y
k
=
b
u
k
Σ
l
b
u
l
=
e
u
k
/τ
Σ
l
e
u
l
/τ
gain b
2.72
equivalent τ = 1/ln b
1.00
winner share
42%
regime
balanced
input logits u
k
output probabilities y
k
sorted gradient
two near-tied
nearly equal
randomize
b → 1 (no gain)
b = e (standard)
b → ∞ (WTA)