A simple note about Chain of Thought (COT)
What is Prompting?
A method to guide language models (LMs) to generate desired outputs by framing the input in a specific way.
- Zero-Shot Promoting: Directly asked to perform a task without being shown any examples.
- One-Shot Promoting: Given one example of the task to guide its response.
- Few-Shot Promoting: Provided with multiple examples to better learn the task pattern.
What is Chain of Thought (CoT) ?
Encourages models to generate intermediate reasoning steps rather than directly outputting the answer.
Enhances performance in tasks requiring multi-step reasoning, like arithmetic or logical problems.
Core idea: Not just output the answer, but the reasoning process.
Zero-Shot-COT
加个think step by step之类的引导思维链,反正比Zero-Shot好
Auto-CoT
预设几个问题类型的思维类,然后把实际问题聚类过去,使用这个思维链
APE
通过自评分自动的探索LLM的最佳COT引导
COT-SC
生成多个答案,选比较一致的结果(多)
思维树
COT-SC是链,这里扩展成树,多个候选项