Is one layer enough? Training a single transformer layer can match full-parameter RL training