---
id: 20260911-T0-12
title: "动态稀疏 MoE 的分布一致性推理：解决固定 top-k 选专家的问题"
title_en: "Distribution-Consistent Inference Tackles Fixed Top-k in Sparse MoE"
url: https://ai.daily.yangsir.net/daily/20260911-T0-12
issue_date: 2026-09-11
publish_date: 2026-09-10T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.09241
---

# 动态稀疏 MoE 的分布一致性推理：解决固定 top-k 选专家的问题

arXiv 论文研究动态稀疏混合专家（MoE）的推理问题。作者指出，MoE 已是在大型基础模型中扩展容量同时保持推理效率的主流方案，但多数模型使用固定的 top-k 专家选择策略，无论输入如何都分配同样数量的专家。论文提出分布一致性推理方法，针对动态稀疏 MoE 的选择策略做改进。若方法有效，可让模型按输入复杂度分配算力，而非固定开销。

## English Version

**Distribution-Consistent Inference Tackles Fixed Top-k in Sparse MoE**

An arXiv paper addresses inference in dynamic sparse Mixture-of-Experts (MoE) models. The authors note that MoE has become a powerful paradigm for scaling capacity while keeping inference efficient in large foundation models, yet most models use a fixed top-k expert selection policy that assigns the same number of experts regardless of input. The paper proposes a distribution-consistent inference method for dynamic sparse MoE. If effective, it would let models allocate compute based on input complexity instead of fixed cost.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.09241)

**详情页**：https://ai.daily.yangsir.net/daily/20260911-T0-12

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*