TNAS: Diagnosing Text-Action Divergence in Reasoning VLAs
November 2, 2026
The Text-Neutral Action-Subspace (TNAS) diagnostic reveals that visual perturbations can trigger physical action changes in reasoning VLAs while leaving natural-language rationales identical. Research shows extremely low norm overlap between action-sensitive directions and text-loss gradients in models like InstructVLA.
HOW THIS AFFECTS YOU
●
builderBe aware that a model's text explanation may not accurately reflect its physical behavior.
●
researcherYou can use TNAS to identify safety-critical decoupling between a model's reasoning and its actions.