[2511.17885] FastMMoE: Accelerating Multimodal Large Language Models through Dynamic Expert Activation and Routing-Aware Token Pruning

Computer Science > Computer Vision and Pattern Recognition arXiv:2511.17885 (cs) [Submitted on 22 Nov 2025 (v1), last revised 20 Mar 2026 (this version, v2)] Title:FastMMoE: Accelerating Multimodal Large Language Models through Dynamic Expert Activation and Routing-Aware Token Pruning Authors:Guoyang Xia, Yifeng Ding, Fengfa Li, Lei Ren, Wei Chen, Fangxiang Feng, Xiaojie Wang View a PDF of the paper titled FastMMoE: Accelerating Multimodal Large Language Models through Dynamic Expert Activation and Routing-Aware Token Pruning, by Guoyang Xia and 6 other authors View PDF HTML (experimental) Abstract:Multimodal large language models (MLLMs) have achieved impressive performance, but high-resolution visual inputs result in long sequences of visual tokens and substantial inference latency. Reducing redundant visual tokens is critical to ease computational/memory burdens while preserving performance, enabling MLLM deployment in resource-constrained or latency-sensitive scenarios. Current visual token pruning methods mainly rely on attention-based redundancy analysis and are tailored to dense architectures. We propose Fast Multimodal Mixture-of-Experts (FastMMoE), a training-free acceleration framework for mixture-of-experts (MoE) based MLLMs, developed from a routing analysis perspective. FastMMoE combines two complementary strategies: (i) expert activation reduction for visual tokens to minimize unnecessary expert computation; and (ii) routing-aware token pruning that leverages...