Luo Fuli says the R&D challenges of MiMo-V2.6 exceed those of DeepSeek R1
BIBIBI
AT A GLANCE
Luo Fuli said that the research and development challenges of MiMo-V2.6 have exceeded those of DeepSeek R1, which she previously participated in.
Article
Luo Fuli: The R&D challenges of MiMo-V2.6 have exceeded those of DeepSeek R1, which I participated in
Beating AI news flash: After the release of MiMo-V2.6, Luo Fuli, head of Xiaomi MiMo, further explained this round of reinforcement learning. She said that she had previously participated in DeepSeek R1, and in her view, the research innovations and engineering challenges behind MiMo-V2.6 have already exceeded those of R1.
She also explained why MiMo uses MixRL and MOPD simultaneously. MixRL mixes code, general-purpose Agent, vision, cybersecurity, and other verifiable tasks together for training in the same round of reinforcement learning; MOPD handles extremely long, difficult-to-verify, or more subjective-reward tasks, training them separately first and then integrating these capabilities back into the main model.
Tasks such as games and 3D have excessively long runtimes and are difficult to evaluate automatically. If they are placed in the same RL round as code and other tasks, they will significantly slow training. MiMo therefore trains these tasks separately and then uses MOPD to merge the learned capabilities back into the main model.
Original link https://m.theblockbeats.info/flash/368367
Key points
01
Luo Fuli previously participated in DeepSeek R1.
02
Luo Fuli says that, in her view, the research innovations and engineering challenges behind MiMo-V2.6 have already exceeded those of DeepSeek R1.
03
MiMo uses MixRL and MOPD simultaneously for reinforcement learning.
04
MixRL trains code, general-purpose Agent, vision, cybersecurity, and other verifiable tasks together in the same round of reinforcement learning.
05
MOPD handles tasks that are excessively long, difficult to verify, or have more subjective rewards by training them separately first, then integrating the capabilities back into the base model.
06
Games and 3D tasks take too long to run once, and it is difficult to automatically determine whether they are correct.
07
MiMo trains these types of tasks separately, then uses MOPD to integrate the learned capabilities back into the base model.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Luo Fuli explained that MiMo-V2.6 processes different types of tasks separately: tasks that are easier to verify are trained together, while tasks that take a long time to run or are difficult to judge for correctness are trained separately first and then merged back into the main model.
Why it matters to readers
This indicates that MiMo-V2.6's reinforcement learning process arranges training methods based on whether tasks are easy to verify and how time-consuming they are to run.
You can think of it this way: easy-to-grade assignments are practiced together, while assignments that are difficult or time-consuming to grade are practiced separately, and the capabilities are combined at the end.
Risks and unknowns
The claim that the research and development challenge “exceeded DeepSeek R1” is Luo Fuli’s personal judgment; there is no independent verification in the evidence.
The evidence does not specify the particular metrics or comparison methodology for MiMo-V2.6 and DeepSeek R1.
The evidence does not further explain the technical implementation details of MixRL and MOPD.
Related Developments
Loading event timeline…
Related concepts
Reinforcement learning
This term is not in the glossary yet. Browse related concepts in the glossary.
