Table-Critic

Code for paper Table-Critic(ACL 2025 Main).

Environment

conda create --name TableCritic python=3.10 -y
conda activate TableCritic
pip install -r requirements.txt 

Command Usages

Arguments

• --dataset_path: path to the dataset
• --result_dir: path to the result directory
• --base_url: base URL of the LLM API
• --model_name: name of the LLM API
• --openai_key: key of the LLM API
• --first_n: number of the first n samples to evaluate, default: -1 means whole dataset
• --n_proc: number of processes to use in multiprocessing, default: 1
• --chunk_size: chunk size used in multiprocessing, default: 1

API setup

Add base_url, model_name, openai_key to both the run_QA.sh and the run_FV.sh file.

Example usages

1. Run the experiment on the WikiTQ dataset

   bash run_QA.sh
   

2. Run the experiment on the TabFact dataset

   bash run_FV.sh
   

Citation

@inproceedings{yu-etal-2025-table,
    title = "Table-Critic: A Multi-Agent Framework for Collaborative Criticism and Refinement in Table Reasoning",
    author = "Yu, Peiying  and
      Chen, Guoxin  and
      Wang, Jingjing",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.853/",
    doi = "10.18653/v1/2025.acl-long.853",
    pages = "17432--17451",
    ISBN = "979-8-89176-251-0"
}