Research map / Multi-agent systems and games / LLM agents

An LLM-based Simulation Framework for Embodied Conversational Agents in Psychological Counseling

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ECAs, an LLM-based framework that expands real counseling case data (D4) into an embodied cognitive memory space grounded in CBT and other psychological theories, and uses it to generate client-agent dialogues released as a public dataset.

Problem

Synthetic mental-health dialogue data generated by humans or LLMs often lacks authenticity and diversity because client psychological processes are implicit and hard to simulate.

Approach

Builds client agents with a structured memory space (life-stage factual memories, CBT core/intermediate beliefs and automatic thoughts, perceptual memories, social networks) generated in a 4-phase LLM pipeline and retrieved context-dependently during dialogue, rather than prompting an LLM with a bare persona.

Key results

• Expert ratings from five licensed counselors: ECAs significantly outperformed human D4 dialogues on necessity, sufficiency, fidelity and consistency (e.g., sufficiency mean difference 1.50 vs human, p<0.001).
• Depression risk macro F1 rose to 0.42 for ECAs vs 0.40 for D4 and 0.35 for GPT-4o; suicide risk macro F1 0.67 vs 0.59.
• LLM-judged quality total score 17.90/20 for ECAs vs 9.28/20 for D4 dialogues.

Limitations

• Evaluation relies on five experts on six client agents and GPT-4o as automated judge, with some comparisons only marginally significant.
• Dataset is built from Chinese D4 data and relies on synthetic client memories generated by an LLM.
• Automated judge (GPT-4o) and the generator (GPT-4o) are the same model family, which may bias results.

Experimental setup

Datasets: D4

Benchmarks: D4

Metrics: necessity sufficiency fidelity consistency macro precision macro recall macro F1 weighted quality score

Baselines: GPT-4o D4

Builds on: D4 Cognitive Behavioral Therapy Generative Agents

Closest papers by meaning

• TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots (NeurIPS 2026 · similarity 0.87)
• Thousand Voices of Trauma: A Large-Scale Synthetic Dataset for Modeling Prolonged Exposure Therapy Conversations (NeurIPS 2025 · similarity 0.85)
• OdysSim: Building Foundation Models for Human Behavior Simulation (NeurIPS 2026 · similarity 0.85)
• Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.85)
• AgentMental: An Interactive Multi-Agent Framework for Explainable and Adaptive Mental Health Assessment (AAAI 2026 · similarity 0.85)
• SocialSim: Towards Socialized Simulation of Emotional Support Conversation (AAAI 2025 · similarity 0.85)