从ACL 2024录用论文看混合专家模型（MoE）最新研究进展

文章预览

作者 |  杨远航 整理 | PaperWeekly ACL 2024 论文放榜，扫了下，SMoE（稀疏混合专家）的论文不算多，这里就仔细梳理一下，包括动机、方法、有趣的发现，方便大家不看论文也能了解的七七八八，剩下只需要感兴趣再看就好。 下面是列表，顺序大抵是个人兴趣程度排序。 1. DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models 2. Harder Tasks Need More Experts: Dynamic Routing in MoE Models  3. XMoE: Sparse Models with Fine-grained and Adaptive Expert Selection 4. HyperMoE: Towards Better Mixture of Experts via Transferring Among Experts 5. Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models  6. Multimodal Instruction Tuning with Conditional Mixture of LoRA 未完待续，大概还遗漏了一二三四篇，后续再加上 2024 年的一些 MoE 论文： 1. Let the Expert Stick to His Last: Expert-Specialize ………………………………