interpretability
fact
neutral
Vision Transformers' geometric evolution of internal representations during training remains insufficiently understood, with existing analyses primarily focusing on attention mechanisms rather than representation geometry
Machine Learning27 Jul 2026