AIAny
Icon for item

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

Transforms scientific code repositories into executable, agent-learnable environments that support task generation, execution, and scientific verification. Agent-guided repository transformation produces verified interaction trajectories used to train the PhAI-IDE model family. Intended for research on agent learning, scientific-code repair, and training RL/SFT models.

Introduction

Scientific software encodes decades of domain knowledge but is fragmented by bespoke toolchains, implicit conventions, and specialized correctness checks. ScienceIDE addresses this bottleneck by converting public scientific repositories into programmable, sandboxed environments that agents can use to propose tasks, execute code, and verify scientific outcomes—turning passive code artifacts into repeatable learning experiences.

Key Findings
  • Agents can autonomously transform repositories into executable environments with expert-defined acceptance criteria, enabling systematic task generation and verified execution — so what: this produces grounded, reproducible training data rather than heuristic synthetic tasks.
  • Verified interaction trajectories derived from these environments can be used to train models (PhAI-IDE-72B/9B/4B) and yield measurable gains in scientific-code repair and select general-purpose benchmarks — so what: scientific execution experience transfers to code, reasoning, and knowledge improvements.
  • The infrastructure supports supervised fine-tuning, reinforcement learning, and evaluation on a common substrate — so what: it standardizes disparate scientific projects into a shared training/evaluation bed for agent learning.
Who it's for and tradeoffs

Great fit if you research autonomous scientific agents, need executable benchmarks for scientific code, or want to convert domain repositories into training environments. Look elsewhere if your goal is a lightweight code assistant or one-off script fixes: ScienceIDE focuses on creating validated, reproducible environments, which requires curator effort (expert cases, acceptance criteria) and compute to run/verify scientific workflows.

Where it fits

ScienceIDE sits between dataset/benchmark efforts and autonomous agent toolchains: it is infrastructure for turning real scientific code into verified agent training environments rather than a single model or end-user app. It complements executable benchmarks and corpus efforts by providing environment generation and verification at repository scale.

Information

  • Websitearxiv.org
  • AuthorsHejia Geng, Zesen Huang, Haoyang Li, Wenbin Li, Koutian Wu, Zihan Zhou, Yuanbo Pang, Weihao Liu, Zigong Xu, Zhiping Li
  • Published date2026/09/16

Categories

More Items

Introduces physical rank consistency (PRC) and ActionPiece, a learned action tokenizer that preserves local physical relationships via joint supervision of representation learning and quantization, improving autoregressive vision-language-action policy performance across multiple benchmarks.

Extracts replayable behaviors from working web apps and converts them into verifiable, reference-guided software-engineering tasks to evaluate and diagnose coding agents; includes an automated pipeline that scales to thousands of tasks.

Estimates a model's probability of correctness by leveraging its past graded episodes: XConf recalls similar past cases and prompts the model to reflect and restate confidence, improving calibration and discrimination across reasoning, coding, multimodal QA, and agent tasks at much lower generation cost.