Agentic Clinical Dialogue Survey
Survey · arXiv 2512.01453

Reinventing Clinical Dialogue:
Agentic Paradigms for LLM-Enabled
Healthcare Communication

A first-principles analysis of the paradigm shift from generative text prediction to agentic autonomy in medical AI, with a novel taxonomy along the axes of knowledge source and agency objective.

ArXiv Paper GitHub 🤗 HuggingFace
365+
GitHub Stars
200+
Papers Covered
4
Agent Paradigms
5
Dialogue Types
50+
Datasets

Abstract

Clinical dialogue demands both empathetic fluency and the rigorous precision of evidence-based medicine. While LLMs possess unprecedented linguistic capabilities, their reactive, stateless architecture favors probabilistic plausibility over factual veracity. This limitation has catalyzed a paradigm shift from generative text prediction to agentic autonomy, where the model acts as a reasoning engine with deliberate planning and persistent memory. We introduce a novel taxonomy along the orthogonal axes of knowledge source and agency objective, categorizing methods into four archetypes that reveal how architectural choices balance autonomy and safety across the entire cognitive pipeline.

Evolution of Clinical Dialogue Systems

The evolution from pipeline-based to LLM-based to agentic clinical dialogue systems.

Four Agentic Paradigms

Methods are categorized along two axes: knowledge source (implicit vs. explicit) and agency objective (event cognition vs. goal execution).

Latent Space Clinicians
LSC
Latent Space Clinicians

Leverage the LLM's internal knowledge for creative clinical synthesis, trusting emergent reasoning to function like an experienced assistant.

Emergent Planners
EP
Emergent Planners

Grant the LLM high autonomy to dynamically devise multi-step plans, independently determining necessary steps for complex clinical goals.

Grounded Synthesizers
GS
Grounded Synthesizers

LLMs serve as natural language interfaces to reliable external sources, retrieving and synthesizing from verifiable knowledge.

Verifiable Workflow Automators
VWA
Verifiable Workflow Automators

Agent autonomy is strictly constrained within pre-defined, verifiable clinical workflows, ensuring maximum safety and predictability.

Taxonomy Framework

Taxonomy framework along the axes of knowledge source and agency objective, with cognitive pipeline components for each paradigm.

Key Contributions

1

Novel Taxonomy Framework

A two-axis taxonomy (knowledge source x agency objective) yielding four distinct archetypes for clinical AI agents.

2

Cognitive Pipeline Analysis

In-depth deconstruction of each paradigm across planning, memory, action, collaboration, and evolution.

3

Bridging Theory and Practice

Mapping real-world applications to our taxonomy with systematic benchmark and evaluation reviews.

4

Curated Resource Collection

200+ papers, 50+ datasets across five dialogue types, with tutorials and leading research groups.

Publication Statistics

Publication Statistics

Publication statistics: annual paper count (left) and venue distribution (right).

Survey Scope

Dialogue Datasets

  • QA Dialogue (MedQA, PubMedQA, ...)
  • Task-oriented Dialogue
  • Recommendation Dialogue
  • Supportive Dialogue
  • Hybrid-function Dialogue

Cognitive Pipeline

  • Strategic Planning
  • Memory Management
  • Action Execution
  • Collaboration & Multi-agent
  • Self-evolution & Meta-learning

Methods & Models

  • Latent Space Clinicians (LSC)
  • Emergent Planners (EP)
  • Grounded Synthesizers (GS)
  • Verifiable Workflow Automators (VWA)

Resources

  • Tutorials & Courses
  • Leading Research Groups
  • Benchmarks & Evaluation
  • Open-source Implementations

Challenges & Future Directions

Four core challenge areas: Neuro-Symbolic Cognitive Architecture, Holistic Patient Management, Human-AI Teaming, and High-Stakes Control.

Future Challenges

Open challenges and future research directions organized around four core themes.

Citation

BibTeX
@article{zhi2025reinventing, title = {Reinventing Clinical Dialogue: Agentic Paradigms for LLM Enabled Healthcare Communication}, author = {Zhi, Xiaoquan and Zhao, Hongke and Wu, Likang and Zhao, Chuang and Zhu, Hengshu}, journal = {arXiv preprint arXiv:2512.01453}, year = {2025} }