论文检索

Linggang Kong, Lei Wu, Yunlong Zhang, Xiaofeng Zhong, Wang Zhen, Yongjie Wang, Yao Pan

Despite the groundbreaking advancements made by large language models (LLMs), hallucination remains a critical bottleneck for their deployment in high-stakes domains. Existing classification-based methods mainly rely on static and passive signals from internal states, which often captures the noise and spurious correlations, while overlooking the underlying causal mechanisms. To address this limitation, we shift the paradigm from passive observation to active intervention by introducing CausalGaze, a novel hallucination detection framework based on structural causal models (SCMs). CausalGaze models LLMs’ internal states as dynamic causal graphs and employs counterfactual interventions to disentangle causal reasoning paths from incidental noise, thereby enhancing model interpretability. Extensive experiments across four datasets and three widely used LLMs demonstrate the effectiveness of CausalGaze, especially achieving over 5.2% improvement in AUROC on the TruthfulQA dataset compared to state-of-the-art baselines.

Daeyeop Lee, Hwanjo Yu

Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Large Language Models (LLMs), yet it often incurs substantial computational costs due to “over-reasoning”—the generation of redundant, verbose, or irrelevant steps. While existing reasoning step evaluators effectively detect logical fallacies and factual errors, our analysis reveals a critical blind spot: they fail to penalize “valid but inefficient” reasoning steps that inflate token usage without contributing to the solution. To systematically diagnose this limitation, we introduce RIV-GSM8K, a diagnostic benchmark injected with five distinct types of inefficiencies, including circular reasoning and excessive decomposition. Diagnostic experiments reveal that state-of-the-art evaluators struggle to distinguish these inefficiencies from necessary reasoning. To address this gap, we propose CAID (Context-Aware Information Density), a training-free metric grounded in information theory that identifies low-utility steps. To validate the metric’s practical utility, we apply it within PACE, a post-hoc compression strategy. Additional control experiments show that the gains of PACE are not explained by trivial pruning: compared with random step removal and PRM-based compression baselines, it preserves accuracy at substantially higher compression rates. Empirical results on GSM8K, StrategyQA, and ARC-Challenge demonstrate that PACE reduces token consumption by 31–53% while maintaining accuracy, confirming that CAID successfully distills informational “froth” from reasoning chains without compromising deductive validity.

Bo-Jyun Wang, Ying-Jia Lin, Hung-Yu Kao

Small language models (SLMs), such as BART, can achieve summarization performance comparable to large language models (LLMs) via distillation. However, existing LLM-based ranking strategies for summary candidates suffer from instability, while classical metrics (e.g., ROUGE) are insufficient to rank high-quality summaries. To address these issues, we introduce SCURank, a framework that enhances summarization by leveraging Summary Content Units (SCUs). Instead of relying on unstable comparisons or surface-level overlap, SCURank evaluates summaries based on the richness and semantic importance of information content. We investigate the effectiveness of SCURank in distilling summaries from multiple diverse LLMs. Experimental results demonstrate that SCURank outperforms traditional metrics and LLM-based ranking methods across evaluation measures and datasets. Furthermore, our findings show that incorporating diverse LLM summaries enhances model abstractiveness and overall distilled model performance, validating the benefits of information-centric ranking in multi-LLM distillation.

Tsan Tsai Chan, Varsha Suresh, Anisha Saha, Michael Hahn, Vera Demberg

Vision-language model (VLM) hallucination is commonly linked to imbalanced allocation of attention across input modalities: system, image and text. However, existing mitigation strategies tend towards an image-centric interpretation of these imbalances, often prioritising increased image attention while giving less consideration to the roles of the other modalities. In this study, we evaluate a more holistic, system-mediated account, which attributes these imbalances to functionally redundant system weights that reduce attention to image and textual inputs. We show that this framework offers a useful empirical perspective on the yes-bias, a common form of hallucination in which VLMs indiscriminately respond ‘yes’. Causally redistributing attention from the system modality to image and textual inputs substantially suppresses this bias, often outperforming existing approaches. We further present evidence suggesting that system-mediated attention imbalances contribute to the yes-bias by encouraging a default reliance on coarse input representations, which are effective for some tasks but ill-suited to others. Taken together, these findings firmly establish system attention as a key factor in VLM hallucination and highlight its potential as a lever for mitigation.

Shidong Cao, Hongzhan Lin, Yuxuan Gu, Ziyang Luo, Jing Ma

Chain-of-Thought (CoT) reasoning improves multi-step mathematical problem solving in large language models but remains vulnerable to exposure bias and error accumulation, as early mistakes propagate irreversibly through autoregressive decoding. In this work, we propose DiffCoT, a diffusion-styled CoT framework that reformulates CoT reasoning as an iterative denoising process. DiffCoT integrates diffusion principles at the reasoning-step level via a sliding-window mechanism, enabling unified generation and retrospective correction of intermediate steps while preserving token-level autoregression. To maintain causal consistency, we further introduce a causal diffusion noise schedule that respects the temporal structure of reasoning chains. Extensive experiments on three multi-step CoT reasoning benchmarks across diverse model backbones demonstrate that DiffCoT consistently outperforms existing CoT preference optimization methods, yielding improved robustness and error-correction capability in CoT reasoning.

Yue Wang, Zhi Zhang, Wang Xi, Chengjie Sun, Lili Shan, Bingquan Liu

Chain-of-thought (CoT) often improves multi-step reasoning, but it remains unclear what kind of additional sequential computation longer traces actually enable. We connect CoT to Bennett’s logical depth, separating an answer’s description length from the sequential effort required to derive it, and view a CoT budget of T steps as a qualitative cap on realizable sequential computation. To operationalize realized depth beyond raw length, we introduce Effective Logical Depth (ELD), a deletion-based measure of step necessity under a specified inference interface. Across depth-controlled prefix-sum tasks and GSM8K rationale perturbations, we observe two consistent signatures of a Time-for-Accuracy tradeoff: (i) plateau-to-transition accuracy curves as the budget increases from being below to matching the task’s required depth, and (ii) sparse, position-dependent deletion sensitivity concentrated in early steps for deeper instances. On GSM8K, an Extract interface, where the model reads off the answer from the remaining rationale, remains near-perfect even after prefix deletions, whereas a Repair interface, where the model must re-solve from truncated rationale context, degrades markedly. Moreover, Socratic human rationales are consistently more robust than Main rationales under Repair. These results suggest that longer CoT helps primarily when it enables additional effective sequential computation, and that deletion-based diagnostics can distinguish computational steps from redundant ones.

Dongjie Cheng, Yongqi Li, Zhixin Ma, Hongru Cai, Yupeng Hu, Wenjie Wang, Liqiang Nie, Wenjie Li

Multimodal Large Language Models (MLLMs) are making significant progress in multimodal reasoning. Early approaches focus on pure text-based reasoning. More recent studies have incorporated multimodal information into the reasoning steps; however, they often follow a single task-specific reasoning pattern, which limits their generalizability across various multimodal tasks. In fact, there are numerous multimodal tasks requiring diverse reasoning skills, such as zooming in on a specific region or marking an object within an image. To address this, we propose unified generative multimodal reasoning, which unifies diverse multimodal reasoning skills by generating intermediate images during the reasoning process. We instantiate this paradigm with Omni-R1, a two-stage SFT+RL framework featuring perception alignment loss and perception reward, thereby enabling functional image generation. Additionally, we introduce Omni-R1-Zero, which eliminates the need for multimodal annotations by bootstrapping step-wise visualizations from text-only reasoning data. Empirical results show that Omni-R1 achieves unified generative reasoning across a wide range of multimodal tasks, and Omni-R1-Zero can match or even surpass Omni-R1 on average, suggesting a promising direction for generative multimodal reasoning. The code and checkpoints are attached for reproducibility and subsequent open release.

Susmit Das

Reasoning-oriented language models typically expose explicit reasoning as a long, front-loaded chain of “thinking” tokens before the main output, either always enabled or externally toggled at inference time. Although this can help on arithmetic, coding, and other multi-step tasks, it is costly, weakens claim-level auditability, and does not allow the model to re-trigger explicit reasoning once presentation has begun. In dialogue, these limitations are compounded by weak sensitivity to temporal structure: unless time is explicitly stated in text, standard models treat replies separated by seconds and replies separated by weeks as equivalent. We introduce TIME (Temporally Intelligent Meta-reasoning Engine), a behavioral alignment framework that learns explicit reasoning as a context-triggered control policy rather than a fixed response mode. TIME augments dialogue with optional ISO 8601 ‘<time>‘ tags, tick events that represent silent time passage, and short ‘<think>‘ blocks that may appear anywhere in a response. Using a four-phase curriculum, including a small maximally diverse full-batch alignment stage, we train Qwen3 dense models to invoke brief, in-place reasoning bursts only when contextual cues warrant them, while keeping user-facing output compact. We also introduce TIMEBench, a diagnostic benchmark for evaluating reasoning from temporal cues in dialogue. Across 4B-32B scales, TIME improves TIMEBench scores over the corresponding base Qwen3 models in both thinking and no-thinking modes while reducing explicit reasoning tokens by roughly an order of magnitude. Beyond score improvements, TIME induces a distinct behavioral shift: explicit reasoning becomes more compact and more responsive to contextual cues. Code, training data, and benchmark artifacts are publicly available.

Christopher Adrian Kusuma, Muhammad Reza Qorib, Hwee Tou Ng

Large language models (LLMs) are highly capable of answering questions, but they are often unaware of their own knowledge boundary, i.e., knowing what they know and what they don’t know. As a result, they can generate factually incorrect responses on topics they do not have enough knowledge of, commonly known as hallucination. Rather than hallucinating, a language model should be more honest and respond with "I don’t know" when it does not have enough knowledge about a topic. Many methods have been proposed to improve LLM honesty, but their evaluations lack robustness, as they do not take into account the knowledge that the LLM has ingested during its pretraining. In this paper, we propose a more robust evaluation benchmark dataset for LLM honesty by utilizing Pythia, a truly open LLM with publicly available pretraining data. In addition, we also propose a novel method for harnessing the pretraining data to build a more honest LLM.

Zichen Chen, Jianda Chen, Jiaao Chen, Misha Sra

Current financial benchmarks prioritize large language models (LLMs) for task accuracy and portfolio returns, yet overlook risks arising from multi-agent cooperation, tool-sharing, and real-world financial actions. We introduce M-SAEA, a Multi-agent, Safety-Aware Evaluation Agent that audits LLM teams without fine-tuning, deploying ten probes across four layers: model, workflow, interaction, and system, to yield a continuous risk vector and natural-language rationale. Evaluated across three high-stakes tasks (finance management, webshop automation, transactional services) with six prominent models, M-SAEA (i) identifies unsafe trajectories with minimal false positives, (ii) reveals latent risks (e.g., temporal staleness) that are not addressed by standard metrics, and (iii) provides granular, actionable scores for balancing safety and latency pre-deployment. By quantifying safety as a model-agnostic metric, M-SAEA reorients evaluation from individual tasks to collaborative teams, offering a robust template for risk-first assessment of agentic AI in finance and beyond.

Hao Yang, Jin Wang, Xuejie Zhang

Human visual reasoning typically follows a coarse-to-fine attention process, starting from global scene understanding and gradually focusing on question-relevant regions. However, multimodal large language models may deviate from this pattern due to attention drift and the underutilization of visual evidence, which can lead to hallucinations. To mitigate these issues, this study proposes a Dual-Indicator Guided Contrastive Alignment (DICA), which tracks two information-theoretic indicators during inference: Visual Attention Entropy (VAE), which reflects the concentration of visual attention, and Output Image Correlation (OIC), which measures the dependence of generated outputs on the visual input. An abnormal increase in VAE or a decrease in OIC corresponds to different failure modes, which trigger targeted contrastive alignment to restore visual grounding. Experimental results across multiple benchmarks demonstrate that DICA consistently outperforms existing approaches and substantially reduces hallucinations, highlighting the effectiveness of indicator-driven intervention in improving multimodal inference reliability. The code is publicly available at https://github.com/BGWH123/DICA/.

Youngwon Lee, Seung-won Hwang, Ruofan Wu, Feng Yan, Danmei Xu, Moutasem Akkad, Zhewei Yao, Yuxiong He

We study ambiguous-query disambiguation in retrieval-augmented generation (RAG). Prior Diversify-then-Verify (DtV) pipelines first generate interpretations and then retrieve evidence, often introducing ungrounded queries that cannot be answered from the corpus and requiring costly post-hoc pruning and verification. We propose VerDICT, a novel approach that unifies diversification with verification by integrating retriever relevance and generator answerability feedback early. This not only reduces cascading errors but also enables parallelism. On ASQA, VerDICT improves grounding-aware F1 by an average of 23% over the strongest baselines across multiple LLM backbones.

Abdelrahman Sadallah, Kareem Elozeiri, Mervat Abassy, Rania Elbadry, Mohamed Anwar, Abed Alhakim Freihat, Preslav Nakov, Fajri Koto

Poetry has long been a central art form for Arabic speakers, serving as a powerful medium of expression and cultural identity. While modern Arabic speakers continue to value poetry, existing research on Arabic poetry within Large Language Models (LLMs) has primarily focused on analysis tasks such as interpretation or metadata prediction, e.g., rhyme schemes and titles. In contrast, our work addresses the practical aspect of poetry creation in Arabic by introducing controllable generation capabilities to assist users in writing poetry. Specifically, we present a large-scale, carefully curated instruction-based dataset in Modern Standard Arabic (MSA) and various Arabic dialects. This dataset enables tasks such as writing, revising, and continuing poems based on predefined criteria, including style and rhyme, as well as performing poetry analysis. Our experiments show that fine-tuning LLMs on this dataset yields models that can effectively generate poetry that is aligned with user requirements, based on both automated metrics and human evaluation with native Arabic speakers. The data and the code are available at https://github.com/mbzuai-nlp/instructpoet-ar

Chenkang, Fan Yu, Junjie Nian, Sihan Zhao, Zhuoka Feng, Zijun Yao, Wang Heng, Yu Minshen, Yixin Cao

Scaling test-time compute via Long Chain-of-Thought (Long-CoT) significantly enhances reasoning capabilities, yet extended generation does not guarantee correctness: after an early wrong commitment, models may keep elaborating a self-consistent but incorrect prefix. Through fine-grained trajectory analysis, we identify Thinking Traps, prefix-dominant deadlocks where later reflection, alternative attempts, or verification fails to revise the root error. On a curated subset of DAPO-MATH, 89% of failures exhibit such traps. To solve this problem, we introduce TAAR (Trap-Aware Adaptive Restart), a test-time control framework that trains a diagnostic policy to predict two signals from partial trajectories: a trap index for where to truncate and an escape probability for whether and how strongly to intervene. At inference time, TAAR truncates the trajectory before the predicted trap segment and adaptively restarts decoding; for severely trapped cases, it applies stronger perturbations, including higher-temperature resampling and an optional structured reboot suffix. Experiments on challenging mathematical and scientific reasoning benchmarks (AIME24, AIME25, GPQA-Diamond, HMMT25, BRUMO25) show that TAAR improves reasoning performance without fine-tuning base model parameters.

Savita Bhat, Vasudeva Varma

LLM-based evaluation systems (LLM judges) have emerged as a scalable alternative to expensive human evaluations. Although LLM judges demonstrate 70-80% agreement with human evaluators, their robustness under semantically equivalent prompt variations remains underexplored. Through systematic evaluation of 8 models across 4 NLG tasks using 10 semantically equivalent paraphrases per prompt (~115000 evaluations), we identify a critical accuracy-robustness gap: attribute verifiability affects the robustness more than model choice, with factually verifiable attributes achieving 0.71 accuracy versus 0.19 for subjective attributes. Our investigations discover three key insights: 1) Task structure characteristics influence the robustness and in turn accuracy, 2) Attribute verifiability as the strongest predictor-factually verifiable attribute achieve 0.71 accuracy versus 0.19 for subjective attributes, 3) No single winning model-smallest model (Llama-3.1-8B) exhibits second-best performance, while the strongest model (Llama-4) from the same family significantly lag behind, thus demonstrating that general capability improvements do not necessarily result in evaluation robustness. With these findings, we propose a diagnostic framework grounded in attribute verifiability that enables principled decisions about evaluation automation. Our work establishes new standards for assessing LLM judge reliability beyond simple accuracy metrics.

Han Zhu, Wei Kang, Liyong Guo, Zengwei Yao, Fangjun Kuang, Weiji Zhuang, Zhaoqing Li, Zhifeng Han, Dong Zhang, Xin Zhang 等

Generating spoken dialogue is inherently more complex than monologue text-to-speech (TTS), as it demands both realistic turn-taking and the maintenance of distinct speaker timbres. While existing autoregressive (AR) models have made progress, they often suffer from high inference latency and stability issues. To overcome these limitations, we propose ZipVoice-Dialog, a non-autoregressive (NAR) zero-shot spoken dialogue generation model based on flow-matching. Observing that applying vanilla flow-matching to dialogue generation leads to poor speech intelligibility and turn-taking precision, we introduce two simple yet effective methods to adapt flow-matching architectures for dialogue generation: (1) a curriculum learning strategy to ensure robust speech-text alignment, and (2) speaker-turn embeddings to govern precise speaker turn-taking. Additionally, we introduce dedicated strategies to support stereo dialogue generation.Recognizing the lack of training datasets in this field, we curate and release OpenDialog, the first large-scale (6.8k hours) open-source spoken dialogue dataset derived from in-the-wild speech data. Moreover, for fair and rigorous evaluations, we established a benchmark to comprehensively evaluate dialogue generation models. Experiments demonstrate the effectiveness of the proposed methods and dataset, showing that ZipVoice-Dialog achieves superior performance in inference speed, intelligibility, speaker turn-taking accuracy, and speaker similarity. Our code, model checkpoints, and the OpenDialog dataset are publicly available.

Han Liu, Jiaqing Zhan, Zhichao Chen, Qin Zhang

In real-world applications of natural language processing, it is essential to effectively adapt a pre-trained model to a downstream task. While text classification is undertaken as a downstream task, it is crucial to produce meaningful sentence embedding that is adaptive to the task. In this paper, we explore how to effectively adapt a pre-trained model for extracting meaningful context representations from sentences, and propose an uncertainty-aware contrastive sentence embedding approach that involves addressing language ambiguity and inter-class separability for a text classification task. Specifically, we design an end-to-end strategy for driving the process of learning to transform a word embedding matrix into a contextualized sentence vector and to quantify the representation uncertainty of the sentence, while the word embedding matrix is produced by a pre-trained model without fine-tuning, and a label-wise contrastive learning strategy is designed to enhance intra-class compactness and inter-class separability. The results on public data sets show that a considerable improvement of text classification accuracy is achieved by adopting the proposed approach in comparison with using those state-of-the-art methods.

Piotr Nawrot, Jianing Li, Renjie Huang, Sebastian Ruder, Kelly Marchisio, Edoardo Ponti

Sparse attention offers a promising strategy to extend long-context capabilities in Transformer LLMs, yet its efficiency–accuracy trade-offs remain unclear due to the lack of comprehensive evaluation. We address this gap with the largest-scale empirical analysis to date of training-free sparse attention, evaluating six methods across multiple model families and sizes, sequences up to 128K tokens, and sparsity levels up to 0.95 (i.e., 1/20 attention budget) on nine diverse tasks. We first organise the rapidly evolving landscape of sparse attention methods into a taxonomy along four design axes. Our analysis then yields actionable insights: 1) sparse attention is effective: larger sparse models outperform smaller dense ones at equivalent cost, improving the Pareto frontier; 2) for the training-free methods we study, fine-grained per-query importance estimation during prefilling remains impractical—due to both the cost of estimation and the lack of sparse kernels that translate fine-grained sparsity into wall-clock gains—forcing a task-dependent choice between global-to-token and block-to-block selection. Instead, during decoding, token-to-page selection becomes feasible, enabling better generalisation and higher sparsity tolerance; 3) longer sequences tolerate higher sparsity, suggesting that fixed-budget methods in production are suboptimal. Together, these findings provide practical guidance for deploying sparse attention and methodological recommendations for future evaluations. Our code is available at https://github.com/PiotrNawrot/sparse-frontier.

Fuqiang Niu, Bowen Zhang, Junting Zhu, Qing Liao, Genan Dai, Hu Huang

Social media platforms have become critical arenas for public discourse, yet existing stance detection methods often reduce opinions to surface-level labels, overlooking the conversational evidence behind stance expressions. We introduce Conversational Stance-Cause Pair Detection (CSCPD), a new task that jointly identifies both the stance polarity and its observable contextual evidence within multi-turn conversations. To advance research in this direction, we present Cause-CSD, the first large-scale dataset for CSCPD, spanning 21,048 annotated stance-cause pairs across diverse open-domain, textual, and multimodal discussions. We further propose Stance-Cause Detection Language Model (SCD-LM), a unified language model framework that leverages explicit context reasoning and joint decoding to predict stances and their supporting causes, along with human-readable rationales. Extensive experiments demonstrate that SCD-LM achieves state-of-the-art results on both text-only and multimodal subtasks, significantly outperforming strong baselines, especially for long-range and image-grounded cause detection. Our work advances explainable stance analysis and underpins understanding of public opinion drivers in impactful online settings.

Florian Baud, Feda Almuhisen, Dorian Midou

Numerical data from sensors and time series are widely used in scientific research fields such as nuclear fusion experiments, which generate vast amounts of complex, high-dimensional data. Therefore, efficient numerical data analysis tools are crucial to accelerate experimental research. Large language models (LLMs) have emerged as promising solutions to analyze numerical data with natural language queries. However, LLMs have difficulties treating this type of data as they have been designed for text in the first place. To overcome these limitations, we propose a model-agnostic and data-agnostic agent that processes numerical data by code generation and multimodal reasoning. Our agent demonstrates competitive performance against baselines on benchmark data on numerical data tasks such as sensor data classification and time series understanding. While outperforming them on information retrieval benchmarks, also we have successfully applied our agent in the context of nuclear fusion research, where physicists and Tokamak operators interact with it to plan and analyze fusion experiments.