Dawn Song, this view really hits hard—just looking at the scores can indeed make you miss the real problems.

Research by Berkeley AI and Dawn Song emphasizes that when evaluating intelligent agents, the focus should be on understanding the specific reasons for failure, rather than just the benchmark scores. Long-term failures should be broken down into diagnosable patterns to more accurately locate and analyze where and why the agent fails. The original text does not provide information about specific benchmarks, analysis details, or failure mode classifications.