Cross-Model Representational Alignment via Generalized Procrustes Analysis
October 27, 2026
Researchers aligned activations from five frozen, independently trained models, including transformers and a state-space model, without weight merging. Using Generalized Procrustes Analysis, they achieved 0.91 to 0.93 retrieval accuracy in a shared representational space across diverse architectures.
HOW THIS AFFECTS YOU
●
builderYou can build cross-model tools that operate on a shared embedding space without retraining weights.
●
researcherThis demonstrates that heterogeneous model architectures can be mapped to a unified activation space.