specific method · filed under post-training
Prefix-Conditioned On-Policy Distillation
Prefix-Conditioned On-Policy Distillation is a policy-distillation method that reuses trajectories from teacher rollouts or SFT demonstrations as prefix contexts.1,1 Each assistant-turn prefix initializes a separate student-generated turn that a domain teacher supervises token-level, rather than the student regenerating full rollouts.2
Trade-off Using SFT-demonstration prefixes limits student deviations in long-horizon tasks while the student still generates its own continuation instead of imitating a fixed response.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
Prefix-Conditioned OPD reuses trajectories from teacher rollouts (Teacher-Prefix OPD) or SFT data (SFT-Prefix OPD
— MiMo V2 6 technical reportA 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 reportSFT-PrefixOPD thereforestartseachrolloutfromafixeddemonstrationprefix,limitingdeviationsbeforethesam- pled turn. Demonstrations supply the context, while the student generates its own continuation rather than imitating a fixed response
— MiMo V2 6 technical report
Also called Prefix-Conditioned OPD, SFT-Prefix OPD, SFT-Prefix, Teacher-Prefix, Teacher-Prefix OPD.