EACL 2026 Papers

Building effective tokenizers for multilingual Large Language Models (LLMs) requires careful control over language-specific data mixtures. While a tokenizer’s compression performance critically affects the efficiency of LLM training and inference, existing approaches rely on heuristics or costly large-scale searches to determine optimal language ratios. We introduce \textbf{T}okenizer \textbf{Re}gression for Optimal Data Mi\textbf{X}ture (\text{TReX}), a regression-based framework that efficiently predicts the optimal data mixture for tokenizer training. \text{TReX} trains small-scale proxy tokenizers on random mixtures, gathers their compression statistics, and learns to predict compression performance from data mixtures. This learned model enables scalable mixture search before large-scale tokenizer training, mitigating the accuracy-cost trade-off in multilingual tokenizer design. Tokenizers trained with TReX’s predicted mixtures outperform mixtures based on LLaMA3 and uniform distributions by up to 12% in both in- and out-of-distribution compression efficiency, demonstrating strong scalability, robustness, and practical effectiveness.

We examine, analyze, and compare four representative creativity measures—perplexity, LLM-as-a-Judge, the Creativity Index (CI; measuring n-gram overlap with web corpora), and syntactic templates (detecting repetition of common part-of-speech patterns)—across the diverse creative domains, such as creative writing, unconventional problem-solving, and research ideation. For each domain, we compile datasets with human-aligned creative and uncreative examples and evaluate each metric’s ability to discriminate between the two sets. Our analyses reveal limited consistency both across domains and metrics, as metrics that distinguish creativity in one domain fail in others (e.g., CI correctly distinguishes in creative writing but fails in problem-solving), and different metrics often disagree on the same data points (e.g., CI suggests one set to be more creative, while perplexity indicates the other set to be more creative.) We highlight key limitations, such as perplexity reflecting fluency rather than novelty; LLM-as-a-Judge producing inconsistent judgments under minor prompt variations and exhibiting bias towards particular labels; CI primarily measuring lexical diversity, with high sensitivity to implementation choices; and syntactic templates being ineffective in settings dominated by formulaic language. Our findings underscore the need for more robust, generalizable evaluation frameworks that better align with human judgments of creativity. We release the datasets and evaluation code: https://github.com/lichun-19/creative_eval.

Grammar comprehension is a critical capability for large language models (LLMs) to achieve fluency in a target language. In low-resource settings, such as the case with Arabic, limited availability of high-quality data can lead to significant gaps in grammatical understanding, making systematic evaluation essential. We introduce Nahw, a comprehensive benchmark for Arabic grammar that covers both theoretical knowledge and practical applications, including grammatical error detection, correction, and explanation. We evaluate a range of LLMs on these tasks and find that many models still exhibit substantial deficiencies in Arabic grammar comprehension, with GPT-4o achieving a score of 67% on average over all tasks, while the best performing Arabic model in our experiment (ALLaM-7B) achieving 42%. Our experiments also demonstrate that while fine-tuning with synthetic data can improve performance, it does not match the effectiveness of training on natural, high-quality data.

Large language models (LLMs) encode rich cultural knowledge learned from diverse web-scale data, offering an unprecedented opportunity to model cultural commonsense at scale. Yet this knowledge remains mostly implicit and unstructured, limiting its interpretability and use. We present an iterative, prompt-based framework for constructing a Cultural Commonsense Knowledge Graph (CCKG) that treats LLMs as cultural archives, systematically eliciting culture-specific entities, relations, and practices and composing them into multi-step inferential chains across languages. We evaluate CCKG on five countries with human judgments of cultural relevance, correctness, and path coherence. We find that the cultural knowledge graphs are better realized in English, even when the target culture is non-English (e.g., Chinese, Indonesian, Arabic), indicating uneven cultural encoding in current LLMs. Augmenting smaller LLMs with CCKG improves performance on cultural reasoning and story generation, with the largest gains from English chains. Our results show both the promise and limits of LLMs as cultural technologies and that chain-structured cultural knowledge is a practical substrate for culturally grounded NLP.

Multi-domain image-to-image translation requires grounding semantic differences expressed in natural language prompts into corresponding visual transformations, while preserving unrelated structural and semantic content. Existing methods struggle to maintain structural integrity and provide fine-grained, attribute-specific control, especially when multiple domains are involved. We propose LACE (Language-grounded Attribute-Controllable Translation), built on two components: (1) a GLIP-Adapter that fuses global semantics with local structural features to preserve consistency, and (2) a Multi-Domain Control Guidance mechanism that explicitly grounds the semantic delta between source and target prompts into per-attribute translation vectors, aligning linguistic semantics with domain-level visual changes. Together, these modules enable compositional multi-domain control with independent strength modulation for each attribute. Experiments on CelebA(Dialog) and BDD100K demonstrate that LACE achieves high visual fidelity, structural preservation, and interpretable domain-specific control, surpassing prior baselines. This positions LACE as a cross-modal content generation framework bridging language semantics and controllable visual translation. Code will be publicly available.

Modelling a language model for a multi-lingual scenario includes several potential challenges, among which catastrophic forgetting is the major challenge. For example, small language models (SLM) built for low-resource languages by adapting large language models (LLMs) pose the challenge of catastrophic forgetting. This work proposes to employ a continual learning strategy using parts-of-speech (POS)-based code-switching along with a replay adapter strategy to mitigate the identified gap of catastrophic forgetting while training SLM from LLM. Experiments conducted on vision language tasks such as visual question answering and language modelling task exhibits the success of the proposed architecture.

Artificial intelligence (AI) and large language models (LLMs) in healthcare require advanced clinical skills (CS), yet current benchmarks fail to evaluate these comprehensively. We introduce MedQA-CS, an AI-SCE framework inspired by medical education’s Objective Structured Clinical Examinations (OSCEs), to address this gap. MedQA-CS evaluates LLMs through two instruction-following tasks—LLM-as-medical-student and LLM-as-CS-examiner—designed to reflect real clinical scenarios. Our contributions include developing MedQA-CS, a comprehensive evaluation framework with publicly available data and expert annotations, and providing the quantitative and qualitative assessment of LLMs as reliable judges in CS evaluation. Our experiments show that MedQA-CS is a more challenging benchmark for evaluating clinical skills than traditional multiple-choice QA benchmarks (e.g., MedQA). Combined with existing benchmarks, MedQA-CS enables a more comprehensive evaluation of LLMs’ clinical capabilities for both open- and closed-source LLMs.

Current evaluation of large language models (LLMs) overwhelmingly prioritizes accuracy; however, in real-world and safety-critical applications, the ability to abstain when uncertain is equally vital for trustworthy deployment. We introduce a unified benchmark and evaluation protocol for abstention in medical multiple-choice question answering (MCQA), integrating conformal prediction, adversarial question perturbations, and explicit abstention options. Our systematic evaluation of both open- and closed-source LLMs reveals that even state-of-the-art, high-accuracy models often fail to abstain when uncertain. Notably, providing explicit abstention options consistently increases model uncertainty and safer abstention, far more than input perturbations, while scaling model size or advanced prompting brings little improvement. These findings highlight the central role of abstention mechanisms for trustworthy LLM deployment and offer practical guidance for improving safety in high-stakes applications.

Existing multilingual long-context benchmarks, often based on the popular needle-in-a-haystack test, primarily evaluate a model’s ability to locate specific information buried within irrelevant texts. However, such a retrieval-centric approach is myopic and inherently limited, as successful recall alone does not indicate a model’s capacity to reason over extended contexts. Moreover, these benchmarks are susceptible to data leakage, short-circuiting, and risk making the evaluation a priori identifiable. To address these limitations, we introduce MLRBench, a new synthetic benchmark for multilingual long-context reasoning. Unlike existing benchmarks, MLRBench goes beyond surface-level retrieval by including bAbI-style tasks that test multi-hop inference, aggregation, and epistemic reasoning. Spanning seven languages, we design MLRBench to be parallel, resistant to leakage, and scalable to arbitrary context lengths. Our extensive experiments with an open-weight large language model (LLM) reveal a pronounced gap between high- and low-resource languages, particularly for tasks requiring the model to aggregate multiple facts or predict the absence of information. We also find that, in multilingual settings, LLMs effectively utilize less than 30% of their claimed context length. Although off-the-shelf Retrieval-Augmented Generation helps alleviate this to a certain extent, it does not solve the long-context problem. We open-source MLRBench to enable future research in the improved evaluation and training of multilingual LLMs.

The role of world knowledge has been particularly crucial to predict the discourse connective that marks the discourse relation between two arguments, with language models (LMs) being generally successful at this task. We flip this premise in our work, and instead study the inverse problem of understanding whether discourse connectives can inform LMs about the world. To this end, we present Wugnectives, a dataset of 8,880 stimuli that evaluates LMs’ inferences about novel entities in contexts where connectives link the entities to particular attributes. On investigating 17 different LMs at various scales, and training regimens, we found that tuning an LM to show reasoning behavior yields noteworthy improvements on most connectives. At the same time, there was a large variation in LMs’ overall performance across connective type, with all models systematically struggling on connectives that express a concessive meaning. Our findings pave the way for more nuanced investigations into the functional role of language cues as captured by LMs.We release Wugnectives at https://github.com/kanishkamisra/wugnectives

Large language models remain predominantly English-centric, which limits their utility for underrepresented languages. We help bridge this gap for Hindi with Llama-3-Nanda-10B-Chat (aka Nanda-10B) and Llama-3.1-Nanda-87B-Chat (aka Nanda-87B), forming the Nanda family of open-weight bilingual models (https://github.com/MBZUAI-IFM/Nanda-Family). Our approach integrates: (i) a tokenizer extending Llama’s vocabulary with 20% Hindi-specific tokens, thus halving Hindi tokenization fertility while preserving English efficiency, (ii) Hindi-first parameter-efficient continual pretraining using Llama Pro on a 65B-token corpus spanning Devanagari script, code-mixed, and Romanized Hindi, and (iii) bilingual instruction and safety alignment on a large culturally grounded dataset. The resulting Nanda models outperform open-weight LLMs of comparable size: Nanda-87B yields high generative quality, and Nanda-10B shows competitive general-purpose performance. Nanda-87B demonstrates state-of-the-art performance on summarization, translation, transliteration, and instruction following. Moreover, both models achieve state-of-the-art performance in safety and in cultural knowledge. Our results demonstrate that careful tokenizer design, data curation, and continual pretraining can yield capable and safe LLMs for resource-poor languages without compromising English performance.

Hallucinations remain a persistent challenge for vision–language models (VLMs), which often describe nonexistent objects or fabricate facts. Existing detection methods typically operate after text generation, making intervention both costly and untimely. We investigate whether hallucination risk can instead be predicted before any token is generated by probing a model’s internal representations in a single forward pass. Across a diverse set of vision–language tasks and eight modern VLMs, including Llama-3.2-Vision, Gemma-3, Phi-4-VL, and Qwen2.5-VL, we examine three families of internal representations: (i) visual-only features without multimodal fusion, (ii) vision token representations within the text decoder, and (iii) query-token representations that integrate visual and textual information before generation. Probes trained on these representations achieve strong hallucination-detection performance without decoding, reaching up to 0.93 AUROC on Gemma-3-12B, Phi-4-VL 5.6B, and Molmo 7B. Late query-token states are the most predictive for most models, while visual or mid layer features dominate in a few architectures (e.g., ∼0.79 AUROC for Qwen2.5-VL-7B using visual-only features). These results demonstrate that (1) hallucination risk is detectable pre-generation, (2) the most informative layer and modality vary across architectures, and (3) lightweight probes has the potential to enable early abstention, selective routing, and adaptive decoding to improve both safety and efficiency.

Large Audio-Language Models (LALMs) are often constrained by short audio context windows, even when their text backbones support long contexts, limiting long-form audio understanding. Prior work has introduced context-extension methods (e.g. YaRN) on unimodal LLMs, yet their application to LALMs remains unexplored. First, building on RoPE-based context extension, we introduce Partial YaRN, a training-free, modality-decoupled extension method that modifies only audio token positions, leaving text positions intact to preserve the base LLM’s text capabilities. Second, we propose Virtual Longform Audio Training (VLAT), a training strategy that extends Partial YaRN into a training-time positional augmentation. VLAT simulates diverse audio lengths during training, enabling generalization to inputs far longer than those seen in training. Our experiments on SALMONN and Qwen2-Audio confirm that Partial YaRN outperforms the original models across wide range of settings, and VLAT provides substantial performance improvement on long audio of unseen lengths.

Ancient people translated classical Chinese into Japanese using a system of annotations placed around characters. We abstract this process as sequence tagging tasks and fit them into modern language technologies. The research on this annotation and translation system faces a low resource problem. We alleviate this problem by introducing an LLM-based annotation pipeline and constructing a new dataset from digitized open-source translation data. We show that in the low-resource setting, introducing auxiliary Chinese NLP tasks enhances the training of sequence tagging tasks. We also evaluate the performance of Large Language Models (LLMs) on this task. While they achieve high scores on direct machine translation, our method could serve as a supplement to LLMs to improve the quality of character’s annotation.

Evaluating the quality of children’s utterances in adult-child dialogue remains challenging due to insufficient context-sensitive metrics. Common proxies such as Mean Length of Utterance (MLU), lexical diversity (vocd-D), and readability indices (Flesch-Kincaid Grade Level, Gunning Fog Index) are dominated by length and ignore conversational context, missing aspects of response quality such as reasoning depth, topic maintenance, and discourse planning. We introduce an LLM-as-a-judge framework that first classifies the Previous Adult Utterance Type and then scores the child’s response along two axes: Expansion (contextual elaboration and inferential depth) and Independence (the child’s contribution to advancing the discourse). These axes reflect fundamental dimensions in child language development, where Expansion captures elaboration, clause combining, and causal and contrastive connectives. Independence captures initiative, topic control, decreasing reliance on adult scaffolding through growing self-regulation, and audience design. We establish developmental validity by showing age-related patterns and demonstrate predictive value by improving age estimation over common baselines. We further confirm semantic sensitivity by detecting differences tied to discourse relations. Our metrics align with human judgments, enabling large-scale evaluation. This shifts child utterance assessment from simply measuring length to evaluating how meaningfully the child’s speech contributes to and advances the conversation within its context.

Turn-taking is a fundamental component of human communication and is signalled through complex cues distributed across lexical, temporal, and prosodic information. Full-duplex models of spoken dialogue integrate these information sources to produce impressive turn-taking behaviour. Yet, existing evaluations of their turn-taking capabilities do not address which information sources drive predictions.We present a systematic analysis of the role of lexical-temporal features on the predictability of turn structure by examining PairwiseTurnGPT, a full-duplex model of spoken dialogue transcripts. Through PCA, mixed-effects modelling, and temporal surprisal analysis, we reveal context-dependent patterns: linguistic fluency paradoxically creates overconfidence at intermediate completion points, while turn-shift overlap dominates boundary detection. Our findings uncover where lexical-temporal information suffices and where additional cues become necessary, establishing a deeper understanding of how turn-taking cues are distributed and how to evaluate dialogue systems.

Language models (LMs) have been reported to implicitly encode character-level information, despite not being explicitly provided during training. However, the mechanisms underlying this phenomenon remain largely unexplored. To reveal the mechanisms, we analyze how models acquire character-level knowledge by comparing LMs trained under controlled settings, such as specifying the pre-training dataset or tokenizer, with those trained under standard settings. We categorize the contributing factors into those independent of tokenization. Our analysis reveals that merge rules and orthographic constraints constitute primary factors arising from tokenization, whereas semantic associations of substrings and syntactic information function as key factors independent of tokenization.

Working memory, or the ability to hold and manipulate information in the mind, is a critical component of human intelligence and executive functioning. It is correlated with performance on various cognitive tasks, including measures of fluid intelligence, which encompasses reasoning and problem solving. We use a comprehensive set of classic working memory tasks to estimate the working memory capacity of large language models (LLMs). We find that in most cases, LLMs exceed normative human scores. However, we do not find that the increased capacity of working memory is associated with higher performance on other executive functioning tasks or problem solving benchmarks. These results suggest that LLMs may have deficits in attentional control and cognitive flexibility, which result in difficulties with inhibiting automatic responses and adapting to shifting information. Our findings suggest that reasoning models, although they often do not currently fully compensate for these deficits, may have the potential to do so in the future.

Leveraging a dataset of paired narratives, we investigate the extent to which large language models (LLMs) can reliably separate incoherent and coherent stories.A probing study finds that LLMs’ internal representations can reliably identify incoherent events in narratives. However, this separation disappears by the narrative’s end, and weakens when the differences between coherent and incoherent stories are more subtle. When asked to rate overall coherence of narratives after reading, LLMs generate responses that fail to satisfactorily separate the coherent and incoherent narratives.Reasoning models tested do not eliminate these deficits, indicating that thought strings may not be able to fully address the discrepancy between model internal state and behavior.Additionally, we find that LLMs appear to be more sensitive to incoherence resulting from an event that violates the setting (e.g., a rainy day in the desert) than to incoherence arising from a character violating an established trait (e.g., Mary, a vegetarian, later orders a cheeseburger), suggesting that LLMs may rely more on prototypical world knowledge than building coherence through a meaning-based world model of the narrative setting. Together, our results indicate that LLMs lack robustness in their ability to recognize incoherence in narratives.