Planning and revision are meant to bring generated code closer to the task specification, but a new arXiv preprint argues that this is not always what happens. The paper introduces the term 'interpretive lock-in' for a pattern in which a mistaken interpretation of the task persists across the plan, the implementation, and the tests.
Because the same wrong interpretation survives every stage, revision alone may not be enough to fix the generated code. The abstract suggests that tests can pass even when the semantics of the task are missed, pointing to a limitation in current plan-based code-generation approaches.
The preprint is available on arXiv under identifier 2610.15797.