ReproBridge: Building Reproducibility into ML Research Artifacts
ReproBridge finds that agent-ready ML artifacts improve reproduction rates and cut API spending across four evaluated models. The practical lesson is to preserve setup and execution decisions alongside code.
ReproBridge packages a refined codebase, environment setup, experiment launchers, and a reproduction guide. On **1,029 tasks** from **387 experiments across 36 papers**, refined artifacts raised mean reproduction success from **76.4% to 92.6%** and reduced downstream API spending by **19.7%** on average. For builders running coding agents on research repos, capture the working environment, launch commands, and recovered procedure so later fresh sessions can reuse them. The paper says removing any artifact component reduced success, with the largest losses from removing code and environment details. The results come from a reproduction benchmark and four evaluated models. The material does not say how much the gains transfer to other coding tasks or project types.
No separate implication has been recorded yet.
ReproBridge packages a refined codebase, environment setup, experiment launchers, and a reproduction guide. On **1,029 tasks** from **387 experiments across 36 papers**, refined artifacts raised mean reproduction success from **76.4% to 92.6%** and reduced downstream API spending by **19.7%** on average. For builders running coding agents on research repos, capture the working environment, launch commands, and recovered procedure so later fresh sessions can reuse them. The paper says removing any artifact component reduced success, with the largest losses from removing code and environment details. The results come from a reproduction benchmark and four evaluated models. The material does not say how much the gains transfer to other coding tasks or project types.
ReproBridge supplies benchmark evidence that packaging code, environment details, launchers, and recovered procedures improves reproduction and reduces API spending on research tasks. It makes reusable research setup a concrete harness practice, while leaving transfer to other coding work uncertain.