Disentangling Representation Evolution in Transformers through Directional Decomposition
AI architecture representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path. We then apply th
Reading in Plain English mode — toggle below to switch
AI architecture representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path. We then apply th
This research provides a novel framework for understanding how information flows and transforms within Transformer layers, offering insights into their internal mechanisms. For ML engineers, this deeper understanding can lead to more effective model debugging, targeted architectural improvements, and potentially new methods for model editing or fine-tuning.