Beyond Content Safety: Real-Time Monitoring for Reasoning Vulnerabilities in LLMs

Summary

This paper, posted to arXiv (2603.25412) by Xunguang Wang, Yuguang Zhou, Qingyue Wang, Zongjie Li, Ruixuan Huang, and colleagues, addresses a safety dimension of large language models beyond content safety: the safety of the chain-of-thought (CoT) reasoning process itself. The authors argue that existing work treats reasoning traces as opaque intermediate artifacts, and they instead define reasoning safety as an orthogonal and equally critical requirement — that reasoning trajectories be logically consistent, computationally efficient, and robust against adversarial manipulation. They formalize reasoning safety and introduce a taxonomy of nine categories of unsafe reasoning behaviors. To operationalize this, they build a reasoning safety monitor evaluated on a static benchmark of 450 reasoning chains, achieving 84.88% step-level localization accuracy in identifying where unsafe reasoning occurs and 85.37% accuracy in classifying the type of error. The work aims to enable real-time monitoring of model reasoning processes rather than only checking final outputs, complementing conventional content-safety filtering for LLMs deployed on complex reasoning tasks.