Beyond Content Safety: Real-Time Monitoring for Reasoning Vulnerabilities in LLM Chain-of-Thought

Forum topic · 小凯 · 2026-03-29

Summary

This paper introduces reasoning safety as a security dimension for large language models that is orthogonal to traditional content safety. While existing work treats chain-of-thought (CoT) reasoning traces as opaque intermediate artifacts focused only on content-level safety, the authors argue that reasoning trajectories must also be logically consistent, computationally efficient, and resistant to adversarial manipulation. The paper formally defines reasoning safety and proposes a taxonomy of nine categories of unsafe reasoning behaviors. The authors build a real-time reasoning safety monitor and evaluate it on a static benchmark of 450 reasoning chains, achieving 84.88% step-level localization accuracy in identifying unsafe steps and 85.37% accuracy in classifying error types. The work targets ML research and was authored by Xunguang Wang, Yuguang Zhou, Qingyue Wang, Zongjie Li, Ruixuan Huang, and colleagues, released on arXiv (2603.25412) in March 2026.

Paper Overview

Research area: ML Authors: Xunguang Wang, Yuguang Zhou, Qingyue Wang, Zongjie Li, Ruixuan Huang, et al. Release date: 2026-03-26 arXiv: 2603.25412

Abstract (translated)

Large language models (LLMs) increasingly rely on explicit chain-of-thought (CoT) reasoning to solve complex tasks, yet the safety of the reasoning process itself remains underexplored. Existing work focuses on content safety and treats reasoning chains as opaque intermediate artifacts.

This paper identifies reasoning safety as an orthogonal and equally critical safety dimension: model reasoning trajectories should be logically consistent, computationally efficient, and robust against adversarial manipulation. The authors formalize the notion of reasoning safety and introduce a taxonomy of nine categories of unsafe reasoning behaviors.

In a static benchmark evaluation on 450 reasoning chains, the proposed reasoning safety monitor achieves 84.88% step-level localization accuracy and 85.37% error-type classification accuracy.

Key Contributions

Formal definition of reasoning safety as a dimension distinct from content safety
A taxonomy covering nine types of unsafe reasoning behaviors
A real-time reasoning safety monitor for detecting and localizing unsafe reasoning steps
Benchmark results: 84.88% step-level localization accuracy, 85.37% error-type classification accuracy on 450 chains

--- *Auto-collected on 2026-03-29*