UniPool: Learning Expert-to-Layer Ownership from Brief Global Access
This work studies whether a brief global-access phase can learn a compatible expert-to-layer allocation that is then executed privately, and proposes UniPool-Lock, which gives every layer its own router over one global pool, balances aggregate pool usage, and scores experts with a scale-stable NormRouter.
Minbin Huang, Han Shi, Chuan-Yang Zheng et al.
· 1 citation