This repository is the official implementation of our paper DetermLR: Augmenting LLM-based Logical Reasoning from Indeterminacy to Determinacy, which has been accepted at ACL 2024 main.
pip install -r requirements.txt
Note: We recommend using guidance-0.0.64 for adopting GPT-3.5-turbo or GPT-4.
Please set your openai api key in os.environ["OPENAI_API_KEY"]="YOUR_API_KEY"
.
Make sure that your device is able to connect to OpenAI API.
Quick start for implementing DetermLR to solve four logical reasoning tasks.
python logiqa-determlr.py
python proofwriter-determlr.py
python folio-determlr.py
python logicaldeduction-determlr.py
In addition to determlr, we also include baseline methods (cot, tot, cr) for four logical reasoning tasks.
Directly run the script python {TASK}-{METHOD}.py
to implement them.
For the implementation of our DetermLR, several important arguments are introduced as follows:
-
propnum
: number of generated determinate premises (choices={2,3,4,5}
) -
reasoningnum
: final question number of vote choice {default=4,8,16} -
condition_divide
: whether to divide premise into determinate and indeterminate {default=True} -
con_select
: whether to include premise prioritization {default=True} -
memory
: whether to infer premise generating history or not {default=True} -
useful_judgement
: whether to include useful_judgement validation {default=True} -
global_validation
: whether to include global_validation, not required on ProofWriter.
Partial credit to previous reprostories: Guidance and Cumulative Reasoning.
Please cite the paper and star this repo if you use DetermLR and find it interesting/useful, thanks!
Feel free to contact xuwk266@gmail.com for any questions.
@article{sun2023DetermLR,
title={DetermLR: Augmenting LLM-based Logical Reasoning from Indeterminacy to Determinacy},
author={Sun, Hongda and Xu, Weikai and Liu, Wei and Luan, Jian and Wang, Bin and Shang, Shuo and Wen, Ji-Rong and Yan, Rui},
journal={arXiv preprint arXiv:2310.18659},
year={2023}
}