EACL 2026 Papers

We investigate whether large language models (LLMs) can improve through recursive training on self-generated text, a topic where prior studies report conflicting outcomes: some find evidence of performance gains (i.e., self-improvement), while others observe performance degradation (i.e., model collapse). To clarify this discrepancy, we use the OLMo-2 models as non-toy LLMs and perform multiple rounds of continual pre-training using self-generated text with different prompting strategies and data filtering. Our experiments show that naive recursive self-training does not improve either perplexity or downstream task performance, regardless of model size. These results suggest that model collapse observed in naive recursive training is inherent to the training procedure itself, while self-improvement likely owes its success not to the model’s autonomous refinement but to human-designed, strategic synthetic pipelines that inject external intelligence.

Emotion recognition in multi-speaker conversations faces significant challenges due to speaker ambiguity and severe class imbalance. We propose a novel framework that addresses these issues through three key innovations: (1) a speaker identification module that leverages audio-visual synchronization to accurately identify the active speaker, (2) a knowledge distillation strategy that transfers superior textual emotion understanding to audio and visual modalities, and (3) hierarchical attention fusion with composite loss functions to handle class imbalance. Comprehensive evaluations on MELD and IEMOCAP datasets demonstrate superior performance, achieving 67.75% and 72.44% weighted F1 scores respectively, with particularly notable improvements on minority emotion classes.

Recent work has explored pruning merges from BPE subword tokenisers using corpus data as a signal for which merges to prune. We argue that because a BPE tokeniser contains a rich data structure on top of its vocabulary set, this in itself can be used as a guide to modify its merges such that segmentations become more desirable. We apply this argument to one of those pruning algorithms, BPE-knockout, by introducing a new reification step that suggests new merges by inspecting the effects left by pruning. By alternating both processes iteratively until convergence, we get a new BPE tokeniser, ReBPE, which outperforms the original BPE-knockout algorithm on morphological alignment in all 14 languages tested by over 11% F1 on average.

We introduce ReciFine, the largest human-evaluated, finely annotated recipe dataset to date, designed to advance controllable and trustworthy recipe generation. Existing resources, such as RecipeNLG, extract food items only from ingredient lists, overlooking entities expressed in instructions, including tools, chef actions, food and tool states, and durations, which are crucial for realistic and context-aware generation. To address this limitation, we extend RecipeNLG with finely annotated extraction of over 97 million entities across ten entity types from 2.2 million recipes. We are the first to explore recipe generation with explicit control over multiple entity types, enabling models to generate recipes conditioned not only on ingredients but also on tools, chef actions, cooking durations, and other contextual factors. Large language models fine-tuned or few-shot prompted with ReciFine extractions consistently outperform those trained on ingredient-list data alone across both automatic and human evaluations. ReciFine establishes a foundation for factual, coherent, structured, controllable recipe generation, and we release a human-annotated benchmark to support future evaluation and model development.

Attribute-controlled translation (ACT) seeks to produce translations that satisfy specific constraints on linguistic and stylistic attributes. While careful prompt engineering can enable large language models to perform strongly in this task, its effectiveness is mainly limited to models of very large size. For this reason, in this paper we set to improve the performance of language models of more contained size by leveraging the contrastive nature of ACT tasks with preference optimization, as well as exploiting knowledge distillation with synthetically-generated training samples from larger models. As a resource for this investigation, we also introduce PREF-FAME-MT, a large, contrastive, formality-controlled parallel corpus which has been generated by expanding the existing FAME-MT dataset with synthetic contrastive samples. Experiments conducted over three datasets for formality- and gender-controlled translation with 71 distinct language pairs have demonstrated the effectiveness of the proposed approach at simultaneously improving attribute matching and translation quality. We release all our code and datasets to allow reproduction and expansion of our work.

We introduce VIGiA, a novel multimodal dialogue model designed to understand and reason over complex, multi-step instructional video action plans. Unlike prior work which focuses mainly on text-only guidance, or treats vision and language in isolation, VIGiA supports grounded, plan-aware dialogue that requires reasoning over visual inputs, instructional plans, and interleaved user interactions. To this end, VIGiA incorporates two key capabilities: (1) multimodal plan reasoning, enabling the model to align uni- and multimodal queries with the current task plan and respond accurately; and (2) plan-based retrieval, allowing it to retrieve relevant plan steps in either textual or visual representations. Experiments were done on a novel dataset with rich Instructional Video Dialogues aligned with Cooking and DIY plans. Our evaluation shows that VIGiA outperforms existing state-of-the-art models on all tasks in a conversational plan guidance setting, reaching over 90% accuracy on plan-aware VQA.

Identifying LLM-generated code through watermarking poses a challenge in preserving functional correctness. Previous methods rely on the assumption that watermarking high-entropy tokens effectively maintains output quality. Our analysis reveals a fundamental limitation of this assumption: syntax-critical tokens such as keywords often exhibit the highest entropy, making existing approaches vulnerable to logic corruption. We present STONE, a syntax-aware watermarking method that embeds watermarks only in non-syntactic tokens and preserves code integrity. For rigorous evaluation, we also introduce STEM, a comprehensive metric that balances three critical dimensions: correctness, detectability, and imperceptibility. Across Python, C++, and Java, STONE preserves correctness, sustains strong detectability, and achieves balanced performance with minimal computational overhead. Our implementation is available at https://github.com/inistory/STONE-watermarking.

Large Language Models (LLMs) possess a remarkable capacity to generate persuasive and intelligible language. However, coherence does not equate to truthfulness, as the responses often contain subtle hallucinations. Existing benchmarks are limited by static and narrow questions, leading to limited coverage and misleading evaluations. We present KGHaluBench, a Knowledge Graph-based hallucination benchmark that assesses LLMs across the breadth and depth of their knowledge, providing a fairer and more comprehensive insight into LLM truthfulness. Our framework utilises the KG to dynamically construct challenging, multifaceted questions, whose difficulty is then statistically estimated to address popularity bias. Our automated verification pipeline detects abstentions and verifies the LLM’s response at both conceptual and correctness levels to identify different types of hallucinations. We evaluate 25 frontier models, using novel accuracy and hallucination metrics. The results provide a more interpretable insight into the knowledge factors that cause hallucinations across different model sizes. KGHaluBench is publicly available to support future developments in hallucination mitigation.

Speculative decoding (SD) has proven effective for accelerating LLM inference by quickly generating draft tokens and verifying them in parallel. However, SD remains largely unexplored for Large Vision-Language Models (LVLMs), which extend LLMs to process both image and text prompts. To address this gap, we benchmark existing inference methods with small draft models on 11 datasets across diverse input scenarios and observe scenario-specific performance fluctuations. Motivated by these findings, we propose Test-time Adaptive Batched Ensemble Drafting (TABED), which dynamically ensembles multiple drafts obtained via batch inference by leveraging deviations from past ground truths available in the SD setting. The dynamic ensemble method achieves an average robust walltime speedup of 1.74× over autoregressive decoding and a 5% improvement over single drafting methods, while remaining training-free and keeping ensembling costs negligible through parameter sharing. With its plug-and-play compatibility, we further enhance TABED by integrating advanced verification and alternative drafting methods. Code and custom-trained models are available at https://github.com/furiosa-ai/TABED.

Scientific research is inherently global. However, the vast majority of academic journals are published exclusively in English, creating barriers for non-native-English-speaking researchers. In this study, we leverage large language models (LLMs) to translate published scientific articles while preserving their native JATS XML formatting, thereby developing a practical, automated approach for implementation by academic journals. Using our approach, we translate articles across multiple scientific disciplines into 28 languages. To evaluate translation accuracy, we introduce a novel question-and-answer (QA) benchmarking method and show an average performance of 95.9%, indicating that the key scientific details are accurately conveyed. In a user study, we translate the scientific papers of 15 researchers into their native languages. Interestingly, a third of the authors found many technical terms “overtranslated,” expressing a preference to keep terminology more familiar in English untranslated. Finally, we demonstrate how in-context learning techniques can be used to align translations with domain-specific preferences such as mitigating overtranslation, highlighting the adaptability and utility of LLM-driven scientific translation.

Emotion classification on social media is especially difficult when texts include informal, culturally grounded language like slang. Standard NLP benchmarks often miss these nuances, particularly in low-resource settings. We present SLANG-GraphRAG, a retrieval-augmented framework that integrates a culture-specific slang knowledge graph into large language models via one-shot prompting. Using multiple retrieval strategies, we incorporate slang definitions, regional usage, and conversational context. Our results show that incorporating structured cultural knowledge into the retrieval process leads to significant improvements, improving accuracy by up to 31% and F1 score by 28%, outperforming traditional and unstructured retrieval methods. To better evaluate model behavior, we propose a probabilistic metric that reflects the distribution of human annotations, providing a more nuanced measure of performance. This highlights the value of culturally sensitive applications and more balanced evaluation in subjective NLP tasks.

Understanding regional similarities is crucial for applications such as urban planning, tourism recommendations, business expansion, and disease prevention. While spatial data, including POI distributions, check-in activity, and building footprints, offer valuable insights, existing similarity methods—based on distance metrics, embeddings, or deep metric learning—fail to capture the contextual richness and adapt to heterogeneous spatial data. To overcome these limitations, we introduce a novel similar region search framework that ranks candidate regions based on their similarity to a query region using large language models. To further enhance performance, we fine-tune the model through self-supervised learning by introducing controlled noise into spatial data. This generates similar and dissimilar samples without relying on extensive labeled data. By transforming spatial data into natural language descriptions, our method seamlessly integrates heterogeneous datasets without requiring structural modifications, ensuring scalability across diverse urban contexts. Experiments on multiple real-world city datasets, including cross-city evaluation, demonstrate that our framework significantly outperforms state-of-the-art methods in both accuracy and ranking performance.

Steering vectors are a lightweight method for controlling text properties by adding a learned bias to language model activations at inference time. While predominantly studied for multiple-choice and toy tasks, their effectiveness in free-form generation remains largely unexplored. Moving “Beyond Multiple Choice,” we evaluate steering vectors for controlling topical focus, sentiment, toxicity, and readability in abstractive summaries across the SAMSum, NEWTS, and arXiv datasets. We find that steering effectively controls targeted properties, but high steering strengths consistently induce degenerate repetition and factual hallucinations. Prompting alone preserves summary quality but offers weaker control. Combining both methods yields the strongest control and the most favorable efficacy-quality trade-off at moderate steering strengths. Our work demonstrates that steering vectors face a critical control-quality trade-off in free-form generation, and that hybrid approaches offer best balance in practice.

Question Answering on Tabular Data (or Table Question Answering) has seen tremendous advances with the coming of new generation Large Language Models (LLMs). Despite this, significant challenges still remain to be solved if we are to develop robust enough approaches for general usage. One of these is ambiguity in question answering, which historically has not merited much attention due to the previously limited capabilities of LLMs. In this work, we outlay the main types of ambiguousness inherent to tabular data. Then, we discuss how they are influenced by the way our models interact with the information stored in the tables, and we test the capabilities of some LLMs in detecting them. This work provides an initial ground for a deeper discussion on how to approach ambiguity in Tabular Data in the age of LLMs.

The inherent risk of generating harmful and unsafe content by Large Language Models (LLMs), has highlighted the need for their safety alignment. Various techniques like supervised fine-tuning, reinforcement learning from human feedback, and red-teaming were developed for ensuring the safety alignment of LLMs. However, the robustness of these aligned LLMs is always challenged by adversarial attacks that exploit unexplored and underlying vulnerabilities of the safety alignment. In this paper, we develop a novel black-box jailbreak attack, called BitBypass, that leverages hyphen-separated bitstream camouflage for jailbreaking aligned LLMs. This represents a new direction in jailbreaking by exploiting fundamental information representation of data as continuous bits, rather than leveraging prompt engineering or adversarial manipulations. Our evaluation of five state-of-the-art LLMs, namely GPT-4o, Gemini 1.5, Claude 3.5, Llama 3.1, and Mixtral, in adversarial perspective, revealed the capabilities of BitBypass in bypassing their safety alignment and tricking them into generating harmful and unsafe content. Further, we observed that BitBypass outperforms several state-of-the-art jailbreak attacks in terms of stealthiness and attack success. Overall, these results highlights the effectiveness and efficiency of BitBypass in jailbreaking these state-of-the-art LLMs.

We introduce MOSAIC (Masked Objective with Selective Adaptation for In-domain Contrastive learning), a multi-stage framework for domain adaptation of text embedding models that incorporates joint domain-specific masked supervision. Our approach addresses the challenges of adapting large-scale general-domain text embedding models to specialized domains. By jointly optimizing masked language modeling (MLM) and contrastive objectives within a unified training pipeline, our method enables effective learning of domain-relevant representations while preserving the robust semantic discrimination properties of the original model. We empirically validate our approach on both high-resource and low-resource domains, achieving improvements up to 13.4% in NDCG@10 (Normalized Discounted Cumulative Gain) over strong general-domain baselines. Comprehensive ablation studies further demonstrate the effectiveness of each component, highlighting the importance of balanced joint supervision and staged adaptation.

We present a novel metric for the evaluation of morphological plausibility of subword segmentation.Unlike the typically used morpheme boundary or retrieval F-score, which requires gold segmentation data that is either unavailable or of inconsistent quality across many languages, our approach utilizes morpho-syntactic features.These are available in resources such as Universal Dependencies or UniMorph for a much wider range of languages.The metric works by probabilistically aligning subwords with morphological features through an IBM Model 1.Our experiments show that the metric correlates well with traditional morpheme boundary recall while being more broadly applicable across languages with different morphological systems.

Table modeling has progressed for decades. In this work, we revisit this trajectory and highlight emerging challenges in the LLM era, particularly the paradox of choice: the difficulty of attributing performance gains amid diverse base models and training sets in the context of table instruction tuning. We replicate four table LLMs by instruction-tuning three foundation models on four existing datasets, yielding 12 models. We then evaluate these models across 16 table benchmarks. Our study is the first to quantitatively disentangle the effects of training data and base model selection, revealing that base model choice plays a more dominant role than the training data itself. Generalization and reasoning remain challenging, inviting future effort on table modeling. Based on our findings, we share our thoughts on the future directions for table modeling.