EACL 2026 Papers

Natural language processing (NLP) now shapes many aspects of our world, yet its potential for positive social impact is underexplored. This paper surveys work in “NLP for Social Good" (NLP4SG) across nine domains relevant to global development and risk agendas, summarizing principal tasks and challenges. We analyze ACL Anthology trends, finding that inclusion and AI harms attract the most research, while domains such as poverty, peacebuilding, and environmental protection remain underexplored. Guided by our review, we outline opportunities for responsible and equitable NLP and conclude with a call for cross-disciplinary partnerships and human-centered approaches to ensure that future NLP technologies advance the public good.

Administrative phone tasks drain roughly $1 trillion annually from U.S. healthcare, with over 500 million insurance–benefit verification calls manually handled in 2024. We introduce INSURE-Dial, to our knowledge the first public benchmark for developing and assessing compliance-aware voice agents for phase-aware call auditing with span-based compliance verification. The corpus includes 50 de-identified, AI-initiated calls with live insurance representatives (mean 71 turns/call) and 1,000 synthetically generated calls that mirror the same workflow. All calls are annotated with a phase-structured JSON schema covering IVR navigation, patient identification, coverage status, medication checks (up to two drugs), and agent identification (CRN), and each phase is labeled for Information and Procedural compliance under explicit ask/answer logic. We define two novel evaluation tasks: (1) Phase Boundary Detection (span segmentation under phase-specific acceptance rules) and (2) Compliance Verification (IC/PC decisions given fixed spans). Per-phase scores are strong across small, low-latency baselines, but end-to-end reliability is constrained by span-boundary errors. On real calls, full-call exact segmentation is low, showing a gap between conversational fluency and audit-grade evidence.

Large language models (LLMs) are increasingly used for long-document question answering, where reliable attribution to sources is critical for trust. Existing post-hoc attribution methods work well for extractive QA but struggle in multi-hop, abstractive, and semi-extractive settings, where answers synthesize information across passages. To address these challenges, we argue that post-hoc attribution can be reframed as a reasoning problem, where answers are decomposed into constituent units, each tied to specific context. We first show that prompting models to generate such decompositions alongside attributions improves performance. Building on this, we introduce DecompTune, a post-training method that teaches models to produce answer decompositions as intermediate reasoning steps. We curate a diverse dataset of complex QA tasks, annotated with decompositions by a strong LLM, and post-train Qwen-2.5 (7B and 14B) using a two-stage SFT + GRPO pipeline with task-specific curated rewards. Across extensive experiments and ablations, DecompTune substantially improves attribution quality, outperforming prior methods and matching or exceeding state-of-the-art frontier models.

Test-time compute methods can significantly improve the reasoning capabilities and problem-solving accuracy of large language models. However, these approaches require substantially more computational resources, with most computation wasted on exploring low-diversity branches where the model already exhibits high confidence. We observe that a small subset of uncertain reasoning steps has a disproportionately large impact on final prediction accuracy, and branching at these points tends to yield higher-quality and more diverse candidate reasoning steps. Therefore, we introduce Entropy-Gated Branching: a novel inference technique that dynamically allocates computational resources by selectively expanding prediction sequences only at points of high uncertainty. Our method leverages entropy as a gating mechanism to identify when branching is most beneficial, coupled with an external feedback model to rank and prune candidate branches. Empirical results on mathematical and financial reasoning benchmarks show that this strategy improves accuracy by 22.6% over standard inference while operating 31%-75% faster across math benchmarks than test-time beam search with higher performance. Our results show that dynamic resource allocation during inference can substantially improve both efficiency and effectiveness, offering a more scalable pathway to enhanced LLM reasoning capabilities. We release our code and tools here[https://github.com/JXL884/entropy_gated_branching]

Existing challenges in misinformation exposure and susceptibility vary across demographics, as some populations are more vulnerable to misinformation than others. Large language models (LLMs) introduce new dimensions to these challenges through their ability to generate persuasive content at scale and reinforcing existing biases. Our study introduces PANDORA, a framework that investigates the bidirectional persuasion dynamics between LLMs and humans when exposed to misinformative content. We use a multi-agent LLM framework to analyze the spread of misinformation under persuasion among demographic-oriented LLM agents. Our findings show that demographic factors influence LLM susceptibility, with up to 15 percentage point differences in misinformation correctness across groups. Multi-agent LLMs also exhibit echo chamber behavior, aligning with human-like group polarization patterns. Therefore, this work highlights demographic divides in misinformation dynamics and offers insights for future interventions.

Retrieval Augmented Generation (RAG) is a popular method for injecting up-to-date information into Large Language Model (LLM)-based Natural Language Generation (NLG) systems. While RAG can enhance factual accuracy, its effect on the social biases inherent in LLMs is not well understood. This paper systematically investigates how RAG modulates social biases across three languages (English, Japanese, and Chinese) and four categories (gender, race, age, and religion). By evaluating various generator LLMs on the BBQ benchmark, we analyse how document collections with controlled stereotypical content affect RAG outputs. We find that biases present in the retrieved documents are often significantly amplified in the generated texts, even when the base LLM itself has a low-level of intrinsic bias. These findings raise concerns about the social fairness of RAG systems, underscoring the urgent need for careful bias evaluation before real-world deployment.

ICD coding is the process of mapping unstructured text from Electronic Health Records (EHRs) to standardised codes defined by the International Classification of Diseases (ICD) system. In order to promote trust and transparency, existing explorations on the explainability of ICD coding models primarily rely on attention-based rationales and qualitative assessments conducted by physicians, yet lack a systematic evaluation across diverse types of rationales using consistent criteria and high-quality rationale-annotated datasets specifically designed for the ICD coding task. Moreover, dedicated methods explicitly trained to generate plausible rationales remain scarce. In this work, we present evaluations of the explainability of rationales in ICD coding, focusing on two fundamental dimensions: faithfulness and plausibility—in short how rationales influence model decisions and how convincing humans find them. For plausibility, we construct a novel, multi-granular rationale-annotated ICD coding dataset, based on the MIMIC-IV database and the updated ICD-10 coding system. We conduct a comprehensive evaluation across three types of ICD coding rationales: entity-level mentions automatically constructed via entity linking, LLM-generated rationales, and rationales based on attention scores of ICD coding models. Building upon the strong plausibility exhibited by LLM-generated rationales, we further leverage them as distant supervision signals to develop rationale learning methods. Additionally, by prompting the LLM with few-shot human-annotated examples from our dataset, we achieve notable improvements in the plausibility of rationale generation in both the teacher LLM and the student rationale learning models.

The meaning conveyed by a sentence often depends on the context in which it appears. Despite the progress of sentence embedding methods, it remains unclear as how to best modify a sentence embedding conditioned on its context. To address this problem, we propose Condition-Aware Sentence Embeddings (CASE), an efficient and accurate method to create an embedding for a sentence under a given condition. First, CASE creates an embedding for the condition using an Large Language Model (LLM) encoder, where the sentence influences the attention scores computed for the tokens in the condition during pooling. Next, a supervised method is learnt to align the LLM-based text embeddings with the Conditional Semantic Textual Similarity (C-STS) task. We find that subtracting the condition embedding will consistently improve the C-STS performance of LLM-based text embeddings and improve the isotropy of the embedding space. Moreover, our supervised projection method significantly improves the performance of LLM-based embeddings despite requiring a small number of embedding dimensions.

Event argument extraction (EAE) is a crucial task in information extraction. However, its performance heavily depends on expensive annotated data, making data scarcity a persistent challenge. Data augmentation serves as an effective approach to improving model performance in low-resource settings, yet research on applying LLMs for EAE augmentation remains preliminary. In this study, we pay attention to the boundary sensitivity of EAE and investigate four LLM-based augmentation strategies: argument replacement, adjunction rewriting, their combination, and annotation generation. We conduct comprehensive experiments across four benchmark datasets, employing GPT-4o-Mini and DeepSeek-R1-7B as data generators. Our results show that boundary-aware augmentation consistently leads to greater performance improvements over boundary-agnostic methods. In addition to performance gains, we provide a detailed analysis of augmentation quality from multiple perspectives, including uncertainty reduction, error types, data quality, and data scale. This work offers both empirical evidence and practical guidance for leveraging LLMs to enhance event argument extraction under low-resource conditions.

The promise of LLM watermarking rests on a core assumption that a specific watermark proves authorship by a specific model. We demonstrate that this assumption is dangerously flawed. We introduce the threat of watermark spoofing, a sophisticated attack that allows a malicious model to generate text containing the authentic-looking watermark of a trusted, victim model. This enables the seamless misattribution of harmful content, such as disinformation, to reputable sources. The key to our attack is repurposing watermark radioactivity, the unintended inheritance of data patterns during fine-tuning, from a discoverable trait into an attack vector. By distilling knowledge from a watermarked teacher model, our framework allows an attacker to steal and replicate the watermarking signal of the victim model. This work reveals a critical security gap in text authorship verification and calls for a paradigm shift towards technologies capable of distinguishing authentic watermarks from expertly imitated ones. Our code is available at https://github.com/hsannn/ditto.

Large Language Models (LLMs) exhibit remarkable capabilities, but no single model optimally balances serving quality and deployment cost across diverse tasks. Multi-LLM systems address this challenge through intelligent routing mechanisms that dynamically allocate queries to the most appropriate model. However, existing routing methods suffer from two fundamental limitations: (i) dependence on extensive full-response datasets for training, and (ii) poor scalability when incorporating new models, typically necessitating retraining from scratch. In this paper, we propose SemiRouter, a novel LLM routing framework designed for data-sparse and evolving model environments. Our approach combines a data-efficient training methodology with an adaptive architecture that enables seamless integration of new models under limited supervision. As an extension, we also consider energy footprint as a potential deployment cost in our routing decision. Empirical evaluations demonstrate that our method improves data efficiency, adaptability, and routing accuracy compared to existing approaches, providing a scalable solution for dynamic multi-LLM deployment.

Pretrained Transformer encoders are the dominant approach to sequence labeling. While some alternative architectures-such as xLSTMs, structured state-space models, diffusion models, and adversarial learning-have shown promise in language modeling, few have been applied to sequence labeling, and mostly on flat or simplified tasks. We study how these architectures adapt across tagging tasks that vary in structural complexity, label space, and token dependencies, with evaluation spanning multiple languages. We find that the strong performance previously observed in simpler settings does not always generalize well across languages or datasets, nor does it extend to more complex structured tasks.

Much recent work has shown how cross-linguistic variation is constrained by competing pressures from efficient communication. However, little attention has been paid to the role of the systematicity of forms (regularity), a key property of natural language. Here, we demonstrate the importance of regularity in explaining the shape of linguistic systems by looking at recursive numeral systems. Previous work has argued that these systems optimise the trade-off between lexicon size and average morphosyntatic complexity (Denić and Szymanik, 2024). However, showing that only natural-language-like systems optimise this trade-off has proven elusive, and existing solutions rely on ad-hoc constraints to rule out unnatural systems (Yang and Regier, 2025). Drawing on the Minimum Description Length (MDL) approach, we argue that recursive numeral systems are better viewed as efficient with regard to their regularity and processing complexity. We show that our MDL-based measures of regularity and processing complexity better capture the key differences between attested, natural systems and theoretically possible ones, including “optimal” recursive numeral systems from previous work, and that the ad-hoc constraintsnaturally follow from regularity. Our approach highlights the need to incorporate regularity across sets of forms in studies attempting tomeasure efficiency in language.

Large language models (LLMs) can answer prompts in many languages, despite being trained predominantly on English; yet, the mechanisms driving this generalization remain poorly understood. This work asks: How does an LLM’s ability to align representations of non-English inputs to English impact its performance on natural language understanding (NLU) tasks? We study the role of representation alignment in instance-level task decisions, complementing prior analyses conducted both at the language level and task-independently. We introduce the Discriminative Alignment Index (\DALI) to quantify instance-level alignment across 24 languages other than English and three distinct NLU tasks. Results show that incorrect NLU predictions are strongly associated with lower representation alignment with English in the model’s middle layers. Through activation patching, we show that incorrect predictions in languages other than English can be fixed by patching their parallel English activations in the middle layers, thereby demonstrating the causal role of representation (mis)alignment in cross-lingual correctness.

Large Language Models (LLMs) have achieved impressive performance yet remain inconsistent across languages, often defaulting to high-resource outputs such as English. Existing multilingual alignment methods mitigate these issues through preference optimization but rely on external supervision, such as translation systems or English-biased signal. We propose Multilingual Self-Alignment (MSA), a targeted preference optimization framework that leverages an LLM’s own latent representations as intrinsic supervision signals, rewarding lower-resource language outputs based on their alignment with high-resource (English) counterparts in the “semantic hub”. We further introduce Language-Consistency MSA (LaCoMSA), which augments MSA with a final-layer language-consistency factor to prevent off-target generation. Integrated with Direct Preference Optimization, LaCoMSA improves a Llama 3 8B-based model multilingual win rates by up to 6.8% absolute (55.0% relatively) on X-AlpacaEval and achieves consistent gains across benchmarks and models. Our findings demonstrate that LaCoMSA can serve as an effective and scalable mechanism, opening a new venue toward multilingual self-alignment.

Quantization is an effective technique for reducing the storage footprint and computational costs of Large Language Models (LLMs), but it often results in performance degradation. Existing post-training quantization methods typically use small, English-only calibration sets; however, their impact on multilingual models remains underexplored. We systematically evaluate eight calibration settings (five single-language and three multilingual mixes) across two quantizers (GPTQ, AWQ) on data from 10 different languages. Our findings reveal a consistent trend: non-English and multilingual calibration sets significantly improve perplexity compared to English-only baselines. Specifically, we observe notable average perplexity gains across both quantizers on Llama3.1 8B and Qwen2.5 7B, with multilingual mixes achieving the largest overall reductions of up to 3.52 perplexity gain. Furthermore, our analysis indicates that tailoring calibration sets to the evaluation language yields the largest improvements for individual languages, underscoring the importance of linguistic alignment. We also identify specific failure cases where certain language-quantizer combinations degrade performance, which we trace to differences in activation range distributions across languages. These results highlight that static, one-size-fits-all calibration is suboptimal, and that tailoring calibration data, both in language and diversity, plays a crucial role in robustly quantizing multilingual LLMs.

Large Language Models (LLMs) are increasingly used for knowledge-based reasoning tasks, yet understanding when they rely on genuine knowledge versus superficial heuristics remains challenging. We investigate this question through entity comparison tasks by asking models to compare entities along numerical attributes (e.g., “Which river is longer, the Danube or the Nile?”), which offer clear ground truth for systematic analysis. Despite having sufficient numerical knowledge to answer correctly, LLMs frequently make predictions which contradict this knowledge. We identify three heuristic biases that strongly influence model predictions: entity popularity, mention order, and semantic co-occurrence. For smaller models, a simple logistic regression using only these surface cues predicts model choices more accurately than the model’s own numerical predictions, suggesting heuristics largely override principled reasoning. Crucially, we find that larger models (32B parameters) selectively rely on numerical knowledge when it is more reliable, while smaller models (7-8B parameters) show no such discrimination, which explains why larger models outperform smaller ones even when the smaller models possess more accurate knowledge. Chain-of-thought prompting steers all models towards using the numerical features across all model sizes.

Grammaticalization denotes a diachronic change of the grammatical category from content words to function words. One of the intensively explored directions in this area is to quantify the degree of grammaticalization. There have been a limited number of automated methods for this task and the existing, best-performing method is heavily language- and word-dependent. In this paper, we explore three methods for quantifying the degree of grammaticalization, which are applicable to a wider variety of words and languages. The difficulty here is that training data is not available in the present task. We overcome this difficulty by using Positive-Unlabeled learning (PU-learning) or Cross-Validation-like learning (hereafter, CV-learning). Experiments show that the CV-learning-based method achieves middle to high correlations to human judgments in English deverbal prepositions and Japanese nouns being grammaticalized. With this method, we further explore words possibly being grammaticalized and counterexamples of the unidirectionality hypothesis.

Large language models (LLMs) have emerged as strong contenders in machine translation. Yet, they still struggle to adequately handle discourse phenomena, such as pronoun resolution and lexical cohesion at the document level. In this study, we thoroughly investigate the discourse phenomena performance of LLMs in context-aware translation. We demonstrate that discourse knowledge is encoded within LLMs and propose the use of quality-aware decoding (QAD), specifically minimum Bayes risk decoding, to effectively extract this knowledge, showcasing its superiority over other decoding approaches through comprehensive analysis. Furthermore, we illustrate that QAD enhances the semantic richness of translations and aligns them more closely with human preferences.