👵👴🏻 DemMA: Dementia Multi-Turn Dialogue Agent with Expert-Guided Reasoning and Action Simulation

---

📄 Abstract

Dementia care requires sustained, adaptive, and behaviorally appropriate interaction, yet existing AI systems primarily focus on diagnosis or text-only dialogue generation. They fail to capture key characteristics of dementia communication, including memory impairment, behavioral variability, and nonverbal signals.

We propose DemMA, a clinically grounded large language model (LLM) agent for high-fidelity multi-turn dementia dialogue simulation. DemMA explicitly models patient behavior through three components:

• Structured planning (planner) for internal reasoning
• Natural language generation (utterance) for dialogue
• Nonverbal action prediction (action labels) for behavioral simulation

To support training, we construct a scalable pipeline for generating high-quality synthetic dialogues with persona constraints, memory gating, and expert-guided reasoning. We further introduce a multi-task learning framework that jointly optimizes planning, response generation, and action prediction.

Experiments and human evaluations demonstrate that DemMA produces more clinically consistent, behaviorally realistic, and temporally coherent interactions compared to standard LLM-based approaches.

---

✨ Contributions

• 🧠 Behavior-aware LLM agent for dementia simulation Introduces a unified framework modeling reasoning, language, and behavior.

• 🧬 Clinically grounded persona & memory modeling Captures subtype-specific traits and realistic memory impairments.

• 🎭 Action-level behavioral simulation Extends dialogue generation to nonverbal behaviors (movement, facial, voice).

• ⚙️ Multi-task training paradigm Jointly learns planner, utterance, and action prediction in a single model.

---

🖼️ Framework Overview

DemMA integrates three components: 1. Clinically grounded patient persona formation module; 2. Multi-agent dialogue dataset generation pipeline encompassing memory analysis, planning, language generation, and action simulation; 3. CoT distillation multi-task training workflow for DemMA agent.

---

🚀 Code Usage

This repository contains three main components:

---

1. Data Generation

python src/demma/data_generation.py

• Generates high-quality multi-turn dialogues
• Incorporates persona constraints and memory gating
• Produces action-labeled training data

---

2. Multi-task Training

python src/demma/train_sft.py \
  --train_data_path data/DemMA_planner_labeled_dialogue_corpus.jsonl \
  --base_model Qwen/Qwen3-8B \
  --output_dir checkpoints/planner_true_multitask

• Jointly trains:

  • planner generation
  • utterance generation
  • action prediction

---

3. Inference / Chatbot

python src/demma/inference.py

• Interactive dementia patient simulation

• Outputs:

  • planner state
  • patient utterance
  • predicted actions

---

📚 Citation

If you find this work useful, please cite:

@article{song2026demma,
  title={DemMA: Dementia Multi-Turn Dialogue Agent with Expert-Guided Reasoning and Action Simulation},
  author={Song, Yutong and Wu, Jiang and Sharif, Kazi and Xu, Honghui and Dutt, Nikil and Rahmani, Amir},
  journal={arXiv preprint arXiv:2601.06373},
  year={2026}
}

---

⚠️ Disclaimer

This repository is intended for research purposes only and is not a clinical tool.