I know in practice using transformers with pure RL (rather than using RL to “polish” the model as in RLHF) doesn’t work super good. But conceptually, why not? It’s not obvious to me why it shouldn’t…

Posters: eshear. eshear

4 likes recorded by the source corpus.