问HN:推理是否仅仅是某种迭代的上下文优化?
上下文 = "你是一个有用的模型。你在初始用户输入后评估自己的输出,持续 N 步"
提示 = "用户发来的提示"
上下文 += 提示
对于 _ 在范围(N)内:
上下文 += 评估_llm(上下文)
查看原文
context = "you are a useful model. you evaluate your own output for N steps after the initial user input"<p>prompt = "here comes a prompt by the user"<p>context += prompt<p>for _ in range(N):
context += evaluate_llm(context)