BRANCH-MoE introduces a hierarchical routing architecture for mixture-of-experts (MoE) layers. Instead of a flat router, it places experts at leaves of a binary decision tree and uses an exponential moving average (EMA) score at each internal node to bias routing toward the arrival-weighted mean. This mechanism promotes balanced utilization of both child subtrees without an auxiliary load-balancing loss.
Experiments on four tabular benchmarks (Criteo, Forest Covertype, HIGGS, YearPredictionMSD) show that BRANCH-MoE matches or exceeds flat routers in held-out loss while inducing a topology that reduces co-routing distance between tokens by 12–23% compared to random baselines. The tree structure also provides explicit bounds on cross-device communication when experts are placed by tree prefix.
This is interesting because it addresses two persistent issues in MoE: imbalanced expert utilization and lack of structure in expert assignments. The tree topology offers natural locality for distributed systems and removes the need for auxiliary balancing losses that compete with the primary objective.