SafeCoT: Improving VLM Safety with Minimal Reasoning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ma, Jiachen, Zhou, Zhanhui, Yang, Chao, Lu, Chaochao
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913889029980160
author Ma, Jiachen
Zhou, Zhanhui
Yang, Chao
Lu, Chaochao
author_facet Ma, Jiachen
Zhou, Zhanhui
Yang, Chao
Lu, Chaochao
contents Ensuring safe and appropriate responses from vision-language models (VLMs) remains a critical challenge, particularly in high-risk or ambiguous scenarios. We introduce SafeCoT, a lightweight, interpretable framework that leverages rule-based chain-of-thought (CoT) supervision to improve refusal behavior in VLMs. Unlike prior methods that rely on large-scale safety annotations or complex modeling, SafeCoT uses minimal supervision to help models reason about safety risks and make context-aware refusals. Experiments across multiple benchmarks show that SafeCoT significantly reduces overrefusal and enhances generalization, even with limited training data. Our approach offers a scalable solution for aligning VLMs with safety-critical objectives.
format Preprint
id arxiv_https___arxiv_org_abs_2506_08399
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SafeCoT: Improving VLM Safety with Minimal Reasoning
Ma, Jiachen
Zhou, Zhanhui
Yang, Chao
Lu, Chaochao
Artificial Intelligence
Machine Learning
Ensuring safe and appropriate responses from vision-language models (VLMs) remains a critical challenge, particularly in high-risk or ambiguous scenarios. We introduce SafeCoT, a lightweight, interpretable framework that leverages rule-based chain-of-thought (CoT) supervision to improve refusal behavior in VLMs. Unlike prior methods that rely on large-scale safety annotations or complex modeling, SafeCoT uses minimal supervision to help models reason about safety risks and make context-aware refusals. Experiments across multiple benchmarks show that SafeCoT significantly reduces overrefusal and enhances generalization, even with limited training data. Our approach offers a scalable solution for aligning VLMs with safety-critical objectives.
title SafeCoT: Improving VLM Safety with Minimal Reasoning
topic Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2506.08399