Xiaomi team publishes review and releases a fixed version of MiMo-V2.6

Article

Xiaomi admits it MiMo-V2.6 will “repeat in place”:RL bad habits only get reinforced through training, 90000 USD Fix

Beating AI Flash News, the Xiaomi MiMo team reviewed the tool-call repetition issue after MiMo-V2.6 went live. The model sometimes repeatedly calls the same or highly similar tools, continuously consuming context without making progress on the task. In OpenCode, the shares of replies with repeated tool calls for Flash and Pro once reached 1.02%and 0.54%, respectively.

The problem lay in the reward design for RL. Training mainly rewarded whether the task was ultimately completed correctly, without sufficiently penalizing inefficient behavior during the process. The original rule imposed a penalty only when tool calls in a single round exceeded 32 calls; repeated calls below 32 calls incurred no deduction at all. As the scale of RL expanded, this bad habit was instead continually reinforced. After replaying training at checkpoint Xiaomi found that, in Flash, the proportion of anomalous samples with more than 10 calls in a single round rose from 0 at step 11.1%to 20 at step 24.6%.

The most direct approach was to lower the penalty threshold from 32 calls to 8 calls and rerun approximately 20 MixRL steps, but this was expected to cost 2310000 USD. Xiaomi ultimately trained a RL teacher specifically to correct repeated calls, ran it for 12 steps with approximately 7000 samples, and then used MOPD to merge this capability back into Pro and Flash. The entire fix cost approximately 90000 USD, only about 4%of the full retraining approach, while the other main benchmark remained essentially unchanged.

The fixed MiMo-V2.6-Pro-MOPD and Flash-MOPD weights are now available, and API has also switched to the new version; the call names remain unchanged. Xiaomi will also reset the remaining quota for MiMo Desktop users in the current cycle.

Original link https://m.theblockbeats.info/flash/369327

Key points

The Xiaomi MiMo team reviewed the tool-call repetition issue after MiMo-V2.6 went live.
The model sometimes repeatedly calls the same or highly similar tools, continuously consuming context without making progress on the task.
In OpenCode, the shares of replies with repeated tool calls for Flash and Pro once reached 1.02%and 0.54%。
The original rule imposed a penalty only when tool calls in a single round exceeded 32 calls; 32 repeated calls below this threshold incurred no deduction.
training checkpoint Replay showed that, in Flash, the proportion of anomalous samples with more than 10 calls in a single round rose from step 0 at 11.1%to step 20. 24.6%。
A full rerun of approximately 20 would MixRL steps is expected to cost 2310000 USD; the targeted approach ran for 12 steps using approximately 7000 samples, and the entire fix cost approximately 90000 USD。
The fixed MiMo-V2.6-Pro-MOPD and Flash-MOPD weights are now available, and API has switched to the new version; the call names remain unchanged.
Xiaomi will also reset the remaining quota for MiMo Desktop users in the current cycle.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This issue is similar to the model learning that “as long as the task is ultimately completed correctly, that is enough,” without receiving sufficient penalties for repeatedly doing the same thing along the way. The original training rules imposed no deduction for repeated tool calls of up to 32 calls, so as the scale of RL expanded, this inefficient behavior was further reinforced. Reports say Xiaomi did not conduct the more expensive full retraining, but instead trained a RL teacher specifically to correct repeated calls, then used MOPD to merge the capability into Pro and Flash.

Why it matters to readers

Repeated tool calls waste context and call resources and can also stall tasks. If the targeted fix is effective, it could improve MiMo-V2.6 tool-use efficiency while avoiding the higher cost of full retraining.

New users of API or the newly available weights may encounter fewer instances of the model repeatedly calling the same tool; however, the material provides no long-term results from real-world use of the new version.

Risks and unknowns

The material is a media news brief and does not provide the Xiaomi MiMo team’s original official review; the relevant figures and conclusions still lack independent verification.
It remains unclear what the actual repeated-call rate is after the new version went live.
“The other main benchmark remained essentially unchanged” is not accompanied by specific test items, results, or evaluation methods.

Xiaomi team publishes review and releases a fixed version of MiMo-V2.6

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center