AAA

Official implementation and datasets of the AAAI 2026 Special Track AI for Social Impact paper:

Towards Scalable Web Accessibility Audit with MLLMs as Copilots.

This repository provides the implementation of the GRASP method and accompanying datasets AWA.

📊 Dataset

The AWA Web Accessibility Benchmark consists of four datasets:

• APR – Accessibility-relevant Page Recognition
• CCT – CAPTCHA of Cognitive Tests
• TPS – Triple-representativeness Page Sampling
• CPE – Complete Process Extraction

All datasets are publicly available on Zenodo:

Important

Before running any scripts, please download at least the TPS dataset and place it in the data/TPS directory at the same level as this project.

This repository implements the GRASP method, which uses only the TPS dataset.

For experiments involving APR, CCT, and CPE, please refer to the prompts and run them with any MLLMs as in the Appendix of our paper.

🛠️ Installation

Requirements:

• Python >= 3.8
• CUDA-compatible GPU (recommended)
• Dependencies are listed in:
  • requirements.txt
  • requirements-dev.txt
  • pyproject.toml

Setup:

# Run installation scripts
bash .ci/install-dev.sh
bash .ci/install.sh

# Activate virtual environment
source .env/bin/activate

🚀 Usage

Please run the experiments of GRASP and SDC as follows.

Page Sampling:

Run different clustering approaches for the TPS dataset:

• GRASP with GCN:

  bash scripts/grasp.sh

• GRASP with IGNN:

  bash scripts/ignn.sh

• SDC w & w/o TSNE:

  bash scripts/sdc.sh

Evaluation:

Compute metrics for evaluation:

• GRASP with GCN:

  bash scripts/cos.sh

• GRASP with IGNN:

  bash scripts/ignn_cos.sh

• SDC w & w/o TSNE:

  bash scripts/sdc_cos.sh

🤝 Contributing

Please see CONTRIBUTING.md for guidelines on reporting issues, contributing improvements, and extending the GRASP framework.

📚 Citation

If you find this work useful, please cite our paper:

@inproceedings{aaa,
title={Towards Scalable Web Accessibility Audit with {MLLMs} as Copilots},
author={Ming Gu and Ziwei Wang and Sicen Lai and Zirui Gao and Sheng Zhou and Jiajun Bu},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
year={2026},
url={https://arxiv.org/abs/2511.03471}, 
}