Access DemMA Planner

Access is granted automatically. If you use this model in a research publication, please cite DemMA (Findings of ACL 2026). Paper: https://aclanthology.org/2026.findings-acl.605/

Log in or Sign Up to review the conditions and access this model content.

🧠 DemMA Planner

Expert-guided dementia dialogue, from reasoning to action.
SFT checkpoint Β· Findings of ACL 2026

πŸ“„ Paper Β· PDF Β· πŸ’» Code Β· πŸ€— Dataset

Every turn connects memory, reasoning, and behavior. DemMA brings patient personas, memory profiles, and nonverbal behavior into multi-turn dementia patient simulation. Structured reasoning connects the conversation context with the simulated patient's response and actions. This repository hosts the DemMA Planner SFT checkpoint.

DemMA framework: persona formation, annotated dialogue generation, and multi-task training

The DemMA framework. Click the figure to explore it at full resolution.

✨ What DemMA brings together

🧠 Reason πŸ’¬ Respond 🎭 Act
Use persona, memory, and dialogue context to form a structured plan. Generate the simulated patient's next utterance. Predict accompanying movement, facial-expression, and voice labels.

🧩 From patient profiles to a dialogue agent

  1. Build a persona. Combine dementia characteristics, personality traits, and memory profiles.
  2. Generate annotated dialogues. Connect each turn with planner rationales, patient responses, and behavioral labels.
  3. Train a unified agent. Distill these annotations through multi-task supervised fine-tuning.

Explore the DemMA-Dialog dataset, or visit the official implementation for data generation, training, and inference.

πŸ“š Citation

If you find this model useful in your research, please consider citing our DemMA paper.

Show BibTeX
@inproceedings{song-etal-2026-demma,
  title = "{D}em{MA}: Dementia Multi-Turn Dialogue Agent with Expert-Guided Reasoning and Action Simulation",
  author = "Song, Yutong and Wu, Jiang and Sharif, Kazi Shaharair and Zhang, Pengfei and Huang, Wenjun and Xu, Honghui and Dutt, Nikil and Rahmani, Amir M.",
  booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
  year = "2026",
  publisher = "Association for Computational Linguistics",
  pages = "12439--12470",
  doi = "10.18653/v1/2026.findings-acl.605",
  url = "https://aclanthology.org/2026.findings-acl.605/"
}
Downloads last month
28
Safetensors
Model size
8B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support