Tag
The paper introduces GeoPair, a training-free framework for transformer compression that optimizes cross-layer factorizations while preserving activation geometries, achieving state-of-the-art results across diverse architectures.