EACL 2026 Papers

As a digraphic language, the Persian language utilizes two written standards: Perso-Arabic in Afghanistan and Iran, and Tajik-Cyrillic in Tajikistan. Despite the significant similarity between the dialects of each country, script differences prevent simple one-to-one mapping, hindering written communication and interaction between Tajikistan and its Persian-speaking “siblings”. To overcome this, previously-published efforts have investigated machine transliteration models to convert between the two scripts. Unfortunately, most efforts did not use datasets other than those they created, limiting these models to certain domains of text such as archaic poetry or word lists. A truly usable transliteration system must be capable of handling varied domains, meaning that suck models lack the versatility required for real-world usage. The contrast in domain between data also obscures the task’s true difficulty. We present a new state-of-the-art sequence-to-sequence model for Tajik-Farsi transliteration trained across all available datasets, and present two datasets of our own. Our results across domains provide clearer understanding of the task, and set comprehensive comparable leading benchmarks. Overall, our model achieves chrF++ and Normalized CER scores of 87.91 and 0.05 from Farsi to Tajik and 92.28 and 0.04 from Tajik to Farsi. Our model, data, and code are available at https://github.com/merchantrayyan/ParsTranslit.

Fact-checking on major platforms, such as X, Meta, and TikTok, is shifting from expert-driven verification to a community-based setup, where users contribute explanatory notes to clarify why a post might be misleading. An important challenge here is determining whether an explanation is helpful for understanding real-world claims and the reasons why, which remains largely underexplored in prior research. In practice, most community notes remain unpublished due to slow community annotation, and the reasons for helpfulness lack clear definitions. To bridge these gaps, we introduce the task of predicting both the helpfulness of explanatory notes and the reason for this. We present COMMUNTYNOTES, a large-scale multilingual dataset of 104k posts with user-provided notes and helpfulness labels. We further propose a framework that automatically generates and improves reason definitions via automatic prompt optimization, and integrate them into prediction. Our experiments show that the optimized definitions can improve both helpfulness and reason prediction. Finally, we show that the helpfulness information is beneficial for existing fact-checking systems. The code and the data are available at https://github.com/ruixing76/Helpfulness-FCExp.

With the growing use of large language models (LLMs) as evaluators, their application has expanded to code evaluation tasks, where they assess the correctness of generated code without relying on reference implementations. While this offers scalability and flexibility, it also raises a critical, unresolved question: Can LLM judges fairly and robustly evaluate semantically equivalent code with superficial variations? Functionally correct code often exhibits variations—such as differences in variable names, comments, or formatting—that should not influence its correctness. Yet, whether LLM judges can reliably handle these variations remains unclear. We present the first comprehensive study of this issue, defining six types of potential bias in code evaluation and revealing their systematic impact on LLM judges. Across five programming languages and multiple LLMs, we empirically demonstrate that all tested LLM judges are susceptible to both positive and negative biases, resulting in inflated or unfairly low scores. Moreover, we observe that LLM judges remain vulnerable to these biases even when prompted to generate test cases before scoring, highlighting the need for more robust code evaluation method.

Contemporary advancements in language model reasoning typically require computationally intensive reinforcement learning (RL) and massive datasets, creating barriers for resource-constrained teams. In this work, we demonstrate that high-quality, iterative training on minimal data can rival modern RL approaches. We introduce a resource-efficient framework that combines Direct Preference Optimization (DPO) and Supervised Fine-Tuning (SFT) with selective guidance from larger models, iteratively refining solutions through a “reflect, rewrite, repeat” cycle (R^3). Using Qwen 2.5 7B and Qwen 2.5 Math 7B as base models, our method shows meaningful performance improvements across arithmetic, symbolic and cognitive reasoning benchmarks—including GSM8K (83.1% → 88.6%), AIME’25@10 (20.0% → 30.0%) and LastLetterConcat (40.7% → 53.3%) problems. The model-agnostic nature of our R^3 framework is further demonstrated through substantial improvements when applied to Mistral and LLaMA-based models. Remarkably, these gains are achieved using mere 700 basic arithmetic training samples, in stark contrast to the hundreds of thousands of examples typically required by RL-based systems. Our results suggest that reasoning improvements need not strictly depend on large-scale data. By emphasizing strategically curated training grounded in foundational principles, we achieve competitive generalization with minimal resource overhead. Our R^3 pipeline also generates high-quality SFT data with high-fidelity reasoning traces as byproduct, further enabling scalable and annotation-free fine-tuning. Code is available.[https://github.com/aws-samples/sample-for-reflect-rewrite-repeat]

Instruction pre-training (IPT) has recently emerged as an effective intermediate stage between vanilla pre-training and post-training for large language models (LLMs). However, the optimal design of IPT corpora—such as the balance between raw and instruction-response data, languages, and task categories—remains unclear. We systematically study IPT corpus composition using a bilingual (English and Japanese) and multi-category (coding, general, math, and reasoning) instruction-response dataset. Through extensive IPT experiments across four base models, including both English-centric and bilingual LLMs, we find that: (1) more instruction-response data generally enhances model performance, particularly for models with large VPT budgets; (2) Japanese instruction data can improve English performance through cross-lingual transfer; and (3) the effectiveness of post-training varies across categories: coding performance is largely determined during IPT, while math and reasoning continue to improve during post-training.

Agentic AI has significantly extended the capabilities of large language models (LLMs) by enabling complex reasoning and tool use. However, most existing frameworks are tailored to domains such as mathematics, coding, or web automation, and fall short on geospatial tasks that require spatial reasoning, multi-hop planning, and real-time map interaction. To address these challenges, we introduce MapAgent, a hierarchical multi-agent plug-and-play framework with customized toolsets and agentic scaffolds for map-integrated geospatial reasoning. Unlike existing flat agent-based approaches that treat tools uniformly—often overwhelming the LLM when handling similar but subtly different geospatial APIs—MapAgent decouples planning from execution. A high-level planner decomposes complex queries into subgoals, which are routed to specialized modules. For tool-heavy modules—such as map-based services—we then design a dedicated map-tool agent that efficiently orchestrates related APIs adaptively in parallel to effectively fetch geospatial data relevant for the query, while simpler modules (e.g., solution generation or answer extraction) operate without additional agent overhead. This hierarchical design reduces cognitive load, improves tool selection accuracy, and enables precise coordination across similar APIs. We evaluate MapAgent on four diverse geospatial benchmarks—MapEval-Textual, MapEval-API, MapEval-Visual, and MapQA—and demonstrate substantial gains over state-of-the-art tool-augmented and agentic baselines.

The strong capabilities of recent Large Language Models (LLMs) have made them highly effective for zero-shot re-ranking task. Attention-based re-ranking methods, which derive relevance scores directly from attention weights, offer an efficient and interpretable alternative to generation-based re-ranking methods. However, they still face two major limitations. First, attention signals are highly concentrated a small subset of tokens within a few documents, making others indistinguishable. Second, attention often overemphasizes phrases lexically similar to the query, yielding biased rankings that irrelevant documents with mere lexical resemblance are regarded as relevant. In this paper, we propose ReAttn, a post-hoc re-weighting strategy for attention-based re-ranking methods. It first compute the cross-document IDF weighting to down-weight attention on query-overlapping tokens that frequently appear across the candidate documents, reducing lexical bias and emphasizing distinctive terms. It then employs entropy-based regularization to mitigate over-concentrated attention, encouraging a more balanced distribution across informative tokens. Both adjustments operate directly on existing attention weights without additional training or supervision. Extensive experiments demonstrate the effectiveness of our method.

Large Language Models (LLMs) have achieved impressive capabilities in various context-based text generation tasks, such as summarization and reasoning; however, their applications in intention-based generation tasks remain underexplored. One such example is revision generation, which requires the generated text to explicitly reflect the writer’s actual intentions. Identifying intentions and generating desirable revisions are challenging due to their complex and diverse nature. Although prior work has employed LLMs to generate revisions with few-shot learning, they struggle with handling entangled multi-intent scenarios. While fine-tuning LLMs using intention-based instructions appears promising, it demands large amounts of annotated data, which is expensive and scarce in the revision community. To address these challenges, we propose Intention-Tuning, an intention-adaptive layer-wise LLM fine-tuning framework that dynamically selects a subset of LLM layers to learn the intentions and subsequently transfers their representations to revision generation. Experimental results suggest that Intention-Tuning is effective and efficient on small revision corpora, outperforming several PEFT baselines.

This paper discusses the internal behavior of Transformer language models.Many recent pre-trained models have been reported to exhibit only slight changes in the angular distance between the input and output hidden state vectors in the middle Transformer layers, despite a disproportionately large “jump” in the angular distance occurring in or around the final Transformer layer.To characterize this, we first introduce a quantitative metric for the jump strength around the final layer, and then demonstrate its prevalence across many open-weight models, as well as its amplification throughout pre-training.Assuming such jumps indicate an undesirable property, we propose the jump-suppressing regularizer (JREG) which penalizes this jump during pre-training, thereby encouraging more balanced capability usage across the middle layers.Empirical evaluations of three model sizes of Llama-based models, trained with the proposed JREG method, reveal improved task performance compared to the baseline without altering the model architecture.

Cultural evaluation of large language models has become increasingly important, yet current benchmarks often reduce culture to static facts or homogeneous values. This view conflicts with anthropological accounts that emphasize culture as dynamic, historically situated, and enacted in practice. To analyze this gap, we introduce a four-part framework that categorizes how benchmarks frame culture, such as knowledge, preference, performance, or bias. Using this lens, we qualitatively examine 20 cultural benchmarks and identify six recurring methodological issues, including treating countries as cultures, overlooking within-culture diversity, and relying on oversimplified survey formats. Drawing on established anthropological methods, we propose concrete improvements: incorporating real-world narratives and scenarios, involving cultural communities in design and validation, and evaluating models in context rather than isolation. Our aim is to guide the development of cultural benchmarks that go beyond static recall tasks and more accurately capture the responses of the models to complex cultural situations.

Open Domain Multi-hop Question Answering faces a dual compositionality challenge: reasoning over complex query structures and integrating evidence scattered across contexts. Despite recent advancements in Graph-based Retrieval-Augmented Generation (GraphRAG), persistent limitations in complex reasoning and retrieval inaccuracies continue to constrain the efficacy of multi-hop QA systems. We introduce HiGraAgent, a framework that unifies graph-based retrieval with adaptive reasoning. It constructs a Hierarchical Knowledge Graph (HiGra) with entity alignment, reducing redundancy by 34.5% while preserving expressiveness; employs HiGraRetriever, a hybrid graph-semantic retriever that consistently outperforms the strongest graph-based method across benchmarks; and integrates a dual-agent adaptive reasoning protocol where a Seeker and a Librarian dynamically coordinate retrieval and reasoning. Together, these innovations enable HiGraAgent to achieve 85.3% average accuracy on HotpotQA, 2WikiMultihopQA, and MuSiQue, surpassing the strongest prior system by 11.7%. Our results highlight the importance of reframing multi-hop QA as a problem of adaptive reasoning, offering a more robust and flexible paradigm for complex information seeking.

Prompts are the interface for eliciting the capabilities of large language models (LLMs). Understanding their structure and components is critical for analyzing LLM behavior and optimizing performance. However, the field lacks a comprehensive framework for systematic prompt analysis and understanding. We introduce PromptPrism, a linguistically-inspired taxonomy that enables prompt analysis across three hierarchical levels: functional structure, semantic component, and syntactic pattern. By applying linguistic concepts to prompt analysis, PromptPrism bridges traditional language understanding and modern LLM research, offering insights that purely empirical approaches might miss. We show the practical utility of PromptPrism by applying it to three applications: (1) a taxonomy-guided prompt refinement approach that automatically improves prompt quality and enhances model performance across a range of tasks; (2) a multi-dimensional dataset profiling method that extracts and aggregates structural, semantic, and syntactic characteristics from prompt datasets, enabling comprehensive analysis of prompt distributions and patterns; (3) a controlled experimental framework for prompt sensitivity analysis by quantifying the impact of semantic reordering and delimiter modifications on LLM performance. Our experimental results validate the effectiveness of our taxonomy across these applications, demonstrating that PromptPrism provides a foundation for refining, profiling, and analyzing prompts.

Understanding cultural heritage artifacts such as ancient Greek pottery requires expert-level reasoning that remains challenging for current MLLMs due to limited domain-specific data. We introduce VaseVQA, a benchmark for ancient Greek pottery, primarily vases, consisting of 31,773 images and 67,614 question–answer pairs across seven expert-defined categories, enabling systematic evaluation of expert-level cultural heritage understanding. Using this dataset, we explore effective training strategies for domain-specific reasoning. While supervised fine-tuning improves adaptation to domain knowledge, it struggles with deeper reasoning tasks. We propose VaseVL, which augments SFT with reinforcement learning using verifiable rewards. Experiments show that VaseVL consistently outperforms supervised baselines, especially on reasoning-intensive questions, highlighting the value of targeted reinforcement learning for cultural heritage visual question answering.

This study investigates the internal information flow of large language models (LLMs) while performing chain-of-thought (CoT) style reasoning.Specifically, with a particular interest in the faithfulness of the CoT explanation to LLMs’ final answer, we explore (i) when the LLMs’ answer is (pre)determined, especially before the CoT begins or after, and (ii) how strongly the information from CoT specifically has a causal effect on the final answer.Our experiments with controlled arithmetic tasks reveal a systematic internal reasoning mechanism of LLMs.They have not derived an answer at the moment when input was fed into the model.Instead, they compute (sub-)answers while generating the reasoning chain on the fly.Therefore, the generated reasoning chains can be regarded as faithful reflections of the model’s internal computation.

Direct Preference Optimization (DPO) is a powerful approach for aligning large language models (LLMs) with human preferences by formulating preference learning as a supervised classification problem over pairwise human-labeled outputs, thereby enabling stable and efficient training. We show that DPO inherits bias from confounders (e.g., topic, style, user objectives) that shape data generation and carry through to training, hindering recovery of true human preferences. We address this from a causal perspective, proposing Causal Direct Preference Optimization (CDPO), a general framework that incorporates causal inference principles to mitigate the influence of confounders and sharpen the signal of genuine human preferences. Our approach preserves the tractability of direct optimization while enhancing robustness to spurious correlations and annotation biases. Empirical evaluations on benchmark datasets show that CDPO surpasses DPO-based baselines by achieving unbiased fine-tuning through causal reasoning, confirming the effectiveness of confounder-aware preference optimization.

Activation steering or editing hidden states to control language-model behavior can be framed as a causal mediation problem: inputs induce internal activations, a subset of which act as mediators transmitting targeted behaviors to outputs. We formalize a structural graph over transformer layers and derive front-door—style identification conditions that justify steering through mediating subspaces while preserving non-mediating features, thereby reducing confounding and off-target effects. Within this mediation-first view, we present CAS-BiPO, a sparse mediation steering approach that learns targeted behavioral interventions via regularized training. Empirically, our method achieves 97-100% of dense baseline effectiveness across four behavioral control tasks while using only 10-30% of activation dimensions. Learned masks concentrate 94.3% of steering effects in 26.7% of dimensions, with neurons exhibiting 2.2\times higher activation changes, validating the sparse mediation hypothesis. Our causal framework provides theoretical grounding while CAS-BiPO demonstrates that end-to-end learning of interpretable, reliable interventions is both feasible and advantageous.

Retrieval-augmented generation (RAG) can substantially enhance the performance of LLMs on knowledge-intensive tasks. Various RAG paradigms—including vanilla, planning-based, and iterative RAG—all depend on a robust retriever, yet existing retrievers rely heavily on public knowledge and often falter when faced with domain-specific queries. To address these limitations, we introduce DRAGON, a framework that combines a data-construction modeling approach with a scalable synthetic data-generation pipeline, specifically designed to optimize domain-specific retrieval performance and bolster retriever robustness. To evaluate RAG performance on domain-specific RAGs, we propose DRAGONBench, a benchmark spanning 8 domain-specific document collections across 4 distinct fields and featuring a wide spectrum of query complexities, answerability, and hops. Leveraging DRAGON, we generate a large-scale synthetic dataset—encompassing both single-hop and multi-hop queries—to enrich retriever training. Extensive experiments demonstrate that retrievers trained on this data yield significant performance gains and exhibit strong cross-domain generalization. Moreover, when our optimized retrievers are integrated into vanilla, planning-based, and iterative RAG paradigms, we observe consistent end-to-end improvements in system accuracy.

The performance of large language models (LLMs) tends to degrade for extremely low-resource languages, primarily due to the lack of labeled training data. Despite growing interest, the availability of high-quality natural language processing (NLP) datasets for these languages remains limited. This paper addresses such gap by focusing on Ladin, an endangered Romance language, specifically the Val Badia variant. Leveraging a small set of parallel Ladin–Italian sentence pairs, we create synthetic datasets for sentiment analysis and question answering by translating monolingual Italian data. To ensure linguistic quality, we apply rigorous filtering and back-translation procedures in our method. We further demonstrate that incorporating these synthetic datasets into machine translation training leads to substantial improvements over existing Italian–Ladin translation baselines. Our contributions include sentiment analysis and question answering datasets for Ladin, establishing foundational resources that support broader NLP research and downstream applications for underrepresented languages.