GeoPair Provides Training-Free Transformer Compression via Cross-Layer Factorization
September 21, 2026
GeoPair optimizes cross-layer weight pairings by preserving layer-specific activation geometries rather than using heuristic grouping. This training-free framework utilizes structured sparsity and shared-dictionary factorizations to compress transformers without sacrificing functional fidelity.
HOW THIS AFFECTS YOU
●
builderYou can compress transformer models post-training without the high cost of retraining or fine-tuning.
●
researcherYou can explore more principled ways to reduce parameter counts by respecting activation geometry.