arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2511.09710cs.AI
收藏

回声:当大语言模型代理相互交谈时的身份失败

Echoing: Identity Failures when LLM Agents Talk to Each Other

Salesforce AI Research(Salesforce AI研究)

Sarath Shekkizhar, Romain Cosentino, Adam Earle, Silvio Savarese

AI总结:

研究发现LLM代理在相互对话中出现身份失败现象,即回声,通过实验表明回声率高达70%,并提出结构化响应作为缓解措施。

摘要:

当基于大语言模型(LLM)的代理自主相互交互时,出现了一种新的故障类型,这种故障无法从单个代理的性能中预测出来:代理-代理对话中的行为漂移(AxA)。不同于人类-代理交互,其中人类为对话提供基础并引导对话,AxA缺乏这种稳定信号,使这些故障变得独特。我们研究了一种此类故障,即回声,其中代理们放弃其分配的角色,转而模仿对话伙伴,从而破坏了其预期的目标。通过在66种AxA配置、4个领域(3个事务性领域,1个咨询性领域)和2500多次对话(超过250000次LLM推理)中的实验,我们发现回声在主要的LLM提供商中普遍存在,回声率高达70%,具体取决于模型和领域。此外,我们发现回声在高级推理模型中也是持续存在的,即使这些模型的回声率(32.8%)较高,且不会因推理努力而减少。我们分析了提示和对话动态,表明回声随着交互时间的增长(7+个代理回合)而出现,并非仅仅是次优化实验设计的产物。最后,我们引入了一种协议层面的缓解措施,即有针对性地使用结构化响应,将回声率降低到9%。

英文摘要:

As large language model (LLM) based agents interact autonomously with one another, a new class of failures emerges that cannot be predicted from single agent performance: behavioral drifts in agent-agent conversations (AxA). Unlike human-agent interactions, where humans ground and steer conversations, AxA lacks such stabilizing signals, making these failures unique. We investigate one such failure, echoing, where agents abandon their assigned roles and instead mirror their conversational partners, undermining their intended objectives. Through experiments across $66$ AxA configurations, $4$ domains (3 transactional, 1 advisory), and $2500+$ conversations (over $250000$ LLM inferences), we show that echoing occurs across major LLM providers, with echoing rates as high as $70\%$ depending on the model and domain. Moreover, we find that echoing is persistent even in advanced reasoning models with substantial rates ($32.8\%$) that are not reduced by reasoning efforts. We analyze prompt, conversation dynamics, showing that echoing arises as interaction grows longer ($7+$ agent turns) and is not merely an artifact of sub-optimal experiment design. Finally, we introduce a protocol-level mitigation where targeted use of structured response reduces echoing to $9\%$.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑