Mind Companion: An Embodied Conversational Agent for Process-Based Psychotherapy

Abstract: Access to evidence-based psychotherapy remains limited worldwide, with long waitlists even in high-income regions. Recent advances in LLMs offer potential for scalable mental health support when designed with clinical oversight and safety mechanisms. We present Mind Companion, an LLM-based embodied conversational agent integrating multi-layered psychological analysis with process-based therapy principles. The system performs real-time analysis of client statements across fact extraction, psychological flexibility process detection, emotion recognition, and safety monitoring. Analysis results are stored for supervising clinicians to inform therapeutic planning. Response generation incorporates retrieval-augmented generation from evidence-based therapeutic literature and context-aware prompting. Responses are delivered through an embodied avatar with synchronized speech synthesis and animation. We evaluated three LLM configurations (GPT-4.1-mini, GPT-5.2, Claude Sonnet 4.5) against therapist responses from real therapy sessions using automated LLM-judge assessment and expert evaluation with 11 professional psychotherapists. GPT-5.2 achieved higher ratings than human therapist responses across understanding, interpersonal effectiveness, collaboration, and therapeutic alignment in both evaluations, demonstrating the feasibility of LLM-based conversational agents as tools to complement clinical care.

• The paper presents a scalable, embodied LLM agent integrating fact extraction, ACT process detection, emotion recognition, and safety monitoring for psychotherapy.
• It leverages retrieval-augmented generation from curated ACT literature to generate evidence-based responses, outperforming human therapists in collaboration and alignment.
• Empirical evaluations demonstrate faster response times (<6s) and higher verbal quality indices, highlighting the need for supervised clinical deployment.

Mind Companion: An Embodied LLM Conversational Agent for Process-Based Psychotherapy

Introduction

The persistent shortage of mental health professionals globally and increasing demand for evidence-based psychotherapy necessitate scalable, technology-mediated interventions. Recent LLMs demonstrate expert-level performance on a range of medical tasks, and conversational agents based on these models are now widely investigated for their potential as digital mental health tools. However, autonomous LLM-based mental health agents raise severe safety, efficacy, ethical, and relational challenges, particularly when applied without clinical oversight. “Mind Companion: An Embodied Conversational Agent for Process-Based Psychotherapy” (2606.17789) directly addresses these constraints, proposing an LLM-powered, embodied agent for process-based therapy deployment within supervised clinical workflows.

System Architecture and Methodology

Mind Companion combines four principal real-time psychological analysis streams—fact extraction, psychological flexibility process detection, emotion recognition, and safety monitoring—with advanced response generation leveraging retrieval-augmented generation (RAG) from curated ACT literature. All user analysis occurs in parallel with response generation to maintain dialogue fluidity. Emphasis is placed on modularity, enabling integration into existing clinical settings and robust clinician-in-the-loop supervision.

• Fact Extraction: Structured knowledge representation of client information is maintained, restricted to clinically meaningful, objective patterns across key categories (e.g., life events, living context, medical history), with a strict cap on memory length for scalability.
• ACT Process Detection: Statements are classified according to the six psychological flexibility/inflexibility processes defined in the ACT hexaflex model, with multi-label, intensity-rated outputs supporting process-based therapy alignment.
• Emotion Recognition: A hierarchical, temporally-aware emotion vector representation (instant, session, long-term) is computed per utterance using an LLM.
• Safety Monitoring: Proactive, conservative risk detection classifies statements as safe, crisis, or unsafe-to-interact, with enforced disengagement and resource signposting in high-risk cases.
• Response Generation: Uses RAG to inject retrieved, contextually relevant ACT material (from curated manuals and books) into the LLM prompt, ensuring output is both evidence-based and therapeutically grounded. The approach adopts a two-stage dialogue strategy: initial broad assessment of ACT domains followed by focused intervention on dimensions exhibiting inflexibility.

This architecture is embodied via a cross-platform Unity mobile client, delivering responses through animated avatars with synchronized speech and rich non-verbal behaviors (gesture, facial expression, lip sync).

Evaluation Protocol and Empirical Results

Evaluation involved both automated LLM-as-judge and blinded expert assessment against real ACT therapy transcripts (German-language), enabling robust benchmarking of agent performance versus human therapists.

• Dataset: 320 high-salience therapy dialogue points (strong emotional valence or arousal) were extracted from 45 sessions (15 clients).
• Models Compared: GPT-5.2, GPT-4.1-mini, and Claude Sonnet 4.5 were benchmarked against real therapist responses.
• Automated Metrics: Five dimensions (understanding, interpersonal effectiveness, collaboration, ineffectiveness (reverse-scored), ACT alignment) were rated on 1-7 scales using an impartial LLM judge, calibrated by ACT situational judgment exemplars.

Key Empirical Findings:

• All LLM systems significantly outperformed human therapist responses on every dimension—most notably, GPT-5.2 achieved higher mean scores for overall performance (6.53 vs. 4.67), collaboration (6.47 vs. 3.87), and ACT alignment (0.94 vs. 0.65), with $p<0.001$.
• Expert clinicians (N=11) also rated GPT-5.2 higher than human therapist responses across all metrics, especially in collaboration, understanding, and adherence to ACT principles.
• Notably, the models produced substantially longer and more structured responses than clinicians (GPT-5.2 mean 83.72 words vs. human mean 30.32), which may have influenced qualitative ratings.
• There was a strong correlation between expert and LLM judge rankings, though the LLM judge systematically assigned higher absolute scores.
• Real-time latency was kept low (<6s end-to-end for GPT-5.2 and Claude Sonnet 4.5).

Theoretical and Practical Implications

Mind Companion demonstrates that when applied in a supervised, embodied context with process-oriented frameworks, advanced LLMs can generate therapist-like, or even higher-rated, responses according to established verbal quality rubrics. The integration of RAG from domain literature enforces therapeutic alignment and significantly reduces hallucinations and off-protocol behavior. Modular, persistent knowledge tracing (facts, process themes, emotion trajectories) enhances both immediate therapeutic guidance and longitudinal clinical oversight.

However, superiority in isolated verbal response scoring does not imply functional equivalence or superiority in longitudinal clinical outcomes, alliance formation, or crisis management. The methodology, emphasizing single-turn response comparison in scripted contexts, cannot capture the subtleties of relationship-building, non-verbal communication, nor real-world client engagement and retention. In practical deployment, the agent’s conservative safety mechanisms and requirement for clinical triggers maintain essential layers of risk mitigation absent from unsupervised systems.

Future Directions

Progress towards real-world adoption requires full-session and longitudinal studies, integration of facial/avatar presentation effects, adversarial safety testing (especially for hallucinations, crisis/intervention failures, and adversarial users), and real-time therapist-in-the-loop workflow validation. Human factors aspects—such as personalization, alliance rupture/repair, and usability by diverse clinical populations—remain open research domains. Extensions to other psychotherapy models (e.g., DBT, MBCT) and cross-language/cultural transfer are feasible, but will require separate factual and affective knowledge curation pipelines.

Conclusion

Mind Companion operationalizes an advanced, evidence-grounded, embodied agent for process-based psychotherapy that, under expert supervision, consistently produces responses rating higher than human therapists on multidimensional verbal quality indices. The architecture sets a new reference point for LLM-based clinical agents: real-time, modular, controlled for safety, and therapeutically precise. However, the constraints of verbal-only evaluation and single-turn focus necessitate longitudinal, relationship-centered research before these systems can be credibly considered as clinical partners outside the supervised augmentation context. AI-augmented psychotherapy should remain explicitly therapist-centric, leveraging such agents to expand reach, consistency, and analytic granularity within ethical, clinical boundaries.