On MATH-500 and Countdown for both SPG and GDPO, IM leaves mean gradient norm similar or somewhat larger rather than systematically shrinking it, while reducing the coefficient of variation in all four framework–task pairs.

On MATH-500 and Countdown for both SPG and GDPO, IM leaves mean gradient norm similar or somewhat larger rather than systematically shrinking it, while reducing the coefficient of variation in all four framework–task pairs. · CiteArk

测量结果

已评估 0/24 项

获得支持 0有争议或冲突 0尚未评估 24

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
model: SPG · dataset: MATH-500	0.1719 gradient-norm units	—	尚未评估
model: SPG · dataset: MATH-500	0.0798 gradient-norm units	—	尚未评估
model: SPG · dataset: MATH-500	0.46 ratio	—	尚未评估
model: SPG+IM · dataset: MATH-500	0.1707 gradient-norm units	—	尚未评估
model: SPG+IM · dataset: MATH-500	0.0634 gradient-norm units	—	尚未评估
model: SPG+IM · dataset: MATH-500	0.37 ratio	—	尚未评估
model: GDPO · dataset: MATH-500	0.0357 gradient-norm units	—	尚未评估
model: GDPO · dataset: MATH-500	0.0174 gradient-norm units	—	尚未评估
model: GDPO · dataset: MATH-500	0.49 ratio	—	尚未评估
model: GDPO+IM · dataset: MATH-500	0.0414 gradient-norm units	—	尚未评估
model: GDPO+IM · dataset: MATH-500	0.0172 gradient-norm units	—	尚未评估
model: GDPO+IM · dataset: MATH-500	0.42 ratio	—	尚未评估

On MATH-500 and Countdown for both SPG and GDPO, IM leaves mean gradient norm similar or somewhat larger rather than systematically shrinking it, while reducing the coefficient of variation in all four framework–task pairs.

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果