Saved in:
Bibliographic Details
Main Authors: Yamaguchi, Kureha, Etheridge, Benjamin, Arditi, Andy
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2507.03167
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914562152857600
author Yamaguchi, Kureha
Etheridge, Benjamin
Arditi, Andy
author_facet Yamaguchi, Kureha
Etheridge, Benjamin
Arditi, Andy
contents Chat models without chain-of-thought (CoT) reasoning must decide whether to refuse a harmful request before generating their first response token. Reasoning models, by contrast, produce extended chains of thought before their final output, raising a natural question: where in this process does the decision to refuse occur? We investigate this across four open-source reasoning models. We first show that the CoT causally influences refusal outcomes; fixing a specific reasoning trace substantially reduces variance in whether the model ultimately refuses or complies. Zooming into the reasoning trace, we find that in distilled models, subtle differences in the opening sentence of the CoT can fully determine the model's refusal decision, and that these patterns transfer across models distilled from the same teacher. Finally, we extract linear refusal directions from model activations and show that ablating them increases harmful compliance, though less reliably than the same technique achieves on non-reasoning models, and with non-negligible degradation to general capabilities.
format Preprint
id arxiv_https___arxiv_org_abs_2507_03167
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Where Do Reasoning Models Refuse?
Yamaguchi, Kureha
Etheridge, Benjamin
Arditi, Andy
Computation and Language
Artificial Intelligence
Machine Learning
Chat models without chain-of-thought (CoT) reasoning must decide whether to refuse a harmful request before generating their first response token. Reasoning models, by contrast, produce extended chains of thought before their final output, raising a natural question: where in this process does the decision to refuse occur? We investigate this across four open-source reasoning models. We first show that the CoT causally influences refusal outcomes; fixing a specific reasoning trace substantially reduces variance in whether the model ultimately refuses or complies. Zooming into the reasoning trace, we find that in distilled models, subtle differences in the opening sentence of the CoT can fully determine the model's refusal decision, and that these patterns transfer across models distilled from the same teacher. Finally, we extract linear refusal directions from model activations and show that ablating them increases harmful compliance, though less reliably than the same technique achieves on non-reasoning models, and with non-negligible degradation to general capabilities.
title Where Do Reasoning Models Refuse?
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2507.03167