超越行为：视觉-语言-动作模型安全评估的刺激-反应研究
Beyond Behaviour: A Stimulus-Response Safety Evaluation of Vision-Language-Action Models
AI 导读
一项 arXiv 研究在三个模型和两个视频 benchmark 上考察视觉-语言-动作模型是否像人类一样对安全关键刺激产生防御性反应。研究发现，即使模型内部已识别到危险，任务执行仍可能是默认响应。
正文 · 原文
Abstract:Humans exhibit biological sensitivity to safety-critical stimuli that can trigger defensive action. Across three models and two video benchmarks, we investigate whether similar responses emerge in vision-language-action models by examining their internal representations and action tendencies in response to visible hazards and harmful consequences. Our findings suggest that task execution may remain the default response even when danger is internally registered.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.34837 [cs.AI] |
| (or arXiv:2610.34837v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.34837 arXiv-issued DOI via DataCite (pending registration) |
|
| Journal reference: | The Science of Physical AI Safety workshop at CoRL 2026 |
Submission history
From: Vy Vo [view email]
[v1]
Mon, 26 Oct 2026 04:36:21 UTC (4,199 KB)
来源：arXiv cs.AI · arxiv.org