Style, Not Self: LLMs Can't Recognize Their Own Code, and Claude Haiku's Self-Preference Vanishes After Normalization

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2609.30048 tests whether commercial LLMs recognize code they wrote, which bears on self-preferring judges and monitor collusion, using MBPP, HumanEval and DS-1000. Single-solution self-attribution sat at 49-58% balanced accuracy in all 15 model-benchmark combinations, and pairwise 'pick your own' accuracy correlated at r=0.93 with how often the evaluator's own solution was longer. Stripping docstrings, comments, type hints and local names kept Pass@1 intact, left ten of twelve re-tested results at chance, and removed Claude Haiku's self-preference. A trained classifier still separated most normalized pairs, so style leakage stays a real channel.