Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning?

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yin, Qingyu, Leong, Chak Tou, Yang, Linyi, Huang, Wenxuan, Li, Wenjie, Wang, Xiting, Yoon, Jaehong, YunXing, XingYu, Gu, Jinjin
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911196105408512
author Yin, Qingyu
Leong, Chak Tou
Yang, Linyi
Huang, Wenxuan
Li, Wenjie
Wang, Xiting
Yoon, Jaehong
YunXing
XingYu
Gu, Jinjin
author_facet Yin, Qingyu
Leong, Chak Tou
Yang, Linyi
Huang, Wenxuan
Li, Wenjie
Wang, Xiting
Yoon, Jaehong
YunXing
XingYu
Gu, Jinjin
contents Large reasoning models (LRMs) with multi-step reasoning capabilities have shown remarkable problem-solving abilities, yet they exhibit concerning safety vulnerabilities that remain poorly understood. In this work, we investigate why safety alignment fails in reasoning models through a mechanistic interpretability lens. Using a linear probing approach to trace refusal intentions across token positions, we discover a striking phenomenon termed as \textbf{refusal cliff}: many poorly-aligned reasoning models correctly identify harmful prompts and maintain strong refusal intentions during their thinking process, but experience a sharp drop in refusal scores at the final tokens before output generation. This suggests that these models are not inherently unsafe; rather, their refusal intentions are systematically suppressed. Through causal intervention analysis, we identify a sparse set of attention heads that negatively contribute to refusal behavior. Ablating just 3\% of these heads can reduce attack success rates below 10\%. Building on these mechanistic insights, we propose \textbf{Cliff-as-a-Judge}, a novel data selection method that identifies training examples exhibiting the largest refusal cliff to efficiently repair reasoning models' safety alignment. This approach achieves comparable safety improvements using only 1.7\% of the vanilla safety training data, demonstrating a less-is-more effect in safety alignment.
format Preprint
id arxiv_https___arxiv_org_abs_2510_06036
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning?
Yin, Qingyu
Leong, Chak Tou
Yang, Linyi
Huang, Wenxuan
Li, Wenjie
Wang, Xiting
Yoon, Jaehong
YunXing
XingYu
Gu, Jinjin
Artificial Intelligence
Cryptography and Security
Large reasoning models (LRMs) with multi-step reasoning capabilities have shown remarkable problem-solving abilities, yet they exhibit concerning safety vulnerabilities that remain poorly understood. In this work, we investigate why safety alignment fails in reasoning models through a mechanistic interpretability lens. Using a linear probing approach to trace refusal intentions across token positions, we discover a striking phenomenon termed as \textbf{refusal cliff}: many poorly-aligned reasoning models correctly identify harmful prompts and maintain strong refusal intentions during their thinking process, but experience a sharp drop in refusal scores at the final tokens before output generation. This suggests that these models are not inherently unsafe; rather, their refusal intentions are systematically suppressed. Through causal intervention analysis, we identify a sparse set of attention heads that negatively contribute to refusal behavior. Ablating just 3\% of these heads can reduce attack success rates below 10\%. Building on these mechanistic insights, we propose \textbf{Cliff-as-a-Judge}, a novel data selection method that identifies training examples exhibiting the largest refusal cliff to efficiently repair reasoning models' safety alignment. This approach achieves comparable safety improvements using only 1.7\% of the vanilla safety training data, demonstrating a less-is-more effect in safety alignment.
title Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning?
topic Artificial Intelligence
Cryptography and Security
url https://arxiv.org/abs/2510.06036