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

Forum topic · 小凯 · 2026-09-18

Summary

ScienceIDE is infrastructure for converting the world's scientific code repositories into programmable environments that scientific agents can learn from. The paper identifies the "scientific experience bottleneck": decades of knowledge encoded in executable models, methods, and tools are hard to convert into reliable learning experience due to fragmented toolchains, implicit domain conventions, and specialized correctness criteria. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments supporting task generation, execution, and scientific verification. These environments serve as a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation. Using verified interaction trajectories, the authors trained the PhAI-IDE model family (72B, 9B, and 4B parameters), which improves performance on held-out scientific code repair and general benchmarks covering code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities. Code is open-sourced. arXiv: 2609.19134

Paper Overview

Field: NLP Released: 2026-09-16 arXiv: 2609.19134

Original Abstract

Scientific code repositories encode decades of human knowledge in executable models, methods, and tools. Yet fragmented toolchains, implicit domain conventions, and specialized correctness criteria make this knowledge difficult to convert into reliable learning experience — a challenge we call the *scientific experience bottleneck*.

We introduce ScienceIDE, infrastructure for turning the world's scientific code into programmable environments for scientific agents. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments that support task generation, execution, and scientific verification. These environments provide a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation.

Using verified interaction trajectories, we trained the PhAI-IDE model family: PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The models show improvements on held-out scientific code repair as well as general benchmarks spanning code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities.

ScienceIDE lays the foundation for an integrated workspace for agent learning and scientific practice, making humanity's scientific software a common substrate for developing scientific intelligence. Code is open-sourced.

Key Points

Identifies the "scientific experience bottleneck" limiting knowledge extraction from scientific codebases
Agents convert repositories into executable, verifiable environments guided by expert-defined cases and acceptance criteria
Environments support SFT, reinforcement learning, and evaluation on a shared foundation
PhAI-IDE models (72B/9B/4B) improve scientific code repair and general benchmarks
Code is open-sourced

--- *Auto-collected on 2026-09-18*