The Pulse
PhAI Labs Presents ScienceIDE as an Environment for Training Scientific AI Agents
PhAI Labs’ ScienceIDE turns scientific code repositories into executable environments for task generation, execution, verification and evaluation, supporting supervised fine-tuning, reinforcement learning and scientific-code assessment.
AI.info Team ·
PhAI Labs has introduced ScienceIDE, infrastructure for turning scientific code repositories into programmable environments for AI agents. The project is designed to convert software that encodes scientific models, methods and tools into settings where agents can generate tasks, execute code and receive evaluations based on scientific acceptance criteria.
The company describes the motivation as the “scientific experience bottleneck.” Scientific repositories often contain fragmented toolchains, implicit domain conventions and specialized standards for correctness. ScienceIDE packages those elements into executable environments intended to make scientific software usable as learning experience.
The PhAI Labs page, dated September 17, 2026, identifies the project as a technical report and links to a downloadable paper, an arXiv entry, a GitHub repository and model pages on Hugging Face. It also identifies a project page associated with ScienceIDE.
From Scientific Code to Agent Environments
ScienceIDE uses expert-defined scientific cases and acceptance criteria to guide the transformation of repositories into environments for scientific agents. Those environments support task generation, execution and scientific verification.
The design is intended to account for more than whether a program runs or produces syntactically valid output. Scientific software depends on numerical behavior, domain-specific conventions and criteria that determine whether a result is scientifically acceptable. By incorporating those requirements into the environment, the system can evaluate agent interactions against criteria defined for the relevant scientific task.
The project treats scientific code as a source of executable experience rather than simply as a collection of files. Agents can work within an environment derived from a repository, while the surrounding system records the interaction and evaluates the resulting work according to the applicable scientific checks.
A Shared Foundation for Training and Evaluation
According to PhAI Labs, ScienceIDE provides a shared foundation for supervised fine-tuning, reinforcement learning and evaluation. The same type of environment can therefore be used to produce interaction trajectories for training and to assess how models perform on scientific-code tasks.
For supervised fine-tuning, verified interactions can provide examples of how an agent works through a scientific software problem. In reinforcement learning, the environment can supply feedback based on the outcome of the interaction and the relevant acceptance criteria. Evaluation can use scientific tasks to measure whether a model can modify or work with code in ways that satisfy domain-specific requirements.
The project’s emphasis on verification is intended to distinguish scientific success from ordinary code completion. A result may need to satisfy numerical or scientific conditions in addition to compiling or passing a basic software test. The environment provides a way to make those conditions part of the agent’s training and evaluation process.
PhAI-IDE Models
Using verified interaction trajectories generated through ScienceIDE, PhAI Labs says it trained three models: PhAI-IDE-72B, PhAI-IDE-9B and PhAI-IDE-4B.
“Using verified interaction trajectories, we train PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B.”
PhAI Labs reports that the model family shows gains on held-out scientific-code repair tasks and on selected general-purpose benchmarks covering code, reasoning and knowledge. The company presents those results as evidence of positive transfer from experience with scientific software to broader capabilities.
The report’s abstract describes ScienceIDE as infrastructure for scientific agents and says the environments are intended to support task generation, execution and scientific verification. It also presents the system as a foundation for an integrated workspace for agent learning and scientific practice.
Why the Environment Matters
Scientific repositories contain knowledge that is not always written as explicit instructions. Assumptions about numerical precision, execution procedures, input preparation and the meaning of outputs may be distributed across source code, tests, documentation and established research practice.
ScienceIDE’s approach is to make those requirements usable in an executable setting. Expert-defined cases and acceptance criteria establish the conditions under which an agent’s work can be assessed, while the environment provides the machinery for generating tasks, running experiments and checking results.
PhAI Labs’ technical report positions ScienceIDE as both a training infrastructure project and an evaluation environment. Its stated aim is to make scientific software a shared substrate for developing agents capable of learning from and working with research code.