Rethinking with Retrieval Faithful Large Language Model Inference
概
LLM (Large Language Model) + 检索.
Rethinking with retrieval (RR)
-
CoT (Chain of thought) 已经被证明是一种挖掘 LLM 能力的有效方式, 但是, 这种方式还是容易发生 LLM 胡编乱造的情况. RR 就是希望 LLM 在思考的时候能够有一个理性的参照以获得更好的更准确的结果 (上图 (c)).
-
让我们以 "Did Aristotle use a laptop?" 来分析这个流程.
-
首先, CoT 的做法是希望 LLM 将这个问题分解然后一步步的解决, 理想的步骤是:
- "Aristotle died in 322 BC.";
- "The first laptop was invented in 1980.";
- "So the answer is no.".
-
但是, 其中涉及到很多的 '常识', 倘若 LLM 本身并没有很好学习这些知识, 即使整体的步骤是对的, 也有可能回答错误. 在实际中可能会产生如下的几种回答 (加粗的是错误的):
- [R1] "Aristotle died in 2000. The first laptop was invented in 1980. Thus, Aristotle used a laptop. So the answer is yes."
- [R2] "Aristotle died in 322BC. The first laptop was invented in 2000. Thus, Aristotle did not user a laptop. So the answer is no."
- [R2] "Aristotle died in 322BC. The first laptop was invented in 1980. Thus, Aristotle did not user a laptop. So the answer is no."
-
甚至, 有些时候, 回答是对的, 但是所依据的常识是错误的 (R2).
-
所以, 我们需要借助外部的数据库, 比如 Wikipedia. 对于上面的问题, 涉及到两个知识点:
- [K1] Aristotle (384–322 BC) was a Greek philosopher and polymath during the Classical period in Ancient Greece. ...
- [K2] The Epson HX-20, the first laptop computer, was invented in 1980. ...
-
RR 所依赖的 Faithful inference 可以这样定义. 对于某个问题 \(Q\), 产生了一组思维链 \(R_1, R_2, \ldots, R_N\) 以及它们所对应的答案 \(P_1, P_2, \ldots, P_N\), faithful inference 定义为:
\[R^* := \mathop{\arg\max} \limits_{R_i, i=1,2,\ldots, N} \mathbb{I}[P_i = P] f_{\mathcal{KB}}(R_i). \]其中 \(f_{\mathcal{KB}}: R \rightarrow \mathbb{R}^+\) 评估思维链 \(R\) 和外部知识库的吻合程度.
代码
[official]