Google's ReproBridge lifts coding agents' ML experiment reproduction to 92.6%

Google researchers present ReproBridge, a method that packages a successful experiment reproduction into reusable artifacts: cleaned code, environment setup, launch scripts, and a guide.

Across four models and 1,029 tasks from 36 papers, the refined artifacts raised mean reproduction success from 76.4% to 92.6% and cut API spending by 19.7%, according to a summary by researcher Elvis Saravia. Removing the code or the environment setup caused the largest drops in success.