像 GPT 这样的 LLM 能否像人类一样从错误中吸取教训并自我反省?
答案是肯定的,该方法使 HotPotQA 的准确率提高了 33-35%!
Lior⚡: Can LLMs like GPT learn from their mistakes and self-reflect just like humans?
The answer is yes and the method resulted in 33-35% accuracy improvement on HotPotQA!
Paper:
Code:
https://arxiv.org/abs/2303.11366 https://github.com/noahshinn024/reflexion