Let a robot explain its reasoning before it acts, and you've also handed over the steering wheel.
Vision-language-action models turn a camera feed and a spoken instruction into motor commands, and some of them now pause to write out a reasoning chain in plain English before acting: text their creators pitch as an oversight tool, something a person can read, catch errors in, and edit. A new study tests what editing that text actually does to the robot's behavior. Researchers used a method called deterministic entity swap to alter instructions and reasoning chains across forty tasks in all four LIBERO simulation suites, then ran a targeted test on DeepThinkVLA, a policy whose reasoning chain is exposed as plain text. Fed a corrupted instruction paired with the untouched reasoning chain from the clean version, the robot recovered 47.8 percentage points of the success rate it had lost, moving in the predicted direction on all 10 tasks in the suite where the goal is set by language alone.
That is a real finding, not a trivial one: it means the written reasoning isn't just a transcript of a decision already made by the camera image, it's causally steering the robot's actions. That is good news for anyone who wants to correct a robot mid-task. It's also a warning: the same text channel that lets an engineer repair a mistake is a channel that a bad edit, or an injected instruction, can use to send a robot the wrong way.
That mirrors an open debate over language-model chain-of-thought, whether it's faithful reasoning or post-hoc narration, but this result is simulation-only, across four scripted LIBERO suites, so treat 'control surface' as proven in the lab, not yet on a warehouse floor.