Large Language Models Can Self-Correct with Key Condition Verification (EMNLP 2024)

Introduction & Setup

This repository contains the code for the paper Large Language Models Can Self-Correct with Key Condition Verification (Accepted to EMNLP Main 2024). ProCo first prompts an LLM to generate an initial response, then iterates a verify-then-correct process to progressively identify and correct (probably) false responses. Extensive experiments on eight complex reasoning datasets demonstrate the effectiveness and efficiency of ProCo.

• Run run.py to generate the answer to the given question

python run.py --data_index 0

Experimental Results

Citing ProCo

@inproceedings{wu2024proco,
  title={Large Language Models Can Self-Correct with Key Condition Verification}, 
  author={Zhenyu Wu and Qingkai Zeng and Zhihan Zhang and Zhaoxuan Tan and Chao Shen and Meng Jiang},
  booktitle={Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2024},
}

License

This project is licensed under the Apache-2.0 License.