This paper proposes a framework called ReasoningShield to address the problem of harmful content within Chain-of-Thoughts (CoTs) of Large Reasoning Models (LRMs). While the final solution may appear acceptable, we recognize that harmful content can arise in intermediate stages, and propose a lightweight solution that effectively censors CoTs. We define a multi-level taxonomy for CoT censorship tasks, encompassing 10 risk categories and 3 safety levels, and build the first CoT censorship benchmark consisting of 9.2K query and inference trace pairs. Furthermore, we develop a two-stage training strategy that combines stage-by-stage risk analysis with contrastive learning. ReasoningShield outperforms LlamaGuard-4 by 35.6% and GPT-4o by 15.8%, demonstrating effective generalization across diverse inference paradigms, tasks, and novel scenarios.
