论文检索

Runsong Jia, Zhen Fang, Mengjia Wu, Jie Lu, Yi Zhang

Hallucination detection is crucial for large language models (LLMs), as hallucinated content creates significant barriers in applications requiring factual accuracy. Current detection methods mainly depend on internal signals like uncertainty and self-consistency checks, using the model’s pre-trained knowledge to identify unreliable outputs. However, pre-trained knowledge may become outdated and has coverage limitations, especially for specialized or recent information. To address these limitations, retrieval-augmented generation (RAG) has emerged as a promising solution by retrieving relevant evidence at inference time, grounding outputs beyond the model’s parametric knowledge. In this paper, we target a critical and practical learning problem RAG-based hallucination detection (RHD), where RAG is employed to enhance hallucination detection by addressing information updating challenges. To address RHD, we propose a novel method Evidence-Aligned Entity Verification (EAEV), which detects entity-level hallucinations by leveraging RAG to align generated entities with retrieved evidence contexts. Specifically, EAEV evaluates entity-evidence alignment through three complementary dimensions and introduces counterfactual stability analysis to ensure robust alignments under evidence perturbations. Experiments across multiple RAG benchmarks demonstrate that EAEV achieves consistent improvements over existing methods with strong generalization capabilities.

Vallabhaneni Raj Kumar, Ashwin S, Supriya Manna, Niladri Sett, Cheedella V S N M S Hema Harshitha, Kurakula Harshitha, Basina Deepakraj, Anand Kumar Sharma, Tanuj Sarkar, Samanthapudi Shakeer 等

Sentiment analysis for low-resource languages remains challenging in an era where interpretability, human alignment, and fairness are increasingly non-negotiable aspects of modern machine learning systems. These challenges stem both from the scarcity of annotated data and from the resulting difficulty of conducting reliable, human-interpretable analyses that go beyond predictive accuracy. Telugu, one of the primary Dravidian languages with over 96 million speakers, is not an exception. In this work, we first introduce TeSent, a large-scale Telugu sentiment classification dataset annotated with sentiment labels and human-selected rationales from multiple native speakers. This resource enables the study of rationale-based supervision for aligning models with human reasoning in this low-resource setting. We fine-tune five transformer-based models with and without rationale supervision and evaluate them on classification performance, explanation quality, and social bias. To facilitate controlled fairness evaluation, we additionally construct TeEEC, an evaluation corpus for Telugu sentiment analysis. Our results show that incorporating human rationales consistently improves alignment and often leads to holistic gains in predictive performance. We further provide extensive analysis of multi-facade explanation quality and fairness, offering insights into the broader effects of alignment-oriented supervision in resource-scarce language contexts.

Yinuo Wang, Qingjie Li, Wenyao Cui, Qiuchi Li, Zhang Huaping

Test-time scaling (TTS) enhances LLM reasoning capabilities by sampling and aggregating diverse solution trajectories. However, existing approaches often rely on external verifiers and one-shot independent sampling, which results in inefficient budget allocation and underutilizes interim high-quality trajectories. We propose ConMA, a training-free, verifier-free TTS framework that reallocates a fixed inference budget into iterative sample–filter–diversify–select cycles: it filters answer groups based on intrinsic token-probability confidence, enriches candidates through diversity-aware expansion, and employs repeated single-choice selection for multi-stage refinement. Across multiple benchmarks, ConMA consistently improves accuracy under fixed budgets. With a maximum budget of N=64, ConMA boosts Qwen3-4B to 80% accuracy on AIME25, significantly outperforming strong baselines while converging early with only 18 samples on average, substantially reducing inference cost.

Yongshuo Zhang, Jixiong Chen, Kaihe xu, Wei Wei, Shihao Zou

Complex table question answering (TQA) remains challenging, as real-world table, usually designed for human readability with multi-level headers and fragmented hierarchical semantics, largely hindering large language models (LLMs) from accurately aligning conditions, attributes, and values during reasoning. Existing approaches typically rely on handcrafted table linearization or prompts, forcing LLMs to infer header hierarchies, which frequently leads to brittle reasoning and hallucinations. To this end, we propose SMART, a unified framework that explicitly decouples table structure understanding from reasoning execution. SMART consists of three components: Semantic Header Flattening for converting multi-level headers into explicit single-level descriptors, Global Understanding for capturing holistic table–question semantics, and Pseudo-Code-Style Reasoning for structured, step-by-step inference with external validation. Extensive experiments on multiple benchmarks demonstrate that SMART substantially improves both the accuracy and robustness of complex TQA, achieving state-of-the-art performance.

Fuyu Wang, Jiangtong Li, Kun Zhu, Changjun Jiang

Current evaluations of large language models (LLMs) mainly rely on dataset-based generation accuracy. However, generative correctness does not guarantee the discriminative capability required to verify solutions, frequently masking an inability to distinguish valid reasoning from plausible errors. While multi-agent debate inherently entails judgment, we show that uncontrolled context growth and convergence to majority voting introduce significant noise, obscuring intrinsic model judgment. To address these limitations, we propose a progressive argumentation-mining diagnostic framework designed to explicitly control context and isolate discriminative behaviors. Instead of indiscriminate aggregation, our approach distills and retains only the single most well-supported rationale per answer, preventing context dilution while enforcing strict quality-based selection. Applying this framework reveals a fundamental cognitive divergence: models exhibit structural susceptibility to plausible misinformation in knowledge tasks, whereas in reasoning tasks they demonstrate latent discriminative potential that remains fragile under pressure. These findings underscore the fragility of discriminative capabilities, advocating for diagnostic methodologies that prioritize judgment stability over simple generation performance.

Junan Hu, Shudan Guo, Wenqi Liu, Jianhua Yin, Yinwei Wei

Large Language Models demonstrate outstanding performance in many language tasks but still face fundamental challenges in managing the non-linear flow of human conversation. The prevalent approach of treating dialogue history as a flat, linear sequence is misaligned with the intrinsically hierarchical and branching structure of natural discourse, leading to inefficient context utilization and a loss of coherence during extended interactions involving topic shifts or instruction refinements. To address this limitation, we introduce Context-Agent, a novel framework that models multi-turn dialogue history as a dynamic tree structure. This approach mirrors the inherent non-linearity of conversation, enabling the model to maintain and navigate multiple dialogue branches corresponding to different topics. Furthermore, to facilitate robust evaluation, we introduce the Non-linear Task Multi-turn Dialogue (NTM) benchmark, specifically designed to assess model performance in long-horizon, non-linear scenarios. Our experiments demonstrate that Context-Agent enhances task completion rates and improves token efficiency across various LLMs, underscoring the value of structured context management for complex, dynamic dialogues. The dataset and code is available at GitHub.

Yingsong Ning, Fu Zhang, Jingwei Cheng, Jiashun Peng, Xiaoke Wang

Temporal knowledge graph (TKG) forecasting aims to infer future facts from historical observations in time-evolving graphs. Traditional rule-based methods often rely on statistical co-occurrences and extensive path enumeration, suffering from rule sparsity and search-space explosion, while recent LLM-based rule reasoning can produce linguistically plausible rules that are weakly constrained by graph evidence and thus may reflect spurious correlations or violate temporal constraints.To address these challenges, we propose Critic-Guided Rule Induction (CRI), which treats temporal rules as rule hypotheses to be examined and adopts a decoupled Generation-Discrimination pipeline to induce rules that are both high-coverage and high-precision. CRI first mines seed rules and path evidence from the historical graph and uses an LLM-based generator to abstract and generalize them into broader raw rule hypotheses. It then introduces a Fact-Grounded Rule Evaluator to perform fact-grounded discrimination of rule hypotheses from complementary perspectives together with necessary temporal and statistical constraints. Finally, CRI performs symbolic reasoning over the refined rule set to produce forecasts with traceable reasoning evidence. Experiments on three benchmarks show that CRI outperforms strong baselines, achieving state-of-the-art performance on TKG forecasting.

Min Jae Jung, YongTaek Lim, Chaeyun Kim, Junghwan Kim, Kihyun Kim, Minwoo Kim

While Large Language Models (LLMs) are widely used, they remain susceptible to jailbreak prompts that can elicit harmful or inappropriate responses. This paper introduces STAR-Teaming, a novel black-box framework for automated red teaming that effectively generates such prompts. STAR-Teaming integrates a Multi-Agent System (MAS) with a Strategy-Response Multiplex Network and employs network-driven optimization to sample effective attack strategies. This network-based approach recasts the intractable high-dimensional embedding space into a tractable structure, yielding two key advantages: it enhances the interpretability of the LLM’s strategic vulnerabilities, and it streamlines the search for effective strategies by organizing the search space into semantic communities, thereby preventing redundant exploration. Empirical results demonstrate that STAR-Teaming significantly surpasses existing methods, achieving a higher attack success rate (ASR) at a lower computational cost. Extensive experiments validate the effectiveness and explainability of the Multiplex Network. The code is available at https://github.com/selectstar-ai/STAR-Teaming-paper.

Manit Baser, Alperen Yildiz, Dinil Mon Divakaran, Mohan Gurusamy

The static knowledge representations of large language models (LLMs) inevitably become outdated or incorrect over time. While model-editing techniques offer a promising solution by modifying a model’s factual associations, they often produce unpredictable ripple effects, which are unintended behavioral changes that propagate even to the hidden space. In this work, we introduce CLaRE, a lightweight representation-level technique to identify where these ripple effects may occur. Unlike prior gradient-based methods, CLaRE quantifies entanglement between facts using forward activations from a single intermediate layer, avoiding costly backward passes. To enable systematic study, we prepare and analyse a corpus of 11,427 facts drawn from three existing datasets. Using CLaRE, we compute large-scale entanglement graphs of this corpus for multiple models, capturing how local edits propagate through representational space. These graphs enable stronger preservation sets for model editing, audit trails, efficient red-teaming, and scalable post-edit evaluation. In comparison to baselines, CLaRE achieves an average of 62.2% improvement in Spearman correlation with ripple effects while being 2.74× faster, and using 2.85× less peak GPU memory. Besides, CLaRE requires only a fraction of the storage needed by the baselines to compute and preserve fact representations. Our entanglement graphs and corpus are available at https://github.com/manitbaser/CLaRE.

Meghana Sunil, Manikandarajan Venmathimaran, Muthu Subash Kavitha

Recent work shows that large multimodal models (LMMs) can self-improve from unlabeled data via self-play and intrinsic feedback. Yet existing self-evolving frameworks mainly reward final outcomes, leaving intermediate reasoning weakly constrained despite its importance for visually grounded decision making. We propose IREASONER, a self-evolving framework that improves an LMM’s implicit reasoning by explicitly eliciting chain-of-thought (CoT) and rewarding its internal agreement. In a Proposer–Solver loop over unlabeled images, IREASONER augments outcome-level intrinsic rewards with a trajectory-aware signal defined over intermediate reasoning steps, providing learning signals that distinguish reasoning paths leading to the same answer without ground-truth labels or external judges. Starting from Qwen2.5-VL-7B, IREASONER yields up to +2.1 points across diverse multimodal reasoning benchmarks under fully unsupervised post-training. We hope this work serves as a starting point for reasoning-aware self-improvement in LMMs in purely unsupervised settings.

Matthew R. DeVerna, Kai-Cheng Yang, Harry Yaojun Yan, Filippo Menczer

Large language models (LLMs) have raised hopes for automated end-to-end fact-checking, but prior studies report mixed results. As mainstream chatbots increasingly ship with reasoning capabilities and web search tools—and millions of users already rely on them for verification—rigorous evaluation is urgent. We evaluate 15 recent LLMs from OpenAI, Google, Meta, and DeepSeek on more than 6,000 claims fact-checked by PolitiFact, comparing standard models with reasoning- and web-search variants. Standard models perform poorly, reasoning offers minimal benefits, and web search provides only moderate gains, despite fact-checks being available on the web. In contrast, a curated RAG system using PolitiFact summaries improved macro F1 by 233% on average across model variants. These findings suggest that giving models access to curated high-quality context is a promising path for automated fact-checking.

Beibei Zhang, Yanan Lu, Lin Fen, Tongwei Ren

Video Speaking Style Recognition (VSSR) aims to classify conversation videos into different types, significantly facilitating human interaction understanding. Recent approaches explore the potential of large language models (LLM) in VSSR with a training-free process. However, directly integrating all multimodal data yields suboptimal results, since the great redundancy in visual data can overshadow other valuable multimodal information, such as valuable textual dialogues and critical visual clues. To address this, we propose CFMiS (Coarse-to-Fine Multimodal Information Selection), a novel framework for VSSR that dynamically obtain valuable multimodal data via coarse-to-fine selection, enhancing LLM reasoning for VSSR. Specifically, the core of CFMiS are two cascaded modules: 1) a text-dominant modality selection module firstly selects VSSR-required modalities originating from text-based prediction; and 2) if vision is included in the selected modalities, a visual refinement module iteratively collects VSSR-relevant critical visual clues. The former resolves which modality to utilize, while the latter determines which information to adopt from selected modalities, efficiently alleviating information redundancy. Extensive experiments on multiple datasets prove that CFMiS is highly effective for VSSR, outperforming all existing training-free approaches and most training-based methods.

Yuyao Zhang, Hongyu Lu, Jiajie Jin, Hongjin Qian, Shiyu Li, Zhao Yang, Yutao Zhu, Ji-Rong Wen, Zhicheng Dou

Recent advances in large language models (LLMs) have enabled web agents to perform interactive tasks on real-world websites. However, existing agents still suffer from limited robustness, efficiency, and task success, largely due to their lack of structural understanding of websites and the absence of browsing priors in pre-trained models. To address these challenges, this paper proposes the Web Agent Sitemap Protocol (WASP), an agent-oriented sitemap that integrate structured website knowledge into web agents. WASP adopts a dual-granularity design, providing global site-level structure and local page-level semantic and interaction guidance. We also introduce a framework LightASM for constructing such sitemaps by identifying core pages and generating concise semantic summaries and block-level descriptions. Experiments on real-world browsing benchmarks demonstrate that WASP substantially improves the robustness, efficiency, and effectiveness of LLM-based web agents without extra training.

Yanxi Chen, Wenhui Zhu, Xiwen Chen, Zhipeng Wang, Xin Li, Peijie Qiu, Hao Wang, Xuanzhao Dong, Yujian Xiong, Anderson Schneider 等

Although Large Audio-Language Models (LALMs) deliver state-of-the-art (SOTA) performance, they frequently suffer from hallucinations, e.g., generating text not grounded in the audio input. We analyze these grounding failures and identify a distinct taxonomy: Event Omission, False Event Identity, Temporal Relation Error, and Quantitative Temporal Error. To address this, we introduce the AHA (Audio Hallucination Alignment) framework. By leveraging counterfactual hard negative mining, our pipeline constructs a high-quality preference dataset that forces models to distinguish strict acoustic evidence from linguistically plausible fabrications. Additionally, we establish AHA-Eval, a diagnostic benchmark designed to rigorously test these fine-grained reasoning capabilities. We apply this data to align Qwen2.5-Omni. The resulting model, Qwen-Audio-AHA, achieves a 13.7% improvement on AHA-Eval. Crucially, this benefit generalizes beyond our diagnostic set. Our model shows substantial gains on public benchmarks, including 1.3% on MMAU-Test and 1.6% on MMAR, outperforming latest SOTA methods.

Pranav Bhandari, Usman Naseem, Mehwish Nasim

Personality steering in large language models (LLMs) commonly relies on injecting trait-specific steering vectors, implicitly assuming that personality traits can be controlled independently. In this work, we examine whether this assumption holds by analysing the geometric relationships between Big Five personality steering directions. We study steering vectors extracted from two model families (LLaMA-3-8B and Mistral-8B) and apply a range of geometric conditioning schemes, from unconstrained directions to soft and hard orthonormalisation. Our results show that personality steering directions exhibit substantial geometric dependence: steering one trait consistently induces changes in others, even when linear overlap is explicitly removed. While hard orthonormalisation enforces geometric independence, it does not eliminate cross-trait behavioural effects and can reduce steering strength. These findings suggest that personality traits in LLMs occupy a slightly coupled subspace, limiting fully independent trait control.

Wenyuan Zhang, Xinghua Zhang, Haiyang Yu, Shuaiyi Nie, Bingli Wu, Juwei Yue, Tingwen Liu, Yongbin Li

Experience intervention in web agents emerges as a promising technical paradigm, enhancing agent interaction capabilities by providing valuable insights from accumulated experiences. However, existing methods predominantly inject experience passively as global context before task execution, struggling to adapt to dynamically changing contextual observations during agent-environment interaction. We propose **ExpSeek**, which shifts experience toward step-level proactive seeking: (1) estimating step-level entropy thresholds to determine intervention timing using the model’s intrinsic signals; (2) designing step-level tailored experience content. Experiments on Qwen3-8B and 32B models across four challenging web agent benchmarks demonstrate that ExpSeek achieves absolute improvements of 9.3% and 7.5%, respectively. Our experiments validate the feasibility and advantages of entropy as a self-triggering signal, reveal that even a small-scale 4B experience model can significantly boost the performance of larger agent models. The code is released at https://github.com/WYRipple/ExpSeek.

Ying Su, Zheng Mingen, Weili Diao, Haoran Li

Pluralism alignment with AI has the sophisticated and necessary goal of creating AI that can coexist with and serve morally multifaceted humanity. Research towards pluralism alignment has many efforts in enhancing the learning of large language models (LLMs) to accomplish pluralism. Although this is essential, the robustness of LLMs to produce moral content over pluralistic values is still under exploration. Inspired by the astonishing persuasion abilities via jailbreak prompts, we propose to leverage jailbreak attacks to study LLMs’ internal pluralistic values. In detail, we develop a morality dataset with 10.4K instances in two categories: Value Ambiguity and Value Conflict. We further formalize four adversarial attacks with the constructed dataset, to manipulate LLMs’ judgment over the morality questions. We evaluate both the large language models and guardrail models which are typically used in generative systems with flexible user input. Our experiment results show that there is a critical vulnerability of LLMs and guardrail models to these subtle and sophisticated moral-aware attacks.

Micky C. Nnamdi, Benoit Louis Marteau, Yishan Zhong, J. Ben Tamo, May Dongmei Wang

Large Multimodal Models (LMMs) achieve state-of-the-art performance in high-stakes domains like healthcare, yet their reasoning remains opaque. Attention- and saliency-based methods often fail to faithfully represent the model’s decision process, particularly when integrating heterogeneous modalities. We introduce Tree-of-Evidence (ToE), an inference-time search algorithm that frames interpretability as a discrete optimization problem. Rather than relying on soft attention weights, ToE employs lightweight Evidence Bottlenecks that score coarse units of data (e.g., vital-sign windows, report chunks) and performs a beam search to identify the compact evidence set required to reproduce the model’s prediction. We evaluate ToE across six tasks spanning three datasets and two domains, including clinical prediction on MIMIC-IV, cross-center validation on eICU, and non-clinical fault detection on LEMMA-RCA. ToE retains over 98% of full-model AUROC with as few as five evidence units, achieves higher decision agreement and lower fidelity error than LIME, SHAP, saliency, and concept-bottleneck baselines under sparse budgets, and outperforms LLMs up to 70B parameters. ToE therefore provides a practical mechanism for auditing multimodal models.

Juncheng Hu, Jiming Yu, Rui Song, Kedi Lyu, Yingji Li, Zheli Liu

ocial bias in Multimodal Large Language Models (MLLMs) has become an increasingly important concern. Prompt-based approaches offer a lightweight solution for debiasing; however, existing methods rely heavily on handcrafted prompts that are brittle, highly context-sensitive, and difficult to generalize across tasks, bias types, and multimodal settings. In this work, we propose Historical Reflection-Guided Prompt Optimization (HRPO), an adaptive self-debiasing framework for black-box MLLMs that automatically optimizes task-specific debiasing prompts to suppress stereotypical outputs. To mitigate forgetting during prompt optimization, we introduce Historical Contrastive Self-Reflection (HCSR), which performs contrastive reflection over positive and negative optimization histories, enabling the model to retain effective prompts and avoid redundant exploration, thereby improving optimization efficiency. Experiments on three benchmarks involving eight open-source and two closed-source MLLMs, covering ten singular and two intersectional bias types, demonstrate that HRPO achieves strong debiasing performance while offering improved interpretability, generalization, and robustness. Code is available at: https://github.com/liyingji1996/HRPO.

Dominik Macko, Jakub Kopál

Machine-generated text detection, as an important task, is predominantly focused on English in research. This makes the existing detectors almost unusable for non-English languages, relying purely on cross-lingual transferability. There exist only a few works focused on any of Central European languages, leaving the transferability towards these languages rather unexplored. We fill this gap by providing the first benchmark of detection methods focused on this region, while also providing comparison of train-languages combinations to identify the best performing ones. We focus on multi-domain, multi-generator, and multilingual evaluation, pinpointing the differences of individual aspects, as well as adversarial robustness of detection methods. Supervised finetuned detectors in the Central European languages are found the most performant in these languages as well as the most resistant against obfuscation.