ScienceIDE packages scientific repositories as versioned environments with expert-defined cases, executable tasks and private scientific verifiers. The same environments support supervised fine-tuning, reinforcement learning and evaluation. Verified trajectories train the PhAI-IDE family at 4B, 9B and 72B parameters. The authors report gains on held-out scientific-code repair and selected general coding, reasoning and knowledge benchmarks, and release code plus model artifacts.
ScienceIDE turns scientific repositories into training environments for agents
2026-09-17Frontier watchTRIAL
ScienceIDE packages scientific repositories as versioned environments with expert-defined cases, executable tasks and private scientific verifiers. The same environments support supervised fine-tuning, reinforcement learning and evaluation. Verified trajectories train the PhAI-IDE family at 4B, 9B and 72B parameters. The authors report gains on held-out scientific-code repair and selected general coding, reasoning and knowledge benchmarks, and release code plus model artifacts.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For teams building scientific agents, the hard problem is often the environment and verifier rather than the base model. ScienceIDE provides a reusable pattern for converting mature scientific software into executable learning experience. Trial it on one well-tested repository before treating aggregate benchmark gains as evidence of cross-domain scientific ability.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Environment Contract | versioned repository, runtime, expert cases, acceptance criteria and private verifier |
|---|---|
| Training Modes | supervised fine-tuning and reinforcement learning from verified interaction trajectories |
| Models | PhAI-IDE-4B, PhAI-IDE-9B and PhAI-IDE-72B |
| Evaluation | held-out scientific-code repair plus selected public code, reasoning and knowledge benchmarks |
| Artifacts | public code repository and model links |
Technical impact 4
Engineering value 4
Adoption signal 2
Maturity 2
Verification cost 4
Risk 3