ScienceClaw formalizes scientific automation as a fixed‑parameter program self‑evolution problem, unifying task solving, scientific verification, and program updates. The accompanying benchmark, ScienceClaw‑Eval, spans 23 disciplines and uses both sequential streams and independent reset evaluations to measure scientific correctness, evolutionary gain, retention, cross‑dataset transfer, and evolution cost.
The framework repairs executable workflows via multi‑turn interaction, converting re‑execution‑verified failure‑success trajectories into Skill and Operator candidates. An update is retained only when the source‑task replay reproduces the repair and independent scientific tasks show improvement.
The implementation is open‑source at the GitHub repository.
Review