yguo86/ReTabVQA

ReTabVQA: Reasoning-Oriented Table VQA Benchmark

ReTabVQA (Reasoning Table VQA) is a manually curated benchmark designed to evaluate multi-step quantitative reasoning over tabular images. It extends and refines the TableVQA-Bench dataset by introducing complex, compositional question–answer pairs that go beyond direct value lookup.

💡 Motivation

While existing datasets such as TableVQA-Bench focus primarily on direct retrieval questions (e.g., locating a single cell value), they offer limited coverage of reasoning-oriented tasks that require integration, arithmetic, or comparison across multiple data points. ReTabVQA was developed to address this gap — providing a testbed for evaluating reasoning capabilities of vision-language and large language models in table-based visual question answering.

📘 Dataset Overview

Total Table Images: 60

Total QA Pairs: 120

Domains: Financial Reports, Sports Statistics, Survey Results, and Scientific Tables

QA Composition: Each table contains two reasoning-oriented questions involving multi-step calculations, logical comparisons, or numerical synthesis.

🚀 Intended Use

ReTabVQA is suitable for:

Evaluating vision–language models (VLMs) and LLMs on tabular reasoning tasks.

Training models that combine numerical reasoning and visual table understanding.

Benchmarking progress in multimodal compositional reasoning.

📜 Citation

If you use this dataset, please cite:

@misc{yutong2025talenttablevqaaugmented, title={TALENT: Table VQA via Augmented Language-Enhanced Natural-text Transcription}, author={Guo Yutong and Wanying Wang and Yue Wu and Zichen Miao and Haoyu Wang}, year={2025}, eprint={2510.07098}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2510.07098}, }