Tag
该论文提出了一个受扰动控制的医疗LLM智能体基准测试，包含400个多轮episode，评测在缺失、过时、冲突和迟到信息下的升级/拒答决策能力。结果显示模型在干净环境下准确率达0.97，但在不确定性扰动下漏掉32-97%的必要升级，且基于词汇规则的控制器在未见过的同义改写下召回率降至0。
The article discusses the problem of AI agents that fail silently by not escalating when stuck, and suggests implementing explicit checks or circuit breakers to improve reliability in production.
Fortune reports on OpenAI's unusual management system, including a 'friction@' inbox used as an escalation path to maintain startup-like speed, with involvement from Sam Altman and Greg Brockman.
The article identifies a common failure in AI agent handoffs to humans on WhatsApp, where the bot says it will transfer but no human responds, breaking trust. It outlines a solution with mode tracking, history injection, and real task creation.