引用rodoxcasta · 2023-05
“Actually, there's some experimental evidence that GPT4 have a Theory of Mind as good as humans, maybe better.”
称有实验证据表明 GPT-4 的心智理论不亚于人类,甚至可能更好
用少样本思维链等提示显著提升大模型的心智理论得分。
测量 GPT-4 和三个 GPT-3.5 变体(Davinci-2、Davinci-3、GPT-3.5-Turbo)在心智理论(ToM)任务上的表现,并考察情境学习能否提升。使用的提示包括两样本思维链和“逐步思考”指令。结果显示,经 RLHF 训练的模型(除 Davinci-2 外)都能借情境学习提高 ToM 准确率;GPT-4 零样本接近 80%,仍低于测试集上 87% 的人类准确率,而加入提示后所有 RLHF 模型都超过 80%,GPT-4 达到 100%。
在讨论 GPT-4 是否有心智理论、情商高不高时,评论者引它作为“GPT-4 的心智理论表现接近甚至不输人类”的实验证据。
“Actually, there's some experimental evidence that GPT4 have a Theory of Mind as good as humans, maybe better.”
“emotional intelligence seems to be very high by the looks of things”