Open Models Almanac
Methodspost-trainingpolicy distillation

specific method · filed under post-training

Multi-Prefix Multi-Teacher On-Policy Distillation

MOPD2 is an on-policy distillation method that combines capabilities from domain-specific teachers, including tasks that are hard to verify.1 A source trajectory with k assistant turns yields k complete history prefixes, each initializing a separate student-generated turn for token-level distillation by the relevant domain teacher.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. use Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2) to combine capabilities from teachers trained for different tasks, including tasks that are hard to verify — MiMo V2 6 technical report
  2. A source trajectorywith 𝑘assistant-turndecisionpointsyields 𝑘completehistoryprefixes ℎ𝑖,eachofwhich can initialize a separate student-generated turn 𝑦𝑖 for token-level distillation by the relevant domain teacher — MiMo V2 6 technical report
sources
3
models
2
lab adopt it
1
strongest
used

Documented in

Evidence

3 quotes from 3 sources, one card per model, each opening with its strongest.

How sources treat it
used 3

MiMo-V2.6-Pro

Xiaomiadopted2 sources

After mixed RL, we use Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2) to combine capabilities from teachers trained for different tasks

usedunclearin MiMo-V2.6Xiaomi
1 more from this model

After mixed RL, MOPD2 combines autonomous student rollouts with prefix-conditioned single-turn rollouts (Teacher-Prefix and SFT-Prefix)

usedpost trainingin MiMo-V2.6-Pro-RLXiaomi

MiMo-V2.6-Flash

Xiaomiadopted2 sources

After mixed RL, we use Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2) to combine capabilities from teachers trained for different tasks

usedunclearin MiMo-V2.6Xiaomi
1 more from this model

MOPD2 combines autonomous student rollouts with prefix-conditioned single-turn rollouts (Teacher-Prefix and SFT-Prefix)

usedpost trainingin MiMo-V2.6-Flash-RLXiaomi
↑