From Inheritance to Saturation: Disentangling the Evolution of Visual Redundancy for Architecture-Aware MLLM Inference Acceleration

TL;DR

This paper introduces HalfV, a framework for accelerating high-resolution multimodal large language models by disentangling visual redundancy into universal and architecture-dependent components, achieving significant efficiency gains.

Contribution

The paper proposes a universal three-stage inference lifecycle and a novel framework, HalfV, that effectively reduces visual redundancy across diverse architectures, improving inference efficiency.

Findings

HalfV retains 96.8% performance on Qwen25-VL at 4.1× FLOPs speedup.

It achieves superior efficiency-performance trade-offs compared to state-of-the-art methods.

The approach decouples intrinsic and architecture-dependent visual redundancies for better generalization.

Abstract

High-resolution Multimodal Large Language Models (MLLMs) face prohibitive computational costs during inference due to the explosion of visual tokens. Existing acceleration strategies, such as token pruning or layer sparsity, suffer from severe "backbone dependency", performing well on Vicuna or Mistral architectures (e.g., LLaVA) but causing significant performance degradation when transferred to architectures like Qwen. To address this, we leverage truncated matrix entropy to uncover a universal three-stage inference lifecycle, decoupling visual redundancy into universal Intrinsic Visual Redundancy (IVR) and architecture-dependent Secondary Saturation Redundancy (SSR). Guided by this insight, we propose HalfV, a framework that first mitigates IVR via a unified pruning strategy and then adaptively handles SSR based on its specific manifestation. Experiments demonstrate that HalfV achieves superior efficiency-performance trade-offs across diverse backbones. Notably, on Qwen25-VL, it retains 96.8\% performance at a 4.1$\times$ FLOPs speedup, significantly outperforming state-of-the-art baselines. Our code is available at https://github.com/civilizwa/HalfV.