ADRD-Bench Benchmark

Paper: https://arxiv.org/abs/2602.11460

ADRD-Bench is a benchmark specifically designed for Alzheimer’s Disease and Related Dementias (ADRD) to rigorously evaluate large language models (LLMs) in both clinical knowledge and daily caregiving contexts.

---

Overview

ADRD-Bench consists of two complementary components:

ADRD Unified QA

A curated and unified question set synthesized from seven established medical QA benchmarks, designed to evaluate LLMs’ ADRD-related clinical knowledge.

• Total questions: 1,352
• Data access note: Due to original dataset licensing restrictions, we do not redistribute the full question texts. Instead, we provide:
  • the source dataset name
  • the original index of each question within its source benchmark

This design is for reproducibility and transparency, while respecting all original data usage policies.

ADRD Daily Caring QA

A novel benchmark focusing on real-world dementia caregiving scenarios, derived from a well-validated collaborative care program’s caregiver education materials.

• Total questions: 149
• Question formats:
  • Multiple-choice questions: 29
  • True / False questions: 120
• Design guidance: developed in collaboration with an experienced clinician
• Focus areas: daily care, communication, safety, behavioral symptoms, and caregiver decision-making

This component helps mitigate the lack of practical caregiving context in existing ADRD-related benchmarks.

Repository Structure

ADRD-Bench/
├── ADRD_Caregiving_QA/
│   ├── ADRD_Caregiving_Multiple_Choice.json
│   └── ADRD_Caregiving_True_or_False.json
├── ADRD_Unified_QA/
│   ├── ADRD_Unified.json
├── README.md
└── LICENSE

Data Usage & Ethics

• ADRD Daily Caring QA is released for research and evaluation purposes only
• ADRD Unified QA complies with all original dataset licenses by providing indices only
• This benchmark is not intended for direct medical decision-making

Contact

Project Team: IIRL-ND

For questions, collaborations, or access inquiries, please open an issue or contact the authors.

Citation

If you find ADRD-Bench useful in your research, please cite:

@misc{zhao2026adrdbenchpreliminaryllmbenchmark,
      title={ADRD-Bench: A Preliminary LLM Benchmark for Alzheimer's Disease and Related Dementias},
      author={Guangxin Zhao and Jiahao Zheng and Malaz Boustani and Jarek Nabrzyski and Meng Jiang and Yiyu Shi and Zhi Zheng},
      year={2026},
      eprint={2602.11460},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.11460}
}