ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents

ProgramGym. Scaling Training Environments for Long-Horizon Program Reproduction Agents. By Chengyi Yang and twelve others, from Xiamen University, Washington University in Saint Louis, and two other institutions. Published on the ninth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Coding agents are more and more often asked to build programs from scratch, the way a software engineer would. That is a different job from resolving an issue inside an existing repository, or generating a repository from a given requirements document. Here the agent has to discover what is required and decide how the system should be designed. Two things hold this back. Training environments that force that kind of discovery, at scale and with verification, remain scarce. And existing agent workflows cannot sustain the long-horizon development involved, in which the agent first explores a reference program and only then implements.

ProgramGym answers both with two parts. Program-Env builds scalable training environments for long-horizon program reproduction agents. Program-Agent is a workflow designed to carry that long-horizon work through.

Program-Env constructs verifiable training environments from repositories, using a feedback-driven loop. The loop repairs two things: the runtime sandbox builds, and the quality of the black-box tests. With this pipeline the authors obtain one thousand six hundred and forty-eight training environments, spanning C, C++, Go and Rust. Each environment is verifiable, which is what makes it usable for training: an agent's reproduction of a program can be checked against black-box tests rather than judged by eye. And because the tests treat the program as a black box, they reward matching its behaviour, whatever internal design the agent chooses.

Program-Agent combines a role-isolated workflow, in which analysis, implementation and review are kept separate, with revision management that handles compilation checks and recovery from failure. The separation of roles means that the step which reads and understands the reference program is distinct from the step which writes code, and both are distinct from the step which reviews it. Revision management then keeps a long build recoverable when a compile breaks or an attempt fails.

Experiments on ProgramBench show that each part improves program reproduction on its own. Fine-tuning on trajectories gathered in the constructed environments raises a twenty-seven billion parameter Qwen model from thirty-four point eight seven percent to forty-two point eight four percent.

Program-Agent then adds more on top. With it, the base model reaches fifty-five point seven one percent, and the fine-tuned model reaches fifty-six point two five percent. Both of those figures exceed some flagship models.

The two halves of ProgramGym are complementary. One supplies the practice, environments where requirements have to be found by exploring a program rather than read off a document. The other supplies the discipline to keep a long build on track. The authors say they will release ProgramGym to support further research on agents that build programs from scratch.