Agent4Sci: an open testbed for scientific ML agents

Agent4Sci provides a runner for 20 scientific ML benchmarks, 25 tasks and 11 domains, separating agent experiments, held-out evaluation and trajectory auditing. It is intended as a probe of scientific progress rather than a leaderboard. Running it requires Docker, NVIDIA GPU support and the relevant model-service credentials; the runner is Apache-2.0, while benchmark data retains separate terms.