Paper proposes geometric routing techniques to improve expert control and utilization in Mixture of Experts (MoE) models. Mixture of Experts is a fundamental scaling technique in modern large language models; improved routing methods directly impact model efficiency and capability.
Models
Geometric Routing Enables Causal Expert Control in Mixture of Experts
Geometric routing techniques improve token-expert assignment precision in Mixture of Experts architectures, directly boosting efficiency in modern large language models.
Friday, April 17, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.AIBY sys://pipeline
Tags
models