MAESTRO Prunes Bad Experts in MoE Models Using Routing-Aware Markov Analysis

arXiv·medium signal

Sparsely-activated Mixture-of-Experts models are inference-efficient but keep their full expert banks resident in memory, creating a deployment bottleneck. MAESTRO (a Markov-chain approach) argues existing structured pruning — designed for dense transformers — scores expert importance with local heuristics blind to the interdependent nature of MoE routing, and instead prunes experts with routing-awareness. Useful for teams trying to shrink open MoE checkpoints for memory-constrained deployment.