Researchers have built a new training ground for AI coding agents that forces them to study a program before rebuilding it from scratch.
The system, called ProgramGym, pairs a feedback-driven environment builder with a role-isolated agent workflow covering analysis, implementation, and review. The environment builder, Program-Env, converts existing repositories into 1,648 verifiable training setups across C, C++, Go, and Rust, repairing broken sandbox builds and weak tests along the way. The agent side, Program-Agent, adds revision management so agents can catch compile failures and fix them before moving on. On the team's ProgramBench evaluation, fine-tuning a Qwen3 model on trajectories from these environments pushed its score from 34.87% to 42.84%, and adding the full Program-Agent workflow lifted both the base and fine-tuned models to 55.71% and 56.25%, enough to beat some flagship models, per the paper.
Most coding-agent benchmarks test bug fixes inside an existing codebase or scaffold a project from a spec. This one asks agents to reverse-engineer intent from a working reference program, closer to what a new hire does when inheriting a system with no documentation. Scaling that to 1,648 verifiable environments, rather than a handful of hand-built ones, is the part that actually matters for training.
The paper reports one number that doesn't parse: a model size listed as 'Qwen3.8-27B', which matches no published Qwen configuration. That's a reminder to read benchmark tables from preprint papers with some skepticism until the code and weights actually ship.