MoR-MLLM：面向高效多模态大模型的递归混合框架
原标题：MoR-MLLM: Mixture of Recursions for Efficient Multimodal Large Language Models
AI 摘要
该论文提出 MoR-MLLM，一种基于 Mixture-of-Recursions（MoR）框架的计算稀疏多模态大语言模型。它通过自适应逐 token 递归，动态调整递归深度，将更多计算分配给视觉或语言上困难的 token，并跳过简单 token 的冗余操作。为稳定多模态场景下的递归稀疏训练，作者设计了三阶段 MoR-Tuning 策略和熵正则化损失。实验表明，相比近期先进的小型 MLLM，MoR-MLLM 在大幅降低训练内存和计算复杂度的同时，在多种视觉-语言任务上保持高性能。
正文节选
MoR-MLLM: Mixture of Recursions for Efficient Multimodal Large Language Models Abstract Multimodal Large Language Models (MLLMs) have demonstrated remarkable reasoning capabilities across vision and language tasks. However, their massive computational and memory demands hinder real-world deployment. While recent efforts reduce costs by employing lightweight language backbones, existing paradigms remain computation-dense due to their static sparsity and depth allocation, which cannot adapt to the