CodeMidas: Scaling Agentic Coding RL Environments from Code Itself
CodeMidas turns existing codebases into 5,545 executable RL coding tasks; GRPO-trained MiMo-V2.5 gains up to 17% on ProgramBench.
CodeMidas is an agentic pipeline that converts implemented functionality in existing codebases into executable RL environments using only source code as task-specific input, with agents formulating behavioral specifications, constructing execution-grounded tests, and filtering candidates via rollouts. The resulting dataset contains 5,545 training tasks from 3,185 open-source codebases spanning 23 programming languages and 15 technical domains. Training MiMo-V2.5 with GRPO improves all five benchmarks, including DeepSWE +11.7%, ProgramBench +17%, and Terminal-Bench v2.1 +8.5%, with trajectory analysis showing more codebase exploration and self-verification.
- 5,545 RL tasks from 3,185 codebases, 23 languages, 15 domains
- Uses only source code, avoiding reliance on issues and commits
- MiMo-V2.5 gains up to +17% (ProgramBench) after GRPO training
- Trained agents explore codebases more and self-verify more diversely
From the original article
1 min read at huggingface.co
Training capable coding agents via reinforcement learning (RL) requires diverse tasks with reliable verifiers. Open-source codebases offer a rich source of such tasks, while existing methods typically rely on development artifacts such as issues and commits, limiting the range of tasks that can be extracted.
By Bowen Ye, Lei Li, Shicheng Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.