EACL 2026 Papers

How capable are large language models (LLMs) in the domain of taxation? Although numerous studies have explored the legal domain, research dedicated to taxation remains scarce. Moreover, the datasets used in these studies are either simplified, failing to reflect the real-world complexities, or not released as open-source. To address this gap, we introduce PLAT, a new benchmark designed to assess the ability of LLMs to predict the legitimacy of additional tax penalties. PLAT comprises 300 examples: (1) 100 binary-choice questions, (2) 100 multiple-choice questions, and (3) 100 essay-type questions, all derived from 100 Korean court precedents. PLAT is constructed to evaluate not only LLMs’ understanding of tax law but also their performance in legal cases that require complex reasoning beyond straight forward application of statutes. Our systematic experiments with multiple LLMs reveal that (1) their baseline capabilities are limited, especially in cases involving conflicting issues that require a comprehensive understanding (not only of the statutes but also of the taxpayer’s circumstances), and (2) LLMs struggle particularly with the “AC” stages of “IRAC” even for advanced reasoning models like o3, which actively employ inference-time scaling.

Large language models have simplified the production of personalized translations reflecting predefined stylistic constraints. However, these systems still struggle when stylistic requirements are implicitly represented by a set of examples, such as texts produced by a specific human translator. In this work, we explore various strategies for personalizing automatically generated translations when few examples are available, with a focus on the challenging domain of literary translation. We begin by determining the feasibility of the task and how style information is encoded within model representations. Then, we evaluate various prompting strategies and inference-time interventions for steering model generations towards a personalized style, with a particular focus on contrastive steering with sparse autoencoder (SAE) latents to identify salient personalization properties. We demonstrate that contrastive SAE steering yields robust style conditioning and translation quality, resulting in higher inference-time computational efficiency than prompting approaches. We further examine the impact of steering on model activations, finding that layers encoding personalization properties are impacted similarly by prompting and SAE steering, suggesting a similar mechanism at play.

Document-Level Zero-Shot Relation Extraction (DocZSRE) aims to predict unseen relation labels in text documents without prior training on specific relations. Existing approaches rely on Large Language Models (LLMs) to generate synthetic data for unseen labels, which poses challenges for low-resource languages like Malaysian English. These challenges include the incorporation of local linguistic nuances and the risk of factual inaccuracies in LLM-generated data. This paper introduces Document-Level Zero-Shot Relation Extraction with Entity Side Information (DocZSRE-SI) to address limitations in the existing DocZSRE approach. The DocZSRE-SI framework leverages Entity Side Information, such as Entity Mention Descriptions and Entity Mention Hypernyms, to perform ZSRE without depending on LLM-generated synthetic data. The proposed low-complexity model achieves an average improvement of 11.6% in the macro F1-Score compared to baseline models and existing benchmarks. By utilising Entity Side Information, DocZSRE-SI offers a robust and efficient alternative to error-prone, LLM-based methods, demonstrating significant advancements in handling low-resource languages and linguistic diversity in relation extraction tasks. This research provides a scalable and reliable solution for ZSRE, particularly in contexts like Malaysian English news articles, where traditional LLM-based approaches fall short.

As vision-language models (VLMs) are deployed globally, their ability to understand culturally situated knowledge becomes essential. Yet, existing evaluations largely assess static recall or isolated visual grounding, leaving unanswered whether VLMs possess robust and transferable cultural understanding. We introduce ‘BLEnD-Vis‘, a multimodal, multicultural benchmark designed to evaluate the robustness of everyday cultural knowledge in VLMs across linguistic rephrasings and visual modalities. Building on the BLEnD dataset, ‘BLEnD-Vis‘ constructs 313 culturally grounded question templates spanning 16 regions and generates three aligned multiple-choice formats: (i) a text-only baseline querying from Region \rightarrow Entity, (ii) an inverted text-only variant (Entity \rightarrow Region), and (iii) a VQA-style version of (ii) with generated images. The resulting benchmark comprises 4,916 images and over 21,000 multiple-choice questions (MCQ) instances, validated through human annotation. ‘BLEnD-Vis‘ reveals significant fragility in current VLM cultural knowledge; models exhibit performance drops under linguistic rephrasing. While visual cues often aid performance, low cross-modal consistency highlights the challenges of robustly integrating textual and visual understanding, particularly in lower-resource regions. ‘BLEnD-Vis‘ thus provides a crucial testbed for systematically analysing cultural robustness and multimodal grounding, exposing limitations and guiding the development of more culturally competent VLMs. Code is available at https://github.com/Social-AI-Studio/BLEnD-Vis.

Understanding how large language models (LLMs) process compositional linguistic structures is integral to enhancing their reliability and interpretability. We present Constituent-Aware Pooling (CAP), a methodology grounded in compositionality, mechanistic interpretability, and information theory that intervenes in model activations by pooling token representations into linguistic constituents at various layers. Experiments across eight models (124M-8B parameters) on inverse definition modelling, hypernym and synonym prediction reveal that semantic composition is not localised to specific layers but distributed across network depth. Performance degrades substantially under constituent-based pooling, particularly in early and middle layers, with larger models showing greater sensitivity. We propose an information-theoretic interpretation: transformers’ training objectives incentivise deferred integration to maximise token-level throughput, resulting in fragmented rather than localised composition. These findings highlight fundamental architectural and training constraints requiring specialised approaches to encourage robust compositional processing.

Document-level Joint Entity and Relation Extraction (JERE) benchmarks such as DocRED, Re-DocRED, and DocGNRE suffer from pervasive False Negatives (FN), undermining training and evaluation. In this paper, we introduce SiftingLogic – a training-free annotation pipeline that leverages LLMs with user-specifiable reasoning, enriched inverse/co-occurring relation schemas, and novel entity-level constraints to systematically address FN gaps. Applying SiftingLogic and our enriched schema of inverse and co-occurring relations, we add 29,580 verified triplets to Re-DocRED (train/dev, +27%) and over 9,700 verified triplets to DocGNRE test (+49.89%), yielding the enhanced Re^2-DocRED dataset. Beyond English datasets, we also apply our SiftingLogic to REDFM Mandarin test set, resulting in a significant increase in triplets from 663 to 1,391 (+109.8%) demonstrating our pipeline’s generalisability across languages and datasets. Our experiments show that recall scores of models trained on existing public datasets drop notably on our revised splits, whereas our enriched training set mitigates this, underscoring persistent FN gaps and motivating our proposed SiftingLogic and Re^2-DocRED. To facilitate further research and reproducibility of our work, the Re^2-DocRED dataset is released at https://github.com/klassessg/re2-docred.

Human cognition, driven by complex neurochemical processes, oscillates between imagination and reality and learns to self-correct whenever such subtle drifts lead to hallucinations or unsafe associations. In recent years, Large Language Models (LLMs) have demonstrated remarkable performance in a wide range of tasks. However, they still lack human cognition to balance factuality and safety. Bearing the resemblance, we argue that both factual and safety failures in LLMs arise from a common underlying issue, “representational misalignment” in their latent activation space. We hypothesize that an external network, trained to understand the fluctuations, can selectively intervene in the model to regulate falsehood into truthfulness and unsafe output into safe output without fine-tuning the LLM’s parameters. Reflecting the hypothesis, we propose ARREST (Adversarial Resilient Regulation Enhancing Safety and Truth), a unified framework that identifies and corrects drifted features, engaging both soft and hard refusals in addition to factual corrections. Our empirical results show that ARREST not only regulates misalignment but is also more versatile compared to the Reinforcement Learning from Human Feedback (RLHF)-aligned models in generating soft refusals due to adversarial training. We make our codebase available at https://github.com/sharanya-dasgupta001/ARREST.

Large language models (LLMs) continue to achieve impressive performance on reasoning benchmarks, yet it remains unclear how their predictions capture semantic consistency between sentences. We investigate the important open question of whether word-level surprisal correlates with sentence-level contradiction between a premise and a hypothesis. Specifically, we compute surprisal for hypothesis words across a diverse set of experimental variants, and analyze its association with contradiction labels over multiple datasets and open-source LLMs. Because modern LLMs operate on subword tokens and can not directly produce reliable surprisal estimates, we introduce a token-to-word decoding algorithm that extends theoretically grounded probability estimation to open-vocabulary settings. Experiments show a consistent and statistically significant positive correlation between surprisal and contradiction across models and domains. Our analysis also provides new insights into the capabilities and limitations of current LLMs. Together, our findings suggest that surprisal can localize sentence-level inconsistency at the word level, establishing a quantitative link between lexical uncertainty and sentential semantics. We plan to release our code and data upon publication.

Vision-Language Models excel in multi-modal tasks but often hallucinate objects or exhibit linguistic bias by over-repeating object names, especially in complex multi-object scenes. Existing methods struggle with multi-object grounding because language priors frequently dominate visual evidence, causing hallucinated or biased objects to produce attention distributions or similarity scores nearly indistinguishable from those of real objects. We introduce SHAPLENS, a Shapley value–based attribution framework using Kernel SHAP and multi-layer fusion to detect hallucinated and biased objects. Evaluated on ADE and COCO datasets across four leading VLMs, SHAPLENS improves hallucination detection accuracy by 8–12% and F1 by 10–14% over the best baselines. It also achieves up to 6% higher bias detection performance across three distinct bias types on a curated HQH benchmark and exhibits minimal degradation (<0.03%) across partial and perturbed contexts.

Ghosting, the ability to predict a user’s intended text input for inline query auto-completion, is an invaluable feature for modern search engines and chat interfaces, greatly enhancing user experience. By suggesting completions to incomplete queries (or prefixes), ghosting aids users with slow typing speeds, disabilities, or limited language proficiency. Ghosting is a challenging problem and has become more important with the ubiquitousness of chat-based systems like ChatGPT, Copilot, etc. Despite the increasing prominence of chat-based systems utilizing ghosting, this challenging problem of Chat-Ghosting has received little attention from the NLP/ML research community. There is a lack of standardized benchmarks and relative performance analysis of deep learning and non-deep learning methods. We address this through an open and thorough study of this problem using four publicly available dialog datasets: two human-human (DailyDialog and DSTC7-Ubuntu) and two human-bot (Open Assistant and ShareGPT). We experiment with various existing query auto-completion methods (using tries), n-gram methods and deep learning methods, with and without dialog context. We also propose a novel entropy-based dynamic early stopping strategy. Our analysis finds that statistical n-gram models and tries outperform deep learning based models in terms of both model performance and inference efficiency for seen prefixes. For unseen queries, neural models like T5 and Phi-2 lead to better results. Adding conversational context leads to significant improvements in ghosting quality, especially for Open-Assistant and ShareGPT. We make code and data publicly available at https://github.com/blitzprecision/Chat-Ghosting.

Large language models (LLMs) demonstrate superior reasoning capabilities compared to small language models (SLMs), but incur substantially higher costs. We propose COllaborative REAsoner (COREA), a system that cascades an SLM with an LLM to achieve a balance between accuracy and cost in complex reasoning tasks. COREA first attempts to answer questions using the SLM, which outputs both an answer and a verbalized confidence score. Questions with confidence below a predefined threshold are deferred to the LLM for more accurate resolution. We introduce a reinforcement learning-based training algorithm that aligns the SLM’s confidence through an additional confidence calibration reward. Extensive experiments demonstrate that our method jointly improves the SLM’s reasoning ability and confidence calibration across diverse datasets and model backbones. Compared to using the LLM alone, COREA reduces cost by 21.5% and 16.8% on out-of-domain math and non-math datasets, respectively, with only an absolute pass@1 drop within 2%.

We introduce BeDiscovER (Benchmark of Discourse Understanding in the Era of Reasoning Language Models), an up-to-date, comprehensive suite for evaluating the discourse-level knowledge of modern LLMs. BeDiscovER compiles 5 publicly available discourse tasks across discourse lexicon, (multi-)sentential, and documental levels, with in total 52 individual datasets. It covers both extensively studied tasks such as discourse parsing and temporal relation extraction, as well as some novel challenges such as discourse particle disambiguation (e.g., just), and also aggregates a shared-task on Discourse Relation Parsing and Treebanking for multilingual and multi-framework discourse relation classification. We evaluate open-source LLMs: Qwen3 series, DeepSeek-R1, and frontier reasoning model GPT-5-mini on BeDiscovER, and find that state-of-the-art models exhibits strong performance in arithmetic aspect of temporal reasoning, but they struggle with long-dependency reasoning and some subtle semantic and discourse phenomena, such as rhetorical relation classification.

Parameter-Efficient Fine-Tuning (PEFT) has become a key strategy for adapting large language models, with recent advances in sparse tuning reducing overhead by selectively updating key parameters or subsets of data. Existing approaches generally focus on two distinct paradigms: layer-selective methods aiming to fine-tune critical layers to minimize computational load, and data-selective methods aiming to select effective training subsets to boost training. However, current methods typically overlook the fact that different data points contribute varying degrees to distinct model layers, and they often discard potentially valuable information from data perceived as of low quality. To address these limitations, we propose Gradient-aligned Sparse Tuning (GAST), an innovative method that simultaneously performs selective fine-tuning at both data and layer dimensions as integral components of a unified optimization strategy. GAST specifically targets redundancy in information by employing a layer-sparse strategy that adaptively selects the most impactful data points for each layer, providing a more comprehensive and sophisticated solution than approaches restricted to a single dimension. Experiments demonstrate that GAST consistently outperforms baseline methods, establishing a promising direction for future research in PEFT strategies.

Large-scale vision–language models (LVLMs) have achieved remarkable progress on various reasoning tasks. However, most studies focus on natural photographic images and pay limited attention to multi-panel visual narratives such as comics. This leaves a clear gap in our understanding of how well LVLMs perform chronological reasoning across comic panels. To address this, we introduce ChrOMIC, a new benchmark dataset for chronological reasoning in multi-panel comics. It covers six types of reasoning questions and spans both Western and Japanese comic styles. To ensure high-quality annotations, we customized a human–AI collaborative annotation process tailored to the characteristics of the two comic styles. We further introduce three core tasks: Description Reordering and Panel Reordering, which jointly assess models’ ability to understand chronological order in panel sequences, and Multiple-Choice Question Answering (MCQA), which evaluates narrative-level reasoning. We evaluate a range of open-source and commercial LVLMs on ChrOMIC, and find that even the leading models struggle with panel-based chronological reasoning. Further analysis reveals key limitations, including weak visual action understanding and frequent hallucinations in fine-grained visual interpretation.

Large language models (LLMs) have demonstrated strong performance across diverse natural language processing tasks. However, their performance varies significantly across different prompts, requiring careful engineering for consistent results. Manual prompt engineering requires substantial human effort and suffers from limited reproducibility. In contrast, automatic prompt optimization methods reduce manual effort but often depend on costly autoregressive generation, resulting in substantial latency overheads. To address these limitations, we present low-latency prompt optimization (LLPO), a novel framework that reframes prompt engineering as a classification problem. LLPO classifies structured prompt fields from user input through multi-task classification and populates a predefined template to generate an optimized system prompt with minimal latency. In LLM-based automatic evaluations across four question-answering benchmarks, LLPO improves answer quality by up to 26.5% in ∆win rate compared to prior automatic prompt optimization methods, while reducing latency by up to 1,956 times. Human evaluation shows that LLPO receives the highest proportion of top-ranked responses. Furthermore, we analyze the contribution of each structured prompt field to performance, highlighting the robustness of our framework.

Retrieving the biological impacts of protein-protein interactions (PPIs) is essential for target identification (Target ID) in drug development. Given the vast number of proteins involved, this process remains time-consuming and challenging. Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) frameworks have supported Target ID; however, no benchmark currently exists for identifying the biological impacts of PPIs. To bridge this gap, we introduce the RAG Benchmark for PPIs (RAGPPI), a factual question-answer benchmark of 4,420 question-answer pairs that focus on the potential biological impacts of PPIs. Through interviews with experts, we identified criteria for a benchmark dataset, such as a type of QA and source. We built a gold-standard dataset (500 QA pairs) through expert-driven data annotation. We developed an ensemble auto-evaluation LLM that incorporates expert labeling characteristics, average fact–abstract similarity (F_1), and low-similarity fact counts (F_2), enabling the construction of a silver-standard dataset (3,720 QA pairs). We are committed to maintaining RAGPPI as a resource to support the research community in advancing RAG systems for drug discovery QA solutions.

Multi-modal large language models (MLLMs) have recently shown impressive capabilities but are also highly vulnerable to jailbreak attacks. While white-box methods can generate adversarial visual inputs via gradient-based optimization, such approaches fail in realistic black-box settings where model parameters are inaccessible. Zeroth-order (ZO) optimization offers a natural path for black-box attacks by estimating gradients from queries, yet its application to MLLMs is challenging due to sequence-conditioned objectives, limited feedback, and massive model scales. To address these issues, we propose Zer0-Jack, the first direct black-box jailbreak framework for MLLMs based on ZO optimization. Zer0-Jack focuses on generating malicious images and introduces a patch-wise block coordinate descent strategy that stabilizes gradient estimation and reduces query complexity, enabling efficient optimization on billion-scale models. Experiments show that Zer0-Jack achieves 98.2% success on MiniGPT-4 and 95% on the Harmful Behaviors Multi-modal dataset, while directly jailbreaking commercial models such as GPT-4o. These results demonstrate that ZO optimization can be effectively adapted to jailbreak large-scale multi-modal LLMs. Codes are provided here.

Offensive language detection poses a significant challenge in modern social spaces, necessitating advanced solutions. Online media platforms have been known to escalate acts of violence and broader conflicts, and thus, an automated system to help counter offensive content is essential. Traditional NLP models have typically dominated the field of hate speech detection, but require careful model design and extensive tuning. Moreover, a notable resource gap exists for addressing offensive languages, particularly those transcribed in non-native scripts, such as Roman Urdu and Urdu. This study explores the potential of pre-trained LLMs in using prompt-based methods using different transcriptions of the Urdu language, particularly their efficacy in detecting offensive content in diverse linguistic contexts. Our study employs state-of-the-art open-source LLMs, including advanced variants of Llama, Qwen, Lughaat, and proprietary GPT-4, which are evaluated through prompting strategies in different under-resourced languages. Our findings show that pre-trained LLMs achieve performance comparable to traditional fine-tuned benchmarks in detecting hateful and offensive content.

User simulation has long played a vital role in computer science due to its potential to support a wide range of applications. Language, as the primary medium of human communication, forms the foundation of social interaction and behavior. Consequently, simulating conversational behavior has become a key area of study. Recent advancements in large language models (LLMs) have significantly catalyzed progress in this domain by enabling high-fidelity generation of synthetic user conversation. In this paper, we survey recent advancements in LLM-based conversational user simulation. We introduce a novel taxonomy covering user granularity and simulation objectives. Additionally, we systematically analyze core techniques and evaluation methodologies. We aim to keep the research community informed of the latest advancements in conversational user simulation and to further facilitate future research by identifying open challenges and organizing existing work under a unified framework.