When coding models edit each other’s work, they tend to over-edit, and this paper shows that training for minimal correct diffs works better than stricter prompts.

A stricter prompt telling the model to make only minimal edits did not fix this consistently.

The researchers instead post-trained Olmo3 7B with 2 signals: keep the edit small, but still build and pass tests. CROCODIL roughly halved its edit distance across implementations from every model, while improving build and all-test pass rates on every foreign implementor tested.

The study is limited to Rust function edits. Still, if a team mixes coding models, it should benchmark cross-model editing and measure unnecessary diff size, not just whether the final code passes.

– arxiv. org/abs/2609.03894

Title: "CROCODIL: Cross-Model Code Editing with LLMs"