New Router-Free MoE Approach

teortaxesTex · x · 2026-07-10

The post discusses a **Routing-Free MoE** concept. Researchers argue that traditional MoE routing suffers from a scale mismatch—a tiny router must compress each expert's capability into a single score, TopK forces all tokens into the same hardcoded sparsity, and softmax discards absolute confidence. This approach from **LMU Munich + UCLA** simply eliminates the router. The method involves: - Each expert featuring its own small gate that projects inputs into the expert's representation space; - The projection magnitude determines the expert's confidence; - A learnable per-expert bias is subtracted, activating experts that exceed a global threshold. This allows each expert to "decide for itself" without central coordination, enabling adaptive token-level sparsity. Training utilizes a unified load-balancing loss that combines expert balancing and token balancing.

Original post →

More from Research

Research channel →