SAFARI: Scaling Long Horizon Agentic Fault Attribution via Active Investigation
SAFARI: 通过主动调查扩展长时域智能体故障归因
机构 * Department of Engineering Sciences(工程科学系) ; Applied Mathematics, Northwestern University(应用数学,西北大学)
专题命中 规划决策 :agentic(title,abstract);agent(abstract);autonomous agent(abstract);multi-agent(abstract)
AI总结 提出SAFARI框架,用工具增强的诊断循环替代线性上下文加载,结合短期记忆解耦诊断准确性与上下文限制,在Who&When和TRAIL GAIA数据集上分别提升20%和19%,并在超出上下文窗口5倍时保持0.58精度。
Comments Published at the Second Workshop on Agents in the Wild: Safety, Security, and Beyond (AIWILD) at ICML 2026