CodeMidas builds 5,545 verified coding RL tasks from source code alone, no issues or commits needed

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

Training coding agents with RL needs diverse tasks with reliable verifiers, and existing pipelines lean on development artifacts like issues and commits, which caps the range of extractable tasks. CodeMidas instead takes source code as its only task-specific input and spends agentic compute at every construction stage: agents explore implemented functionality to write behavioral specifications, build tests grounded in executing the original code, then filter candidates through execution checks and repeated solution rollouts. The result is 5,545 training tasks drawn from over 3,100 repositories, and it was the second most upvoted paper on HuggingFace's September 21 daily list at 103 upvotes.