Expert-Choice Routing
Routing scheme in mixture-of-experts and mixture-of-recursions where each expert selects its top-k tokens, rather than each token selecting its top-k experts.
Expert-choice routing is an assignment scheme for mixture-of-experts and related architectures in which each expert independently selects the top-k tokens it will process. It is the inverse of token-choice routing, where each token picks its own top-k experts. Because every expert picks a fixed number of tokens, the load is automatically balanced and no auxiliary load-balancing loss is needed.
The scheme was formalized by Zhou et al. in their 2021 paper, 'Mixture-of-Experts with Expert Choice Routing.' It generalizes to variable recursion depth: in Mixture-of-Recursions, each depth level picks the tokens that should be processed at that depth, producing a separation between tokens that loop once, twice, or three times.
Expert-choice routing trades some flexibility (some tokens may be chosen by multiple depths, others by none) for predictable compute. The scheme is useful in inference scenarios where the cost of uneven routing outweighs the benefit of fully token-adaptive assignment.
Sources
- Mixture-of-Experts with Expert Choice Routing
arXiv · Oct 13, 2021
- Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
arXiv · Jul 14, 2025
- Sebastian Raschka: OpenAI Astra and Recurrent Depth / Looped Transformers (with expert-choice routing diagram)
X (Twitter) · Sep 2, 2026
Research this in Signals
Signals turns a topic into a sourced research record you can inspect and rerun. Your first scan is free, and this one starts with Expert-Choice Routing already loaded, so edit it or scan as is.