Tracing the Representation Geometry of Language Models from Pretraining to Post-training

Open in new window