Tracing the Representation Geometry of Language Models from Pretraining to Post-training | Read Paper on Bytez