Large Language Models are Better Reasoners with Self-Verification

This is the official implementation of Large Language Models are Better Reasoners with Self-Verification.

(EMNLP 2023 Findings)

Demo

https://github-production-user-asset-6210df.s3.amazonaws.com/64484703/250267125-281968bf-a8a2-438f-838b-d189c2501d40.mp4 https://user-images.githubusercontent.com/70048414/232352935-55c6bf7c-3958-406e-8610-0913475a0b05.mp4

Installation

Make sure you have Python>=3.8 installed on your machine.

pip install torch==1.8.2+cu111 torchtext==0.9.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
pip install tqdm transformers sklearn pandas numpy sentencepiece openai

Set your OpenAI API key

# https://beta.openai.com/account/api-keys
export OPENAI_API_KEY=(YOUR OPENAI API KEY)

Set arguments.

model=CODEX # {"gpt3", "gpt3-medium", "gpt3-large", "gpt3-xl", "CODEX", "CODEX-001"}. "codex" is the smallest model.
dataset=multiarith # We can use other datasets. See help for the details.
api_time_interval=4.0 # Caution. The API allows users request API up to 20 times in a minutes, otherwise errors happen.

Quick Start

Demo

python demo.py

Self-Verification (our proposal)

python main.py --method=verifier_cot --model=${model} --dataset=${dataset}

CoT

# MultiArith and GSM8K are currently available.
python main.py --method=few_shot_cot --model=${model} --dataset=${dataset}

Method

1. Forward Reasoning, the LLM generates candidate thought chains and conclusions for a given problem text;
2. Backward Verification, we use the LLM to verify whether the conditions meet the candidate conclusions and rank the candidate conclusions based on a verification score.

Cite

@misc{weng2023large,
      title={Large Language Models are Better Reasoners with Self-Verification}, 
      author={Yixuan Weng and Minjun Zhu and Fei Xia and Bin Li and Shizhu He and Kang Liu and Jun Zhao},
      year={2023},
      eprint={2212.09561},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}