SLERP (Spherical Linear Interpolation)

Interpolates between two normalized model weight vectors $\theta_0, \theta_1$ along the geodesic arc on the hypersphere rather than a straight chord (preserving weight-norm statistics across transformer layers): $\text{SLERP}(\theta_0, \theta_1; t) = \frac{\sin((1-t)\Omega)}{\sin\Omega}\theta_0 + \frac{\sin(t\Omega)}{\sin\Omega}\theta_1, \quad \cos\Omega = \frac{\langle \theta_0, \theta_1\rangle}{\|\theta_0\|\|\theta_1\|}$