EACL 2026 Papers

Interpretability in black-box dense retrievers remains a central challenge in Retrieval-Augmented Generation (RAG). Understanding how queries and documents semantically interact is critical for diagnosing retrieval behavior and improving model design. However, existing dense retrievers rely on static embeddings for both queries and documents, which obscures this bidirectional relationship. Post-hoc approaches such as re-rankers are computationally expensive, add inference latency, and still fail to reveal the underlying semantic alignment. To address these limitations, we propose Interpretable Modular Retrieval Neural Networks (IMRNNs), a lightweight framework that augments any dense retriever with dynamic, bidirectional modulation at inference time. IMRNNs employ two independent adapters: one conditions document embeddings on the current query, while the other refines the query embedding using corpus-level feedback from initially retrieved documents. This iterative modulation process enables the model to adapt representations dynamically and expose interpretable semantic dependencies between queries and documents. Empirically, IMRNNs not only enhance interpretability but also improve retrieval effectiveness. Across seven benchmark datasets, applying our method to standard dense retrievers yields average gains of +6.35% nDCG, +7.14% recall, and +7.04% MRR over state-of-the-art baselines. These results demonstrate that incorporating interpretability-driven modulation can both explain and enhance retrieval in RAG systems.

Multilingual models are widely used for machine translation (MT). However, their effectiveness for extremely low-resource languages (ELRLs) depends critically on how related languages are incorporated during fine-tuning. In this work, we study the role of language mixing directionality, linguistic relatedness, and script compatibility in ELRL translation. We propose SrcMix, a simple source-side mixing strategy that combines related ELRLs during fine-tuning while constraining the decoder to a single target language. Compared to its target-side counterpart TgtMix, SrcMix improves performance by +3 ChrF++ and +5 BLEU in high-resource to ELRL translations, and by +5 ChrF++ and +12 BLEU in mid-resource to ELRL translations. We also release the first Angika MT dataset and provide a systematic comparison of LLM (Aya-101) and NMT (mT5-Large) models under ELRL settings, highlighting the importance of directional mixing and linguistic compatibility.

While large language models (LLMs) excel at generating structured data, such as code, their ability to precisely manipulate it based on instructions remains relatively under-explored. Regular expressions (regexes), critical in practice, are challenging to manipulate. Crucially, the correctness of transformations can be mathematically verified, making them exceptionally well-suited for measuring the symbolic reasoning of LLMs. We introduce Query4Regex, a new benchmark for evaluating verifiable transformations on regexes. Our benchmark tests two query formats: natural language instructions and a program-like domain-specific language (DSL) that specifies the sequence of operations. We evaluate a range of LLMs, verifying semantic correctness through rigorous deterministic finite automata (DFA) equivalence testing. Our empirical studies reveal: 1) the formal DSL significantly outperforms natural language, achieving up to 6.74%p accuracy gains on average. 2) Performance for both formats degrades sharply as compositional complexity increases, highlighting a core challenge in multi-step reasoning. 3) Models often generate plausible but unparsable outputs. Even among parsable outputs, semantic errors remain common, making failures difficult to detect without formal verification. Query4Regex provides a robust framework for analyzing the gap between LLMs’ linguistic fluency and their symbolic reasoning, paving the way for more reliable and verifiable manipulation of formal languages. Our code is available at https://github.com/peer0/Query4Regex.

Hallucinations generated by Large Language Models (LLMs) pose significant challenges for their application to low-resource languages. We present Multi-Hall-SA, a cross-lingual benchmark for hallucination detection spanning English and four low-resource South African languages: isiZulu, isiXhosa, Sepedi, and Sesotho. Derived from government texts, this benchmark categorizes hallucinations into four types aligned with established taxonomies of factual errors: temporal shifts, entity errors, numerical inaccuracies, and location mistakes. Human validation confirms the quality and cross-lingual alignment of our synthetically generated hallucinations. Our cross-lingual alignment methodology enables direct performance comparison between high-resource and low-resource languages, revealing notable gaps in detection capabilities. Evaluation across four state-of-the-art models shows they detect up to 23.6% fewer hallucinations in South African languages compared to English. Knowledge augmentation reduces this disparity, decreasing cross-lingual performance gaps by 59.4% on average. Beyond introducing a validated resource for low-resource languages, Multi-Hall-SA provides a framework for evaluating and improving factual reliability across linguistic boundaries, advancing more inclusive and equitable AI development.

Time series data is ubiquitous across various domains, including manufacturing, finance, and healthcare. High-quality annotations are essential for effectively understanding time series and facilitating downstream tasks. However, obtaining such annotations is challenging, particularly in mission-critical domains. In this paper, we propose TESSA, a multi-agent system designed to automatically generate both general and domain-specific annotations for time series data. TESSA introduces two agents: a general annotation agent and a domain-specific annotation agent. The general agent captures common patterns and knowledge across multiple source domains, leveraging both time-series-wise and text-wise features to generate general annotations. Meanwhile, the domain-specific agent utilizes limited annotations from the target domain to learn domain-specific terminology and generate targeted annotations. Extensive experiments on multiple synthetic and real-world datasets demonstrate that TESSA effectively generates high-quality annotations, outperforming existing methods.

Reinforcement learning (RL) has re-emerged as a natural approach for training interactive LLM agents in real-world environments. However, directly applying the widely used Group Relative Policy Optimization (GRPO) algorithm to multi-turn tasks exposes notable limitations, particularly in scenarios requiring long-horizon reasoning. To address these challenges, we investigate more stable and effective advantage estimation strategies, especially for multi-turn settings. We first explore Proximal Policy Optimization (PPO) as an alternative and find it to be more robust than GRPO. To further enhance PPO in multi-turn scenarios, we introduce turn-PPO, a variant that operates on a turn-level MDP formulation, as opposed to the commonly used token-level MDP. Our results on the WebShop and Sokoban datasets demonstrate the effectiveness of turn-PPO, both with and without long reasoning components.

Large Language Models (LLMs) often default to overly cautious and vague responses when handling sensitive topics, sacrificing helpfulness for safety. Existing evaluation frameworks lack systematic methods to identify and address specific weaknesses in responses to sensitive topics, making it difficult to improve both safety and helpfulness simultaneously. To address this, we introduce FINEST, a FINE-grained response evaluation taxonomy for Sensitive Topics, which breaks down helpfulness and harmlessness into errors across three main categories: Content, Logic, and Appropriateness. Experiments on a Korean-sensitive question dataset demonstrate that our score- and error-based improvement pipeline, guided by FINEST, significantly improves the model responses across all three categories, outperforming refinement without guidance. Notably, score-based improvement—providing category-specific scores and justifications—yields the most significant gains, reducing the error sentence ratio for Appropriateness by up to 33.09%. This work lays the foundation for a more explainable and comprehensive evaluation and improvement of LLM responses to sensitive questions.

In the age of advanced large language models (LLMs), the boundaries between human and AI-generated text are becoming increasingly blurred. We address the challenge of segmenting mixed-authorship text, that is identifying transition points in text where authorship shifts from human to AI or vice-versa, a problem with critical implications for authenticity, trust, and human oversight. We introduce a novel framework, called Info-Mask for mixed authorship detection that integrates stylometric cues, perplexity-driven signals, and structured boundary modeling to accurately segment collaborative human-AI content. To evaluate the robustness of our system against adversarial perturbations, we construct and release an adversarial benchmark dataset Mixed-text Adversarial setting for Segmentation (MAS), designed to probe the limits of existing detectors. Beyond segmentation accuracy, we introduce Human-Interpretable Attribution (HIA) overlays that highlight how stylometric features inform boundary predictions, and we conduct a small-scale human study assessing their usefulness. Across multiple architectures, Info-Mask significantly improves span-level robustness under adversarial conditions, establishing new baselines while revealing remaining challenges. Our findings highlight both the promise and limitations of adversarially robust, interpretable mixed-authorship detection, with implications for trust and oversight in human-AI co-authorship.

Knowledge Graph-enhanced Large Language Models (KG-Enhanced LLMs) integrate the linguistic capabilities of LLMs with the structured semantics of Knowledge Graphs (KGs), showing strong potential in knowledge-intensive reasoning tasks. However, existing methods typically adopt query-driven iterative reasoning from a local perspective, which limits their ability to capture semantically distant but crucial information, leading to dual bottlenecks in efficiency and accuracy for complex multi-hop tasks. To address this issue, we propose MIAoG, a multi-view instructed adaptive reasoning of LLM on KG, which is designed to overcome the limitations of local exploration by enabling LLMs to plan, evaluate, and adapt reasoning paths from a global perspective. Instead of query-anchored exploration, MIAoG first prompts the LLM to generate a multi-view instruction set that outlines diverse potential reasoning paths and explicitly specifies global reasoning intentions to guide the model toward coherent and targeted reasoning. During reasoning, MIAoG integrates a real-time introspection mechanism that evaluates the alignment between the current path and the instructions, adaptively pruning inconsistent trajectories to enhance global consistency while maintaining efficiency. Extensive experiments on multiple public datasets show that MIAoG achieves state-of-the-art performance in KG-enhanced LLM reasoning, particularly excelling in complex multi-hop scenarios.

Large Language Models (LLMs) are increasingly being considered for high-stakes decision-making, yet their application in statistical risk analysis remains largely underexplored. A central challenge in this domain is enabling LLMs to effectively leverage historical data. To address this, we propose novel methods for extracting key information from raw data and translating it into structured contextual input within the LLM prompt. Applying our methods to a case study of power outage risk assessment, we demonstrate that this contextualization strategy significantly improves the LLM’s performance in risk assessment tasks. While the LLM’s prediction performance still does not match that of a standard machine learning model, the LLM-based approach offers distinct advantages in versatility and interpretability. These findings demonstrate a new paradigm for contextualizing data to support risk assessment.

We introduce ChemComp, the first chemistry-focused benchmark for evaluating compositional multi-hop reasoning in large language models (LLMs). Our automated pipeline constructs benchmarks from proprietary or public data by integrating generative reasoning models, chemical named-entity recognition, and external knowledge bases to build knowledge graphs. Applied to recent chemistry literature, this approach minimizes overlap with LLM pretraining data. The resulting dataset comprises 1,188 multi-hop questions, refined through domain-expert feedback and robust evaluation protocols.Using ChemComp, we systematically compare LLM performance with and without retrieval augmentation, including an idealized gold-context scenario. Our results show that even state-of-the-art models struggle with compositional reasoning: retrieval significantly improves accuracy, yet reasoning errors persist even under perfect retrieval. These findings highlight the limitations of current LLMs and the critical role of retrieval-augmented methods in scientific reasoning. Furthermore, our pipeline is generalizable with fine-tuning, enabling the creation of challenging multi-hop reasoning benchmarks across domains and proprietary datasets.

Restoring power distribution networks after disruptions demands rapid, reliable coordination across repair crews, mobile power sources, and switching actions under strict constraints. Classical optimization yields high-quality plans but can be slow, while reinforcement learning often requires feeder-specific training and careful reward shaping. We recast restoration as language-conditioned planning: a large language model generates high-level restoration plans over a compact pre-validated catalogue of feasible actions. This constrained generation design makes decisions reliably, scalably, and interpretably, and allows for real-time human-in-the-loop decision-making while requiring no topology-specific setup or retraining. Our method achieves near-mixed-integer-linear programming performance on the IEEE 13-node standard power distribution feeder and outperforms a time-capped MILP solver on the IEEE 33-node standard feeder by around 13%, while using less than 1% of its wall-clock runtime.

Large language models (LLMs) provide excellent performance, but their practical deployment is limited by the substantial compute and memory demands of large models and the latency of auto-regressive decoding. To mitigate these inefficiencies, block pruning reduces the number of executed transformer blocks, effectively lowering latency while preserving architectural coherence. However, existing methods typically rely on representation similarity or computationally expensive sensitivity analyses to estimate block importance, thereby neglecting task-aware model behavior. To address this limitation, we introduce Task-aware Block Pruning (TaBP), a novel approach that directly captures task-specific inference dynamics by quantifying block-level uncertainty from the statistics of each block’s early-exited output distribution on a calibration dataset. Since output distributions reflect the model’s confidence and decision uncertainty conditioned on downstream tasks, these statistics provide a principled signal for identifying blocks that are less critical for task performance. Extensive experiments demonstrate that TaBP preserves downstream task performance while substantially reducing inference latency and computational cost, without relying on cost-heavy sensitivity analyses. To facilitate reproducibility and further research, we release our implementation of TaBP on GitHub.

Recent studies apply psychometric questionnaires to Large Language Models (LLMs) to assess high-level psychological constructs such as values, personality, moral foundations, and dark traits. Although prior work has raised concerns about possible data contamination from psychometric inventories, which may threaten the reliability of such evaluations, there has been no systematic attempt to quantify the extent of this contamination. To address this gap, we propose a framework to systematically measure data contamination in psychometric evaluations of LLMs, evaluating three aspects: (1) item memorization, (2) evaluation memorization, and (3) target score matching. Applying this framework to 21 models from major families and four widely used psychometric inventories, we provide evidence that popular inventories such as the Big Five Inventory (BFI-44) and Portrait Values Questionnaire (PVQ-40) exhibit strong contamination, where models not only memorize items but can also adjust their responses to achieve specific target scores.

Modern human labor is characterized by specialization; we train for years and develop particular tools that allow us to perform well across a variety of tasks. Similarly, specialized AI agents with task-specific tools or architectures often fail to generalize beyond their intended scope. In this work, we ask: can agents achieve generalizability across diverse domains with a small, but well-chosen set of general tools? We propose OpenHands-Versa, a single-agent system with a modest number of general tools like code execution, search engine, web browser and multimodal file viewer, for three practical domains: software engineering, deep research, and web browsing. Notably, OpenHands-Versa demonstrates superior or competitive performance over task-specific specialized agents on three challenging benchmarks: SWE-Bench Multimodal, GAIA, and The Agent Company, with absolute improvements in success rate of 9.1, 1.3, and 9.1 points, respectively. Thus, our single-agent system can achieve strong generalization indicating that specialist agents for these domains provide no practical benefit. Furthermore, we find that specialist multi-agent systems do not generalize beyond their intended scope. These findings establish OpenHands-Versa as a strong baseline for future research.

Access control is a cornerstone of secure computing, yet large language models often blur role boundaries by producing unrestricted responses. We study role-conditioned refusals, focusing on the LLM’s ability to adhere to access control policies by answering when authorized and refusing when not. To evaluate this behavior, we created a novel dataset that extends the Spider and BIRD text-to-SQL datasets, both of which have been modified with realistic PostgreSQL role-based policies at the table and column levels. We compare three designs: (i) zero or few-shot prompting, (ii) a two-step generator-verifier pipeline that checks SQL against policy, and (iii) LoRA fine-tuned models that learn permission awareness directly. Across multiple model families, explicit verification (the two-step framework) improves refusal precision and lowers false permits. At the same time, fine-tuning achieves a stronger balance between safety and utility (i.e., when considering execution accuracy). Longer and more complex policies consistently reduce the reliability of all systems. We release RBAC-augmented datasets and code.

Test-time compute has emerged as a promising paradigm that enables small language models (SLMs) to achieve large language model (LLM)-level capabilities by allocating additional compute for explicit reasoning during inference. Two common approaches are beam search and Best-of-N sampling. Beam search improves reasoning quality by scoring and optimizing token sequences using Process Reward Models (PRMs), but can incur non-trivial computational overhead and latency. In contrast, Best-of-N executes all reasoning trajectories without PRM guidance, often wasting compute on low-quality trajectories that may have gone astray early in the generation process. To address both inefficiencies, we propose THROW (THink haRd Only When needed)—a hybrid inference pipeline that combines the diversity of Best-of-N with the reasoning trajectory optimization of beam search. THROW introduces a selective branch truncation and expansion mechanism: it generates shorter initial trajectories than Best-of-N and evaluates them using PRMs to classify each query as “easy” or “hard.” Based on this classification, THROW applies branch truncation for easy queries, mimicking Best-of-N, and PRM-guided branch expansion for hard ones, similar to beam search. Evaluations on MATH500, AMC23, and AIME24 demonstrate that THROW achieves 1.54× and 14.38× latency speedups and 35.7% and 80.4% token reductions on average while preserving high reasoning accuracy compared to Best-of-N and Beam Search, respectively.