Selected work

Representative Papers

LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models

Sep 30, 2025

Emerging safety risks in multi-turn multimodal (MMT) dialogue with vision-language models (VLMs) involve malicious intent distributed stealthily across turns and modalities, evading detection by single-turn or single-modality safety auditing. Method: This work formally defines the MMT dialogue safety problem; introduces MMDS—the first fine-grained, evidence-annotated, safety-specific dataset for MMT dialogue; proposes a Monte Carlo Tree Search–based multimodal red-teaming framework to automatically generate cross-turn, cross-modal harmful dialogues; and designs a risk assessment model integrating multimodal contextual modeling and policy-aware reasoning for joint input-response safety evaluation. Contribution/Results: The proposed LLaVAShield framework achieves significant improvements over strong baselines on multi-turn content moderation tasks and maintains state-of-the-art performance under dynamic safety policies.

0 citationsRead paper

D$^2$HScore: Reasoning-Aware Hallucination Detection via Semantic Breadth and Depth Analysis in LLMs

Sep 15, 2025

Large language models (LLMs) frequently generate factually incorrect content—so-called hallucinations—hindering their deployment in high-stakes domains such as finance and healthcare. To address this, we propose D²HScore, a training-free, annotation-free hallucination detection framework. Unlike prior approaches, D²HScore operates entirely on internal model dynamics: it jointly quantifies intra-layer semantic dispersion and inter-layer semantic drift across LLM transformer layers, while leveraging attention weights to identify conceptually critical tokens—enabling highly interpretable detection. Grounded in the intrinsic multi-layer architecture and autoregressive decoding behavior of LLMs, D²HScore requires no external data, fine-tuning, or supervision. Extensive evaluation across five open-source LLMs and five standard hallucination benchmarks demonstrates that D²HScore consistently outperforms existing unsupervised baselines, exhibits strong cross-model generalization, and enables plug-and-play deployment—establishing a novel, practical paradigm for trustworthy LLM evaluation.

0 citationsRead paper

Unveiling Trust in Multimodal Large Language Models: Evaluation, Analysis, and Mitigation

Aug 21, 2025

Existing evaluations of multimodal large language models (MLLMs) are fragmented and overlook modality-specific risks, hindering holistic trustworthiness assessment. Method: We introduce MultiTrust-X—the first comprehensive benchmark covering authenticity, robustness, safety, fairness, and privacy—grounded in a novel three-dimensional analytical framework. It systematically identifies two previously unrecognized vulnerability classes: “multimodal risks” and “cross-modal impacts,” exposing limitations and unintended side effects of current mitigation strategies. The benchmark integrates 32 tasks, 28 datasets, and 8 mainstream mitigation methods to evaluate over 30 MLLMs. Additionally, we propose Reasoning-Enhanced Safety Alignment (RESA), a novel alignment technique that improves overall trustworthiness without compromising model performance. Contribution/Results: MultiTrust-X establishes an empirical foundation and a new paradigm for trustworthy multimodal intelligence, enabling rigorous, standardized, and modality-aware evaluation of MLLM trustworthiness.

0 citationsRead paper

Done Is Better than Perfect: Unlocking Efficient Reasoning by Structured Multi-Turn Decomposition

May 26, 2025

Large reasoning models (LRMs) suffer from high time-to-first-token (TTFT) and low inference efficiency due to lengthy implicit chain-of-thought (CoT) reasoning. To address this, we propose MinD (Structured Multi-turn Decomposition), a novel paradigm that explicitly decomposes CoT into controllable, interruptible multi-turn interactions—each generating one reasoning unit and its corresponding answer. Our contributions include: (1) the first explicit, structured, turn-controllable multi-turn reasoning framework; (2) real-time user intervention capability (e.g., pause/resume); and (3) the first application of reinforcement learning (specifically GRPO) to jointly optimize both the number of reasoning turns and answer correctness. MinD integrates supervised fine-tuning (SFT), LLM-based rewriting distillation, multi-turn formatted data construction, and the R1-Distill architecture. Evaluated on MATH, MinD reduces TTFT and output token count by ~70%, while maintaining competitive performance on MATH-500 and AIME24.

0 citationsRead paper

Understanding Pre-training and Fine-tuning from Loss Landscape Perspectives

May 23, 2025

This work investigates the capability evolution mechanisms of large language models (LLMs) during pretraining and fine-tuning from a loss landscape perspective, specifically addressing whether fine-tuning degrades pretrained foundational capabilities. Method: We propose a dual-perspective loss landscape analysis framework—characterizing both best-case and worst-case capability basins—and integrate direction-sensitive modeling, high-dimensional optimization theory, and overparameterization analysis. Contribution/Results: We theoretically prove, for the first time, that the size of the best-case basin bounds both the extent of the worst-case basin and input robustness. We reveal that overparameterization significantly expands capability basins—by up to 5×—and establish a geometric criterion: remaining within any capability basin guarantees preservation of the corresponding capability. Experiments quantitatively validate the correlation between basin expandability and robustness boundaries, providing an interpretable geometric foundation for safe and controllable fine-tuning.

0 citationsRead paper

Recent publications

Latest Papers

LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models

Sep 30, 2025

Emerging safety risks in multi-turn multimodal (MMT) dialogue with vision-language models (VLMs) involve malicious intent distributed stealthily across turns and modalities, evading detection by single-turn or single-modality safety auditing. Method: This work formally defines the MMT dialogue safety problem; introduces MMDS—the first fine-grained, evidence-annotated, safety-specific dataset for MMT dialogue; proposes a Monte Carlo Tree Search–based multimodal red-teaming framework to automatically generate cross-turn, cross-modal harmful dialogues; and designs a risk assessment model integrating multimodal contextual modeling and policy-aware reasoning for joint input-response safety evaluation. Contribution/Results: The proposed LLaVAShield framework achieves significant improvements over strong baselines on multi-turn content moderation tasks and maintains state-of-the-art performance under dynamic safety policies.

0 citationsRead paper

D$^2$HScore: Reasoning-Aware Hallucination Detection via Semantic Breadth and Depth Analysis in LLMs

Sep 15, 2025

Large language models (LLMs) frequently generate factually incorrect content—so-called hallucinations—hindering their deployment in high-stakes domains such as finance and healthcare. To address this, we propose D²HScore, a training-free, annotation-free hallucination detection framework. Unlike prior approaches, D²HScore operates entirely on internal model dynamics: it jointly quantifies intra-layer semantic dispersion and inter-layer semantic drift across LLM transformer layers, while leveraging attention weights to identify conceptually critical tokens—enabling highly interpretable detection. Grounded in the intrinsic multi-layer architecture and autoregressive decoding behavior of LLMs, D²HScore requires no external data, fine-tuning, or supervision. Extensive evaluation across five open-source LLMs and five standard hallucination benchmarks demonstrates that D²HScore consistently outperforms existing unsupervised baselines, exhibits strong cross-model generalization, and enables plug-and-play deployment—establishing a novel, practical paradigm for trustworthy LLM evaluation.

0 citationsRead paper

Unveiling Trust in Multimodal Large Language Models: Evaluation, Analysis, and Mitigation

Aug 21, 2025

Existing evaluations of multimodal large language models (MLLMs) are fragmented and overlook modality-specific risks, hindering holistic trustworthiness assessment. Method: We introduce MultiTrust-X—the first comprehensive benchmark covering authenticity, robustness, safety, fairness, and privacy—grounded in a novel three-dimensional analytical framework. It systematically identifies two previously unrecognized vulnerability classes: “multimodal risks” and “cross-modal impacts,” exposing limitations and unintended side effects of current mitigation strategies. The benchmark integrates 32 tasks, 28 datasets, and 8 mainstream mitigation methods to evaluate over 30 MLLMs. Additionally, we propose Reasoning-Enhanced Safety Alignment (RESA), a novel alignment technique that improves overall trustworthiness without compromising model performance. Contribution/Results: MultiTrust-X establishes an empirical foundation and a new paradigm for trustworthy multimodal intelligence, enabling rigorous, standardized, and modality-aware evaluation of MLLM trustworthiness.

0 citationsRead paper

Done Is Better than Perfect: Unlocking Efficient Reasoning by Structured Multi-Turn Decomposition

May 26, 2025

Large reasoning models (LRMs) suffer from high time-to-first-token (TTFT) and low inference efficiency due to lengthy implicit chain-of-thought (CoT) reasoning. To address this, we propose MinD (Structured Multi-turn Decomposition), a novel paradigm that explicitly decomposes CoT into controllable, interruptible multi-turn interactions—each generating one reasoning unit and its corresponding answer. Our contributions include: (1) the first explicit, structured, turn-controllable multi-turn reasoning framework; (2) real-time user intervention capability (e.g., pause/resume); and (3) the first application of reinforcement learning (specifically GRPO) to jointly optimize both the number of reasoning turns and answer correctness. MinD integrates supervised fine-tuning (SFT), LLM-based rewriting distillation, multi-turn formatted data construction, and the R1-Distill architecture. Evaluated on MATH, MinD reduces TTFT and output token count by ~70%, while maintaining competitive performance on MATH-500 and AIME24.

0 citationsRead paper

Understanding Pre-training and Fine-tuning from Loss Landscape Perspectives

May 23, 2025

This work investigates the capability evolution mechanisms of large language models (LLMs) during pretraining and fine-tuning from a loss landscape perspective, specifically addressing whether fine-tuning degrades pretrained foundational capabilities. Method: We propose a dual-perspective loss landscape analysis framework—characterizing both best-case and worst-case capability basins—and integrate direction-sensitive modeling, high-dimensional optimization theory, and overparameterization analysis. Contribution/Results: We theoretically prove, for the first time, that the size of the best-case basin bounds both the extent of the worst-case basin and input robustness. We reveal that overparameterization significantly expands capability basins—by up to 5×—and establish a geometric criterion: remaining within any capability basin guarantees preservation of the corresponding capability. Experiments quantitatively validate the correlation between basin expandability and robustness boundaries, providing an interpretable geometric foundation for safe and controllable fine-tuning.

0 citationsRead paper