ByteDance Seed ties long-context swings to chunked KV cache 'phase sensitivity'
ByteDance's Seed team says in a new research paper that large language models' performance swings on very long inputs stem mainly from "phase sensitivity" introduced by chunked KV cache compression. To cut memory use in long-context inference, the method compresses consecutive token windows into fewer entries at a fixed stride, creating a new positional coordinate: a token's "phase" relative to the compression window boundary. Experiments show the same information becomes much harder or easier to retrieve at different phases, with long-context retrieval accuracy gaps of up to 40 percentage points in some large open-source models.
It suggests models with strong average long-context scores may still hide periodic weak spots, so averages alone are a shaky guide to long-context reliability.
DeepSeek 长文本为何突然“抽风”?字节 Seed 团队揭开 AI 性能波动谜团

AIbase基地
发布于AI新闻资讯 · 1 分钟阅读 · Oct 9, 2026
19
字节跳动旗下 Seed 团队在 最新 的研究论文中,揭示了大语言模型在处理超长文本时出现性能波动的技术原因。研究团队指出,这一现象主要源于分块 KV 缓存压缩技术所带来的“相位敏感性”。