Xiaomi Mitigates MiMo-V2.6 Tool-Call Repetition via MOPD for $90k

Following the release of MiMo-V2.6, tool-call repetition emerged as a significant user experience issue in agentic settings like MiMo Desktop and MiMo Code. Internal evaluations confirmed a response-level repetition rate exceeding 0.05%, with specific harnesses like OpenCode showing rates as high as 1.02% for the Flash model. The problem manifested as the model issuing identical or highly similar tool calls repeatedly, consuming context and time without making progress.

Diagnosing the RL Reward Blind Spot

The team traced the issue to a blind spot in the reinforcement learning (RL) reward structure. While a penalty existed for tool-call flooding (more than 32 calls in a single turn), it was too permissive. Analysis of training logs showed that while the flooding penalty was rarely triggered early on, the share of examples with more than 8 tool calls per turn rose steadily from RL step 0 to step 20. This behavior was amplified during training because calls below the 32-call threshold were not penalized, eventually leading to severe flooding and repetition.

Lowering the penalty threshold to 8 calls in a separate experiment successfully suppressed flooding but required restarting 20 steps of MixRL, with an estimated cost of $2.31 million. Furthermore, this environment-level fix generalized poorly, reducing repetition rates from 13.45% to 3.83% but failing to eliminate the problem entirely.

MOPD-Based Solution

To address the issue efficiently, Xiaomi adopted a Multi-teacher On-Policy Distillation (MOPD) approach. They first trained a specialized single-turn RL teacher using internally collected repetition examples. This teacher, trained for only 12 steps on approximately 7,000 examples, reduced replay repetition to zero on both training and held-out sets. The teacher learned to assign high probability to the end-of-turn token rather than continuing to generate tool calls.

The specialized teacher was then merged into the main model via MOPD. The final training run cost approximately $90,000, which is only 4% of the estimated cost of the MixRL alternative. Post-MOPD, repetition rates for both MiMo-V2.6-Pro and MiMo-V2.6-Flash dropped substantially across various agent harnesses and context lengths, while overall benchmark performance remained stable.

Release and User Impact

Xiaomi has open-sourced the latest MOPD checkpoints on Hugging Face, with model names carrying the MOPD suffix. The updated models, mimo-v2.6-pro and mimo-v2.6-flash, became available on the API platform on September 25 at 06:00 (UTC+8). As an apology for the earlier issues, remaining quotas in the current usage window for all MiMo Desktop users will be reset.