●builderThis offers a way to reduce memory bottlenecks in long-context MoE inference without breaking the routing logic required for accurate generation.
●researcherThe use of Fisher information to quantify routing sensitivity provides a new framework for studying MoE compression.