1 paper touches this idea.
Papers
One Transformer Layer Can Match Full RL Training
Related concepts
Ideas that show up alongside Parameter-Efficient Training.