AI 导读
Agnes 2.5 Pro Beta 的 AA-Omniscience 分数从 -25 提升到 -11，但这一提升来自弃答而非准确率改善。它只尝试回答了 45% 的问题，而 Agnes 2.5 Pro Alpha 为 94%；其幻觉率从 88% 降至 33%，但 AA-Omniscience 准确率从 33% 腰斩至 17%。
正文
Agnes 2.5 Pro Beta's AA-Omniscience score improves from -25 to -11, however the improvement comes from abstention rather than accuracy improvement. It attempted only 45% of questions compared to 94% for Agnes 2.5 Pro Alpha, and while its hallucination rate improves from 88% to 33%, its AA-Omniscience Accuracy halves from 33% to 17%.
来源：@ArtificialAnlys · x.com