Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs

A new research paper, "Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs," shows that visual perturbations can make reasoning vision-language-action models act differently while their natural-language rationales remain unchanged, unlike prior attacks that corrupt the rationale and leave visible evidence in the text. The work identifies this text-action divergence in embodied models that pair physical control with supervisor-readable rationales, and proposes diagnosing and steering it. The finding matters because it undermines rationale-based monitoring as a safety check for embodied AI systems.

Reasoning vision-language-action VLA models pair physical control with natural-language rationales that supervisors can read to monitor embodied safety. Prior attacks on VLA reasoning corrupt the rationale itself and leave visible evidence in the text. We show that visual perturbations can instead