Pith. sign in

REVIEW 11 cited by

ZipVL: Efficient Large Vision-Language Models with Dynamic Token Sparsification

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2410.08584 v2 pith:VVK5Q4DM submitted 2024-10-11 cs.CV cs.AI

classification cs.CVcs.AI
keywords attentionphaseimportantlvlmstokensbottleneckcachedecoding
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase and the memory bottleneck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs through a dynamic ratio allocation strategy of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyper-parameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform sparse attention mechanism solely on those important tokens, reducing the latency in the prefill phase. Tokens deemed less important will be discarded to reduce KV cache size, alleviating the memory bottleneck in the decoding phase. Our experiments demonstrate that ZipVL can accelerate the prefill phase by 2.3$\times$ and improve decoding throughput by 2.8$\times$, with a minimal accuracy reduction of only 0.5\% on VQAv2 benchmark over LLaVA-Next-13B model, effectively enhancing the generation efficiency of LVLMs.

Discussion (0). Sign in to comment.

Forward citations

Cited by 11 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SpecEyes: Accelerating Agentic Multimodal LLMs via Speculative Perception and Planning

    cs.CV 2026-03 conditional novelty 7.0 of 10

    A small tool-free MLLM plus answer-separability gating bypasses agentic tool loops for many queries, yielding 1.1–3.35× speedup with preserved or higher accuracy.

  2. SlimVLM: Sensitivity-aware Dynamic Structured Pruning with Adaptive Visual Token Selection for Efficient Vision-Language Models

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Visual-token attention filtering improves structured pruning of vision-language models, keeping 94% of average benchmark accuracy after removing 20% of parameters.

  3. Seeing the End at Step Zero: Accelerating Diffusion MLLMs via MLP Sparsity-Aware Truncation

    cs.AI 2026-07 conditional novelty 6.0 of 10

    A diffusion multimodal LLM's early-layer MLP sparsity at denoising step 0 pinpoints the semantic boundary, enabling one-shot suffix truncation that raises throughput up to ~31x.

  4. Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval

    cs.IR 2026-07 conditional novelty 6.0 of 10

    Object-aware soft merging of post-projector visual tokens preserves MaxSim-selectable evidence, yielding >93% token reduction and higher R@1 than full ColPali on Flickr30K and MSCOCO.

  5. BVS: Bayesian Visual Search with Multimodal Large Language Model for Fine-grained Perception

    cs.CV 2026-07 conditional novelty 6.0 of 10

    BVS combines early-stop attention rollout priors with a scale-aware non-stationary kernel and GP-UCB to locate tiny objects in UHR images more accurately and with fewer MLLM queries than prior visual-search methods.

  6. DATE: Dynamic Absolute Time Enhancement for Long Video Understanding

    cs.CV 2025-09 conditional novelty 6.0 of 10

    DATE combines inference-time timestamp token injection with a caption-rewritten, temporally regularized CLIP sampling strategy to improve absolute time reasoning and event localization in long videos.

  7. METEOR: Multi-Encoder Collaborative Token Pruning for Efficient Vision Language Models

    cs.CV 2025-07 conditional novelty 6.0 of 10

    METEOR is a three-stage token pruning framework that reduces visual tokens in multi-encoder MLLMs by 76% with only a 0.3% average accuracy drop.

  8. VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning

    cs.CV 2025-07 conditional novelty 6.0 of 10

    A vision-language model learns via reinforcement learning when to upscale a low-resolution image, cutting visual tokens roughly in half while preserving accuracy on most benchmarks.

  9. Streamline Without Sacrifice -- Squeeze out Computation Redundancy in LMM

    cs.CV 2025-05 conditional novelty 6.0 of 10

    ProxyV introduces proxy vision tokens that take over expensive attention and feed-forward computation in later layers of decoder-only multimodal models, cutting FLOPs by 25-46% while retaining or improving accuracy on...

  10. OmniFocus: Query-Guided Modality-Balanced Token Compression for Omni-Modal Large Language Models

    cs.LG 2026-07 conditional novelty 5.5 of 10

    Query-guided, modality-symmetric token compression preserves audio-visual evidence better than audio-guided budgets at 25–35% retention on Qwen2.5-Omni.

  11. Video-XL-2: Towards Very Long-Video Understanding Through Task-Aware KV Sparsification

    cs.CV 2025-06 conditional novelty 5.0 of 10

    Video-XL-2 cuts long-video inference cost with chunked pre-filling and query-gated dense-or-sparse KV reloading, reporting half the FLOPs and a third less decoding memory at roughly equal benchmark scores.

Pith tools