SaRO: Enhancing LLM Safety through Reasoning-based Alignment

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Mou, Yutao, Luo, Yuxiao, Zhang, Shikun, Ye, Wei
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908316955836416
author Mou, Yutao
Luo, Yuxiao
Zhang, Shikun
Ye, Wei
author_facet Mou, Yutao
Luo, Yuxiao
Zhang, Shikun
Ye, Wei
contents Current safety alignment techniques for large language models (LLMs) face two key challenges: (1) under-generalization, which leaves models vulnerable to novel jailbreak attacks, and (2) over-alignment, which leads to the excessive refusal of benign instructions. Our preliminary investigation reveals semantic overlap between jailbreak/harmful queries and normal prompts in embedding space, suggesting that more effective safety alignment requires a deeper semantic understanding. This motivates us to incorporate safety-policy-driven reasoning into the alignment process. To this end, we propose the Safety-oriented Reasoning Optimization Framework (SaRO), which consists of two stages: (1) Reasoning-style Warmup (RW) that enables LLMs to internalize long-chain reasoning through supervised fine-tuning, and (2) Safety-oriented Reasoning Process Optimization (SRPO) that promotes safety reflection via direct preference optimization (DPO). Extensive experiments demonstrate the superiority of SaRO over traditional alignment methods.
format Preprint
id arxiv_https___arxiv_org_abs_2504_09420
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SaRO: Enhancing LLM Safety through Reasoning-based Alignment
Mou, Yutao
Luo, Yuxiao
Zhang, Shikun
Ye, Wei
Computation and Language
Current safety alignment techniques for large language models (LLMs) face two key challenges: (1) under-generalization, which leaves models vulnerable to novel jailbreak attacks, and (2) over-alignment, which leads to the excessive refusal of benign instructions. Our preliminary investigation reveals semantic overlap between jailbreak/harmful queries and normal prompts in embedding space, suggesting that more effective safety alignment requires a deeper semantic understanding. This motivates us to incorporate safety-policy-driven reasoning into the alignment process. To this end, we propose the Safety-oriented Reasoning Optimization Framework (SaRO), which consists of two stages: (1) Reasoning-style Warmup (RW) that enables LLMs to internalize long-chain reasoning through supervised fine-tuning, and (2) Safety-oriented Reasoning Process Optimization (SRPO) that promotes safety reflection via direct preference optimization (DPO). Extensive experiments demonstrate the superiority of SaRO over traditional alignment methods.
title SaRO: Enhancing LLM Safety through Reasoning-based Alignment
topic Computation and Language
url https://arxiv.org/abs/2504.09420