Polymorphism Is Rotation: Operational Mechanistic Interpretability from a Two-Layer Transformer to Pythia-70m
Researchers have discovered a phenomenon called "polymorphism" in independently trained transformers, where they compute the same function but use different internal coordinate systems that are rotated versions of each other. This rotation, which is uniformly random within SO(d_model), makes the internal representations unintelligible between models. However, a single matrix multiplication using an orthogonal Procrustes fit can align these bases, allowing for the transfer of feature dictionaries and steering vectors between models without retraining. AI
IMPACT Reveals that independently trained models can compute identical functions through rotated internal representations, suggesting potential for cross-model transferability of learned features.