SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs

• URL: http://arxiv.org/abs/2506.05344v2
• Date: Sat, 05 Jul 2025 15:40:51 GMT
• Title: SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs
• Authors: Jiahui Wang, Zuyan Liu, Yongming Rao, Jiwen Lu,
• Abstract summary: We investigate how Multimodal Large Language Models (MLLMs) process visual inputs by analyzing their attention mechanisms.<n>We reveal a surprising sparsity phenomenon: only a small subset of attention heads in LLMs actively contribute to visual understanding.<n>We introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores.
• Score: 74.2538340966038
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than 5%) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers 1.38x real-time acceleration and 52% memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM.

Related papers

• CrossLMM: Decoupling Long Video Sequences from LMMs via Dual Cross-Attention Mechanisms [16.41418610688371]
  We introduce CrossLMM, which substantially reduces visual token quantity with minimal performance degradation.<n>We also introduce a text-to-visual cross-attention mechanism, for which the text tokens are enhanced through interaction with the original visual tokens.<n>Our approach achieves comparable or superior performance across diverse video-based Large Language Models benchmarks.
  arXiv  Detail & Related papers  (2025-05-22T17:59:53Z)
• RedundancyLens: Revealing and Exploiting Visual Token Processing Redundancy for Efficient Decoder-Only MLLMs [38.34856927170692]
  We propose a training-free framework for analyzing trained Multimodal Large Language Model (MLLM)<n>It consists of Probe-Activated Dynamic FFN and Hollow Attention, which enable adjustable reductions in computations for visual tokens.<n>Experiments demonstrate substantial, structured, and clustered redundancy unique to decoder-only MLLMs.
  arXiv  Detail & Related papers  (2025-01-31T11:09:16Z)
• OLA-VLM: Elevating Visual Perception in Multimodal LLMs with Auxiliary Embedding Distillation [95.78870389271832]
  The standard practice for developing contemporary MLLMs is to feed features from vision encoder(s) into the LLM and train with natural language supervision.<n>We propose OLA-VLM, the first approach distilling knowledge into the LLM's hidden representations from a set of target visual representations.<n>We show that OLA-VLM boosts performance by an average margin of up to 2.5% on various benchmarks, with a notable improvement of 8.7% on the Depth task in CV-Bench.
  arXiv  Detail & Related papers  (2024-12-12T18:55:18Z)
• Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings [69.35226485836641]
  Excessive use of visual tokens in existing Multimoal Large Language Models (MLLMs) often exhibits obvious redundancy and brings in prohibitively expensive computation.<n>We propose a simple yet effective method to improve the efficiency of MLLMs, termed dynamic visual-token exit (DyVTE)<n>DyVTE uses lightweight hyper-networks to perceive the text token status and decide the removal of all visual tokens after a certain layer.
  arXiv  Detail & Related papers  (2024-11-29T11:24:23Z)
• Mono-InternVL: Pushing the Boundaries of Monolithic Multimodal Large Language Models with Endogenous Visual Pre-training [48.455597568212944]
  We present Mono-InternVL, a novel monolithic MLLM that seamlessly integrates a set of visual experts via a multimodal mixture-of-experts structure.<n>In particular, EViP is designed as a progressive learning process for visual experts, which aims to fully exploit the visual knowledge from noisy data to high-quality data.
  arXiv  Detail & Related papers  (2024-10-10T17:59:22Z)
• Treat Visual Tokens as Text? But Your MLLM Only Needs Fewer Efforts to See [37.7015406019386]
  Multimodal Large Language Models (MLLMs) treat visual tokens from visual encoders as text tokens.<n>As token counts grow, the quadratic scaling of computation in LLMs introduces an efficiency bottleneck.<n>In this study, we investigate the redundancy in visual computation at both the parameter and computational pattern levels within LLaVA.
  arXiv  Detail & Related papers  (2024-10-08T16:13:24Z)
• Bridging LLMs and KGs without Fine-Tuning: Intermediate Probing Meets Subgraph-Aware Entity Descriptions [49.36683223327633]
  Large Language Models (LLMs) encapsulate extensive world knowledge and exhibit powerful context modeling capabilities.<n>We propose a novel framework that synergizes the strengths of LLMs with robust knowledge representation to enable effective and efficient KGC.<n>We achieve a 47% relative improvement over previous methods based on non-fine-tuned LLMs and, to our knowledge, are the first to achieve classification performance comparable to fine-tuned LLMs.
  arXiv  Detail & Related papers  (2024-08-13T10:15:55Z)
• F-LMM: Grounding Frozen Large Multimodal Models [53.8059045627934]
  We present F-LMM -- grounding frozen off-the-shelf LMMs in human-AI conversations.<n>It is based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention mechanism of well-trained LMMs.<n>It achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks.
  arXiv  Detail & Related papers  (2024-06-09T15:14:26Z)
• DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models [28.019592576500113]
  This study examines the projector module by interpreting the vision-language semantic flow within MLLMs. Our findings reveal that compressive projectors abstract visual patches into a limited set of semantic concepts, such as objects or attributes, resulting in a 'double abstraction' phenomenon. We propose the key insight of 'Decouple Compression from Abstraction (DeCo) that is compressing the visual token number at the patch level by projectors.
  arXiv  Detail & Related papers  (2024-05-31T16:31:38Z)
• RAR: Retrieving And Ranking Augmented MLLMs for Visual Recognition [78.97487780589574]
  Multimodal Large Language Models (MLLMs) excel at classifying fine-grained categories. This paper introduces a Retrieving And Ranking augmented method for MLLMs. Our proposed approach not only addresses the inherent limitations in fine-grained recognition but also preserves the model's comprehensive knowledge base.
  arXiv  Detail & Related papers  (2024-03-20T17:59:55Z)
• From CLIP to DINO: Visual Encoders Shout in Multi-modal Large Language Models [36.41816380074965]
  We investigate the effectiveness of different vision encoders within Large Language Models (MLLMs) Our findings reveal that the shallow layer features of CLIP offer particular advantages for fine-grained tasks such as grounding and region understanding. We propose a simple yet effective feature merging strategy, named COMM, that integrates CLIP and DINO with Multi-level features Merging.
  arXiv  Detail & Related papers  (2023-10-13T02:41:55Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.