Reasoning Structure Matters for Safety Alignment of Reasoning Models

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.

Key Contributions

Identifies that safety risks in large reasoning models stem from their reasoning structure prioritizing task-solving over safety
Proposes AltTrain: a 3-step reasoning structure (problem understanding → harmfulness assessment → conditional reasoning) that enables safe reasoning
Achieves strong safety alignment using only 1K training examples via supervised fine-tuning, requiring ~60 minutes on a single A6000 for 8B models