CodeMidas is a pipeline that turns implemented functionality in open‑source repositories into executable RL environments using only the source code as task input. At every stage of environment construction the pipeline allocates agentic compute: (1) agents explore the codebase to extract behavioral specifications; (2) tests are built from the execution results of the original code; (3) candidate tasks are filtered through execution checks and repeated solution rollouts. The process yields 5,545 training tasks drawn from 3,185 repositories, covering 23 programming languages and 15 technical domains. Training MiMo‑V2.5 with GRPO on this dataset improves all five benchmarks: issue repair (DeepSWE) +11.7%, whole‑program synthesis (ProgramBench) +17%, and terminal interaction (Terminal‑Bench v2.1) +8.5%. Ablation studies show that increasing the number of high‑quality tasks boosts performance. Trajectory analysis reveals that the RL‑trained agent explores codebases more aggressively and performs more diverse self‑verification. These results demonstrate that source code can serve as a scalable foundation for building RL environments that enhance coding agents across a wide range of software tasks.
Review