MoPD
A technique for merging and injecting the capabilities of lightweight specialist models into a main model. It transfers a teacher model's behavioral preferences to the main model without full retraining, substantially reducing the training cost associated with correcting model behavior.
Timeline (last 90 days)
Xiaomi MiMo's fix involved training a lightweight repetition-specific reinforcement learning teacher—approximately 7,000 samples over 12 steps—and then merging it into the main model via MoPD, at a cost of approximately 4% of a full retraining.
MoPD (Model Parameter Diffusion/Merging) is a technique for injecting the capabilities of a lightweight specialized model into a main model, enabling the transfer of behavioral preferences learned by the teacher model without full retraining.
Xingchen 4.0 conducts multi-expert reinforcement learning training separately in the Agent, Coding, and Reasoning directions, and integrates the capabilities of different experts through MoPD.
MoPD (Mixture of Policy Distillation) merges expert policies independently trained for different tasks into a single unified model through distillation.
All Facts (4)
Xiaomi MiMo's fix involved training a lightweight repetition-specific reinforcement learning teacher—approximately 7,000 samples over 12 steps—and then merging it into the main model via MoPD, at a cost of approximately 4% of a full retraining.
50%UnverifiedMoPD (Model Parameter Diffusion/Merging) is a technique for injecting the capabilities of a lightweight specialized model into a main model, enabling the transfer of behavioral preferences learned by the teacher model without full retraining.
50%UnverifiedXingchen 4.0 conducts multi-expert reinforcement learning training separately in the Agent, Coding, and Reasoning directions, and integrates the capabilities of different experts through MoPD.
50%UnverifiedMoPD (Mixture of Policy Distillation) merges expert policies independently trained for different tasks into a single unified model through distillation.
50%