Robots that explain their thinking may not be explaining it honestly.
A new paper from researchers studying embodied AI finds a significant gap between what Vision-Language-Action models say they're doing and what's actually driving their decisions. The work distinguishes between "functional" reasoning — where a verbal chain-of-thought improves task performance — and "faithful" reasoning, where that chain-of-thought genuinely reflects the model's internal process. In a human evaluation of a state-of-the-art autonomous driving model, the team found inconsistent coupling between reasoning quality and actual trajectory improvement, meaning a robot can narrate sensible-sounding logic while something else entirely governs its movement.
This matters because interpretability in robotics is not an academic nicety — it's the prerequisite for debugging failures, auditing edge cases, and ultimately deploying systems in environments where a wrong turn has physical consequences. If a VLA model's verbal output is decorative rather than causal, the entire chain-of-thought mechanism flatters engineers into false confidence about what they understand.
To address the gap, the researchers built a learned critic they call Pinocchio, which scores a model's observation grounding and stepwise coherence, then uses that score as a reward signal during reinforcement-learning post-training. The result: faithfulness improved 4% over state-of-the-art alignment baselines and 18% over trajectory-error baselines, and on out-of-distribution scenarios the post-trained planner responded to rare counterfactual situations at 1.6 times the rate of the baseline policy.
The naming is on the nose, but the finding isn't subtle — the AI field has spent years praising chain-of-thought as a transparency tool, and this research suggests that in embodied settings, the chain may be mostly ornamental until someone builds a mechanism to hold it accountable.