VLM-Pruner：在高效 VLM 离心式词元剪枝范式中为空间稀疏性引入缓冲机制
VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm
提出免训练的视觉词元剪枝范式：用离心式由近及远的选择配合空间稀疏缓冲准则，避免保留词元过度稀疏而漏掉目标区域，可直接套用到任意 VLM 的视觉编码器输出上做推理加速。
原文依据 we propose VLM-Pruner, a training-free token pruning algorithm that explicitly balances redundancy and spatial sparsity
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论视觉语言模型（VLM）在图像理解任务上表现优异，但大量视觉词元带来显著计算开销，阻碍其在移动设备上的部署。许多剪枝方法仅依赖词元重要性，因而忽视词元间冗余，保留了大量重复词元并浪费容量。尽管已有一些冗余感知方法被提出，它们往往忽略视觉词元之间的空间关系，这可能导致保留词元的选择过于稀疏，无法充分覆盖目标物体所在区域。为解决上述局限，我们提出 VLM-Pruner，一种免训练的词元剪枝算法，显式地平衡冗余性与空间稀疏性。具体而言，我们引入离心式词元剪枝范式，实现由近及远的选择。此外，空间稀疏性缓冲（BSS）准则延迟对空间上较远词元的选择，从而优先保留细粒度的物体细节。为进一步提升效率，词元选择采用并行贪心策略。为缓解剪枝带来的信息损失，我们将被丢弃词元中的显著信息有选择地融合进保留词元。综合对比表明，VLM-Pruner 在五种 VLM、多种剪枝比例下始终优于其他竞争方法，并带来端到端推理加速。代码已发布于 https://github.com/Casey-bit/VLMPruner。
英文原文摘要 1503 字符
Vision-language models (VLMs) excel at image understanding tasks, but the large number of visual tokens imposes significant computational costs, hindering deployment on mobile devices. Many pruning methods rely solely on token importance and thus overlook inter-token redundancy, retaining numerous duplicated tokens and wasting capacity. Although some redundancy-aware approaches have been proposed, they often ignore the spatial relationships among visual tokens. This can lead to overly sparse selections of retained tokens that fail to adequately cover the regions of target objects. To address these limitations, we propose VLM-Pruner, a training-free token pruning algorithm that explicitly balances redundancy and spatial sparsity. Specifically, the centrifugal token pruning paradigm is introduced for near-to-far selection. Moreover, the Buffering for Spatial Sparsity (BSS) criterion delays the selection of spatially distant tokens, thereby prioritizing the preservation of fine-grained object details. To further improve efficiency, a parallel greedy strategy is adopted for token selection. To mitigate information loss from pruning, we selectively fuse salient information from the discarded tokens into the retained ones. Comprehensive comparisons demonstrate that VLM-Pruner consistently outperforms other competitive methods across five VLMs under various pruning ratios, while delivering an end-to-end inference speedup. The code is available at https://github.com/Casey-bit/VLMPruner.