From Directions to Regions: Decomposing Activations in Language Models via Local Geometry - Explained Simply | ArXiv Explained