Attention-Guided Safety Filter for Vision-Language-Action Models — 'Your Model Already Knows'

arXiv·medium signal

Vision-Language-Action (VLA) models drive strong end-to-end robotic manipulation but lack robust runtime safety filtering. This paper argues a VLA model's own internal attention signals can guide a safety filter without bolting on a separate external safety model, lowering the cost of guardrails. For builders of embodied and robotic agents, it points toward cheaper in-model safety enforcement rather than a separate verification stack.