New preprint: Is One Layer Enough? Training a single transformer layer can match full-parameter RL training.