Program-Env
The component of the ProgramGym framework responsible for training environment construction. It automatically builds verifiable training environments from real code repositories, employing a feedback-driven loop mechanism to ensure sandbox execution quality and the validity of black-box tests.
Timeline (last 90 days)
The ProgramGym framework consists of two core components: Program-Env, which is responsible for building the training environment, and Program-Agent, which is responsible for executing development tasks.
Program-Env can automatically construct verifiable training environments from real code repositories.
Program-Env adopts a feedback-driven loop mechanism to address two engineering challenges: fixing runtime sandbox construction issues and improving black-box test quality.
After fine-tuning using trajectories generated from building environments with Program-Env, the performance of the Qwen3.8-27B model improved from 34.87% to 42.84%.
All Facts (4)
The ProgramGym framework consists of two core components: Program-Env, which is responsible for building the training environment, and Program-Agent, which is responsible for executing development tasks.
50%UnverifiedProgram-Env can automatically construct verifiable training environments from real code repositories.
50%UnverifiedProgram-Env adopts a feedback-driven loop mechanism to address two engineering challenges: fixing runtime sandbox construction issues and improving black-box test quality.
50%UnverifiedAfter fine-tuning using trajectories generated from building environments with Program-Env, the performance of the Qwen3.8-27B model improved from 34.87% to 42.84%.
50%