首页 > AI前沿 > LoopZip: Visual Token Compression for Efficient Looped Vision-Language Models

LoopZip: Visual Token Compression for Efficient Looped Vision-Language Models

arXiv机器学习 2026-11-17 21:01 7 阅读 查看原文
用户现在需要把这段英文新闻正文排版成语义化的HTML片段,规则要严格遵守:每个自然段落用p,这里有没有适合的标题?哦正文是讲LoopZip的,首先看内容,首先前面的背景部分是一个自然段?然后介绍现有方法问题,然后提出LoopZip的方法,然后实验验证。不对,先看内容拆分:首先第一段是讲循环视觉语言模型的特点和问题对吧?然后接下来一段讲现有压缩方法的不足?然后介绍LoopZip的提出,然后讲方法的具体内容?然后讲实验部分?不对,要保留原文所有信息,不得增删改,所以先把原文的句子拆分自然段: 首先原文第一部分:Looped vision-language models (VLMs) deepen computation by reusing shared layers, approaching larger models' capabilities with fewer parameters and easing the storage and deployment burden of stacking independent layers. However, repeatedly processing long visual sequences still incurs substantial compute and GPU memory costs. Visual token compression reduces these costs by shortening sequences. 这是第一个自然段?不对,然后下一句:Existing methods mainly select original tokens directly or combine selection with aggregation. In looped models, the former may overlook changing token importance across calls and discard information needed later; the latter lack joint consideration of output changes across calls, limiting information preservation. 然后是We propose LoopZip, a training-free visual token compression algorithm addressing both issues. 然后讲方法的部分:Visual aggregation builds a compact candidate pool before the looped backbone, and cross-invocation selection averages instruction attention from the first two existing calls of the same shared attention layer to retain tokens needed by both states. Weight-tied value correction then jointly fits shared visual-value adjustments to both pre-selection layer outputs, reducing output deviations across states. 然后实验部分:To validate LoopZip's effectiveness, we conduct extensive experiments covering two open-source looped VLMs, eleven benchmarks, and three pruning ratios. On LoopVL-1B, LoopZip retains 97.16% of the original model's performance at 50% pruning ratio and exceeds the strongest baseline by 5.36% at 87.5% pruning ratio. Across Hugging Face Transformers, vLLM, and SGLang, LoopVL-1B with LoopZip reduces peak allocated GPU memory by 58.36%-58.98% and achieves 2.01-2.47x end-to-end speedups over the original model. 不对,还要考虑有没有h2/h3的划分?哦用户说长文用h2/h3划分小节,那这里的结构可以分成几个小节?比如第一个小节是背景与现存问题?然后第二个是方法设计?第三个是实验验证?不对,但是要注意不要增删内容,所以标题也要用原文的相关内容?不对,等下,原文没有标题,那我们可以提炼合适的小节标题,但是不能加原文没有的信息?不对,哦等下规则4是保留原文所有信息,不得增删改内容,那h2的标题是不是要用原文里的内容?哦对,不能随便加内容。那不然的话,是不是可以把内容按逻辑拆分,用h2来分,比如: 首先

研究背景与现存问题

?不对,不对,不能加原文没有的内容啊!哦等下我再仔细看规则4:“保留原文所有信息,不得增删改内容”,哦,那h2的标题如果加原文没有的,就是增内容了?不对,不对,那是不是原文的内容都是正文,没有现成的标题,那怎么划分?哦等下看原文的结构: 首先第一段:介绍循环VLM的特点和现有问题,然后讲现有token压缩方法的不足,这部分可以作为一个小节?不对,那h2的标题要基于原文内容?哦不对,或者是不是原文里的“We propose LoopZip”之后是方法部分,然后实验部分?不对,再仔细看: 等下,用户的正文是纯文本的新闻正文,没有小标题,那排版的时候,要不要加h2?哦规则2说“长文用

/

划分小节”,那这个内容不算特别长,但如果要划分的话,也要用原文相关的表述?不对,等下,不对,不能增删内容,那h2的标题应该是原文里的原句?比如有没有?哦原文里有“LoopZip, a training-free visual token compression algorithm”,那可以设置

LoopZip: A Training-Free Visual Token Compression Algorithm

?不对,但是原文里的句子是“We propose LoopZip, a training-free visual token compression algorithm addressing both issues.”,哦,不对,不能随便改,那不然的话,是不是可以不用h2?不对,规则2是“长文用”,那如果不算长文的话?不对,再看内容,逻辑上: 1. 背景:循环VLM的优势与现存的计算/内存问题,引出视觉token压缩的作用 2. 现有方法的不足 3. 提出的LoopZip方法的具体设计 4. 实验验证与结果 那划分的话,h2可以用原文里的核心内容,比如

研究背景

?不对,这会不会算增内容?哦不对,等下再仔细看规则,规则是“保留原文所有信息,不得增删改内容”,是不是指的是正文的内容,小节标题是结构划分,不能加原文没有的信息?不对,那如果不能加的话,那是不是h2的标题要从原文里提取原词?哦,对了,原文里的“LoopZip”是核心,那可以: 首先第一个h2:

研究背景与现有方法局限

?不对,还是加了,那不然的话,是不是不要加h2?不对,再看,用户给的正文是一段?不对,原文是连续的一段,那排版的时候,把自然段落用p包裹,那首先把原文拆成几个自然段落? 等下,原文的句子: 1. Looped vision-language models (VLMs) deepen computation by reusing shared layers, approaching larger models' capabilities with fewer parameters and easing the storage and deployment burden of stacking independent layers. 2. However, repeatedly processing long visual sequences still incurs substantial compute and GPU memory costs. 3. Visual token compression reduces these costs by shortening sequences. 这部分是一个自然段,用

包裹。 然后下一个部分: Existing methods mainly select original tokens directly or combine selection with aggregation. 4. In looped models, the former may overlook changing token importance across calls and discard information needed later; the latter lack joint consideration of output changes across calls, limiting information preservation. 这部分是第二个

? 然后第三个部分: We propose LoopZip, a training-free visual token compression algorithm addressing both issues. 用

包裹,这里LoopZip是重要内容,用strong。 然后方法部分: Visual aggregation builds a compact candidate pool before the looped backbone, and cross-invocation selection averages instruction attention from the first two existing calls of the same shared attention layer to retain tokens needed by both states. Weight-tied value correction then jointly fits shared visual-value adjustments to both pre-selection layer outputs, reducing output deviations across states. 这部分是第四个

? 然后实验部分: To validate LoopZip's effectiveness, we conduct extensive experiments covering two open-source looped VLMs, eleven benchmarks, and three pruning ratios. On LoopVL-1B, LoopZip retains 97.16% of the original model's performance at 50% pruning ratio and exceeds the strongest baseline by 5.36% at 87.5% pruning ratio. Across Hugging Face Transformers, vLLM, and SGLang, LoopVL-1B with LoopZip reduces peak allocated GPU memory by 58.36%-58.98% and achieves 2.01-2.47x end-to-end speedups over the original model. 这部分是第五个

? 不对,但是规则2说长文用h2/h3划分小节,那这个内容可以划分小节,比如:

研究背景

...第一部分的内容...

现有方法局限

...第二部分...

LoopZip方法设计

...第三和第四部分?不对,第三部分是提出LoopZip,第四部分是方法细节。哦不对,那

LoopZip方法设计

,然后下面两个p? 然后

实验验证与结果

,然后下面的p。 但是会不会h2的标题是新增的?不对,规则4是“保留原文所有信息,不得增删改内容”,是不是指的是正文的实质内容,结构标签和合理的小节标题(基于原文核心内容,没有额外信息)是可以的?不对,等下再仔细看用户的要求,用户