跳到正文
arXiv cs.AI· Zefeng Cai, Zerui Cai·· 9 天前AI 评分28

稀疏事件-KV 的记忆契约:全局计算,本地物化

Compute Globally, Materialize Locally: The Memory Contract of Sparse Event-KV

AI 导读

研究提出"语义物化"现象:在长程智能体的 KV cache 中,下游事件的缓存行可作为独立可服务的计算视图,即使产生它的输入已被丢弃。在 Qwen3-8B 上,刻意构造的不含答案的事件可将 donor-aligned 恢复率从 6% 提升至 51%,而被动采集长期对话中的自然提及未检测到优势。紧凑状态可存活,较大载荷衰减至随机水平,且是否写入取决于措辞而非语义。

正文 · 原文

View PDF HTML (experimental)

Abstract:Long-horizon agents increasingly reuse their KV cache as memory: a serving system keeps a subset of cached entries and drops the rest. Eviction and episodic-memory schemes therefore rest on a premise rarely tested directly, that a retained event is still informative once the observations that produced it are gone. We test it by omitting one earlier observation from what is served, across otherwise identical agent histories. Among items sensitive to that observation, the answer overwhelmingly follows the omitted value, though no served span says which value is correct. We call this semantic materialization: a downstream event's cached rows act as an independently servable view of computation whose inputs are gone. It can also be written on purpose. A deliberately phrased, answer-free event raises donor-aligned recovery from 6% to 51% on Qwen3-8B without ever naming the value, whereas passively harvesting natural mentions from long-term dialog yields no detected advantage. What such a row carries is specific and bounded. Compact state survives, larger payloads decay toward chance, and whether a construction writes at all turns on phrasing rather than on meaning alone, so two phrasings the model comprehends equally well can diverge sharply. The result is a memory contract for sparse event-KV serving: what to write, where it lands, and what survives once the source is gone. For anyone who evicts the corollary is that dropping a source event and observing no accuracy loss does not show the source was unnecessary.
Subjects: Artificial Intelligence (cs.AI)
Cite as: arXiv:2607.23693 [cs.AI]
  (or arXiv:2607.23693v2 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2607.23693

arXiv-issued DOI via DataCite

Submission history

From: Zerui Cai [view email]
[v1] Sun, 26 Jul 2026 14:37:53 UTC (167 KB)
[v2] Wed, 7 Oct 2026 20:13:59 UTC (203 KB)

来源:arXiv cs.AI · arxiv.org