[Summary] Mini-vec2vec: Scaling Universal Geometry Alignment with Linear Transformation
TL;DR Mini-vec2vec is an unsupervised framework for aligning different embedding spaces without paired data. It transforms vectors between models (e.g., BERT to T5) using a linear assumption and iterative refinement. This approach achieves alignment quality comparable to the popular CycleGAN-based vec2vec method while reducing computational overhead. Motivation Deep learning models often converge toward a geometric structure representing the underlying data manifold. However, direct comparison across models is usually impossible without expensive supervised re-training or ground-truth pairs....