Xiaomi's MiMo-V2.6 diagnoses and fixes tool-call duplication, cutting fix cost to 4% of MixRL approach with MOPD
Original title:Diagnosing and Mitigating Tool-Call Repetition in MiMo-V2.6
AI summary
Xiaomi recounted the tool-call duplication issue after MiMo-V2.6's release, with a response-level duplication rate above 0.05%, distinguishing normal parallel calls, call flooding and call duplication. Replaying RL checkpoints showed the flooding rate rose from 11.1% to 24.6% during training, and the 32-call threshold penalty was too lenient; directly lowering the threshold would require restarting 20 steps of MixRL at an estimated cost of $2.31 million, and internal testing reduced the duplication rate only from 13.45% to 3.83%.
Key Takeaways
- Xiaomi recounted the tool-call duplication issue after MiMo-V2.
- 6's release,with a response-level duplication rate above 0.
- 05%,distinguishing normal parallel calls,call flooding and call duplication.
- Replaying RL checkpoints showed the flooding rate rose from 11.
- 6% during training,and the 32-call threshold penalty was too lenient
Quick Facts
- Published
- Updated
- Source
- 小米 MiMo:官网发布与博客
- Category
- Papers & Research
- AI score
- 63
原文给出完整的归因实验和修复路径,还公开了成本对比,读者可借鉴其诊断 RL 训练中涌现行为的方法。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Xiaomi's MiMo-V2.6 diagnoses and fixes tool-call duplication, cutting fix cost to 4% of MixRL approach with MOPD. September 27, 2026. https://new.nat6.net/en/news/cmuk25h00156vro9h173e1zjz/