ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments

The authors introduce ScienceIDE, an infrastructure designed to convert scientific code repositories into programmable environments for scientific agents. Guided by expert-defined acceptance criteria, agents transform repositories into executable environments supporting task generation, execution, and scientific verification. Using verified interaction trajectories, the authors train the PhAI-IDE model family in 72B, 9B, and 4B sizes. The model family achieves performance gains in held-out scientific-code repair and across general-purpose benchmarks in code, reasoning, and knowledge, demonstrating positive transfer from scientific experience to broader capabilities.

ScienceIDE transforms scientific code repositories into programmable environments supporting task generation, execution, and verification.

The PhAI-IDE model family includes 72B, 9B, and 4B variants trained using verified interaction trajectories.

The models show performance gains in held-out scientific-code repair and across general-purpose benchmarks in code, reasoning, and knowledge.