Reasoning Structure Matters for Safety Alignment of Reasoning Models

TL;DR

This paper shows that the reasoning structure in large reasoning models influences safety risks and introduces AltTrain, a simple supervised finetuning method to improve safety alignment by modifying reasoning structures.

Contribution

It proposes AltTrain, a practical method for safety alignment that requires only supervised finetuning to alter reasoning structures in LRMs.

Findings

AltTrain achieves strong safety alignment across various models and tasks.

It requires only 1K supervised examples and no complex RL training.

The method generalizes well across reasoning, QA, summarization, and multilingual tasks.

Abstract

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.