CodeMidas: Scaling Agentic Coding RL Environments from Code Itself
CodeMidas derives verified coding-agent RL tasks directly from working source, widening training data beyond issues and commits while improving repair, construction, and terminal benchmarks.
CodeMidas turns existing implementations into executable RL environments using source code as the only task-specific input. Its agents derive specifications, build execution-grounded tests, and filter tasks through checks and repeated solution attempts, producing **5,545 tasks from 3,185 repositories**.
For coding-agent training, treat working code as a source of behavioral tasks even when issue and commit histories are sparse. MiMo-V2.5 trained with GRPO improved across all five tested benchmarks, including **+17% on ProgramBench** and **+11.7% on DeepSWE**.
CodeMidas turns existing implementations into executable RL environments using source code as the only task-specific input. Its agents derive specifications, build execution-grounded tests, and filter tasks through checks and repeated solution attempts, producing **5,545 tasks from 3,185 repositories**. For coding-agent training, treat working code as a source of behavioral tasks even when issue and commit histories are sparse. MiMo-V2.5 trained with GRPO improved across all five tested benchmarks, including **+17% on ProgramBench** and **+11.7% on DeepSWE**. The abstract reports that more high-quality tasks helped and that trained agents explored and self-verified more, but it does not establish how well the pipeline handles underspecified, unsafe, or hard-to-execute repositories.
CodeMidas broadens executable task generation beyond repositories with usable issue or PR histories by deriving behavior from implementations themselves, and supplies training gains rather than only a data-pipeline proposal. It complements history-derived environments, while leaving task validity, unsafe repositories, execution feasibility, and supervision quality as unresolved filters rather than solved properties.