ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments
The paper proposes infrastructure that turns scientific code repositories into executable training and evaluation environments for AI agents.
ScienceIDE uses expert-defined scientific cases and acceptance criteria to convert repositories into environments where agents can generate tasks, execute code, and verify scientific correctness. The authors use verified interaction trajectories from those environments to train PhAI-IDE models at 72B, 9B, and 4B sizes. They report gains on held-out scientific-code repair and selected broader benchmarks in code, reasoning, and knowledge. HF Daily Papers' note
ScienceIDE uses expert-defined scientific cases and acceptance criteria to convert repositories into environments where agents can generate tasks, execute code, and verify scientific correctness. The authors use verified interaction trajectories from those environments to train PhAI-IDE models at 72B, 9B, and 4B sizes. They report gains on held-out scientific-code repair and selected broader benchmarks in code, reasoning, and knowledge. HF Daily Papers' note
score 5