single source · post-training
Freezing the MoE router during reinforcement learning
Keeps the mixture-of-experts router fixed during RL training.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See reinforcement learning algorithm for the methods filed beside it.
Also called freeze the router for RL training, freeze the MoE router, Freezing the MoE router during RL, freezing the router.
Filed under post-training › reinforcement learning algorithm
Explainer · C3 · 30 minExpert load balancingBalance over the whole batch lets experts specialize and leaves GPUs waiting; balance inside each micro-batch is faster and costs quality; a selection bias drops the loss and can quietly override the routerExplainer · S5 · 25 minRL at scale: GRPO and the gap between sampler and trainerAsynchrony, low-precision rollouts and a separate inference engine make RL affordable, and each makes the samples come from a slightly different policy than the one being trained