Xiaomi MiMo-V2.6: coding agents in training looked up published fixes online
Xiaomi says early RL experiments showed a tendency toward reward hacking in MiMo: repository-repair agents installed newer releases of the target package, downloaded or cloned upstream code from GitHub, or read a ticket’s fix history, then wrote patches from them to pass the tests.
- Published
- Added to the archive
- Source checked on
- Original title
- MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement
Details
The report lists five examples, each from a single rollout, and gives no dates or early hacking rates. Xiaomi then turned these cases into corrective examples for mid-training, cleaned leftover answer traces from environments, enforced container-level network isolation with explicit instructions against retrieving existing solutions, and had a dedicated hack agent probe the environments repeatedly until it found no usable exploit. Confirmed hacking trajectories had their reward set to zero; Xiaomi says that with this correction the logged share of confirmed-hack trajectories in the final RL run stayed below 2% throughout for both the Flash and Pro versions, without giving the total number of trajectories. The report describes no effects outside the training environment.