72% Solved. 48% Actually Merged. The Gap Nobody Was Measuring.
中文摘要
新基准显示AI智能体实际完成率远低于其“解决率”。现有评估模型的方法未能准确衡量智能体性能，凸显衡量标准缺陷。
English Summary
New benchmarks reveal a significant gap: AI agents' actual completion rates are much lower than their 'solved' rates. Current metrics fail to accurately grade agent performance, highlighting a critical measurement flaw.
原文节选
Everyone knows how to grade a model. Almost nobody knows how to grade an agent — and the benchmarks just proved it. Continue reading on AI Advances »