EACL 2026 Papers

Large audio-language models increasingly operate on raw speech inputs, enabling more seamless integration across domains such as voice assistants, education, and clinical triage. This transition, however, introduces a distinct class of vulnerabilities that remain largely uncharacterized. We examine the security implications of this modality shift by designing a text-to-audio jailbreak that embeds disallowed directives within a narrative-style audio stream. The attack leverages an advanced instruction-following text-to-speech (TTS) model to exploit structural and acoustic properties, thereby circumventing safety mechanisms primarily calibrated for text. When delivered through synthetic speech, the narrative format elicits restricted outputs from state-of-the-art models, including Gemini 2.0 Flash, achieving a 98.26% success rate that substantially exceeds text-only baselines. These results highlight the need for safety frameworks that jointly reason over linguistic and paralinguistic representations, particularly as speech-based interfaces become more prevalent.

Large language models (LLMs) excel in English but still struggle with complex reasoning in many low-resource languages (LRLs). Existing methods align LLMs with multilingual encoders, such as LangBridge and MindMerger, raising the accuracy for mid and high-resource languages, yet large performance gap remains for LRLs. We present MERLIN, a model-stacking framework that iteratively refines in 2-stages based on a curriculum strategy (from general to specific where general is bilingual bitext and specific is task-specific data) and adapts only a small set of DoRA weights. On the AfriMGSM benchmark MERLIN improves exact-match accuracy by +12.9 pp over MindMerger and outperforms GPT-4o-mini by 15.2 pp. It also yields consistent gains on MGSM and MSVAMP (+0.9 and +2.8 pp), demonstrating effectiveness across both low and high-resource settings.

Text anonymization is a critical task for enabling research and development in high-stakes domains containing private data, like medicine, law, and social services. While much research has focused on redacting sensitive content from text, substantially less work has focused on what to replace redacted content with, which can enhance privacy and becomes increasingly important with greater levels of redaction. In this work, we formulate predicting replacements for sensitive spans as a research task with principled use-inspired evaluation criteria. We further propose a multi-token completion method for accomplishing this task that is designed to preserve consistency with low compute requirements, thus facilitating practitioners to anonymize data locally before sharing it externally. Human and automated annotations demonstrate that our approach produces more realistic text and better preserves utility than alternative infilling methods and differentially private mechanisms across multiple domains without retraining. Overall, our work explores the under-studied task of what to replace redacted content with and contributes grounded evaluations capturing utility, facilitating future work.

Recent advances in self-supervised learning (SSL) have led to powerful speech representation models, yet their robustness in real-world conversational settings remains largely untested. Most existing benchmarks focus on clean, single-speaker, single-channel audio, failing to reflect the complexities of natural human interaction—where background noise, reverberation, and overlapping speech are the norm. To bridge these critical gaps, we present the Conversational Speech Processing Benchmark (CSPB), a new benchmark designed to assess the robustness of SSL speech models in realistic conversational scenarios. CSPB is constructed from four multi-party datasets—AMI, AliMeeting, MMCSG, and DiPCo—and supports both single-channel and multi-channel evaluation. By releasing CSPB as an open-source toolkit, we aim to establish a unified framework for evaluating and advancing robust, spatially-aware self-supervised speech models.

Understanding emotion from speech requires sensitivity to both lexical and acoustic cues. However, it remains unclear whether large audio language models (LALMs) genuinely process acoustic information or rely primarily on lexical contents. We present LISTEN (Lexical vs. Acoustic Speech Test for Emotion in Narratives), a controlled benchmark designed to disentangle lexical reliance from acoustic sensitivity in emotion understanding. Across evaluations of six state-of-the-art LALMs, we observe a consistent lexical dominance. Models predict “neutral” when lexical cues are neutral or absent, show limited gains under cue alignment, and fail to classify distinct emotions under cue conflict. In paralinguistic settings, performance approaches chance. These results indicate that current LALMs largely “transcribe” rather than “listen,” relying heavily on lexical semantics while underutilizing acoustic cues. LISTEN offers a principled framework for assessing emotion understanding in multimodal models.

A common solution for mitigating outdated or incorrect information in Large Language Models (LLMs) is to provide updated facts in-context or through knowledge editing. However, these methods introduce knowledge conflicts when the knowledge update fails to overwrite the model’s parametric knowledge, which propagate to faulty reasoning. Current benchmarks for this problem, however, largely focus only on single knowledge updates and fact recall without evaluating how these updates affect downstream reasoning. In this work, we introduce Tʀᴀᴄᴋ (Testing Reasoning Amid Conflicting Knowledge), a new benchmark for studying how LLMs propagate new knowledge through multi-step reasoning when it conflicts with the model’s initial parametric knowledge. Spanning three reasoning-intensive scenarios (WIKI, CODE, and MATH), Tʀᴀᴄᴋ introduces multiple, realistic conflicts to mirror real-world complexity. Our results on Tʀᴀᴄᴋ reveal that providing updated facts to models for reasoning can worsen performance compared to providing no updated facts to a model, and that this performance degradation exacerbates as more updated facts are provided. We show this failure stems from both inability to faithfully integrate updated facts, but also flawed reasoning even when knowledge is integrated. Tʀᴀᴄᴋ provides a rigorous new benchmark to measure and guide future progress on propagating conflicting knowledge in multi-step reasoning.

Recent advances in multimodal large language models (MLLMs) offer a promising approach for natural language-based scene change queries in virtual reality (VR). Prior work on applying MLLMs for object state understanding has focused on egocentric videos that capture the camera wearer’s interactions with objects. However, object state changes may occur in the background without direct user interaction, lacking explicit motion cues and making them difficult to detect. Moreover, no benchmark exists for evaluating this challenging scenario. To address these challenges, we introduce ObjChangeVR-Dataset, specifically for benchmarking the question-answering task of object state change. We also propose ObjChangeVR, a framework that combines viewpoint-aware and temporal-based retrieval to identify relevant frames, along with cross-view reasoning that reconciles inconsistent evidence from multiple viewpoints. Extensive experiments demonstrate that ObjChangeVR significantly outperforms baseline approaches across multiple MLLMs.

Curriculum learning—organizing training data from easy to hard—has improved efficiency across machine learning domains, yet remains underexplored for language model pretraining. We present the first systematic investigation of curriculum learning in LLM pretraining, with over 200 models trained on up to 100B tokens across three strategies: vanilla curriculum learning, pacing-based sampling, and interleaved curricula, guided by six difficulty metrics spanning linguistic and information-theoretic properties. We evaluate performance on eight benchmarks under three realistic scenarios: limited data, unlimited data, and continual training. Our experiments show that curriculum learning consistently accelerates convergence in early and mid-training phases, reducing training steps by 18-45% to reach baseline performance. When applied as a warmup strategy before standard random sampling, curriculum learning yields sustained improvements up to 3.5%. We identify compression ratio, lexical diversity (MTLD), and readability (Flesch Reading Ease) as the most effective difficulty signals. Our findings demonstrate that data ordering—orthogonal to existing data selection methods—provides a practical mechanism for more efficient LLM pretraining.

Tokenization is a critical component of language model pretraining, yet standard tokenization methods often prioritize information-theoretical goals like high compression and low fertility rather than linguistic goals like morphological alignment. In fact, they have been shown to be suboptimal for morphologically rich languages, where tokenization quality directly impacts downstream performance. In this work, we investigate morphologically-aware tokenization for Latin, a morphologically rich language that is medium-resource in terms of pretraining data, but high-resource in terms of curated lexical resources – a distinction that is often overlooked but critical in discussions of low-resource language modeling. We find that morphologically-guided tokenization improves overall performance on four downstream tasks. Performance gains are most pronounced for out of domain texts, highlighting our models’ improved generalization ability. Our findings demonstrate the utility of linguistic resources to improve language modeling for morphologically complex languages. For low-resource languages that lack large-scale pretraining data, the development and incorporation of linguistic resources can serve as a feasible alternative to improve LM performance.

Multilingual domain adaptation (ML-DA) enables large language models (LLMs) to acquire domain knowledge across languages. Despite many methods, how domain knowledge is acquired within a language and transferred across languages remains, leading to suboptimal performance, particularly in low-resource settings.This work examines the learning dynamics of LLMs during ML-DA. Because prior ML-DA studies often train and evaluate on datasets with mismatched knowledge coverage, we propose AdaXEval, an adaptive evaluation method that constructs multiple-choice QA datasets from the same bilingual domain corpus used for training, thereby enabling direct analysis of multilingual knowledge acquisition.Through continual training of LLMs with diverse data recipes, we track how LLMs acquire domain facts and pinpoint the loss shielding mechanism behind the knowledge memorization and generalization in domain adaptation. Our experiments on multilingual LLMs reveal that cross-lingual transfer remains challenging.The code is released.

Model steering, which involves intervening on hidden representations at inference time, has emerged as a lightweight alternative to finetuning for precisely controlling large language models. While steering efficacy has been widely studied, evaluations of whether interventions alter only the intended property remain limited, especially with respect to unintended changes in behaviors related to the target property. We call this notion specificity. We propose a framework that distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts). We study two safety-critical use cases: steering models to reduce overrefusal and faithfulness hallucinations, and show that while steering achieves high efficacy and largely maintains general and control specificity, it consistently fails to preserve robustness specificity. In the case of overrefusal steering, for example, all steering methods reduce overrefusal without harming general abilities and refusal on harmful queries; however, they substantially increase vulnerability to jailbreaks. Our work provides the first systematic evaluation of specificity in model steering, showing that standard efficacy and specificity checks are insufficient, because without robustness evaluation, steering methods may appear reliable even when they compromise model safety.

Multimodal Large Language Models (MLLMs) show progress across many visual–language tasks; however, their capacity to evaluate artistic expression remains limited: aesthetic concepts are inherently abstract and open-ended, and multimodal artwork annotations are scarce. We introduce KidsArtBench, a new benchmark of over 1k children’s artworks (ages 5-15) annotated by 12 expert educators across 9 rubric-aligned dimensions, together with expert comments for feedback. Unlike prior aesthetic datasets that provide single scalar scores on adult imagery, KidsArtBench targets children’s artwork and pairs multi-dimensional annotations with comment supervision to enable both ordinal assessment and formative feedback. Building on this resource, we propose an attribute-specific multi-LoRA approach – where each attribute corresponds to a distinct evaluation dimension (e.g., Realism, Imagination) in the scoring rubric – with Regression-Aware Fine-Tuning (RAFT) to align predictions with ordinal scales. On Qwen2.5-VL-7B, our method increases correlation from 0.468 to 0.653, with the largest gains on perceptual dimensions and narrowed gaps on higher-order attributes. Our results show that educator-aligned supervision and attribute-aware training yield pedagogically meaningful evaluations and establish a rigorous testbed for sustained progress in educational AI. We release data and code with ethics documentation.

Despite the impressive performance of vision-language models (VLMs) on downstream tasks, their ability to understand and reason about causal relationships in visual inputs remains unclear. Robust causal reasoning is fundamental to solving complex high-level reasoning tasks, yet existing benchmarks often include a mixture of reasoning questions, and VLMs can frequently exploit object recognition and activity identification as shortcuts to arrive at the correct answers, making it challenging to truly assess their causal reasoning abilities. To bridge this gap, we introduce VQA-Causal and VCR-Causal, two new benchmarks specifically designed to isolate and rigorously evaluate VLMs’ causal reasoning abilities. Our findings reveal that while VLMs excel in object and activity recognition, they perform poorly on causal reasoning tasks, often only marginally surpassing random guessing. Further analysis suggests that this limitation stems from a severe lack of causal expressions in widely used training datasets, where causal relationships are rarely explicitly conveyed. We additionally explore fine-tuning strategies with hard negative cases, showing that targeted fine-tuning can improve model’s causal reasoning while maintaining generalization and downstream performance. Our study highlights a key gap in current VLMs and lays the groundwork for future work on causal understanding. We will release the code upon acceptance.

Enterprise documents, such as forms and reports, embed critical information for downstream applications like data archiving, automated workflows, and analytics. Although generalist Vision Language Models (VLMs) perform well on established document understanding benchmarks, their ability to conduct holistic, fine-grained structured extraction across diverse document types and flexible schemas is not well studied. Existing Key Entity Extraction (KEE), Relation Extraction (RE), and Visual Question Answering (VQA) datasets are limited by narrow entity ontologies, simple queries, or homogeneous document types, often overlooking the need for adaptable and structured extraction. To address these gaps, we introduce ExStrucTiny, a new benchmark dataset for structured Information Extraction (IE) from document images, unifying aspects of KEE, RE, and VQA. Built through a novel pipeline combining manual and synthetic human-validated samples, ExStrucTiny covers more varied document types and extraction scenarios. We analyze open and closed VLMs on this benchmark, highlighting challenges such as schema adaptation, query under-specification, and answer localization. We hope our work provides a bedrock for improving generalist models for structured IE in documents.

This paper investigates gender biases exhibited by LLM-based virtual assistants when providing educational recommendations, focusing on minimal gender indicators. Experimenting on Italian, a language with grammatical gender, we demonstrate that simply changing noun and adjective endings (e.g., from masculine “-o” to feminine “-a”) significantly shifts recommendations. More specifically, we find that LLMs i) recommend STEM disciplines less for prompts with feminine grammatical gender and ii) narrow down the set of disciplines recommended to prompts with masculine grammatical gender; these effects persist across multiple commercial LLMs (from OpenAI, Anthropic, and Google). We show that grammatical gender cues alone trigger substantial distributional shifts in educational recommendations, and up to 76% of the bias exhibited when using prompts with proper names is already present with grammatical gender markers alone.Our findings highlight the need for robust bias evaluation and mitigation strategies before deploying LLM-based virtual assistants in student-facing contexts and the risks of using general purpose LLMs for educational applications, especially in languages with grammatical gender.

Large language models are being rapidly applied across many fields such as healthcare, finance, transportation, and energy, among many others. These applications often involve analyzing time-series data alongside contextual information in the form of natural language to support informed decisions. However, current time-series models are limited in their ability to perform reasoning that involves both time-series and their textual content. In this work, we address this gap by introducing Chat-TS, a large language model (LLM) based framework designed to support reasoning over time series and textual data. Unlike traditional models, Chat-TS integrates time-series tokens into LLMs’ vocabulary, enhancing its reasoning ability over both modalities without compromising core natural language capabilities. To support learning and evaluation, we contribute new datasets: the TS Instruct Training Dataset (pairing diverse time-series data with relevant text instructions and responses for instruction tuning), the TS Instruct Question and Answer (QA) Gold Dataset (multiple-choice questions to evaluate multimodal reasoning), and a TS Instruct Quantitative Probing Set (a small subset of TS Instruct QA reasoning tasks alongside math and decision-making questions for LLM evaluation). We design a training strategy to preserve the inherent reasoning capabilities of LLMs while augmenting them for time-series reasoning. Experiments show that Chat-TS achieves state-of-the-art performance in multimodal reasoning tasks by maintaining strong natural language proficiency while improving time-series reasoning.

Membership inference attacks (MIAs) reveal whether specific data was used to train machine learning models, serving as important tools for privacy auditing and compliance assessment. Recent studies have reported that MIAs perform only marginally better than random guessing against large language models, suggesting that modern pre-training approaches with massive datasets may be free from privacy leakage risks. Our work offers a complementary perspective to these findings by exploring how examining LLMs’ internal representations, rather than just their outputs, may provide additional insights into potential membership inference signals. Our framework, memTrace, follows what we call neural breadcrumbs extracting informative signals from transformer hidden states and attention patterns as they process candidate sequences. By analyzing layer-wise representation dynamics, attention distribution characteristics, and cross-layer transition patterns, we detect potential memorization fingerprints that traditional loss-based approaches may not capture. This approach yields strong membership detection across several model families achieving average AUC scores of 0.85 on popular MIA benchmarks. Our findings suggest that internal model behaviors can reveal aspects of training data exposure even when output-based signals appear protected, highlighting the need for further research into membership privacy and the development of more robust privacy-preserving training techniques for large language models.

Memorization has been shown to greatly inflate Large Language Models’ (LLMs) performance on domains such as math and logic, where success should primarily rely on applying generalizable reasoning rules. In many real-world applications, however, memorization is not meant to be eliminated but selectively constrained—for example, in story understanding, where background knowledge must be integrated with narrative context. Drawing on the cognitive science distinction between “verbatim” (exact recall) and “gist” (semantic abstraction) memorization, we propose a two-tier framework for analyzing how LLMs reason under different degrees of memory access. The Inductive (prompt-guided) Setting softly steers models to reason through selective, context-relevant recall, while the Restrictive Setting imposes stronger constraints by limiting verbatim memory access. Evaluating GPT-4o, LLaMA3.3-70B, and DeepSeek V3 on six character-centric story understanding benchmarks, we find up to a 45.2% accuracy drop under the Restrictive Setting, revealing strong dependence on surface recall. By contrast, the Inductive Setting maintains performance, indicating that prompting can align LLMs toward memorization-constrained reasoning.

As multilingual large language models become more widely used, ensuring their safety and fairness across diverse linguistic contexts presents unique challenges. While existing research on machine unlearning has mainly focused on monolingual settings, typically English, multilingual environments introduce additional complexities due to cross-lingual knowledge transfer and biases embedded in both pretraining and fine-tuning data. In this work, we address the problem of multilingual unlearning using the Aya-Expanse 8B model under two settings: (1) data unlearning and (2) concept unlearning. We extend benchmarks for factual knowledge and stereotypes into ten languages through translation—English, French, Arabic, Japanese, Russian, Farsi, Korean, Hindi, Hebrew, and Indonesian—spanning five language families and varying resource levels. Our experiments show that unlearning in high-resource languages tends to be more stable, with asymmetric transfer observed between typologically related languages. Moreover, analysis of linguistic distances reveals that syntactic similarity is the most predictive factor of cross-lingual unlearning effects.