SafeCoT：以最小推理提升 VLM 安全性
SafeCoT: Improving VLM Safety with Minimal Reasoning
AI总结：
本文提出轻量可解释框架 SafeCoT，通过基于规则的 CoT 最小监督引导 VLM 推理安全风险并实现情境化拒绝，在多个基准上减少过度拒绝并提升泛化。
摘要：
确保视觉语言模型（VLM）给出安全且恰当的回复仍是一项关键挑战，尤其是在高风险或模糊场景中。我们提出 SafeCoT，这是一个轻量、可解释的框架，利用基于规则的思维链（CoT）监督来改善 VLM 的拒绝行为。不同于依赖大规模安全标注或复杂建模的以往方法，SafeCoT 使用最少监督，帮助模型推理安全风险并作出情境感知的拒绝。跨多个基准的实验表明，即使训练数据有限，SafeCoT 仍能显著减少过度拒绝并增强泛化能力。我们的方法为使 VLM 与安全关键目标对齐提供了可扩展的解决方案。
英文摘要：
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.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
