Larger baselines build a stronger, more nearly zero-valued first-position attention sink, while an explicit sink logit moves parked mass to an exact zero-valued phantom and restores the first-position read norm; the per-key norm gate leaves much of the baseline sink in place.

Larger baselines build a stronger, more nearly zero-valued first-position attention sink, while an explicit sink logit moves parked mass to an exact zero-valued phantom and restores the first-position read norm; the per-key norm gate leaves much of the baseline sink in place. · CiteArk

测量结果

已评估 0/14 项

获得支持 0有争议或冲突 0尚未评估 14

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
arm: baseline · tier: 124M	0.06 fraction	—	尚未评估
arm: baseline · tier: 124M · denominator: ordinary read norm	0.42 ratio	—	尚未评估
arm: sink logit · tier: 124M	0.005 fraction	—	尚未评估
arm: sink logit · tier: 124M · denominator: ordinary read norm	0.99 ratio	—	尚未评估
arm: norm gate · tier: 124M	0.037 fraction	—	尚未评估
arm: norm gate · tier: 124M · denominator: ordinary read norm	0.65 ratio	—	尚未评估
arm: baseline · tier: 350M	0.11 fraction	—	尚未评估
arm: baseline · tier: 350M · denominator: ordinary read norm	0.36 ratio	—	尚未评估
arm: sink logit · tier: 350M	0.007 fraction	—	尚未评估
arm: sink logit · tier: 350M · denominator: ordinary read norm	0.98 ratio	—	尚未评估
arm: sink logit · tier: 350M	0.53 fraction	—	尚未评估
arm: sink logit · tier: 10M	0.13 fraction	—	尚未评估

Larger baselines build a stronger, more nearly zero-valued first-position attention sink, while an explicit sink logit moves parked mass to an exact zero-valued phantom and restores the first-position read norm; the per-key norm gate leaves much of the baseline sink in place.

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果