AI as Normal Technology· Sayash Kapoor·· 2026-02-24精选AI 评分70
AI as Normal Technology 团队发布论文《Towards a Science of AI Agent Reliability》
New Paper: Towards a science of AI agent reliability
AI 导读
Sayash Kapoor、Arvind Narayanan 与 Stephan Rabanser 等人发布 66 页论文《Towards a Science of AI Agent Reliability》，将智能体可靠性拆解为一致性、鲁棒性、可预测性、安全性四维共 12 项指标。
推荐理由
论文把智能体可靠性拆成12个维度并实测14个模型，显示18个月内准确率大涨而可靠性仅小幅提升。
来源：AI as Normal Technology · normaltech.ai