CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

An agentic pipeline from Xiaomi's LLM Core team that turns functionality already implemented in open-source codebases into executable RL environments for coding agents, with the source code as its only task-specific input. Earlier pipelines seed tasks from development records, such as issues and pull requests (SWE-rebench V2, daVinci-Env), commits (R2E-Gym), existing tests (SWE-smith) or documentation, so their coverage ends where those records end. In CodeMidas, agents handle every construction stage. One selects functionality with a public entry point and observable behavior, removes the implementation, and writes a behavioral task statement. Another writes tests whose expected values come from running the original code, then replaces assertions that pin incidental details (exact wording, ordering, internal structure) with behavioral checks. A task must fail in two fresh containers with the stripped codebase and pass in four with the reference solution. Three post-rollout filters then drop tasks with exploitable leakage (probed by adversarial rollouts), tasks whose tests a reviewing agent finds at odds with the statement, and tasks that a frontier model always or never solves.

The result is 5,545 training tasks from 3,185 codebases in 23 programming languages and 15 technical domains (Python 21.4%, TypeScript 18.3%, Go 16.2%; median reference solution 142 lines). Training MiMo-V2.5 on them with GRPO (binary execution reward, 32 rollouts per task, up to 500 turns per rollout) improves all five external benchmarks, among them DeepSWE v1.1 (10.0% to 21.7%), ProgramBench's almost-solved rate (4.5 to 21.5) and Terminal-Bench v2.1 (63.7% to 72.2%). Task count and task quality both matter: pools of 1k, 3k and 5.5k tasks reach 17.57, 19.05 and 21.70 on DeepSWE, and the filtered 5.5k set beats an 8k sample built without cleaning or filtering by 4.59 points there (even the 3k subset beats it). During RL the agent reads more before its first edit (27.2 to 40.1 read/search calls) and runs more varied checks after its last. The MiMo-V2.6 technical report names CodeMidas as the source-code-driven path among its coding RL task pipelines. Sixteen of the 19 authors list LLM Core, Xiaomi, several jointly with Peking University, HKU or Renmin University; first author Bowen Ye did the work as a Xiaomi intern, and the co-corresponding authors are Tong Yang (Peking University) and Luo Fuli (Xiaomi). The paper links no code or data.