Debate Training Reduces Reward Hacking in RLAIF
We demonstrate that RL finetuning an LLM using debate, a two-player adversarial game between a generator and a critic adjudicated by a weaker LLM judge, reduces reward hacking compared to a reinforcement learning from AI feedback (RLAIF) baseline. Reward hacking is a central obstacle in RLAIF: as training progresses, t...
Zachary Kenton, Lili Janzer, Rory Greig et al.
· 1 citation