TableDART Gating Network Checkpoint
This repository provides the trained gating network checkpoint for TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding.
TableDART is a training-efficient framework that dynamically routes each table-query pair through the most appropriate reasoning path: Text-only, Image-only, or Fusion, while keeping all pretrained expert models frozen.
🔍 Overview
Modeling semantic and structural information from tabular data remains a core challenge for effective table understanding.
Existing LLM-based approaches face several limitations:
- Table-as-Text methods flatten tables into text sequences, losing structural cues.
- Table-as-Image methods preserve layout but struggle with precise semantics.
- Static multimodal methods process all modalities for every query, introducing redundancy and potential cross-modal conflicts.
- Most approaches require expensive fine-tuning of large LLMs or multimodal models.
Our Solution: TableDART addresses these limitations through:
- Reusing pretrained single-modality expert models (kept frozen, plug-and-play)
- Learning only a lightweight 2.59M-parameter MLP gating network
- Dynamically selecting the optimal path for each table-query pair (instance-level)
- Introducing an LLM agent that mediates cross-modal knowledge integration when needed
This design avoids full LLM/MLLM fine-tuning, reduces computational redundancy, and maintains strong efficiency-performance trade-offs.
🚀 Performance
Across 7 benchmarks, TableDART:
- Achieves state-of-the-art results on 4/7 benchmarks among open-source models
- Outperforms the strongest baseline by +4.02% accuracy on average
- Maintains significant computational efficiency gains
📦 What This Checkpoint Contains
This Hugging Face model includes:
- The trained MLP gating network checkpoint
⚠️ Note: This checkpoint does not include the pretrained text or image expert models. Please load those separately according to the official repository instructions.
🛠 Code and Usage
Full training scripts, inference pipelines, and reproduction details are available at our Github Repository:
https://github.com/xiaobo-xing/TableDART
📄 Paper
📚 Citation
If you find TableDART helpful, please cite our paper and consider starring the repository.
ICLR 2026 Version
1@inproceedings{xing2026tabledart,
2 title={Table{DART}: Dynamic Adaptive Multi-Modal Routing for Table Understanding},
3 author={Xiaobo Xing and Wei Yuan and Tong Chen and Quoc Viet Hung Nguyen and Xiangliang Zhang and Hongzhi Yin},
4 booktitle={The Fourteenth International Conference on Learning Representations},
5 year={2026},
6 url={https://openreview.net/forum?id=4aZTiLH3fm}
7}
ArXiv Version
1@misc{xing2025tabledartdynamicadaptivemultimodal,
2 title={TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding},
3 author={Xiaobo Xing and Wei Yuan and Tong Chen and Quoc Viet Hung Nguyen and Xiangliang Zhang and Hongzhi Yin},
4 year={2025},
5 eprint={2509.14671},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2509.14671}
9}
