English(EN) Pixel Decodability Is Not a Compression Signal: Causally Evaluating Importance Proxies for Visual KV-Cache Eviction

新论文发现视觉 KV 缓存保留与任务无关

作者 PulseAugur 编辑部 · [1 个来源] · 2026-09-14 04:00

一项新的研究论文挑战了视觉语言模型中的视觉键值（KV）缓存保留任务相关信息的假设。研究发现，保留在 KV 缓存中的视觉内容的数量在很大程度上与任务无关，并且与信息是否被因果用于回答问题无关。虽然注意力机制与因果利用显示出微弱的相关性，但像素可解码的可重构性被证明是 KV 缓存压缩的一个糟糕信号，一个模型保留的任务无关内容比另一个模型多 2.7 倍。 AI

影响挑战了当前关于 KV 缓存效率的假设，并为优化视觉语言模型性能提供了新的方向。

排序理由研究论文发布在 arXiv 上，详细介绍了视觉语言模型中视觉 KV 缓存的发现。[lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.CV 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 23 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 研究论文发布在 arXiv 上，详细介绍了视觉语言模型中视觉 KV 缓存的发现。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: paper, infra; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Breaking (< 6h); Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

arXiv cs.CV TIER_1 English(EN) · Chenyu Zhou, Qiliang Jiang, Shuning Wu, Xu Zhou · 2026-09-14 04:00

像素可解码性并非压缩信号：因果评估视觉 KV 缓存驱逐的重要性代理

arXiv:2609.13012v1 Announce Type: new Abstract: Vision-language models retain a substantial amount of pixel-decodable visual content in their visual key-value cache. We show, in our setting, that this retention is task-inert: across our preregistered tests, how much a unit retain…

本文如何被排名

报道来源 [1]

像素可解码性并非压缩信号：因果评估视觉 KV 缓存驱逐的重要性代理

相关实体

相关话题