Training one transformer layer can match full-model reinforcement learning

Original: What if most RL gains come from 1 transformer layer?

Source: x.com ↗