Luo Fuli: The R&D challenges of MiMo-V2.6 have surpassed those of DeepSeek R1, which I was involved in.

Beating AI News Flash: Luo Fuli, head of Xiaomi MiMo, further explained this round of reinforcement learning after the release of MiMo-V2.6. She said she had previously worked on DeepSeek R1, and in her view, the research innovations and engineering challenges behind MiMo-V2.6 have already surpassed those of R1.

She also explained why MiMo uses both MixRL and MOPD simultaneously. MixRL mixes verifiable tasks such as code, general Agent, vision, and cybersecurity into the same round of reinforcement learning for joint training; MOPD is responsible for handling ultra-long, hard-to-verify, or relatively subjective-reward tasks, training them separately first, and then integrating these capabilities back into the main model.

Tasks such as gaming and 3D take too long to run in a single pass, and it is difficult to automatically judge right from wrong. If they were placed in the same round of RL as tasks like code, they would significantly slow down training. MiMo therefore trains such tasks separately, and then uses MOPD to merge the learned capabilities back into the main model.