Xiaomi's MiMo-V2.6 paper describes agent-run reinforcement learning for coding models

Xiaomi's MiMo-V2.6 paper says its reinforcement learning loop for coding agents now lets the agents build tasks, audit tests, grade answers and detect cheating, with humans setting the compute budget and rules.

According to the paper, a grader agent that rewards cleaner patches over reward-hacking fixes stopped drift toward longer runs and workarounds such as swallowed exceptions.

The paper reports that MiMo-V2.6-Pro's DeepSWE score rose from 58.4 to 72.6 over $2.6M of RL compute, and that the score was still climbing when training stopped.