ProgramGym Builds 1,648 Verifiable Environments for Long-Horizon Coding Agents
ProgramGym · hf · 2026-10-12
Researchers from Xiamen University and others introduce ProgramGym, a framework for scalable training environments for coding agents that build programs from scratch.
- Program-Env: constructs verifiable environments from repositories via a feedback-driven loop repairing sandbox builds and black-box test quality, yielding 1,648 environments across C, C++, Go, and Rust
- Program-Agent: role-isolated analysis/implementation/review workflow with revision management for compilation checks and failure recovery
Fine-tuning on trajectories from these environments lifts Qwen3.8-27B from 34.87% to 42.84% on ProgramBench; Program-Agent further raises base and SFT models to 55.71% and 56.25%, exceeding some flagship models. The framework will be open-sourced.
More from coding & agent
- Editing video purely with chat and cloud agents: a full workflow breakdown — mattyp · 2026-10-12
- What if every query had its own harness? — amankhan · 2026-10-12
- MCP Server for Images: Compress, Convert, and Optimize a Whole Site's Images Safely — Able-Place · 2026-10-12
- Frontier models now handle day-long tasks; the bottleneck is your 'bandwidth', says new long-form analysis — dotey · 2026-10-12
- Developer has AI agents build profiling tools and benchmarks to find system bottlenecks — tristanbob · 2026-10-12
- Unlinked: open-source Rust project clones most of Simulink, with MATLAB-to-Rust compiler — rickasaurus · 2026-10-12