Model Interpolation
Interpolation is the most fundamental operation in Model Morphing: blending two checkpoints' weights into a new set of weights, parameterized by how much of each source to keep.
Linear vs. spherical
Two checkpoints can be blended two ways:
- Linear interpolation - a straight-line blend,
(1 − t) * A + t * B. Simple, but a straight line through high-dimensional weight space doesn't always correspond to a meaningful intermediate model - weight space isn't guaranteed to behave like an ordinary flat space along that path. - Spherical interpolation (SLERP) - a blend along the shorter arc of a hypersphere connecting the two weight vectors, which tends to better respect weight space's actual geometry, particularly when the two source checkpoints are far apart.
Both are exposed as morph-DAG node types (see Model Morphing); which one is appropriate depends on how similar the two source checkpoints already are - closer sources make the linear/spherical distinction matter less.
The interpolation coefficient
The coefficient t (0 = entirely the first source, 1 = entirely the second) is a single scalar you set exactly, mid-operation, the same way a modal transform tool lets you type an exact numeric value rather than only dragging - precision matters here because a small change in t can correspond to a meaningfully different resulting model.
t doesn't have to be uniform across the whole model - a morph DAG can apply a different coefficient per layer or module, which is how a task-vector-style merge (see Model Morphing) achieves "blend the parts that matter, leave the rest close to the base" rather than one global blend ratio.
Previewing before materializing
Because a morph is an expression DAG, not committed weights until materialized, trying several interpolation coefficients is cheap - adjust t, inspect a preview of the result's statistics or a quick diff against either source, and only materialize once a coefficient looks right.
Related resources
- Model Morphing - the DAG/materialize model interpolation operates within.
- SVD, PCA and Effective Rank - relevant when an interpolation is effectively low-rank by construction.
- Merge Conflicts - interpolation as one possible resolution for a numerically resolvable conflict.