Researchers introduced CodeMidas, a pipeline turning implemented code into RL environments. It uses agents to explore code, formulate specs, build tests and validate tasks. The dataset covers 5,545 tasks across 3,185 codebases in 23 languages. Training MiMo-V2.5 on it improved performance on five benchmarks. Issue repair scores rose 11.7% on DeepSWE. Whole-program construction scores rose 17% on ProgramBench. Terminal work scores rose 8.5% on Terminal-Bench v2.1. Ablations show more high-quality tasks improve results. Trajectory analysis shows better codebase exploration and self-verification.
What changed
Existing methods relied on development artifacts like issues and commits to extract tasks.
What it unlocks
Creating executable RL environments directly from source code without additional artifacts.
- 5,545 training tasks
- 3,185 open-source codebases
- 23 programming languages
- 15 technical domains
Sources