Researchers from Xiaomi, working with Peking University, the University of Hong Kong and Renmin University of China, posted a paper on CodeMidas on Sept. 18, 2026. It turns existing code into practice tasks for AI coding agents, AI that carries out programming tasks on its own. The agents train on those tasks with reinforcement learning, learning by trial and error, with a reward for getting it right. Training of this kind often stalls for lack of fresh tasks that can actually be run and checked, and CodeMidas builds them from code alone, with no bug reports, proposed code changes or edit histories needed.
CodeMidas is not a coding chatbot, and it does not collect past bug reports from GitHub. It is a factory for practice tasks. AI agents inspect a feature a codebase already has, remove the core code behind it, write a description of what that code should do, and build tests by running the original version. A task is kept only if it fails while the code is missing and passes once the original is restored, checked in sealed-off test environments.
The finished set has 5,545 tasks drawn from 3,185 open-source codebases. They cover 23 programming languages and 15 technical fields. Before a task is kept, it must fail twice from its starting state and pass four times with the original code, and a final filter after practice runs screens for leaked answers, agreement between the checking programs and a mix of passes and failures.
The authors report that training MiMo-V2.5, Xiaomi's AI model, on those tasks improved its results on all five outside coding tests they ran. On one, DeepSWE, its pass rate more than doubled, from 10.0% to 21.7%. The comparison is the same model before and after this training, under the same test settings.
The authors also report that quality beat quantity: 3,000 carefully filtered tasks trained a better model than a sample of about 8,000 unfiltered ones on three tests, while the full set of 5,545 did best later in training. In their analysis of the agent's working steps, it explored the code more before editing and checked its work in more varied ways afterward, and runs in which it wrote and ran its own checks tended to pass more often.
Xiaomi's MiMo web page on this research says little about the method, and the team has not said whether it will release the tasks. The gains come from the authors' own tests, and nobody outside the team has repeated them yet.