EACL 2026 Papers

Aligning Large Language Models (LLMs) with human preferences is crucial, but standard methods like Reinforcement Learning from Human Feedback (RLHF) are often complex and unstable. In this work, we propose a new, simpler approach that recasts alignment through the lens of Maximum Marginal Likelihood (MML) estimation. Our new MML-based Preference Optimization (MMPO) maximizes the marginal log-likelihood of a preferred text output, using the preference pair as samples for approximation, and forgoes the need for both an explicit reward model and entropy maximization. We theoretically demonstrate that MMPO implicitly performs preference optimization, producing a weighted gradient that naturally up-weights chosen responses over rejected ones. Across models ranging from 135M to 8B parameters, we empirically show that MMPO: 1) is more stable with respect to the hyperparameter compared to alternative baselines, and 2) achieves competitive or superior preference alignment while better preserving the base model’s general language capabilities. Through a series of ablation experiments, we show that this improved performance is indeed attributable to MMPO’s implicit preference optimization within the gradient updates.

Text-to-Visualization (Text2Vis) systems translate natural language queries over tabular data into concise answers and executable visualizations. While closed-source LLMs generate functional code, the resulting charts often lack semantic alignment and clarity—qualities that can only be assessed post-execution. Open-source models struggle even more, frequently producing non-executable or visually poor outputs. Although supervised fine-tuning can improve code executability, it fails to enhance overall visualization quality, as traditional SFT loss cannot capture post-execution feedback. To address this gap, we propose RL-Text2Vis, the first reinforcement learning framework for Text2Vis generation. Built on Group Relative Policy Optimization (GRPO), our method uses a novel multi-objective reward that jointly optimizes textual accuracy, code validity, and visualization quality using post-execution feedback. By training Qwen2.5 models (7B and 14B), RL-Text2Vis achieves a 22% relative improvement in chart quality over GPT-4o on the Text2Vis benchmark and boosts code execution success from 78% to 97% relative to its zero-shot baseline. Our models significantly outperform strong zero-shot and supervised baselines and also demonstrate robust generalization to out-of-domain datasets like VIS-Eval and NVBench. These results establish GRPO as an effective strategy for structured, multimodal reasoning in visualization generation. We release our code at https://github.com/vis-nlp/RL-Text2Vis.

Logical fallacies are common in public communication and can mislead audiences; fallacious arguments may still appear convincing despite lacking soundness, because convincingness is inherently subjective. We present the first computational study of how emotional framing interacts with fallacies and convincingness, using large language models (LLMs) to systematically change emotional appeals in fallacious arguments. We benchmark eight LLMs on injecting emotional appeal into fallacious arguments while preserving their logical structures, then use the best models to generate stimuli for a human study. Our results show that LLM-driven emotional framing reduces human fallacy detection in F1 by 14.5% on average. Humans perform better in fallacy detection when perceiving enjoyment than fear or sadness, and these three emotions also correlate with significantly higher convincingness compared to neutral or other emotion states. Our work has implications for AI-driven emotional manipulation in the context of fallacious argumentation.

Language identification (LID) is a critical step in curating multilingual LLM pretraining corpora from web crawls. While many studies on LID model training focus on collecting diverse training data to improve performance, low-resource languages – often limited to single-domain data, such as the Bible – continue to perform poorly. To resolve these class imbalance and bias issues, we propose a novel supervised contrastive learning (SCL) approach to learn domain-invariant representations for low-resource languages. We show that our approach improves LID performance on out-of-domain data for low-resource languages by 3.2 percentage points, while maintaining its performance for the high-resource languages.

Inspired by their success in image synthesis, diffusion models offer a flexible, iterative alternative to rigid left-to-right text generation. However, a fundamental training-inference discrepancy hinders their performance: models are trained on corrupted ground-truth tokens, but at inference time they must denoise inputs corrupted from their own predictions. To bridge this gap, we propose a unified framework. First, Deeper Self-Prediction (DSP) is a multi-step training objective that teaches robust self-correction by forcing the model to denoise its own intermediate outputs. Second, UCB-guided Decoding is a principled inference algorithm that frames token re-masking as a multi-armed bandit problem, using the Upper Confidence Bound (UCB) to balance exploration and exploitation. Experiments on text generation tasks demonstrate consistent improvements over existing diffusion baselines. The framework achieves higher faithfulness and coherence according to both automatic metrics and LLM-as-a-Judge evaluations.

Recent advances in LLMs offer new opportunities for scalable, interactive mental health assessment, but excessive querying burdens users and is inefficient for real-world screening across transdiagnostic symptom profiles. We introduce MAQuA, a multi-outcome modeling and adaptive question-asking framework for simultaneous, multidimensional mental health screening. Combining multi-outcome modeling on language responses with item response theory (IRT) and factor analysis, MAQuA selects the questions with most informative responses across multiple dimensions at each turn to optimize diagnostic information, improving accuracy and potentially reducing response burden. Empirical results on a novel dataset reveal that MAQuA reduces the number of assessment questions required for score stabilization by 50–87% compared to random ordering (e.g., achieving stable depression scores with 71% fewer questions and eating disorder scores with 85% fewer questions). MAQuA demonstrates robust performance across both internalizing (depression, anxiety) and externalizing (substance use, eating disorder) domains, with early stopping strategies further reducing patient time and burden. These findings position MAQuA as a powerful and efficient tool for scalable, nuanced, and interactive mental health screening, advancing the integration of LLM-based agents into real-world clinical workflows.

In continual learning, our aim is to learn a new task without forgetting what was learned previously. In topic models, this translates to learning new topic models without forgetting previously learned topics. Previous work either considered Dynamic Topic Models (DTMs), which learn the evolution of topics based on the entire training corpus at once, or Online Topic Models, which are updated continuously based on new data but do not have long-term memory. To fill this gap, we propose the Continual Neural Topic Model (CoNTM), which continuously learns topic models at subsequent time steps without forgetting what was previously learned. This is achieved using a global prior distribution that is continuously updated. In our experiments, CoNTM consistently outperformed the dynamic topic model in terms of topic quality and predictive perplexity while being able to capture topic changes online. The analysis reveals that CoNTM can learn more diverse topics and better capture temporal changes than existing methods.

This paper presents a mechanistic investigation of how large language models (LLMs) generate contrastive sentiments. We define this task as transforming the sentiment of a given text (e.g., from positive to negative) while making minimal changes to its content. We identify two core mechanisms: (1) a preservation mechanism that maintains the sentiment of the input text, primarily mediated by specific attention heads, and (2) a sentiment transformation mechanism, which integrates a representation of the target sentiment label with the original valenced words using a circuit containing both MLP and attention layers. Building on these findings, we propose and validate a novel mechanistic intervention. By modifying key attention heads, we steer the LLM toward more effective contrastive generation, increasing the sentiment flip rate without sacrificing the minimality of changes. Our work not only deepens the understanding of the mechanisms underlying contrastive sentiment generation in LLMs, but also introduces a promising new direction to steer LLM behavior via targeted, mechanistic interventions.

Large language models (LLMs) have shown strong capabilities as task-solving agents across interactive domains. However, in complex environments, these agents may need to rely on auxiliary guidance to reduce the search space or make up for limited domain-specific knowledge. Such guidance includes human-provided manuals and demonstrations, retrieved examples from memory or external tools, high-level heuristics, and agent-acquired knowledge from prior interactions. However, this guidance may be imperfect. For example, due to changes in the environment, ambiguous or simplified language, or retrieval errors from external sources, guidance can be incomplete, outdated, or contextually mismatched, potentially causing errors or failures during task execution. To address this, we introduce MIRAGE, a benchmark for MeasurIng Robustness of LLM Agents under Imperfect GuidancE. MIRAGE includes procedurally generated environments in navigation, cooking, and gaming, where both the environment and the auxiliary guidance vary in fidelity and relevance. We further extend MIRAGE to realistic web tasks via WebArena, using noisy or underspecified instructions extracted from demonstrations. Our findings reveal critical failure modes in current LLM agents and motivate future work on improving their robustness under imperfect guidance.

Integrating external knowledge into Large Language Models (LLMs) iscrucial for many real-world applications, yet current methods like Retrieval-Augmented Generation (RAG) face limitations with broad, multi-source queries, while long-context models are computationally prohibitive.We introduce CacheNotes: Task-Aware Key-Value Cache Compression. Given a task description and a corpus, CacheNotes first generates a sequence of Compression-Planning-Tokens (CPTs), an offline task-focused distillation pass that identifies and organizes key information from the corpus. These CPTs are then used to guide a one-time compression of the corpus into a compact, reusable KV cache, which is then used alone at inference time to efficiently answer diverse, reasoning-intensive queries, eliminating repeated retrieval or context expansion.Experiments on LongBench show that, on Question-Answering tasks at a 20\times compression, CacheNotes outperforms RAG by over 8 F1 points and reduces latency by over 4\times. On RULER, it surpasses previous query-agnostic compression methods by 55 points, narrowing the gap to query-aware compression approaches. Additional results on real-world enterprise and synthetic datasets demonstrate its strong performance on multi-hop and broad-coverage queries.

In this paper, we study the expansion of pluralia tantum, i.e., defective nouns which lack a singular form, like scissors. We base our work on an annotation framework specifically developed for the study of lexicalization of pluralia tantum, namely Lexicalization profiles. On a corresponding hand-annotated testset, we show that the OpenAI and DeepSeek models provide useful annotators for semantic, syntactic and sense categories, with accuracy ranging from 51% to 89%, averaged across all feature groups and languages. Next, we turn to a large-scale investigation of pluralia tantum. Using dictionaries, we extract candidate words for Italian, Russian and English and keep those for which the changing ratio of singular and plural form is evident in a corresponding reference corpus. We use an LLM to annotate each instance from the reference corpora according to the annotation framework. We show that the large amount of automatically annotated sentences for each feature can be used to perform in-depth linguistic analysis. Focusing on the correlation between an annotated feature and the grammatical form (singular vs. plural), patterns of morpho-semantic change are noted.

AI-text detectors achieve high accuracy on in-domain benchmarks, but often struggle to generalize across different generation conditions such as unseen prompts, model families, or domains. While prior work has reported these generalization gaps, there are limited insights about the underlying causes. In this work, we present a systematic study aimed at explaining generalization behavior through linguistic analysis. We construct a comprehensive benchmark that spans 6 prompting strategies, 7 large language models (LLMs), and 4 domain datasets, resulting in a diverse set of human- and AI-generated texts. Using this dataset, we fine-tune classification-based detectors on various generation settings and evaluate their cross-prompt, cross-model, and cross-dataset generalization. To explain the performance variance, we compute correlations between generalization accuracies and feature shifts of 80 linguistic features between training and test conditions. Our analysis reveals that generalization performance for specific detectors and evaluation conditions is significantly associated with linguistic features such as tense usage and pronoun frequency.

Existing datasets for multimodal table understanding, such as MMTab, primarily provide short factual answers without explicit multi-step reasoning supervision. Models trained on these datasets often generate brief responses that offers insufficient accuracy and limited interpretability into how these models arrive at the final answer. We introduce CoReTab, a code-driven reasoning framework that produces scalable, interpretable, and automatically verifiable annotations by coupling multi-step reasoning with executable Python code. Using the CoReTab framework, we curate a dataset of 115K verified samples averaging 529 tokens per response and fine-tune open-source MLLMs through a three-stage pipeline. We evaluate the resulting model trained on CoReTab across 17 MMTab benchmarks spanning table question answering, fact verification, and table structure understanding. Our model achieves significant gains of +6.2%, +5.7%, and +25.6%, respectively, over MMTab-trained baselines, while producing transparent and verifiable reasoning traces. These results establish CoReTab as a robust and generalizable supervision framework for improving multi-step reasoning in multimodal table understanding.

People increasingly rely on Large Language Models (LLMs) for moral advice, which may influence humans’ decisions. Yet, little is known about how closely LLMs align with human moral judgments. To address this, we introduce the Moral Dilemma Dataset, a benchmark of 1,618 real-world moral dilemmas paired with a distribution of human moral judgments consisting of a binary evaluation and a free-text rationale. We treat this problem as a pluralistic distributional alignment task, comparing the distributions of LLM and human judgments across dilemmas. We find that models reproduce human judgments only under high consensus; alignment deteriorates sharply when human disagreement increases. In parallel, using a 60-value taxonomy built from 3,783 value expressions extracted from rationales, we show that LLMs rely on a narrower set of moral values than humans. These findings reveal a pluralistic moral gap–a mismatch in both the distribution and diversity of values expressed. To close this gap, we introduce Dynamic Moral Profiling (DMP), a Dirichlet-based sampling method that conditions model outputs on human-derived value profiles. DMP improves alignment by 64.3% and enhances value diversity, offering a step toward more pluralistic and human-aligned moral guidance.

Fine-tuning Large Language Models (LLMs) on sensitive datasets carries a substantial risk of unintended memorization and leakage of Personally Identifiable Information (PII), which can violate privacy regulations and compromise individual safety. In this work, we systematically investigate a critical and underexplored vulnerability: the exposure of PII that appears only in model inputs, not in training targets. Using both synthetic and real-world datasets, we design controlled extraction probes to quantify unintended PII memorization and study how factors such as language, PII frequency, task type, and model size influence memorization behavior. We further benchmark four privacy-preserving approaches including differential privacy, machine unlearning, regularization, and preference alignment, evaluating their trade-offs between privacy and task performance. Our results show that post-training methods generally provide more consistent privacy-utility trade-offs, while differential privacy achieves strong reduction in leakage in specific settings, although it can introduce training instability. These findings highlight the persistent challenge of memorization in fine-tuned LLMs and emphasize the need for robust, scalable privacy-preserving techniques.

The accelerating growth of scientific publications has intensified the need for scalable, trustworthy systems to synthesize knowledge across diverse literature. While recent retrieval-augmented generation (RAG) methods have improved access to scientific information, they often overlook citation graph structure, adapt poorly to complex queries, and yield fragmented, hard-to-verify syntheses. We introduce SciRAG, an open-source framework for scientific literature exploration that addresses these gaps through three key innovations: (1) adaptive retrieval that flexibly alternates between sequential and parallel evidence gathering; (2) citation-aware symbolic reasoning that leverages citation graphs to organize and filter supporting documents; and (3) outline-guided synthesis that plans, critiques, and refines answers to ensure coherence and transparent attribution. Extensive experiments across multiple benchmarks such as QASA and ScholarQA demonstrate that SciRAG outperforms prior systems in factual accuracy and synthesis quality, establishing a new foundation for reliable, large-scale scientific knowledge aggregation.

Claim verification is a core module in automated fact-checking systems, tasked with determining claim veracity using retrieved evidence. This work presents KG-CRAFT, a novel knowledge graph-based contrastive reasoning method that enhances automatic claim verification by LLMs. Our approach first constructs a knowledge graph from claims and associated reports, then formulates contextually relevant contrastive questions based on the knowledge graph structure. These questions guide the distillation of evidence-based reports, which are synthesised into a concise summary for veracity assessment. Extensive evaluations on two real-world datasets (LIAR-RAW and RAWFC) demonstrate that our method achieves a new state-of-the-art in predictive performance. Comprehensive analyses validate in detail the effectiveness of our knowledge graph-based contrastive reasoning approach in improving LLMs’ fact-checking capabilities.

Speculative generation has emerged as a promising technique to accelerate inference in large language models (LLMs) by leveraging parallelism to verify multiple draft tokens simultaneously. However, the fundamental limits on the achievable speedup remain poorly understood. In this work, we establish the first “tight” lower bounds on the runtime of any deterministic speculative generation algorithm. This is achieved by drawing a parallel between the token generation process and branching random walks, which allows us to analyze the optimal draft tree selection problem. We prove, under basic assumptions, that the expected number of tokens successfully predicted per speculative iteration is bounded as \mathbb{E}[X] \leq (\mu + \mu_{(2)})\log(B )/\mu^2 + O(1), where B is the verifier’s batch size, \mu is the expected entropy of the verifier’s output distribution, and \mu_{(2)} is this entropy’s second moment. This result provides new insights into the limits of parallel token generation, and could guide the design of future speculative decoding systems. Empirical evaluations on Llama models validate our theoretical predictions, confirming the tightness of our bounds in practical settings.

Large Language Models exhibit implicit personalities in their generation, but reliably controlling or aligning these traits to meet specific needs remains an open challenge. The need for effective mechanisms for behavioural manipulation of the model during generation is a critical gap in the literature that needs to be fulfilled. Personality-aware LLMs hold a promising direction towards this objective.However, the relationship between these psychological constructs and their representations within LLMs remains underexplored and requires further investigation. Moreover, it is intriguing to understand and study the use of these representations to steer the models’ behaviour. We propose a novel pipeline that extracts hidden state activations from transformer layers using the Big Five Personality Traits (Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism), which is a comprehensive and empirically validated framework to model human personality applies low-rank subspace discovery methods, andidentifies trait-specific optimal layers across different model architectures for robust injection. The resulting personality-aligned directions are then operationalised through a flexible steering framework with dynamic layer selection, enabling precise control of trait expression in LLM outputs. Our findings reveal that personality traits occupy a low-rank shared subspace, and that these latent structures can be transformed into actionable mechanisms for effective steering through careful perturbations without impacting the fluency, variance and general capabilities, helping to bridge the gap between psychological theory and practical model alignment.