The study compares nine attention arms that separately or jointly supply an exact or partial zero option and a value-path filter; scalar per-key value gates are mathematically decomposed into renormalized routing plus a per-query attenuation factor. · CiteArk
The study compares nine attention arms that separately or jointly supply an exact or partial zero option and a value-path filter; scalar per-key value gates are mathematically decomposed into renormalized routing plus a per-query attenuation factor.
来源：fixed-paper:PDF p. 3, Sections 3.2–3.3; Table 4 on p. 13
