Why it matters

DyT is emerging norm alternative. Understanding shapes future.

Advertisement

The architecture

y = tanh(alpha * x) * gain + bias.

Alpha: learnable scale.

DyT structureLearn alphaper-channel scaletanh(alpha * x)bounded outputGain + biasaffineRecent 2025 research; adoption early but promising
DyT computation.
Advertisement

How it works end to end

Bounded output via tanh.

No statistics computation (mean, var).

Learnable scale.