LLMs Claim to Find Bugs in Bug-Free Programs, and a Steering Vector Controls the Urge to Edit

arXiv 2609.10123·high signalXBlueskyLinkedInCopy link

arXiv 2609.10123 (2026-09-09) ran LLMs as blind iterative bug-fixers across multiple models and repair environments and found they consistently report bugs in entirely correct programs, with the rate of damage to correct code exceeding the rate of repair to buggy code. Iterating long enough reaches a pseudo-bug-fixing cycle where the same edit is added and removed forever. Mechanistic probing located a steering vector controlling editing propensity, implying the model carries an internal 'buggy code' representation that gets falsely activated. If you run an autonomous repair loop, this is the argument for a hard stopping condition rather than a confidence threshold.