The paper shows that only a small subset of attention heads in vision-language models is responsible for selecting critical visual tokens. By pruning tokens based on similarity before LLM reasoning and then applying head‑aware pruning during reasoning, the proposed ProViP framework achieves high task performance with significant speedups. Experiments on LLaVA‑1.5‑7B demonstrate 95.9% performance retention and a 1.62× inference speedup at an 88.9% pruning ratio.
By Chaofang Ma, Lin Jiang, Carol Jingyi Li, Xingyu Liu, Zeyu Li, Jiang Xu, Wei Zhang
arXiv:2609.10346v1 Announce Type: new
Abstract: Multimodal large language models (MLLMs) process hundreds or thousands of visual tokens per image, incurring prohibitive inference costs. While existin...
By Haiji Liang, Pengfei Zhou, Zhenglin Wan, Wei Wang, Yang You, Wangbo Zhao
arXiv:2608. 07193v1 Announce Type: new Abstract: Visual-token pruning can substantially reduce the inference cost of multimodal large language models (MLLMs), yet existing methods largely rely on fixed, handcrafted heuristics and costly expert trial and error.
By Zhen Liu, Wenli Huang, Wei Song, Yuhan Liu, Zhiqin Yang, Jingwen Fu
LearnPrunerV2 is a fully learnable token‑pruning framework for Vision‑Language Models that replaces heuristic pruning with lightweight modules in every backbone layer. These modules estimate token importance by combining visual saliency and query relevance via a shadow token, revealing a consistent staircase redundancy pattern across layers. The method introduces Budget‑Deterministic Pruning to set fixed per‑layer budgets, achieving state‑of‑the‑art efficiency while preserving 99.8% of performance even after pruning 77.8% of visual tokens on LLaVA‑v1.5.
By Rinyoichi Takezoe, Yaqian Li, Weiren Wang, Zihao Bo, Huihong Shi, Anzhou Hou, Mo Guang, Jiale Yan, Kaiwen Long
QCPruner is a training‑free visual token pruning method that conditions both token selection and representation on the query via bilateral utility weighting. It fuses keyword‑matched query anchors with cross‑modal cues to compute a nonnegative facility‑location objective that is monotone and submodular, guaranteeing a (1‑1/e) greedy approximation. Across multiple multimodal large language models, QCPruner consistently outperforms existing pruning methods, achieving over 96% of unpruned performance even with very few tokens retained.
By Shengli He (Guizhou University), Yongchao Liang (Guizhou University), Roumeng He (Shanghai Ocean University), Junjie Zeng (Guizhou University), Jiyuan He (Guizhou University), Can Wu (Guizhou University), Li Zheng (Guizhou University)
arXiv:2609.05916v1 Announce Type: cross
Abstract: Large vision-language models (LVLMs) achieve strong multimodal understanding, but the hundreds to thousands of visual tokens they process impose subs...
By Yichen Guo, Tinghao Wang, Qizhe Zhang, Lingbei Meng, Yuan Zhang, Jiajun Cao, Hao Jiang, Chenwei Wu, Jixian Wu, Sixiang Chen, Tao Luo, Hongyang Cheng, Kai Tang, Chenxi Li, Renyuan Li, Xiande Huang, Wenya Wang, Shanghang Zhang
SheafPrune introduces a sheaf‑guided graph propagation method for pruning visual tokens in multimodal large language models. It constructs modality‑specific maps and a sparse text‑visual graph to transfer relevance, then propagates and degree‑corrects scores on a visual graph before selecting a hard top‑k subset. Evaluated across four backbones, ten benchmarks, and seven pruning ratios, SheafPrune consistently outperforms baselines, achieving up to a 5.15× prefill speedup and 84.0% FLOPs reduction on LLaVA‑NeXT‑8B while retaining 10% of tokens.
By Xiaoyang Ma, Chi Wang, Hong Sun
Raincheck is a framework that defers multimodal prefill for large language models, allowing initial token selection to be decoupled from later information needs. It preserves unselected tokens and progressively prefills them during decoding, using unused compute capacity and scheduling gaps. Experiments on Qwen3-VL-8B and Qwen3.5-9B show that Raincheck recovers a significant portion of quality lost to hard pruning and reduces first-turn time-to-first-token by up to 24.1% with minimal impact on token generation speed.
By Wenjie Huang, Jun Zhang, Hongyaoxing Gu, Haoyang Huang, Tianqi Xu, Xinyue Hao, Xinyi Hu, Xinle Du, Jun Dai, Shuang Ge, Chen Lei, Yue Li, Mingcheng Wan
arXiv:2607. 23913v1 Announce Type: new Abstract: Modern vision-language models (VLMs) increasingly rely on dynamic or high-resolution visual encoding, producing thousands of visual tokens that substantially increase downstream language-model inference cost.
By Jun Ling, Tao Huang, Junzhuo Liu, Bowen Tang, Peng Wang
arXiv:2610.33688v1 Announce Type: new
Abstract: Visual token pruning substantially reduces the inference cost of multimodal large language models (MLLMs), yet existing methods are mostly training-fre...
By Peiran Yin, Hao Wu, Zihao Yang, Haozhe Hu, Xudong Wang, Xiaoyu Shen
SinkPruner is a training‑free framework that prunes visual tokens for multimodal large language models by first removing high‑norm redundant tokens with a visual sanitizer and then selectively keeping tokens that align with the text query using a text‑guided pruner. The coarse‑to‑fine design reduces attention sink and dispersion, enabling an 89% token reduction while preserving 96.5% of LLaVA‑1.5’s performance and 91.8% of Qwen2.5‑VL’s performance across twelve image‑language and four video‑language benchmarks. The visual sanitizer also improves existing pruning methods, showing strong transferability.
By Shiyu Li, Zi-Yuan Hu, Shijia Huang, Yanyang Li, Yiwu Zhong, Liwei Wang
LearnPrunerV2 introduces a fully learnable token pruning framework for Vision‑Language Models, embedding lightweight modules in every backbone layer to estimate token importance by combining visual saliency and query relevance via a shadow token. It identifies a consistent staircase redundancy pattern across layers and employs Budget‑Deterministic Pruning to set fixed per‑layer budgets, yielding a hardware‑friendly pruning schedule. Experiments show that LearnPrunerV2 outperforms prior methods, preserving 99.8% of performance while removing 77.8% of visual tokens on LLaVA‑v1.5, thus achieving a superior accuracy‑efficiency trade‑off.