研究显示辩论能提升人类判断准确率，但弱AI模型无法自我纠错

arXiv 论文《Humans Correct Their Judgments Through Debate, but Weaker AI Models Do Not》研究 AI Safety via Debate 协议能否让人类与 LLM 监督者在存在先验信念的情况下接近真相。