1 article
New research identifies how reward model uncertainty and diversity collapse undermine LLM alignment via reinforcement learning.