Blake Aaron Richards

We present NuCLR, a self-supervised framework that learns high-quality, population-aware neuron-level embeddings directly from spike train data using a spatio-temporal transformer and tailored contrastive loss.

While loss decreases monotonically during LLM training, the representations undergo distinct geometric phases across pretraining and post-training, which in turn determine when and how the model acquires memorization or generalization capabilities.