跳到正文
arXiv cs.AI· Hejia Zhang, Zhongming Yu, Yimin Ding, Wei-Ting Chen, Hengjia Yu, Wentao Ni, Yijun Liu, Cong Xu, Xin Chen, Matrix Yao, Fan Zhao, Jishen Zhao·· 56 分钟前AI 评分46

多智能体记忆不一致:LLM Agent 中事实冲突为何消失

Multi-Agent Memory Inconsistency: Where Factual Conflicts Disappear in LLM Agents

AI 导读

一项 arXiv 研究提出针对多智能体 LLM 记忆不一致的基准与证据流分析工具,基于 Harvey LAB 文档审查任务构建可控事实冲突。在三个模型汇总、生成开销相当的条件下,委托多智能体使两条冲突论断的检索率从 50.2% 升至 68.7%,但冲突检出率从 16.2% 降至 13.2%。两条论断被拆分到不同 worker 时检出率更低,worker 向 lead 的交接是改进关键。

正文 · 原文

Authors:Hejia Zhang, Zhongming Yu, Yimin Ding, Wei-Ting Chen, Hengjia Yu, Wentao Ni, Yijun Liu, Cong Xu, Xin Chen, Matrix Yao, Fan Zhao, Jishen Zhao

View PDF HTML (experimental)

Abstract:Multi-agent delegation lets LLM systems distribute evidence collection across workers with separate contexts. This creates a consistency challenge: facts retrieved by workers may be missing from the evidence available to the lead for a joint conclusion. We introduce a benchmark and evidence-flow profiler for studying multi-agent memory inconsistency through controlled factual conflicts in Harvey LAB document-review tasks. Pooled across three models with comparable generation spending, requiring delegation increases retrieval of both conflicting claims from 50.2% to 68.7% relative to single-agent execution, yet conflict detection falls from 16.2% to 13.2%. Within this setting, workers report both claims for only 35.2% of retrieved pairs, and pooled matched comparisons show higher detection than single-agent execution when one worker reads both claims but lower detection when the claims are split across workers. Worker-to-lead handoffs are a central target for improving conflict detection in multi-agent LLMs.
Subjects: Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
Cite as: arXiv:2611.22245 [cs.AI]
  (or arXiv:2611.22245v1 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2611.22245

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Hejia Zhang [view email]
[v1] Tue, 17 Nov 2026 05:47:01 UTC (1,612 KB)

来源:arXiv cs.AI · arxiv.org