TokenTown visualizes a transformer language model as an isometric city, with each district representing a stage of the model. A convoy transports a hidden state through various processes, including tokenization, vector casting, and layer processing, ultimately producing a probability distribution for token selection.

Training a single transformer layer can achieve performance comparable to full-parameter reinforcement learning (RL) training. This finding suggests that RL adaptation may not require uniform updates across all layers of large language models.