So far, scientists have relied on positive reinforcement learning to train LLMs, but the opposite seems to be giving much ...