The Multi-Teacher On-Policy Distillation (MOPD) paradigm, introduced by Xiaomi's MiMo team in June 2026, addresses the challenge of integrating multiple expert capabilities into a single large language model without performance degradation. MOPD decouples the development of domain-specific expert teachers and uses an on-policy routing mechanism for the student model to receive dense, token-level supervision from relevant frozen teachers. This approach has been rapidly adopted by industry players like Nvidia for its Nemotron models and by other leading AI labs, leading to numerous variations aimed at further improving its stability and effectiveness. AI
IMPACT Accelerates multi-expert integration in LLMs, enabling parallel development and improved stability for complex capabilities.
RANK_REASON The item describes the formal introduction of a new LLM post-training paradigm (MOPD) by a major AI lab (Xiaomi MiMo team), including its technical details, paper citation, and initial adoption by other leading entities. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- arXiv:2606.30406
- Cascade RL
- DeepSeek V4
- MiMo
- MiMo-V2-Flash
- Mix RL
- NeMo Gym
- Nemotron 3.5
- Nemotron 3 Ultra
- Nvidia
- On-Policy Distillation
- Param-Merge
- Qwen3-30B-A3B
- reinforcement learning from human feedback
- Wenhan Ma
- Xiaomi
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →