PCoMoE: Shifting MoE Inference from Monolithic Expert Selection to Fine-Grained Path Composition

Ziyan Gan, Fangxin Liu, Chenyang Guan, Junjie Wang, Ning Yang +8 more
9/1/2026
cs.CL

Official code: gzyyy0/PCoMoE (0 stars, last commit 2026-09-02).

Abstract

Mixture-of-Experts (MoE) architectures scale Large Language Model (LLM) capacity efficiently by activating a sparse subset of experts per token. However, modern MoE inference remains heavily constrained by the rigid, whole-expert abstraction. Existing frameworks manage, schedule, or prune experts as atomic execution units, which fixes the optimization boundary too early and leaves fine-grained intra-expert computational redundancy underexplored. In this work, we present PCoMoE, a path-compositional execution framework that shifts MoE inference from coarse-grained expert selection to fine-grained path composition. PCoMoE incorporates a path-level formulation of expert computation, a compatibility-aware layer-wise pruning strategy to suppress low-value path combinations, and a hardware-friendly execution engine to exploit reusable sub-expert structures under strictly bounded overheads. Experimental results demonstrate that PCoMoE achieves up to a 1.31x end-to-end inference speedup while enhancing model accuracy by 10%. The code is available at https://github.com/gzyyy0/PCoMoE

View on arXivView PDF

Code Implementations(1)

gzyyy0/PCoMoEOfficial100%

PCoMoE | EMNLP 2026 Main Conference | Fine-grained path composition for efficient MoE inference

00Aug 21, 20261 months ago

Similar Papers

SoftMoE: Soft Differentiable Routing for Mixture-of-Experts in LLMs

86% match2026

ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference

86% match2026

ExpertFlow: Optimized Expert Activation and Token Allocation for Efficient Mixture-of-Experts Inference

86% match2024

Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts

85% match2026

Let the Expert Stick to His Last: Expert-Specialized Fine-Tuning for Sparse Architectural Large Language Models

85% match1 repo7512024

DynaNDE: Dynamic Near-Data Expert Scheduling for Batched MoE Inference

85% match2026

Cite this paper

@article{gan2026pcomoe,
  title  = {PCoMoE: Shifting MoE Inference from Monolithic Expert Selection to Fine-Grained Path Composition},
  author = {Ziyan Gan and Fangxin Liu and Chenyang Guan and Junjie Wang and Ning Yang and Haomin Li and Xiang Li and Siran Yang and Jiamang Wang and Lin Qu and Zongwu Wang and Li Jiang and Haibing Guan},
  year   = {2026},
  eprint = {2609.01024},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2609.01024v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2609.01024.svg)](https://codeofpaper.com/paper/2609.01024)

Discussion