Training beats stricter prompts for smaller Rust code edits, a post says
A post summarizing the CROCODIL paper says researchers post-trained Olmo3 7B using two signals: keep edits small, but still build successfully and pass tests.
TLDR
According to a post summarizing “CROCODIL: Cross-Model Code Editing with LLMs,” stricter prompts did not consistently stop coding models from over-editing one another’s work. The post says CROCODIL roughly halved Olmo3 7B’s edit distance across implementations from every model, while improving build and all-test pass rates when editing implementations from every other model tested. It notes that the study was limited to Rust function edits.
Combined views
733
2 Sources, first seen 23d ago