AI 导读
Step 5 Preview 以 600B 总参数量在 AA-Omniscience 上达到 42% 准确率，领先 GLM-5.3（max，34%，753B），落后于 Kimi K3（max，48%，2.8T），略高于事实回忆随参数量扩展的趋势。限制因素在于行为而非知识：它在 43% 的尝试中产生幻觉，而非拒绝回答，这使其 AA-Omniscience Index 停留在 16
正文
Step 5 Preview reaches 42% accuracy on AA-Omniscience at 600B total parameters, ahead of GLM-5.3 (max, 34%, 753B) and behind Kimi K3 (max, 48%, 2.8T), slightly above the trend of factual recall scaling with parameter count. The limiter is behavior rather than knowledge: it hallucinates on 43% of attempts rather than declining to answer, which holds its AA-Omniscience Index at 16
来源：@ArtificialAnlys · x.com