TACL 2026 论文列表

Language Identification (LID) is a key component of multilingual NLP pipelines and multilingual corpus construction. LID errors can create spurious text–language associations, harming downstream model quality, especially for low-resource languages and fine-grained taxonomies involving closely related languages and dialects. We propose LIBE, a representation-based approach that embeds input texts and natural-language descriptions into a shared semantic space with a bi-encoder, enabling LID via similarity search. This formulation supports data-efficient learning and the addition of new languages through their descriptions. To improve discrimination in ambiguous or low-resource cases, we also introduce an optional cross-encoder re-ranking stage. Across multiple benchmarks, LIBE achieves competitive performance with strong LID systems when initialized from pretrained multilingual encoders, while requiring substantially less task-specific LID supervision in several adaptation settings. The hybrid bi-encoder– cross-encoder formulation further improves robustness in challenging open-set scenarios, achieving up to 90% F1 with as few as five labeled examples. Code and data are available at https://github.com/cafoscari-nlp/LIBE.

Medical multimodal large language models (MLLMs) can perform well on existing medical visual question answering (MedVQA) benchmarks, but their training data often does not match clinical diagnosis. Most supervision is organized around isolated images or short QA pairs, leaving two structures weakly specified: how evidence leads to a decision, and how views, series, modalities, and patient context from the same case are linked. We introduce MultiViewDx, a partly physician-validated multimodal instruction dataset for evidence-linked multi-view medical imaging diagnosis. Multi-ViewDx uses the clinical case as the supervision unit. It links imaging studies with patient context, normalizes heterogeneous reports into an evidence-linked workflow (evidence → findings → differential discussion → diagnosis), and uses a unified image-text retriever to constrain instruction synthesis to source-supported evidence. It covers X-ray, CT, MRI, ultrasound, histopathology, and other clinical visual sources. We fine-tune MultiViewDx-8B-AN and evaluate it on both existing MedVQA benchmarks and real-world case-based diagnostic reasoning. Across four MedVQA benchmarks, it achieves the best average accuracy among compared systems (79.0%), outperforming HuatuoGPT-Vision-34B (66.7%) and Claude3-Opus (55.7%). Beyond Med-VQA, on JAMA Clinical Challenge cases, it receives the strongest overall rating under a physician-designed rubric for key clinical points, diagnostic inference, and evidence grounding. Controlled ablations and clinician evaluation show that both case-level multi-view organization and evidence-linked reasoning targets contribute to the gain.1

Hallucinations in large multimodal models (LMMs), where responses appear correct but are factually wrong, limit their reliability in practical applications. This paper focuses on hallucinations in video understanding, a more dynamic and challenging domain than static modalities. We present HAVEN, a comprehensive benchmark with 6K questions spanning three dimensions: hallucination causes, aspects, and question formats. Then, we systematically evaluate 18 LMMs and analyze seven influential factors affecting hallucinations, including video duration, model size, and reasoning capabilities. Additionally, we explore two reasoning-based strategies to reduce hallucinations in LMMs: (1) text-based reasoning via rejection fine-tuning on reasoning datasets and segment-weighted Direct Preference Optimization, and (2) video-based reasoning through a tool-augmented frame-level operations on video optimized via supervised tool learning and Hallucination-aware Group Relative Policy Optimization (HGRPO). Extensive experiments and analyses demonstrate the effectiveness. Remarkably, video-based reasoning strategy improves the Qwen2.5-VL-7B-Instruct by 10.16% in accuracy and reduces the bias score by 4.89%.

Task-oriented dialogue (ToD) systems aim to help users accomplish goals through natural language interaction. Beyond task success, effective ToD systems must also maintain positive emotional interaction and accurately convey information in inherently noisy and ambiguous conversational environments. Recent advances in large language models (LLMs) have substantially improved conversational fluency and contextual understanding. However, although emotion has been incorporated into existing ToD systems, its integration is typically confined to limited roles, restricting its influence on overall dialogue behaviour and user experience. To address this, we propose a novel end-to-end framework that combines fully lexicalised representations with an LLM backbone for improved semantic accuracy and employs online reinforcement learning with short-term emotional and long-term task-success rewards for joint optimisation of emotional interaction and task success. We further develop a challenging and realistic simulation environment that enables systematic evaluation across diverse ToD set-ups. Experiments show that incorporating coarse affective signals improves task success and sentiment rating in aggregate. Furthermore, integrating emotional interaction throughout the dialogue system using our proposed framework results in more pronounced improvements in both task success and user emotional experience.1

Across all fields, respected experts cite their sources. While large language models (LLMs) excel at synthesizing information, they do not reliably cite sources, making it difficult to verify provided information. In contrast, search engines make sources readily accessible but burden users with synthesizing information. To rigorously examine the interplay between verifiability and utility, we build on classic theories of abstraction to propose an extractive-abstractive spectrum for cited generations; we define five operating points that span the extreme endpoints of extractive search engine outputs and abstractive LLM generations without reliable citation. We perform an analysis of the drawbacks and benefits of seven systems across the spectrum in the QA settings of search queries, information simplification, multi-hop reasoning, and health advice. As outputs become more abstractive, our human evaluations find that the proportion of properly cited sentences decreases by as much as 50%. Simultaneously, perceived utility improves by as much as 200% and users take up to 3 times as long to verify cited information. We recommend operating points for specific task settings and our failure analysis informs methods for citation identification, information retrieval, and multi-operating point systems.

Large language models learn word meanings through distributional semantics (i.e., patterns in text), and distribution does not always capture information about taxonomy and function, which are essential to human representations of word meaning. However, sound and spelling often conveys such information, and language models decompose many words into subword tokens, which may identify category markers. For example, GPT-5 segments fluorine and bromine into fluor + ine and brom + ine, which share the category marker ine. We provide evidence that, across 18 languages, shared tokens predict greater semantic priming effects in humans (e.g., people recognize bromine faster after reading fluorine), beyond what distributional similarity explains. Furthermore, across seven languages, shared tokens boost semantic priming beyond what shared morphemes explain. This suggests that subword tokens could help language models arrive at human-like representations of word meanings. However, we also provide evidence that the trend towards larger vocabularies of tokens obscures category markers, and even when language models have access to subword tokens, they underestimate the semantic relatedness of words that share tokens.

The effectiveness of in-context learning (ICL) varies significantly across models and tasks, posing challenges for practitioners to determine when ICL reliably improves performance. Current evaluation approaches, reliant on performance changes after applying ICL, suffer from low reliability, poor attribution, and impracticality in data-insufficient scenarios. We propose the Learning-to-Context Slope (LCS), a novel metric that quantifies ICL effectiveness by modeling the slope between learning gain (loss decrease from demonstrations) and contextual relevance (demonstration-input relevance). LCS addresses key limitations of performance-based metrics: (i) it captures loss changes even when outputs are incorrect; (ii) its formulation attributes ICL failures to weak contextual alignment or strong output calibration; (iii) it minimizes reliance on labeled data via synthetic evaluation. Experiments show that LCS can better reflect ICL effectiveness when performance change fails, demonstrating its effectiveness.

This study explores the capacity of Large Language Models (LLMs) to perform tasks requiring Theory of Mind (ToM), a critical component of pragmatic language understanding. Although previous work suggests that LLMs may exhibit emergent ToM abilities, this research examines whether such capabilities genuinely involve reasoning about beliefs or merely reflect the reliance on shallow statistical cues. Through a series of controlled experiments featuring indirect speech acts and verbal irony, we assess how belief contexts influence LLM interpretations. The results reveal that, although LLMs occasionally succeed in decoding communicative intentions, their performance is not attributable to human-like ToM reasoning. This work underscores the limitations of LLMs in simulating humanlike ToM and offers insight into their interpretive biases, contributing to a deeper understanding of their linguistic capabilities.1

Weight-space model merging combines independently fine-tuned checkpoints without access to the original training data. While merging has shown promise in multitask settings, its behavior in multilingual generative systems remains underexplored. We systematically study weight-space merging for multilingual machine translation by fully fine-tuning language models on large-scale bilingual corpora and evaluating representative merging strategies across shared-source, shared-target, and bidirectional consolidation settings. Our experiments reveal a strong directional asymmetry. Merging is comparatively more effective when models share a target language, improving multilingual coverage over the base model, but it still fails to preserve the peak performance of language-specific checkpoints. In contrast, when target languages differ, performance degrades sharply, especially in shared-source and bidirectional settings. To explain this behavior, we analyze internal representations and find that fine-tuning does not create disjoint language-specific sub-networks. Instead, independently fine-tuned models activate largely overlapping neurons while reshaping upper-layer target-generation representations into incompatible geometries. These findings suggest that multilingual merging failures arise from target-side geometric misalignment within shared computational units, challenging the assumptions underlying standard weight-space merging for multilingual translation. We make the code publicly available at https://github.com/babangain/mt-model-merging.

While prior work has highlighted that systematic nation-level bias in large language models (LLMs) can pose operational risks for international relations (IR) applications, many existing evaluations still lack a clearly specified unbiased reference (ground truth), limiting fully quantitative and cross-setting measurement. Using an expanded corpus of 722 official UN Security Council (UNSC) records, we measure nation-level bias with five complementary tests, thereby extending prior approaches to cover a broader set of bias expression channels: opinion ranking (WOW), association sorting (AK+-), vote simulation (PVS), opposition attribution (F-Oppose), and framing (Headliner). To enable grounded bias evaluation, we explicitly define an unbiased status for each test; in particular, for WOW and AK+-where ground truth is difficult to specify, we propose an action-based pseudo unbiased status derived from observable voting behavior to construct a quantifiable baseline. Benchmarking 4 LLMs (GPT-4o-mini, Llama 3.3-70B, Mistral-22B-Small, and Qwen3-30B) across the five permanent members of the UNSC, we find a test-agnostic macro-pattern: relative to the reference baselines defined for each test, model outputs lean toward the opposition (red) side for China and Russia and toward the supportive (blue) side for the U.K., the U.S., and France. These results underscore the need to audit IR-facing LLM deployments with grounded, multi-test benchmarks tied to explicit reference points.

Large language models (LLMs) remain vulnerable to jailbreaking attacks despite their impressive capabilities. Investigating these vulnerabilities is crucial for developing robust safety mechanisms. Existing attacks primarily bypass LLM safeguards by introducing additional context or adversarial tokens, while leaving the core harmful intent intact. In this paper, we introduce ISA (Intent Shift Attack), which conceals malicious intent from LLMs through subtle edits. More specifically, we establish a taxonomy of intent transformations and leverage them to generate attacks that may be misperceived by LLMs as benign requests. Unlike prior methods relying on complex tokens or lengthy context, our approach requires only limited modifications to the original request, yielding natural, fluent, and seemingly harmless prompts. Extensive experiments on both open-source and commercial LLMs demonstrate that ISA effectively induces safety failures, achieving high attack success rates through simple linguistic transformations. More critically, fine-tuning models solely on benign data reformulated with ISA templates elevates the success rates to nearly 100%. For defense, we evaluate existing methods and demonstrate their limited effectiveness against ISA, while exploring both training-free and training-based mitigation strategies. Our findings reveal fundamental challenges in intent inference for LLM safety and underscore the need for more effective defenses. Our code and datasets are available at https://github.com/NJUNLP/ISA.

We introduce Guideline-Centered Annotation Methodology (GCAM), a novel methodology designed to report the annotation guidelines associated with each data instance. GCAM addresses four key limitations of the standard application of the prescriptive annotation methodology by reducing the information loss during annotation, ensuring adherence to guidelines, and enabling the efficient reuse of annotated data across multiple tasks that rely on the same guidelines. We evaluate GCAM with a focus on text classification tasks through (i) a human annotation study and (ii) an experimental evaluation with several machine learning models. Our results highlight the advantages of GCAM from multiple perspectives, guaranteeing a transparent evaluation of the successful application of the prescriptive paradigm and enabling a fine-grained model error analysis.

Parameter-efficient fine-tuning (PEFT) methods like LoRA have become crucial for adapting large language models under computational constraints. While LoRA focuses on dense weight updates, recent sparse methods propose updating only selected portions of pretrained weights. However, these sparse approaches predominantly use element-wise parameter selection, overlooking inherent structured patterns in LLM weight distributions. We introduce StructTune1, a PEFT method that uses entire rows or columns as basic units for identifying task-relevant sparse components. Our analysis reveals LLaMA models exhibit distinct structured patterns in pretrained weights, with fine-tuning updates showing structured amplification rather than uniform changes. We develop a sensitivity-based framework to efficiently identify task-relevant weight structures without full fine-tuning. Experiments on mathematical reasoning and code generation demonstrate StructTune’s structured sparse updates outperform both dense LoRA and element-wise sparse methods while maintaining computational efficiency. Ablation studies validate our sensitivity-based selection strategy, confirming the efficacy of exploiting structured weight patterns in PEFT.

Routing Large and Small Language Models (LLMs SLMs) is commonly framed as query-level difficulty prediction, yet lightweight routers are often unreliable and stronger evaluators introduce non-trivial overhead. We propose VERTICAL ROUTING, a stage-level collaboration framework that avoids monolithic difficulty prediction by allocating the large model to critical subtasks and delegating the remaining generation to a small model. VERTICAL ROUTING instantiates two templates: (i) domain-specific templates that decompose a task into ordered stages with criticality scores, and (ii) a robust default template that follows a prefix-first prior for general queries. Under a token budget, we allocate large-model capacity to the most critical stages, or to a budgeted prefix under the default template. Experiments show that VERTICAL ROUTING outperforms the strongest baseline (RouterDC) by 2.0 points in the averaged metric, while significantly reducing token usage by 48.9% and large-model output share by 41.7%. Additionally, it enhances stability by lowering the standard deviation by 87.5%. These results highlight its advantages in accuracy, efficiency, and robustness.1

Memes, as a widely used mode of online communication, often serve as vehicles for spreading harmful content. However, limitations in data accessibility and the high costs of dataset curation hinder the development of robust meme moderation systems. To address this challenge, in this work, we introduce a first-of-its-kind dataset – TOXICTAGS consisting of 6,300 real-world meme-based posts annotated in two stages: (i) binary classification into toxic and normal, and (ii) fine-grained labelling of toxic memes as hateful, dangerous, or offensive. A key feature of this dataset is that it includes collaborative tags associated with the original posts, enhancing the context of each meme. In addition, we propose a novel entropy-guided multi-tasking framework – STEMTOX – that leverages these collaborative tags alongside visual and textual inputs within a robust classification framework. Experimental results show that incorporating these tags substantially enhances the performance of state-of-the-art VLMs in toxicity detection tasks. Our contributions offer a novel and scalable foundation for improved content moderation in multi-modal online environments. We have made our code1 and dataset2 publicly available for research purposes. Warning: Contains potentially toxic contents.

Group Relative Policy Optimization (GRPO) is a promising policy-based approach for Large Language Model alignment, yet its performance is often limited by training instability and suboptimal convergence. In this paper, we identify and analyze two main GRPO issues: (i) the token-level penalization, where valuable tokens shared across different responses receive contradictory feedback signals, leading to conflicting gradient updates that can reduce their likelihood; and (ii) the policy collapse, where negatively rewarded completions may penalize confident responses and shift model decisions toward unlikely tokens, destabilizing training process. To address these issues we introduce GTPO (Group-relative Trajectory-based Policy Optimization), which prevents conflicting gradients on valuable tokens by skipping negative updates while amplifying positive ones and filters out completions whose entropy exceeds a provable threshold, to prevent policy collapse. By omitting KL-divergence regularization, GTPO eliminates the reference model while achieving superior stability and performance over GRPO and DAPO. Extensive evaluations across GSM8K, MATH, R1, AIME 2024–2025, AMC 2023, and MMLU validate these gains. The code is available here.1.

This paper introduces a general methodology through which a population of autonomous agents can converge on a linguistic convention that enables them to refer to arbitrary entities in their environment. The linguistic convention emerges in a decentralised manner through local communicative interactions between pairs of agents drawn from the population. The emergent convention consists of associations between symbolic labels (word forms) and subsymbolic concept representations (word meanings) that are grounded in a continuous feature space. We confirm the generality and scalability of the method through its evaluation on a wide and diverse selection of 37 publicly available datasets. Through a range of experiments, we demonstrate the robustness of the method against perceptual variation, including in heteromorphic populations, as well as the ability of the emergent conventions to self-adapt to changes in the environment.

Reliable assessment of the abilities of large audio language models (LALMs) is essential to advancing the state of the art. As benchmarks rapidly evolve to incorporate complex reasoning and subjective tasks, they increasingly necessitate open-ended responses from LALMs. We present Open-ended Response Correctness Assessment (ORCA)—a reliable and lightweight model-based approach for answer correctness and disagreement modeling. We employ a three-stage annotation pipeline combining human judgment, structured feedback, and human-AI correction, yielding 9,663 annotations across 3,699 question-answer pairs from 15 LALMs on three audio understanding and reasoning benchmarks (achieving a Krippendorff’s alpha of 0.82). Our experiments employing curriculum learning show that ORCA models achieve a Spearman correlation of 0.91 with average human correctness ratings on seen benchmarks and generalize to unseen benchmarks with a score of 0.85, outperforming several LLM judge baselines including Gemini 2.5 Flash. Furthermore, we demonstrate that ORCA’s predicted variance correlates strongly with human disagreement, allowing it to effectively identify problematic benchmark items.

Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose topic judge, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.1