Diagnosing and Mitigating Tool-Call Repetition in MiMo-v2.6 (mimo.xiaomi.com)

The recent release of MiMo-V2.6 has highlighted the issue of tool-call repetition, where the AI model frequently issues identical or similar commands, leading to inefficiencies and frustrating user experiences. Internal evaluations revealed a response-level repetition rate exceeding 0.05%, with specific agent settings like OpenCode showing rates as high as 1.02%. This redundancy not only wastes computation but also stalls progress, as models fall into loops of unnecessary calls without advancing in tasks. To address this, the team implemented a Multi-teacher On-Policy Distillation (MOPD) approach, training a specialized RL teacher that drastically reduced repetition rates to zero on both training and validation datasets after just 12 training steps. This solution was efficient and cost-effective, completing the necessary adjustments at approximately $90,000—just 4% of the anticipated MixRL costs. The updated models, available on Hugging Face and through the API, demonstrate the importance of rigorous diagnostic techniques to tackle issues that can severely impact AI real-world applications, reinforcing the need for ongoing improvements in reinforcement learning methodologies.