CoT进阶
- 一:Self Consistency
- 1.1 方法简介
- 1.2 实验
- 1.3 结果
- 二:Least-to-most
- 2.1 方法简介
- 2.2 示例
- 2.3 结果
一:Self Consistency
题目: SELF-CONSISTENCY IMPROVES CHAIN OF THOUGHT REASONING IN LANGUAGE MODELS
机构:Google Brain, ICLR 2023
论文: https://arxiv.org/pdf/2203.11171.pdf
任务: 对于复杂问题而言,往往可以从多条推理路径得到最终的答案,因此将原来的CoT贪心解码进行优化,提出一种Self Consistency的解码算法
特点: sample-and-marginalize,投票,能够避免CoT的解码的局部最优以及输出重复,可以视作一种“self-ensemble”,无需训练/标注/微调,很容易与现存的采样算法,比如 temperature sampling, top-k sampling,nucleus sampling即插即用。
前置相关工作:CoT
1.1 方法简介
- 利用CoT prompting大模型
- 将CoT中的贪心解码替换为采样生成一组推理路径
- 答案一致性投票
关于NLG的各种采样算法:Greedy Search (Maximization),Beam Search,Temperature Sampling,Top-K Sampling,Top-P Sampling (Nucleus sampling),可以参见:
- https://www.cnblogs.com/miners/p/14950681.html
- THE CURIOUS CASE OF NEURAL TEXT DeGENERATION
1.2 实验
- Arithmetic Reasoning
- Commonsense and Symbolic Reasoning
- SELF-CONSISTENCY HELPS WHEN CHAIN-OF-THOUGHT HURTS PERFORMANCE
- Comparison to Sample-and-Rank
- Comparison to Beam Search
- Comparison to Ensemble-based Approaches
- Self-Consistency is Robust to Sampling Strategies and Scaling
- Self-Consistency Improves Robustness to Imperfect Prompts
- Self-Consistency Works for Non-Natural-Language Reasoning Paths and Zero-shot CoT
1.3 结果
二:Least-to-most
题目: LEAST-TO-MOST PROMPTING ENABLES COMPLEX REASONING IN LARGE LANGUAGE MODELS
机构:Google Brain, ICLR 2023
论文: https://arxiv.org/pdf/2205.10625.pdf
任务: 为了克服CoT在easy-to-hard示例学习中的泛化性
方法:将复杂的问题分解为一系列的更简单的子问题,然后一个接一个解决,每一个待解决的子问题,都会被上一个已经解决好的子问题的答案促进
特点: 方法中的两个阶段都是通过几次提示(few-shot prompting)来实现的,因此在任何阶段都不需要训练或微调
前置相关工作:CoT,self consistency
2.1 方法简介
为了解决easy-to-hard的泛化性问题,提出Least-to-most prompting方法,它包含两个阶段:
- 将一个复杂的问题,分解为一序列简单的子问题
- 依次解决这些子问题,每一个待解决的子问题,都需要历史已经解决的子问题的答案来促进
2.2 示例
该论文在SYMBOLIC MANIPULATION,COMPOSITIONAL GENERALIZATION,MATH REASONING进行了实验,这儿展示MATH REASONING的示例以及结果
Least-to-most样例:
CoT样例: