Reasoning Structure Matters for Safety Alignment of Reasoning Models
For developers of large reasoning models, this work addresses a critical safety vulnerability by showing that reasoning structure itself is the root cause, and offers a practical, lightweight fix.
Large reasoning models (LRMs) often generate harmful responses to malicious queries due to their reasoning structure. The proposed AltTrain method alters this structure via supervised finetuning with only 1K examples, achieving strong safety alignment and robust generalization across reasoning, QA, summarization, and multilingual tasks.
Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks but often generate harmful responses to malicious user queries. This paper investigates the underlying cause of these safety risks and shows that the issue lies in the reasoning structure itself. Based on this insight, we claim that effective safety alignment can be achieved by altering the reasoning structure. We propose AltTrain, a simple yet effective post training method that explicitly alters the reasoning structure of LRMs. AltTrain is both practical and generalizable, requiring no complex reinforcement learning (RL) training or reward design, only supervised finetuning (SFT) with a lightweight 1K training examples. Experiments across LRM backbones and model sizes demonstrate strong safety alignment, along with robust generalization across reasoning, QA, summarization, and multilingual setting.