coTherapist: A Behavior-Aligned Small Language Model to Support Mental Healthcare Experts

TL;DR

The paper tackles the shortage of mental health clinicians by developing coTherapist, a lightweight, therapist-aligned AI assistant built on a 1B parameter model. It combines domain-adaptive pretraining, LoRA-based style tuning, retrieval-augmented generation, and an agentic reasoning pipeline to produce clinically grounded and empathetic responses. The authors introduce the psychotherapy knowledge corpus PsyKC and the Therapist Behavior Rating Scale (T-BARS) to rigorously assess therapist-like behavior, reporting improvements in automatic metrics, therapist alignment, and expert evaluations, including safety. They demonstrate feasible edge deployment with privacy-preserving, on-premise operation, suggesting a scalable path for digital mental health tools while acknowledging licensing, scope, and validation limitations that require broader multi-site testing.

Abstract

Access to mental healthcare is increasingly strained by workforce shortages and rising demand, motivating the development of intelligent systems that can support mental healthcare experts. We introduce coTherapist, a unified framework utilizing a small language model to emulate core therapeutic competencies through domain-specific fine-tuning, retrieval augmentation, and agentic reasoning. Evaluation on clinical queries demonstrates that coTherapist generates more relevant and clinically grounded responses than contemporary baselines. Using our novel T-BARS rubric and psychometric profiling, we confirm coTherapist exhibits high empathy and therapist-consistent personality traits. Furthermore, human evaluation by domain experts validates that coTherapist delivers accurate, trustworthy, and safe responses. coTherapist was deployed and tested by clinical experts. Collectively, these findings demonstrate that small models can be engineered to exhibit expert-like behavior, offering a scalable pathway for digital mental health tools.

Figures (3)

• Figure 1: Illustrative interaction with coTherapist. The figure visualizes the end-to-end pipeline triggered by a mental healthcare expert's query, highlighting the stages of planning, retrieval, internal reasoning, and self-refinement. The final output demonstrates how the system generates a response that is grounded in cited sources and aligned with a professional therapeutic tone, despite being powered by a lightweight small language model.
• Figure 2: System overview of the proposed coTherapist framework. (A) Data Curation: We compile a high-quality clinical corpus from psychotherapy textbooks, university lecture notes, lecture videos, etc. The data is cleaned, segmented, and indexed to cover major evidence-based therapies used in MHx training and practice. (B) Experiment Design: The model is trained through continued pretraining for domain tone, LoRA fine-tuning for therapist communication style, and retrieval-augmented generation to provide clinically accurate knowledge. An agentic reasoning step supports structured and context-aware responses. (C) Evaluation: Outputs are evaluated using both traditional NLG metrics and the T-BARS behavioral framework. Across all settings, the full coTherapist model demonstrates the highest therapist-like alignment and is preferred in human evaluation.
• Figure 3: Analysis of personality alignment and its impact on clinical performance. (A) Big Five profiles indicate that coTherapist exhibits higher Agreeableness and Conscientiousness with lower Neuroticism compared to TG-RAG, consistent with established therapist traits. (B) Correlation analysis confirms that these specific personality dimensions are positively associated with higher scores in the Behavioral Style Alignment and Relational Competence pillars of the T-BARS framework.