CodeMidas: Scaling Agentic Coding RL Environments from Code Itself
CodeMidas turns existing source code directly into executable RL training environments for coding agents.
The paper says the pipeline uses code itself as the task-specific input, rather than relying on issues or commits. It generated 5,545 training tasks from 3,185 open-source codebases across 23 languages and 15 domains. Training MiMo-V2.5 on those tasks improved results on all five tested benchmarks, including DeepSWE, ProgramBench, and Terminal-Bench v2.1. The authors also report that more high-quality tasks helped performance and that the trained agent explored codebases and self-verified more effectively. HF Daily Papers' note
The paper says the pipeline uses code itself as the task-specific input, rather than relying on issues or commits. It generated 5,545 training tasks from 3,185 open-source codebases across 23 languages and 15 domains. Training MiMo-V2.5 on those tasks improved results on all five tested benchmarks, including DeepSWE, ProgramBench, and Terminal-Bench v2.1. The authors also report that more high-quality tasks helped performance and that the trained agent explored codebases and self-verified more effectively. HF Daily Papers' note
score 5