PulseAugur
中
实时 08:45:07
English(EN) When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions

LLM代理通过KV缓存泄露客户目标,影响跨客户决策

研究人员发现大型语言模型(LLM)代理存在一个重大问题,即来自一个客户交互的缓存信息会不当影响为另一客户做出的决策,尤其是在这些客户目标冲突时。在对Olmo-3-7B-Instruct模型进行的实验中,共享的对话记忆导致22%的决策偏离了当前客户的目标。该研究 pinpointed 了KV缓存中存储这些客户目标表征的具体层和注意力头,并发现消除这些表征可以从因果上转移模型行为,使其更符合预期客户,这强调了在自主LLM代理中实现关系特定状态隔离的必要性。 AI

影响 强调了多客户LLM代理的一个关键安全问题,需要改进状态隔离机制。

排序理由 学术论文,详细介绍了关于LLM代理行为的一项新发现。[lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

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

LLM代理通过KV缓存泄露客户目标,影响跨客户决策

本文如何被排名

Signal score
16 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
学术论文,详细介绍了关于LLM代理行为的一项新发现。[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, safety
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]

  1. arXiv cs.AI TIER_1 English(EN) · Saloni Gupta, S\'ebastien Krier, Glenn Matlin ·

    当目标跨越关系泄露:缓存的目标表征如何影响跨客户决策

    arXiv:2610.17365v1 Announce Type: new Abstract: As LLM agents take on more autonomous roles, they may interact with multiple clients and counterparties whose goals and incentives conflict. If an agent maintains shared inference state across these interactions, information associa…