ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents
ProgramGym turns real repositories into verifiable training environments for agents that must reproduce whole programs.
The paper says its pipeline repairs sandbox builds and black-box tests, producing 1,648 environments across C, C++, Go, and Rust. Its Program-Agent separates analysis, implementation, and review, with revision management for compile checks and recovery. On ProgramBench, fine-tuning Qwen3.8-27B with these environments raised performance from 34.87% to 42.84%, and Program-Agent pushed base and SFT models to 55.71% and 56.25%. The authors say ProgramGym will be released for further coding-agent research. HF Daily Papers' note
The paper says its pipeline repairs sandbox builds and black-box tests, producing 1,648 environments across C, C++, Go, and Rust. Its Program-Agent separates analysis, implementation, and review, with revision management for compile checks and recovery. On ProgramBench, fine-tuning Qwen3.8-27B with these environments raised performance from 34.87% to 42.84%, and Program-Agent pushed base and SFT models to 55.71% and 56.25%. The authors say ProgramGym will be released for further coding-agent research. HF Daily Papers' note
score 5