Recent research into RL training reveals that updating just one transformer layer can replicate the performance gains of full-parameter training. This discovered structural pattern shows that AI model improvements concentrate in specific middle layers rather than across the entire stack.