LinMU: Multimodal Understanding Made Linear

TL;DR

LinMU introduces a linear-complexity multimodal understanding model that maintains high performance while significantly reducing computational costs, enabling efficient processing of high-resolution images and long videos.

Contribution

The paper proposes a novel M-MATE block and a three-stage distillation framework to transform pre-trained VLMs into linear-complexity models without sacrificing accuracy.

Findings

LinMU matches the performance of teacher models on multiple benchmarks.

Reduces Time-To-First-Token by up to 2.7 times.

Improves token throughput by up to 9.0 times on long videos.

Abstract

Modern Vision-Language Models (VLMs) achieve impressive performance but are limited by the quadratic complexity of self-attention, which prevents their deployment on edge devices and makes their understanding of high-resolution images and long-context videos prohibitively expensive. To address this challenge, we introduce LinMU (Linear-complexity Multimodal Understanding), a VLM design that achieves linear complexity for the language model decoder without using any quadratic-complexity modules while maintaining the performance of global-attention-based VLMs. LinMU replaces every self-attention layer in the language model decoder with an M-MATE block: a dual-branch module that combines a bidirectional state-space model for global context (Flex-MA branch) with localized Swin-style window attention (Local-Swin branch) for adjacent correlations. To transform a pre-trained VLM into the LinMU architecture, we propose a three-stage distillation framework that (i) initializes both branches with self-attention weights and trains the Flex-MA branch alone, (ii) unfreezes the Local-Swin branch and fine-tunes it jointly with the Flex-MA branch, and (iii) unfreezes the remaining blocks and fine-tunes them using LoRA adapters, while regressing on hidden states and token-level logits of the frozen VLM teacher. On MMMU, TextVQA, LongVideoBench, Video-MME, and other benchmarks, LinMU matches the performance of teacher models, yet reduces Time-To-First-Token (TTFT) by up to 2.7$\times$ and improves token throughput by up to 9.0$\times$ on minute-length videos. Ablations confirm the importance of each distillation stage and the necessity of the two branches of the M-MATE block. The proposed framework demonstrates that state-of-the-art multimodal reasoning can be achieved without quadratic attention, thus opening up avenues for long-context VLMs that can deal with high-resolution images and long videos.