RT-Transformer: The Transformer Block as a Spherical State Estimator
Provides a theoretical grounding for Transformer architecture, potentially influencing future model design for ML practitioners.
The paper shows that the Transformer block's core components (attention, residual connections, normalization) emerge naturally from a geometric estimation problem on the hypersphere, unifying them under a single principled framework.
We show that the core components of the Transformer block -- attention, residual connections, and normalization -- arise naturally from a single geometric estimation problem. Modeling the latent state as a direction on the hypersphere, with noise defined in the tangent plane at the current estimate, yields a precision-weighted directional inference procedure in which attention aggregates evidence, residual connections implement incremental state updates, and normalization retracts the updated state back onto the hypersphere. Together, these components follow from the geometry of the estimation problem rather than being introduced as independent architectural choices.