If It's Not Buggy, Don't Fix It: On the Dynamics of Iterative Bug-fixing with LLMs

RESUMO

LLM-based automated program repair tools are increasingly used during code review. Across multiple models and repair environments, LLMs consistently claim to detect bugs in entirely bug-free programs, and the rate of correct repair of buggy programs is less than the rate of damage to correct programs. Iterative use often reaches a pseudo-bug-fixing cycle where the same changes are added and removed repeatedly; mechanistic probing reveals a steering vector controlling editing propensity that is falsely activated to induce pseudo-bug fixing.