RADAR: Relative Angular Divergence Across Representations

In three linesRADAR is a geometric metric for estimating cross-domain transferability in foundation models. It analyzes layer-wise representation evolution by measuring angular alignments and distance changes. Evaluated on cross-lingual sentiment classification and cross-domain image classification, RADAR achieves competitive performance against existing transferability metrics.

Summary generated by Claude — human-verified