arXiv subjects

Minghui Pan

Publications and source records attributed to Minghui Pan.

4 recordsLinked to original sources

arXiv2026

Tool Specifications Matter: Uncovering and Mitigating Safety Risks in AI Agents

AI agents extend large language models (LLMs) with external tools, enabling them to perform complex tasks and translate model outputs into consequential real-world actions. Yet LLMs often become substantially less safe when deployed as agents, and the source of this degradation remains poorly understood. In this paper, we identify schema-formatted tool specifications as a primary source of agent safety degradation and show, through white-box representation analysis, that they weaken the model's internal refusal signals and contribute to unsafe tool execution. Building on this finding, we propose SafeKeep, an inference-time safeguard that decouples safety judgment from tool execution: it assesses requests using flattened textual tool specifications while retaining the original schema-formatted specifications for execution. Across two representative benchmarks and four LLMs, including both white-box and black-box models, SafeKeep increases the average refusal rate for harmful requests from 23.8% to 70.6% and reduces the average attack success rate under observation-level prompt injection from 25.6% to 2.5%. It also outperforms existing safeguards and preserves task-handling capability. We release the code and data at https://github.com/snowcatsmoking/SafeKeep .

cs.AI↗

arXiv2026

One-Eval: An Agentic System for Automated and Traceable LLM Evaluation

Reliable evaluation is essential for developing and deploying large language models, yet in practice it often requires substantial manual effort: practitioners must identify appropriate benchmarks, reproduce heterogeneous evaluation codebases, configure dataset schema mappings, and interpret aggregated metrics. To address these challenges, we present One-Eval, an agentic evaluation system that converts natural-language evaluation requests into executable, traceable, and customizable evaluation workflows. One-Eval integrates (i) NL2Bench for intent structuring and personalized benchmark planning, (ii) BenchResolve for benchmark resolution, automatic dataset acquisition, and schema normalization to ensure executability, and (iii) Metrics \& Reporting for task-aware metric selection and decision-oriented reporting beyond scalar scores. The system further incorporates human-in-the-loop checkpoints for review, editing, and rollback, while preserving sample evidence trails for debugging and auditability. Experiments show that One-Eval can execute end-to-end evaluations from diverse natural-language requests with minimal user effort, supporting more efficient and reproducible evaluation in industrial settings. Our framework is publicly available at https://github.com/OpenDCAI/One-Eval.

cs.CL↗

arXiv2026

Streaming Hallucination Detection in Long Chain-of-Thought Reasoning

Long chain-of-thought (CoT) reasoning improves the performance of large language models, yet hallucinations in such settings often emerge subtly and propagate across reasoning steps. We suggest that hallucination in long CoT reasoning is better understood as an evolving latent state rather than a one-off erroneous event. Accordingly, we treat step-level hallucination judgments as local observations and introduce a cumulative prefix-level hallucination signal that tracks the global evolution of the reasoning state over the entire trajectory. Overall, our approach enables streaming hallucination detection in long CoT reasoning, providing real-time, interpretable evidence.

cs.AI↗

arXiv2025

Reallocating Attention Across Layers to Reduce Multimodal Hallucination

Multimodal large reasoning models (MLRMs) often suffer from hallucinations that stem not only from insufficient visual grounding but also from imbalanced allocation between perception and reasoning processes. Building upon recent interpretability findings suggesting a staged division of attention across layers, we analyze how this functional misalignment leads to two complementary failure modes: perceptual bias in shallow layers and reasoning drift in deeper layers. To alleviate these issues, we propose Functional Head Identification and Class-Conditioned Rescaling , a lightweight, training-free plugin that identifies perception- and reasoning-oriented heads and adaptively rebalances their layerwise contributions. Our method improves reasoning consistency and visual faithfulness without retraining or any architectural modification. Evaluations across three representative MLRMs and five multimodal reasoning benchmarks show an average 4.2% point gain, with less than 1% additional computation and only 9% baseline latency. Beyond empirical improvements, our study provides an interpretable perspective on regulating cross-layer functional dynamics to enhance the reliability of multimodal reasoning.

cs.AI↗