For modular multiplication at p=97 with a two-layer transformer and Adam, Grokfast reaches 95% accuracy sooner than exact-SVD EGD in both steps and wall-clock time. RSVD rank 36 is closer to Grokfast but remains slower by the printed values; all three accelerated methods finish at 100% validation accuracy, versus 99.1% for vanilla Adam.

For modular multiplication at p=97 with a two-layer transformer and Adam, Grokfast reaches 95% accuracy sooner than exact-SVD EGD in both steps and wall-clock time. RSVD rank 36 is closer to Grokfast but remains slower by the printed values; all three accelerated methods finish at 100% validation accuracy, versus 99.1% for vanilla Adam. · CiteArk

测量结果

已评估 0/22 项

获得支持 0有争议或冲突 0尚未评估 22

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
指标与条件	论文报告	最近可用实测	证据判断
method: Vanilla Adam	55,780 optimization steps	—	尚未评估
method: Vanilla Adam	5,198.7 seconds	—	尚未评估
method: Vanilla Adam	93.2 milliseconds	—	尚未评估
method: Vanilla Adam	99.1%	—	尚未评估
method: Grokfast	910 optimization steps	—	尚未评估
method: Grokfast	61.3 fold	—	尚未评估
method: Grokfast	114 seconds	—	尚未评估
method: Grokfast	45.6 fold	—	尚未评估
method: Grokfast	125.3 milliseconds	—	尚未评估
method: Grokfast	100%	—	尚未评估
method: EGD (SVD)	1,170 optimization steps	—	尚未评估
method: EGD (SVD)	47.7 fold	—	尚未评估

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果