ACL 2026 Papers

While Large Language Models (LLMs) have increasingly assisted in historical tasks such as text processing, their capacity for professional-level historical reasoning remains underexplored. Existing benchmarks primarily assess basic knowledge breadth or lexical understanding, failing to capture the higher-order skills—such as evidentiary reasoning—that are central to historical research. To fill this gap, we introduce ProHist-Bench, a novel benchmark anchored in the Chinese Imperial Examination (Keju) system—a comprehensive microcosm of East Asian political, social, and intellectual history spanning over 1,300 years. Developed through deep interdisciplinary collaboration, ProHist-Bench features 400 challenging, expert-curated questions across eight dynasties, accompanied by 10,891 fine-grained evaluation rubrics. Through a rigorous evaluation of 18 LLMs, we reveal a significant proficiency gap: even state-of-the-art LLMs struggle with complex historical research questions. We hope ProHist-Bench will facilitate the development of domain-specific reasoning LLMs, advance computational historical research, and further uncover the untapped potential of LLMs. We release ProHist-Bench at https://github.com/inclusionAI/ABench/tree/main/ProHist-Bench.

Post-training adaptation of large language models is commonly achieved through parameter updates or input based methods such as fine-tuning, parameter-efficient adaptation, and prompting. In parallel, a growing body of work modifies internal activations at inference time to influence model behavior, an approach known as steering. Despite increasing use, steering is rarely analyzed within the same conceptual framework as established adaptation methods.In this work, we argue that steering should be regarded as a form of model adaptation. We introduce a set of functional criteria for adaptation methods and use them to compare steering approaches with classical alternatives. This analysis positions steering as a distinct adaptation paradigm based on targeted interventions in activation space, enabling local and reversible behavioral change without parameter updates. The resulting framing clarifies how steering relates to existing methods, motivating a unified taxonomy for model adaptation.

Label-free reinforcement learning enables large language models to improve reasoning capabilities without ground-truth supervision, typically by treating majority-voted answers as pseudo-labels. However, we identify a critical failure mode: as training maximizes self-consistency, output diversity collapses, causing the model to confidently reinforce systematic errors that evade detection. We term this the consensus trap. To escape it, we propose CoVerRL, a framework where a single model alternates between generator and verifier roles, with each capability bootstrapping the other. Majority voting provides noisy but informative supervision for training the verifier, while the improving verifier progressively filters self-consistent errors from pseudo-labels. This co-evolution creates a virtuous cycle that maintains high reward accuracy throughout training. Experiments across Qwen and Llama model families demonstrate that CoVerRL outperforms label-free baselines by 4.7-5.9% on mathematical reasoning benchmarks. Moreover, self-verification accuracy improves from around 55% to over 85%, confirming that both capabilities genuinely co-evolve.

Large Vision–Language Models (LVLMs) have shown strong potential as multilingual Graphical User Interface (GUI) agents, as evidenced by existing GUI benchmarks. However, these benchmarks exhibit two primary limitations: (1) although Perception and Reasoning (P R) capabilities are fundamental for GUI agents, current benchmarks lack fine-grained diagnostics to identify which specific capabilities lead to task failures, hindering targeted improvements; (2) existing benchmarks fail to provide a strictly aligned cross-lingual evaluation environment, introducing confounding factors that prevent isolating the language impact on GUI agent performance. To address these issues, we propose the Multilingual P R GUI Benchmark (MPR-GUI-Bench), featuring strictly aligned environments across six languages and eight fine-grained P R tasks. Our benchmark reveals consistent P R gaps between English and non-English settings, particularly on reasoning-intensive tasks. To leverage the superior English P R capabilities for bridging cross-lingual gaps, we identify layers sensitive to language and propose GUI-XLI, a GUI Cross-Lingual Intervention method that aligns non-English hidden states with their English counterparts at these layers during inference. Experiments show that GUI-XLI effectively reduces the cross-lingual gaps, with an average gain of 6.5% in non-English settings.

Fine-tuning Large Language Models on a political topic will significantly manipulate their political stance on various issues and unintentionally affect their stance on broad topics. While previous studies have proposed this issue, there is still a lack of understanding regarding the internal representations of these stances and the mechanisms that lead to unintended cross-topic generalization. In this paper, we systematically explore the internal mechanisms underlying this phenomenon from a neuron-level perspective and how to mitigate the cross-topic generalization of political fine-tuning. Firstly, we propose Political Neuron Localization through Activation Contrasting (PNLAC) to identify two distinct types of political neurons: general political neurons, which govern stance across multiple political topics, and topic-specific neurons that affect the model’s political stance on individual topics. We find that these political neuron types exist in the middle and later layers across four models and datasets through activation patching experiments. Leveraging these insights, we introduce InhibitFT, an inhibition-based fine-tuning method that effectively mitigates the cross-topic stance generalization. Experimental results demonstrate the robustness of the identified neuron types across various models and datasets and show that InhibitFT significantly reduces the cross-topic stance generalization by 20% on average while preserving topic-specific performance. Moreover, we demonstrate that selectively inhibiting only 5% of neurons is sufficient to effectively mitigate the cross-topic stance generalization.

Wikipedia’s perceived high quality and broad language coverage have established it as a fundamental resource in NLP. However, in recent years, such assumptions of high quality have become the subject of scrutiny in low-resource and multilingual contexts. In this study, we subject the entirety of non-English Wikipedia to a data filtering procedure typically reserved for noisy web-text — a process which removes a large percentage of the collection’s data. In analysing the removed data, we reveal numerous systematic quality issues, such as script and language contamination, repeated template and placeholder articles, and a high concentration of bot-generated content. We consolidate these findings into a 4-level quality ranking of Wikipedia, which shows strong correspondence with alternative quality measures and heuristics. Lastly, we evaluate the downstream impact of quality filtering in three practical language modelling scenarios, showing that models trained on filtered data largely match or outperform those trained on raw Wikipedia, with the largest gains observed for lower-quality language editions. Ultimately, our experiments serve as a first step in establishing quality-aware best practices for Wikipedia utilization in NLP, laying groundwork that can inform future dataset creation and curation efforts.

Retrieval-Augmented Generation (RAG) enhances the response capabilities of language models by integrating external knowledge sources. However, document chunking as an important part of RAG system often lacks effective evaluation tools. This paper first analyzes why existing RAG evaluation benchmarks are inadequate for assessing document chunking quality, specifically due to evidence sparsity. Based on this conclusion, we propose HiCBench, which includes manually annotated multi-level document chunking points, synthesized evidence-dense question answer(QA) pairs, and their corresponding evidence sources. We also propose HiChunk, a hierarchical document structuring framework using fine-tuned LLMs and the Auto-Merge retrieval algorithm to enhance retrieval quality. Experiments demonstrate that HiCBench effectively evaluates the impact of different chunking methods across the entire RAG pipeline. Moreover, HiChunk achieves better chunking quality within reasonable time consumption, thereby enhancing the overall performance of RAG systems. Source code is available at https://github.com/TencentCloudADP/hichunk.

The rapid advancement of large language models (LLMs) has driven the deployment of LLM-based AI tutors on online learning platforms. This widespread adoption highlights an urgent need for systematic benchmarks to evaluate their tutoring capabilities. However, existing evaluations predominantly focus on isolated, short-term interactions, overlooking the inherently long-term nature of learning. To bridge this gap, we introduce LongTutor, a benchmark for long-term personalized tutoring grounded in formative assessment theory. Built from expert-annotated real-world learning logs, LongTutor evaluates LLMs across three progressive tasks: historical evidence acquisition, knowledge state diagnosis, and adaptive teaching action. Our experiments reveal a critical capability mismatch: while LLMs excel at evidence acquisition, they struggle to effectively leverage long-term history for accurate diagnosis and adaptive teaching. To enable scalable benchmark expansion, we further propose an automated generator–verifier pipeline, paving the way toward truly long-term AI tutoring systems.

Supervised fine-tuning (SFT) on chain-of-thought (CoT) trajectories demonstrations is a common approach for enabling reasoning in large language models. Standard practices typically only retain trajectories with correct final answers (positives) while ignoring the rest (negatives). We argue that this paradigm discards substantial supervision and exacerbates overfitting, limiting out-of-domain (OOD) generalization. Specifically, we surprisingly find that incorporating negative trajectories into SFT yields substantial OOD generalization gains over positive-only training, as these trajectories often retain valid intermediate reasoning despite incorrect final answers. To understand this effect in depth, we systematically analyze data, training dynamics, and inference behavior, identifying 22 recurring patterns in negative chains that serve a dual role: they moderate loss descent to mitigate overfitting during training and boost policy entropy by 35.67% during inference to facilitate exploration. Motivated by these observations, we further propose Gain-based LOss Weighting (GLOW), an adaptive, sample-aware scheme that exploits such distinctive training dynamics by rescaling per-sample loss based on inter-epoch progress. Empirically, GLOW efficiently leverages unfiltered trajectories, yielding a 5.51% OOD gain over positive-only SFT on Qwen2.5-7B and boosting MMLU from 72.82% to 76.47% as an RL initialization. Code is available at Github.

Speculative decoding accelerates autoregressive generation by letting draft tokens bypass full verification, but conventional frameworks suffer from frequent false rejections, particularly when draft models produce semantically correct but lexically divergent outputs. In this paper, we present Calibrated Speculative Decoding (CSD), a training-free framework that recovers valid tokens discarded by standard verification. Guided by the principle of “Frequency-Guided Candidate Selection and Probability-Guarded Acceptance,” CSD incorporates two lightweight modules: Online Correction Memory, which aggregates historical rejections to propose recurring divergence patterns as rescue candidates, and Semantic Consistency Gating, which verifies candidate admissibility using probability ratios instead of exact token matching. Our evaluation across diverse large language models demonstrates that CSD outperforms existing methods, achieving a peak throughput speedup of 2.33x. CSD preserves model accuracy across all tasks while further boosting performance on complex reasoning datasets. These results establish CSD as a highly effective, lightweight solution for practical LLM deployments.

Natural Language to Excel Formula (NL2Formula) translates user intent into executable spreadsheet formulas. However, current models often produce near-miss outputs—formulas that parse correctly yet fail at execution due to an incorrect function, operator, or reference. Through a systematic error analysis, we find that these errors repeatedly arise from a small set of structural decision points, motivating the need for typed error supervision rather than general error signals. To this end, we introduce an abstract syntax tree (AST)-based error taxonomy that organizes common error modes by the kind of decision that goes wrong in the parse tree. Building on this taxonomy, we propose Error-Aware Contrastive Few-Shot Learning (ECFL), an error-aware framework that unifies training and inference around typed error supervision. During offline training, ECFL mines near-miss errors, assigns error types under the taxonomy, and builds error-aware contrastive demonstrations for fine-tuning. During online inference, a lightweight predictor estimates likely error types and triggers targeted retrieval of contrastive demonstrations to guide single-pass decoding. Experiments show ECFL improves Exact Match (EM) by 6.4 points over supervised fine-tuning (SFT) and matches self-consistency (SC@5) accuracy at substantially lower inference cost.

Large language models (LLMs) have achieved strong performance in general machine translation, yet their ability in culture-aware scenarios remains poorly understood. To bridge this gap, we introduce CanMT, a Culture-Aware Novel-Driven Parallel Dataset for Machine Translation, together with a theoretically grounded, multi-dimensional evaluation frame work for assessing cultural translation quality. Leveraging CanMT, we systematically evaluate a wide range of LLMs and translation systems under different translation strategy constraints. Our findings reveal substantial performance disparities across models and demonstrate that translation strategies exert a systematic influence on model behavior. Further analysis shows that translation difficulty varies across types of culture-specific items, and that a persistent gap remains between models’recognition of culture-specific knowledge and their ability to correctly operationalize it in translation outputs. In addition, incorporating reference translations is shown to substantially improve evaluation reliability in LLM-as-a-judge, underscoring their essential role in assessing culture-aware translation quality. The corpus and code are available at CanMT.

Fine-tuning pretrained LLMs has proven effective for reaching state-of-the-art performance on specific tasks like machine translation. However, this process often implies sacrificing general-purpose capabilities, such as conversational reasoning and instruction-following, hampering the usefulness of the system in real-world applications requiring a mixture of skills. In this paper, we introduce Tower+, a suite of models designed to deliver strong performance on both translation and multilingual general-purpose text capabilities. We improve the TOwer (Alves et al., 2024) recipe by adding novel stages of preference optimization and reinforcement learning with verifiable rewards, in addition to continued pretraining and supervised fine-tuning. At each stage, we carefully generate and curate data to strengthen performance on translation and general-purpose tasks like coding, mathematics, and instruction-following. We develop models at multiple scales: 2B, 9B, and 72B. Our smaller models often outperform larger general-purpose open-weight and proprietary LLMs (e.g., Llama 3.3 70B, GPT-4o). Our largest model delivers best-in-class translation performance for high-resource languages, and top results on multilingual Arena Hard and IF-MT, a benchmark we introduce for evaluating both translation and instruction-following.

Medical multiple-choice question answering (MCQA) benchmarks implicitly assume that large language models (LLMs) should always commit to an answer. However, in clinical practice, uncertainty is pervasive and abstaining is often the safest action. We introduce MedQAbstain, a benchmark explicitly designed to evaluate medical abstention under uncertainty. MedQAbstain repurposes standard medical MCQA datasets by removing the gold answer and introducing an explicit “I abstain” option, framed as a safety-critical decision with clinical consequences. The benchmark supports systematic analysis across abstention regimes, distractor complexity, and input modalities, and elicits self-reported model confidence to study calibration. Across all settings, we find that state-of-the-art LLMs systematically overcommit, rarely abstaining even when the question itself is hidden. These results reveal a fundamental mismatch between LLM behavior and clinical norms, highlighting abstention as a critical but overlooked dimension of medical decision-making evaluation.

The semantic gap between colloquial user queries and professional legal documents presents a fundamental challenge in Legal Case Retrieval (LCR). Existing dense retrieval methods typically treat LCR as a black-box semantic matching process, neglecting the explicit juridical logic that underpins legal relevance. To address this, we propose GLIER (Generative Legal Inference and Evidence Ranking), a framework that reformulates retrieval as an inference process over latent legal variables. GLIER decomposes the task into two interpretability-driven stages: (1) A Joint Generative Inference module that translates raw queries into latent legal indicators (Charges and Legal Elements), employing a unified sequence-to-sequence strategy where charges and elements are generated jointly to enforce logical consistency; and (2) A Multi-View Evidence Fusion mechanism that aggregates generative confidence with structural and lexical signals for precise ranking. Extensive experiments on LeCaRD and LeCaRDv2 demonstrate that GLIER outperforms strong baselines like SAILER and KELLER. Notably, our framework exhibits exceptional data efficiency, maintaining robust performance even when trained with only 10% of the data.

Large Language Models (LLMs) have emerged as central planners in Vision-and-Language Navigation (VLN), yet their complexity increasingly obscures their internal decision-making. Existing interpretability methods typically isolate temporal criticality from feature salience, creating an alignment gap and failing to account for the behavioral instability of black-box agents. To address this, we propose DEFT, a unified dual-view framework that demystifies agent behavior by jointly analyzing \textit{when} a decision is pivotal and \textit{what} visual evidence grounds it. Featuring a dual-head architecture with a shared latent representation, DEFT employs a \textit{Mask Head} for counterfactual-based criticality detection and an \textit{Action Head} that leverages an ensemble of surrogates to recover robust visual cues. Extensive experiments on MatterPort3D across three LLM-based agents demonstrate that DEFT outperforms baselines in both temporal and feature fidelity. User studies further validate its utility, showing 78% alignment with human intuition.

Modal dependency parsing-the task of identifying a semantic graph that represents who is responsible for an event-centered claim and with what degree of certainty-relies on recognizing source-introducing cues and correctly linking them to their associated content. However, prior work has largely focused on identifying sources only, treating cue expressions and their modal coverage as auxiliary signals. In this work, we propose a structured prediction framework that leverages large language models (LLMs) to explicitly identify source-cue pairs as well as their respective scope, which together define the modal contexts governing downstream source attribution for events. By concentrating learning at the source-cue level and constraining event-level decisions to a small, scope-defined candidate set, our top-down approach enables more efficient inference in long, event-rich documents. Experiments show this approach surpasses prior state-of-the-art results by 3 and 4% for English and Chinese datasets, respectively.

Temporal Knowledge Graph Question Answering (TKGQA) aims to answer temporal questions using knowledge from Temporal Knowledge Graphs (TKGs).Existing LLM-based TKGQA methods typically utilize RAG-based or Agent-based paradigms, yet both struggle to construct reliable temporal evidence chains. RAG-based approaches primarily rely on semantic retrieval to fetch question-relevant contexts but overlook the structural dependencies within TKGs, leading to broken evidence chains, whereas iterative agents are prone to error propagation during multi-step reasoning.To address these limitations, we propose TECQA, a framework designed to construct temporal evidence chains for LLM reasoning. Firstly, TECQA employs structure-guided subgraph retrieval to capture structural dependencies and intermediate reasoning paths. Subsequently, it utilizes a k-nearest temporal neighbor pruning strategy to filter irrelevant noise while strictly preserving the continuous local history surrounding critical events. Finally, the retained temporal neighbors are serialized by temporal proximity to explicitly reconstruct a coherent temporal evidence chain. Extensive experiments on MultiTQ and CronQuestions demonstrate that TECQA achieves state-of-the-art performance, outperforming strong baselines by 45.3% particularly on complex queries. Code is available at https://github.com/SimonsLiu/TECQA.

Recently, activation steering has attracted considerable attention as a low-cost approach to improving the safety of large language models (LLMs). However, most existing methods apply interventions indiscriminately, often causing excessive refusal of benign queries. Although recent works have begun to explore selective intervention, their intervention decisions typically rely on residual stream activations where information is highly entangled, resulting in limited discriminative power and unreliable interventions. To address this issue, we propose FFN-Guided activation steering (FGAS). Motivated by the observation that feed-forward networks (FFNs) in LLMs serve as core modules for knowledge storage, we propose leveraging FFN output activations as more discriminative signals for intervention, since these activations more explicitly reflect the intent of a query. For a given query, FGAS projects the corresponding FFN output activation into a low-dimensional subspace that effectively separates harmful and benign queries, and then makes precise intervention decisions by assessing its similarity to pre-constructed prototype activations representing harmful and benign classes. Extensive experiments demonstrate that FGAS achieves state-of-the-art defense performance against various jailbreak attacks, while nearly preserving the model’s original performance on benign tasks.