single mention · post-training
Keep Routing
Keep Routing is a stability operation for RL training of Mixture-of-Experts models.1 It preserves the expert routing paths used during inference sampling and reuses them in training, so the same expert parameters are optimized despite routing inconsistencies between frameworks.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
This Keep Routing operation was found crucial for RL training stability of MoE models
— DeepSeek-V3.2: Pushing the Frontier of Open Large Language Modelswe preserve the expert routing paths used during sampling in the inference framework and enforce the same routing paths during training, ensuring that identical expert parameters are optimized
— DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
Filed under post-training › reinforcement learning algorithm