CROCODIL paper: training for minimal diffs halves cross-model edit distance on Olmo3 7B
rohanpaul_ai · x · 2026-09-07
Coding models tend to over-edit when modifying other models' code, and stricter prompts alone don't fix it consistently, according to the paper "CROCODIL: Cross-Model Code Editing with LLMs".
Instead, the researchers post-trained Olmo3 7B with two signals: keep the edit small while still building and passing all tests. CROCODIL roughly halved edit distance across every foreign implementation tested, while improving build and all-test pass rates on all of them.
Caveat: the study is limited to Rust function edits. The takeaway for teams mixing coding models: benchmark cross-model editing and measure unnecessary diff size, not just final pass rates. (arxiv.org/abs/2609.03894)
Related event: CROCODIL Curbs LLMs' Over-Editing of Other Models' Code(3 posts)→
More from coding & agent
- Grady Booch revives Weinberg's 1971 quip: software so buggy one woodpecker could destroy civilization — Grady_Booch · 2026-10-23
- One Agent Replaces Half an SDR Team: Full Lead-Research-to-Email Workflow — tibo_maker · 2026-10-23
- Why don't coding agent harnesses ship with their own sandbox? Only amp does — Aryvyo · 2026-10-23
- Dev builds a VST synth with Codex and MCP, lets the AI jam in his studio — Yamapama · 2026-10-23
- AgentsInTheCloud update: use your Claude subscription with built-in coding agents — lucasmeijer · 2026-10-23
- Phone agents vs. secure screens: banking apps treat helpful agents like malware — ghostmalone02 · 2026-10-23