Progressive-Hint Prompting Improves Reasoning in Large Language Models

Figure 1: Progressive-Hint Prompting (PHP) interacts with LLM.

PHP: Simple and Effective for improving LLM reasoning ability.
Chuanyang Zheng, Zhengying Liu, Enze Xie, Zhenguo Li and Yu Li
Preprint Paper

This repository contains the official Pytorch implementation of code for PHP. PHP is a simple, effective and powerful method, as shown in Figure 1. It is can be easily combined with preivious works such as Chain-of-Thought and Self-Consistency, as they are orthogonal.

News

• [06/05/2023]: With PHP and Self-Consistency (K=40), we achieve new SOTA of GSM8K dataset: 96.5!
• [05/11/2023]: We update the code and data of MATH dataset
• [05/02/2023]: We achieve the SOTA performance 53.9% on MATH dataset, which is the currently most challenging reasoning dataset! We will update our paper and the code soon! Thank you for your attention!!!
• [04/25/2023]: We update the dataset!
• [04/21/2023]: We upload the code!

PaperWithCode Leaderboard

We achieve the SOTA performance on AQuA, SVAMP, GSM8K and MATH dataset, as the shown in SVAMP Leaderboard, GSM8K Leaderboard and MATH Leaderboard (Date: May 05, 2023)

We achieve the SOTA performance on GSM8K dataset.

We achieve the SOTA performance on MATH dataset.

Installation

pip install jsonlines
pip install openai

Usage

The code needs to be configued witt your account' secret key which is available on website. Set openai.api_key to its value:

import openai
openai.api_key = "sk-..."

Run

We run the main_clean with the following:

python python main_clean.py --prompt_dir [base prompt] --eng [openAI model] --seed [seed number] --hint [PHP prompt] --dataset [datasetname] --num_test -1 --q1 [ori: standard or CoT, complex: complex CoT] --q2 [ori: standard or CoT, complex: complex CoT] --sample [sample number] --temp [0.0 for greedy, 0.7 for sc]

Or, we can just use the file in bash directory:

bash bash/cot.sh

Result

We Only show Complex CoT with GPT-3.5-Turbo and GPT-4 Here. For more experiments results (such as text-davinci-002 and text-davinci-003), please refer to our Paper.

Table 8: Progressive-Hint Prompting (PHP) with GPT-3.5-Turbo and GPT-4. PHP works better when the model is more powerful.

Citation

@article{zheng2023php,
  title={Progressive-Hint Prompting Improves Reasoning in Large Language Models},
  author={Zheng Chuanyang, Liu Zhengying, Xie Enze, Li Zhenguo, Li Yu},
  journal={arXiv preprint arXiv:2304.09797},
  year={2023}
}