MiMo-V2.6 fixes repeated tool calls for about $90,000, 4% of the cost of restarting its RL run
Synopsis
After the MiMo-V2.6 release, repeated identical tool calls in agentic settings were traced to a reward blind spot in reinforcement learning: correctness was rewarded, but excessive calling below the penalty threshold was not. The team first trained a single-turn RL teacher that learned when to stop, then merged that behavior into the main model through multi-teacher on-policy distillation. Repetition rates for MiMo-V2.6-Pro and MiMo-V2.6-Flash fell substantially across context lengths and agent harnesses, and the full training run cost about $90,000, 4% of the estimated $2.31 million for restarting the MixRL alternative.
Interpretation
In agentic settings such as MiMo Desktop, MiMo Code and OpenCode, MiMo-V2.6 sometimes issued the same or highly similar tool calls repeatedly, consuming time and context without making progress; internal evaluation put the response-level repetition rate above 0.05%. The existing reinforcement-learning pipeline penalized only turns with more than 32 tool calls, leaving excessive calling below that threshold unconstrained. On a fixed replay set, the share of MiMo-V2.6-Flash-RL examples with more than ten tool calls in a single turn rose from 11.1% at RL step 0 to 24.6% at step 20, and under MiMo Code from 30.6% to 41.7%.
The fix is a distillation route: a teacher model is trained to end its turn at the right point, and that stopping behavior is merged into the main model. Lowering the penalty threshold directly suppressed flooding, but the effect appeared only after about 20 training steps and cut the repetition rate on the full internal evaluation set only from 13.45% to 3.83%. Trained for 12 steps on roughly 7,000 internal repetition examples, the teacher reduced replay repetition to zero on both training and held-out sets; the full run that merged it into the main model cost about $90,000.
Perspective
The published repetition metric counts only calls within one turn whose tool name and canonicalized arguments are identical, so it is a lower bound on observable repetition and excludes cross-turn repetition after tool feedback, near-duplicates with slightly different arguments, and calls embedded in scripts under code-mode harnesses. The results come from before-and-after comparisons of MiMo-V2.6-Pro and MiMo-V2.6-Flash on internal evaluation sets and the listed agent harnesses.
Exact within-turn repetition is only a lower bound, and the authors state it is neither a complete repetition rate nor a measure of tool-call flooding. The lower-threshold experiment ran on a small isolated data source, its effect appeared only after about 20 steps, and it did not eliminate the problem on the full internal evaluation set. The stopping-probability analysis rests on one sampled trajectory.