1 paper
Sidharth Mudgal, Jong Lee, Harish Ganapathy +10
KL-regularized reinforcement learning (RL) is a popular alignment framework to control the language model responses towards high reward outcomes. We pose a tokenwise RL objective a…