TL;DR
This paper introduces a multi-agent simulation system designed to train therapists by realistically modeling couple conflict. By utilizing a stateful "sense-plan-act" architecture, the system coordinates two distinct large language model (LLM) agents to simulate the complex "demand-withdraw" cycle across six evolving therapeutic stages. Evaluated by 21 licensed therapists, the system significantly outperformed static, prompt-based baselines in behavioral realism and training efficacy.
What Problem Does This Paper Address?
Couples therapy is notoriously difficult to master because clinicians must manage multi-party, interdependent emotional dynamics rather than just a single patient. Traditional training relies heavily on human role-play, which is resource-intensive, difficult to standardize, and often lacks realistic emotional nuance.
While LLM-based virtual patients have recently emerged for individual counseling training, adapting them to multi-party scenarios like couples therapy introduces a major technical hurdle: behavioral drift. Relying on a single static prompt often causes LLMs to lose role consistency over extended conversations. Furthermore, standard setups fail to capture how partners trigger and react to one another over time—such as the classic demand-withdraw pattern, where one partner's criticism fuels the other's defensive withdrawal.
Methods
To overcome behavioral drift, the authors model the therapy session as a controlled, stateful dynamical system using a Sense-Plan-Act architecture:
- Sense (Detection): The system tracks the dialogue history and ingests the therapist's input.
- Plan (Stage Controller): A finite-state-machine-like controller determines the next logical interaction stage based on clinical theory. The six stages are: Greeting, Problem Raising, Escalation, De-escalation, Enactment, and Wrap-up.
- Act (Output Generator): The active stage triggers specific behavioral instructions for two distinct LLM agents: Alex (the aggressive demander) and Jordan (the defensive withdrawer).
A key innovation is the Agent-to-Agent Interaction Loop. During the Problem Raising and Escalation stages, if one agent uses accusatory language (e.g., "you always"), the system automatically triggers a direct argument loop between the two agents for 3 to 5 turns. This mimics realistic couple fighting without requiring constant therapist prompts, though the therapist can step in to interrupt at any time.
Additionally, the system is multimodal. It maps discrete emotional states to each stage, rendering the agents' responses via text, animated avatars, and synthesized speech (using OpenAI's TTS) with tailored emotional prompts (e.g., adjusting vocal speed and intensity for an "Angry" state).
Datasets
- Alexander Street Counseling and Psychotherapy Transcripts corpus: The authors analyzed 7 anonymized real-world couples therapy transcripts (1,621 conversational turns) to validate their six-stage framework and adapt two authentic training scenarios: one centered on infidelity and the other on coping with severe depression.
- User Study Evaluation Dataset: The authors collected a new dataset of 42 interaction logs (566 turns, 1,571 utterances) generated during their study with 21 practicing therapists to evaluate runtime fidelity.
Key Insights and Ideas
- Stateful Control Beats Pure Prompting: Managing multi-agent interactions via an explicit stage controller prevents the conversational decay typical of LLMs, ensuring the agents remain entrenched in their clinical personas over long horizons.
- Legibility Through Theory: Grounding the state machine in established psychotherapy frameworks (like Emotionally Focused Therapy) makes the system's transitions highly predictable and interpretable for domain experts.
- The Power of Autonomy: Allowing agents to talk directly to each other—rather than just responding to the user—is essential for simulating interpersonal conflict and teaching therapists when and how to execute timely interventions.
Key Results
The system was evaluated by 21 licensed U.S. therapists in a blind, within-subjects experiment comparing the stateful system against a standard prompt-based baseline.
- Technical Fidelity: An LLM-as-judge evaluation showed that the experimental system achieved 70.7% role fidelity (vs. a dismal 4.9% for the baseline) and 83.8% stage fidelity (vs. 63.8%). Human-system agreement on stage transitions was also high ($\kappa = 0.77$).
- Perceived Realism: Therapists rated the experimental system significantly higher on a 5-point Likert scale for both response realism (4.11 vs. 2.86) and overall realism (4.16 vs. 2.71). The realism gap widened most dramatically during the high-conflict Problem Raising, Escalation, and De-escalation stages.
- Training Effectiveness: Experts rated the experimental system as substantially more effective for training novice therapists than the baseline (3.95 vs. 2.62), noting that it successfully replicated the "fast pace" and "defensive postures" seen in real-world practice.
Limitations and Future Work
While promising, the authors note several limitations:
- Simplified Emotion Modeling: The system currently assigns a single discrete emotion per agent per stage, failing to capture complex, blended, or rapidly shifting affects within a single exchange.
- Brief Sessions: The 15-minute simulation sessions are a compressed proxy for real-world couples therapy, which typically unfolds over months.
- Cultural Bias: The underlying LLMs inherit Western-centric assumptions regarding relationship dynamics, communication styles, and gender roles.
Future work aims to expand this design pattern to other high-stakes multi-party domains that require managing competing goals and shifting alignments, such as mediation, multi-party negotiation, and crisis communication.
Summary updated