🤗 Hugging Face   |   📄 arXiv

Table-R1 introduces the first systematic study of inference-time scaling for table reasoning tasks. We develop two post-training strategies: distillation from frontier model reasoning traces and reinforcement learning with verifiable rewards (RLVR). Our 7B-parameter Table-R1-Zero model matches or surpasses GPT-4.1 and DeepSeek-R1 performance while exhibiting strong generalization to out-of-domain datasets.

Overview of the Table-R1.

git clone https://github.com/Table-R1/Table-R1.git

# Install verl framework
cd Table-R1/verl
pip install -e .

cd ..
pip install -r requirements.txt

# Prepare SFT dataset
python data/table-r1-sft.py

# Run SFT training
bash script/table-r1-sft.sh

# Prepare RLVR dataset  
python data/table-r1-zero.py

# Run RLVR training
bash script/table-r1-zero.sh

# Prepare evaluation dataset
python data/table-r1-eval.py

# Run evaluation
bash script/table-r1-eval.sh

• All models are trained using the excellent verl framework

If you find Table-R1 useful in your research, please cite our paper:

@article{yang2025tabler1,
  title={Table-R1: Inference-Time Scaling for Table Reasoning},
  author={Yang, Zheyuan and Chen, Lyuhao and Cohan, Arman and Zhao, Yilun},
  journal={arXiv preprint arXiv:2505.23621},
  year={2025}
}