基于结果的RL难提升思维链忠实度

Somewhat surprising that faithfulness of chain-of-thought doesn't improve much with outcome-based RL

来源：Jan Leike · x.com