Vision-language-action models now reason before they act — generating text-based plans (“pick up the red block, move it left”) before producing motor commands. The chain-of-thought step is supposed to improve robustness by making the reasoning explicit and correctable. If the robot can articulate its plan, the plan can be verified.
The paper (arXiv:2603.12717, March 2026) corrupts the reasoning trace in seven different ways and measures which corruptions break the robot. The result is asymmetric. Substituting object names in the reasoning trace — replacing “red block” with “blue cylinder” — drops success rate by 8.3 percentage points overall and up to 45 points on specific tasks. But reordering sentences, reversing spatial directions, adding noise, and even having a language model generate plausible-but-wrong plans produce minimal impact.
The action decoder ignores the reasoning and reads the nouns. The chain-of-thought is not a plan that the motor system follows step by step — it is a name-resolution layer that tells the motor system which objects to attend to. The logical structure (first do X, then do Y, because Z) is decorative. The entity references (the red block, the left corner) are load-bearing.
The structural lesson: a system that appears to reason from plans may actually reason from labels. The plan is not the control signal — the vocabulary embedded in the plan is. Corrupting the logic preserves performance because the logic was never the operative channel. Corrupting the names destroys performance because the names were always the operative channel. The architecture's apparent reasoning is a side effect of its actual operation: entity grounding.