English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

UniPool: A Globally Shared Expert Pool for Mixture-of-Experts

Forum topic · 小凯 · 2026-05-09

Summary

UniPool is a Mixture-of-Experts (MoE) architecture that replaces the conventional per-layer expert allocation with a single globally shared expert pool accessed by independent per-layer routers. The design is motivated by a routing probe showing that replacing a deeper layer's learned top-k router with uniform random routing drops downstream accuracy by only 1.0-1.6 points across multiple production MoE models, suggesting deep layers do not require isolated expert capacity. UniPool introduces a pool-level auxiliary loss to balance expert utilization across the entire pool and a NormRouter mechanism for sparse, scale-stable routing. Trained on 30 billion tokens from The Pile across five LLaMA-architecture model sizes (182M to 978M parameters), UniPool consistently outperforms matched vanilla MoE baselines, reducing validation loss by up to 0.0386. Slimmed pool variants using only 41.6%-66.7% of the vanilla expert parameter budget match or beat per-layer MoE, showing expert parameters can scale sub-linearly with depth. The arXiv paper is 2505.03485.

Paper Overview

Research area: Machine Learning Authors: Minbin Huang, Han Shi, Chuanyang Zheng Published: 2026-05-06 arXiv: 2505.03485

Summary

Modern Mixture-of-Experts (MoE) architectures allocate expert capacity through a rigid per-layer rule: each transformer layer owns a separate expert set. This convention couples depth scaling with linear expert-parameter growth and assumes that every layer needs isolated expert capacity. However, recent analyses and the authors' routing probe challenge this allocation rule: replacing a deeper layer's learned top-k router with uniform random routing drops downstream accuracy by only 1.0-1.6 points across multiple production MoE models.

Key Contributions

  • UniPool architecture: Expert capacity is treated as a global architectural budget. A single shared expert pool replaces per-layer expert ownership, with each layer accessing the pool through its own independent router.
  • Pool-level auxiliary loss: Balances expert utilization across the entire shared pool, enabling stable and balanced training under the sharing mechanism.
  • NormRouter: Provides sparse and scale-stable routing into the shared expert pool.
  • Experimental Results

  • Trained on 30 billion tokens from The Pile across five LLaMA-architecture model sizes: 182M, 469M, 650M, 830M, and 978M parameters.
  • UniPool consistently outperforms matched vanilla MoE baselines on both validation loss and perplexity, reducing validation loss by up to 0.0386 at these scales.
  • Slimmed-pool UniPool variants using only 41.6%-66.7% of the vanilla expert parameter budget match or outperform per-layer MoE at the tested scales.
  • This demonstrates that expert parameters need not grow linearly with depth under a shared-pool design; they can scale sub-linearly while being more efficient and effective than vanilla MoE.
  • Further analysis shows UniPool's advantages stack with finer-grained expert decomposition.

Significance

The results identify pool size as an explicit depth-scaling hyperparameter, challenging the long-standing assumption that every transformer layer requires its own isolated expert capacity in MoE models.

Tags

#mixture-of-experts#deep-learning#llm#scaling#arxiv#sparsity#model-architecture

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177619661