods to improve robustness typically require changing the fine-tuning process or large-scale data augmentation, methods that are infeasible or cost prohibitive for closed-source models. To address this, we propose strategically selecting the NLI fine-tuning data, prioritising more complex examples or replacing existing training examples with LLM-generated data. Prioritising more complex training examples improves performance on challenging OOD NLI datasets, while training with synthetic data leads to substantial improvements on easier OOD datasets. We find that synthetic examples are often too simple, and by prompting LLMs to create more complex synthetic data we can improve performance on both easy and challenging OOD datasets. Finally, we show that recent autoregressive LLMs are substantially more robust to distributional shifts compared to encoder models, and should be a preferred baseline for future research.

Understanding user intent in online reviews requires modeling not only explicit aspect ratings but also implicit motivations shaped by contextual factors. Existing large language models (LLMs) often lack structured grounding, fail to capture nuanced intent expression. We propose HII-KG, a two-stage Hierarchical Intent Inference framework that first predicts fine-grained aspect ratings and then generates natural language intent statements, guided by contextual subgraphs retrieved from a domain-specific knowledge graph (KG). We first employ parameter-efficient fine-tuning of LLaMA3.1-8B to predict aspect ratings in an instruction-based format. Moreover, we leverage Cypher-aware prompting to generate user intent from KG summaries. Experiments on a online hotel review dataset show that HII-KG consistently outperforms strong LLM and encoder-based baselines in both aspect classification (avg. F1 +4.5%) and intent generation (BLEU +3.3, ROUGE-L +2.9). The results demonstrate that structured KG integration can significantly enhance fluency, contextual relevance, and factual alignment in user intent inference.

Manga, or Japanese comics, is a richly multimodal narrative form that blends images and text in complex ways. Teaching large multimodal models (LMMs) to understand such narratives at a human-like level could help manga creators reflect on and refine their stories. To this end, we introduce two benchmarks for multimodal manga understanding: MangaOCR, which targets in-page text recognition, and MangaVQA, a novel benchmark designed to evaluate contextual understanding through visual question answering. MangaVQA consists of 526 high-quality, manually constructed question-answer pairs, enabling reliable evaluation across diverse narrative and visual scenarios. Building on these benchmarks, we develop MangaLMM, a manga-specialized model finetuned from the open-source LMM Qwen2.5-VL to jointly handle both tasks. Through extensive experiments, including comparisons with proprietary models such as GPT-4o and Gemini 2.5, we assess how well LMMs understand manga. Our benchmark and model provide a comprehensive foundation for evaluating and advancing LMMs in the richly narrative domain of manga.

The paper extends the Data Movement Distance (DMD) – a metric defined to measure the locality in computer memory – to text by defining a normalized version called nDMD. A key feature of nDMD is a new term designed to better characterize low-frequency tokens. By evaluating nDMD on English subset of the M4 dataset and GenAI detection shared task, the paper shows three key findings. First, nDMD is systematically higher in human-written text than in machine-generated text. Second, nDMD-based features not only outperform frequency baselines but also improve overall performance when combined. Finally, the proposed DMD normalization is more effective in distinguishing human and machine text than alternative normalization approaches.

Large Language Models (LLMs) demonstrate impressive capabilities but exhibit inconsistent performance across diverse domains. We propose DFPE (Diverse Fingerprint Ensemble), a novel training-free method that systematically constructs subject-adaptive ensembles by balancing model diversity and competence. DFPE introduces three key innovations: (1) semantic fingerprinting using averaged response embeddings to capture distinct problem-solving patterns, (2) DBSCAN-based clustering with quantile-based competence filtering to ensure diverse yet capable model selection, and (3) exponentially-weighted aggregation adapted to subject-specific performance. Our method’s effectiveness is highlighted on the challenging MMLU-pro benchmark, where DFPE achieves a striking 17.1 percentage point gain over the best single model, reaching 71.4% accuracy. This strong performance is consistent across other standard benchmarks, with significant accuracy improvements of 4.4 points on AGIEval and 2.7 points on MMLU. Our results underscore that a systematic approach to ensemble construction - one that balances diversity, subject-specific competence, and adaptive weighting, can substantially enhance the generalization and robustness of LLMs on multifaceted language understanding tasks.

Training complexity often scales with the size of hyperparameter space for Large Language Models (LLMs). While Direct Preference Optimization (DPO) offers learning stability through reparameterizing the reward function, its regularization against the reference policy can lead to suboptimal outcomes when the reference policy is not optimal. Recent DPO variants address this concern, but at a cost: they introduce additional hyperparameters, reducing feasibility for LLM fine-tuning. To overcome this challenge, we introduce Implicit policy Regularized Preference Optimization (IRPO), which tackles suboptimality while maintaining training simplicity. By treating the winning policy that generated the chosen responses in a pairwise dataset as an implicit policy, IRPO maximizes KL-regularized reward without extra hyperparameters. Then we propose a novel PO algorithm that directly optimizes the IRPO objective by estimating the likelihood ratio between implicit policies. As the winning policy generally outperforms the reference policy, IRPO can effectively address suboptimality. Our experiments show that IRPO significantly outperforms baseline algorithms with the same hyperparameter complexity. Moreover, IRPO demonstrates comparable performance to recent algorithms that rely on a larger number of hyperparameters, offering a practical solution for scalable LLM fine-tuning.

Logical table-to-text generation aims to generate natural language descriptions that fluently and precisely describe the given table with both surface-level and logic-level fidelity. Although large language models (LLMs) have demonstrated strong capabilities in plain text, their proficiency in interpreting and reasoning tabular data is still limited. In this paper, we are the first to comprehensively explore the performance of various LLMs in the logical table-to-text generation task. However, we find that existing LLMs are difficult to achieve satisfactory results in this task. Even worse, existing prompt strategies cannot cope with complex non-chain logical reasoning scenarios on tables. To address the challenges mentioned above, we constructed a new table-related instruction dataset called LogicTableInstruct and instruction-tuned the open-source LLM on this dataset, resulting in the specialized LLM (LogicTableLLaMA-3.1-8B) for table-related tasks. We also introduced a novel reasoning method, Logic Tree-of-Program (LogicToP), to improve the logical reasoning ability of the LLMs on tables. Our extensive experiments on various LLMs demonstrated that LogicToP can effectively improve the performance of LLMs on this task. Our LogicTableLLaMA-3.1-8B model in the 5-shot LogicToP setting achieves state-of-the-art results on the Logic2Text dataset. The code and data will be released at https://github.com/FXLP/LogToP to boost future work on table-related tasks.

Large language models demonstrate limited capability in proficiency-controlled sentence simplification, particularly when simplifying across large readability levels. We propose a framework that decomposes complex simplifications into manageable steps through dynamic path planning, semantic-aware exemplar selection, and chain-of-thought generation with conversation history for coherent reasoning. Evaluation on five languages across two benchmarks shows our approach improves simplification effectiveness while reducing computational steps. Human evaluation confirms the fundamental trade-off between simplification effectiveness and meaning preservation. Notably, even human annotators struggle to agree on semantic preservation judgments, highlighting the inherent complexity of this task. Our work shows that while step-by-step simplification improves control, preserving semantic fidelity during extensive simplification remains an open challenge.

High quality summarization data remains scarce in under-represented languages. However, historical newspapers, made available through recent digitization efforts, offer an abundant source of untapped, naturally annotated data. In this work, we present a novel method for collecting naturally occurring summaries via Front-Page Teasers, where editors summarize full length articles. We show that this phenomenon is common across seven diverse languages and supports multi-document summarization. To scale data collection, we develop an automatic process, suited to varying linguistic resource levels. Finally, we apply this process to a Hebrew newspaper title, producing HEBTEASESUM, the first dedicated multi-document summarization dataset in Hebrew.

Question generation plays an important role in educational applications, enabling automated assessment and reading comprehension support. Attribute-controlled question generation aims to produce questions that fit predefined characteristics such as difficulty, focus, or coverage. Existing methods predominantly rely on supervised fine-tuning, which often fails to impose a strong adherence to attribute values, resulting in weak coupling between prompt specifications and model outputs. We introduce Odds-Ratio Steerable Optimization (ORSO), a framework designed to enhance attribute sensitivity in question generation models. Building upon preference-based learning techniques without requiring human-curated preference sets, ORSO employs input-level perturbations to create contrastive training signals. Empirical evaluations on both exhaustive and expert-validated attribute configurations indicate that ORSO performs better in enforcing attribute conformity while maintaining output quality. These results argue for the benefits of explicit attribute-aware optimization in controllable question generation tasks.

Large reasoning models (LRMs) increasingly rely on step-by-step Chain-of-Thought (CoT) reasoning to improve task performance, particularly in high-resource languages such as English. While recent work has examined final-answer accuracy in multilingual settings, the thinking traces themselves, i.e., the intermediate steps that lead to the final answer, remain underexplored. In this paper, we present a comprehensive study of multilingual CoT reasoning, evaluating three key dimensions: performance, consistency, and faithfulness. We begin by measuring language compliance, answer accuracy, and answer consistency when LRMs are explicitly instructed or prompt-hacked to think in a target language, revealing strong language preferences and divergent performance across languages. Next, we assess crosslingual consistency of thinking traces by interchanging them between languages. We find that the quality and effectiveness of thinking traces vary substantially depending on the prompt language. Finally, we adapt perturbation-based techniques – i.e., truncation and error injection – to probe the faithfulness of thinking traces across languages, showing that models rely on traces to varying degrees. We release our code and data to support future research.

Deep neural network classifiers for dysarthria impairment severity face limitations regarding interpretability and treatment guidance. To overcome these, we introduce CLINIC-GENIE, an explainable two-stage framework consisting of: (1) CLINIC, a dysarthria severity classification model combining acoustic and speech embeddings with Clinically Explainable Acoustic Features (CEAFs); and (2) GENIE, a module translating CEAFs and their Shapley values into intuitive natural language explanations via a large language model. CLINIC achieved a balanced accuracy of 0.952 (17.3% improvement over using CEAFs alone), and certified speech-language pathologists rated explanations from CLINIC-GENIE with an average fidelity score of 4.94, confirming enhanced clinical utility.