Concepts

Candidate Model

The model whose internal representations you want to study.

Reference Model

A model used as an external comparison target. In repralign, a reference model is typically used in one of two roles:

• semantic reference
• generation-oriented reference

Layer-Wise Alignment

For each chosen candidate layer, repralign extracts features, compares them against one or more reference features, and writes the resulting scores as a curve across depth.

Pooling

Many hooked layers produce token-level or spatial features. repralign reduces these to sample-by-feature matrices using one of:

• cls
• mean_tokens
• flatten_mean

Metrics

CKA

Linear CKA measures whether two feature spaces preserve similar pairwise structure across samples.

CKNNA

CKNNA in repralign measures agreement between top-k nearest-neighbor structure induced by two feature spaces.