Researchers from PhAI Labs published ScienceIDE on Hugging Face Papers on Sep 16, with Ling Yang submitting it on Sep 17, as infrastructure for turning scientific code repositories into programmable environments for scientific agents. The paper, listed as #2 Paper of the day with 68 upvotes, says scientific code repositories contain decades of human knowledge in executable models, methods, and tools, but fragmented toolchains, implicit domain conventions, and specialized correctness criteria make that knowledge hard to convert into reliable learning experience.

ScienceIDE defines that barrier as the “scientific experience bottleneck” and uses expert-defined scientific cases and acceptance criteria to transform repositories into executable environments. The environments support task generation, execution, and scientific verification, giving agents a shared foundation for supervised fine-tuning (training on labeled examples), reinforcement learning (learning from rewards), and evaluation.

The authors used verified interaction trajectories to train three models: PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The paper says the model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, which the authors describe as evidence of positive transfer from scientific experience to broader capabilities.

ScienceIDE is presented as a step toward an integrated workspace for agent learning and scientific practice, where scientific software becomes a shared substrate for developing scientific intelligence. The Hugging Face page links code at https://github.com/aitofound/ScienceIDE and lists a model collection at https://huggingface.co/collections/AItonomy/scienceide-model-series; it also shows 0 models, 0 datasets, and 0 Spaces citing the paper at the time of the page snapshot.

Researchers from PhAI Labs published ScienceIDE on Hugging Face Papers on Sep 16, with Ling Yang submitting it on Sep 17, as infrastructure for turning scientific code repositories into programmable environments for scientific agents. The paper, listed as #2 Paper of the day with 68 upvotes, says scientific code repositories contain decades of human knowledge in executable models, methods, and tools, but fragmented toolchains, implicit domain conventions, and specialized correctness criteria make that knowledge hard to convert into reliable learning experience.

ScienceIDE defines that barrier as the “scientific experience bottleneck” and uses expert-defined scientific cases and acceptance criteria to transform repositories into executable environments. The environments support task generation, execution, and scientific verification, giving agents a shared foundation for supervised fine-tuning (training on labeled examples), reinforcement learning (learning from rewards), and evaluation.

The authors used verified interaction trajectories to train three models: PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The paper says the model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, which the authors describe as evidence of positive transfer from scientific experience to broader capabilities.

ScienceIDE is presented as a step toward an integrated workspace for agent learning and scientific practice, where scientific software becomes a shared substrate for developing scientific intelligence. The Hugging Face page links code at https://github.com/aitofound/ScienceIDE and lists a model collection at https://huggingface.co/collections/AItonomy/scienceide-model-series; it also shows 0 models, 0 datasets, and 0 Spaces citing the paper at the time of the page snapshot.