ScienceIDE Turns the World's Scientific Codebases into Agent-Learnable Environments
Hejia Geng · hf · 2026-09-17
ScienceIDE is infrastructure that converts the world's scientific code repositories into programmable, executable environments for scientific agents. The authors identify a "scientific experience bottleneck": decades of knowledge encoded in code is hard to turn into reliable learning experience due to fragmented toolchains, implicit conventions, and specialized correctness criteria.
Key points:
- Guided by expert-defined cases and acceptance criteria, agents transform repositories into executable environments supporting task generation, execution, and scientific verification;
- These environments underpin supervised fine-tuning, RL, and evaluation;
- Training on verified interaction trajectories yields the PhAI-IDE-72B/9B/4B family.
The models show gains on held-out scientific-code repair and selected general benchmarks in code, reasoning, and knowledge, evidencing positive transfer from scientific experience. Code is open-sourced.
More from coding & agent
- Diorama gives OpenAI Codex coding agents a visual office you can watch work in real time — davidfromkansas · 2026-09-17
- Code-first, UI on top: building bespoke brand design tools with AI — floguo · 2026-09-17
- Study of 7 models across Claude Code, Codex, Pi: harness barely affects success but swings cost — DavideCrapis · 2026-09-17
- AI trading bot built with Jev is down 85%, owner shrugs it off — generativist · 2026-09-17
- Redditor's 3-Day SoL-Pi Test: Memory Objects Save ~12k Tokens Per Tool Run — Garblyx · 2026-09-17
- Reviewing AI code through Steve Jobs' lens: unseen internals deserve beauty too — sergeykarayev · 2026-09-17