LLMs Make More and Larger Edits on Code Written by a Different LLM

arXiv 2609.03894·medium signalXBlueskyLinkedInCopy link

arXiv 2609.03894 identifies a failure mode created by teams that mix assistants: because different models are trained on different data and carry different stylistic preferences, any one model's edits are frequently applied to foreign code another model wrote, and the paper finds models make more, often excessive, edits in that setting. CROCODIL is a post-training framework combining a similarity reward that penalizes large changes with an execution reward scoring build and test success, multiplied so the policy shrinks edit size without dropping task success. This is the direct multi-model counterpart to the same day's over-editing result, and it names a cost of assistant-switching that teams currently absorb invisibly in review.