Z.Y. Su

Enhancing on-policy reinforcement learning can paradoxically reduce the diversity of successful behaviors, leading to a trade-off that challenges future trainability.

Language models are increasingly doing their real work in the "invisible" latent space, not the tokens we see.