Amazon Science

A better training method for reinforcement learning with human feedback - Amazon Science

Contrasting training pairs with large reward differences mitigate spurious correlations and improve performance of direct-alignment algorithms by as much as 20%–40%.