reinforcement-training

ProgramGym introduces a framework with Program-Env and Program-Agent for building scalable, verifiable training environments for long-horizon program reproduction agents, achieving notable performance gains when fine-tuning Qwen3.8-27B on the generated environments.