In-Context VLA argues robots should consume grounded language rather than generate free-form reasoning

The preprint “In-Context VLA” argues that free-form chain-of-thought can degrade low-level vision-language-action control through latency, ungrounded reasoning and conflicting objectives. Its proposed post-training and tool-use approach supplies grounded language context to the controller rather than asking the controller to generate unrestricted text.

Impact: the design supports a clean separation between high-level planning and safety-critical control. It remains preliminary research; the authors’ performance results have not been independently reproduced or peer reviewed.