On prompt tokens, Qwen3-VL’s selected routing weights are substantially flatter than the text model’s: its top three hold less of the selected-eight mass, and the selected eight hold less of the full 128-way softmax mass.

On prompt tokens, Qwen3-VL’s selected routing weights are substantially flatter than the text model’s: its top three hold less of the selected-eight mass, and the selected eight hold less of the full 128-way softmax mass. · CiteArk

测量结果

已评估 0/6 项

获得支持 0有争议或冲突 0尚未评估 6

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
printedValue: 49.8	49.8%	—	尚未评估
printedValue: 59.8	59.8%	—	尚未评估
printedValue: 50.2	50.2%	—	尚未评估
printedValue: 40.2	40.2%	—	尚未评估
printedValue: 15.5	15.5%	—	尚未评估
printedValue: 44.3	44.3%	—	尚未评估

On prompt tokens, Qwen3-VL’s selected routing weights are substantially flatter than the text model’s: its top three hold less of the selected-eight mass, and the selected eight hold less of the full 128-way softmax mass.

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果