coTherapist: A Behavior-Aligned Small Language Model to Support Mental Healthcare Experts Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates DOI: 1

Prottay Kumar Adhikary, Reena Rawat, Tanmoy Chakraborty · 2026

Read coTherapist: A Behavior-Aligned Small Language Model to Support Mental Healthcare Experts Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates DOI: 1 by Prottay Kumar Adhikary, Reena Rawat, Tanmoy Chakraborty (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-10246.txt.

Introduction

Prottay Kumar Adhikary, Reena Rawat, and Tanmoy Chakraborty build an aid for mental health pros. Demand for care is high. The pool of pros is thin. An aid that can back a pro could ease the load. It must act with the care a pro would show.

The aid is named coTherapist. It uses a small language model at its core. The model is tuned on care tasks. It can fetch facts and plan steps like an agent. The aim is help for the pro, not a swap for the pro. The paper sets out how the aid is built and checked.

Catalog record

CREATORProttay Kumar Adhikary, Reena Rawat, Tanmoy Chakraborty
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-10246

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

coTherapist: A Behavior-Aligned Small Language Model to Support Mental Healthcare Experts Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates DOI: 10.1145/3774904.3792988 ISBN: 979-8-4007-2307-0/2026/04 CCS: Human-centered computing Web-based interaction CCS: Computing methodologies Natural language processing CCS: Applied computing Health care information systems Prottay Kumar Adhikary, Reena Rawat, Tanmoy Chakraborty 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. † † cc-license: by 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. 1. Introduction Digital mental health platforms are increasingly utilizing Artificial Intelligence (AI) to bridge gaps in care and support Mental Health Experts (MHx). Large Language Models (LLMs) have recently demonstrated a surprising ability to provide critical advice ( 18 ) . While these systems promise better accessibility and can offer support that users find clearer and less biased than human therapists ( 21 ; 69 ) , naive implementation can still lead to unhelpful, script-like errors or harmful outputs ( 19 ) . To address this within the emerging field of computational psychiatry ( 8 ) , it is essential to define an AI’s “behavior” not merely as text generation, but as its observable communication patterns ( 32 ) . For a model to truly resemble a skilled therapist rather than a generic chatbot, it must master three specific components: style imitation involving tone and paraphrasing ( 48 ) , a structured instruction-following capability ( 59 ) , and a conceptual reasoning framework based on evidence-based therapy logic ( 61 ) . AI as a Therapeutic Assistant Rather than attempting to replace human clinicians, we envision AI serving as a supportive assistant that helps trained practitioners deliver better care ( 5 ) . In clinical practice, therapists often consult manuals, worksheets, or peers to guide their interventions ( 53 ) , but they generally avoid relying on the open internet due to the high risk of misinformation. Similarly, an AI could fulfill this role by instantly answering a therapist’s query about treatment techniques, including specific book sections for verification. It can also provide the rationale for a diagnosis or suggest a gentle way to rephrase a question asked to a client ( 72 ) . By doing so, it can ease the therapist’s cognitive load and provide quick access to consolidated expertise, all while the therapist remains in charge of the client’s care ( 25 ) . This human-AI collaboration approach, which we term coTherapy , aligns with the broader goals of digital well-being, leveraging web-based platforms to improve mental health support without removing the human element ( 38 ) . Challenges with Small Language Models Transforming a general Small Language Model (SLM) into an effective clinical tool requires the model to adopt core counseling skills like reflective listening and structured frameworks such as Cognitive Behavioral Therapy (CBT) ( 6 ) . While generic SLMs are proficient at following technical rules, they often fall short in relational subtleties, producing formulaic empathy rather than the nuanced validation provided by human experts ( 63 ) . To bridge this gap, we need both behavioral style alignment and conceptual alignment , enabling the AI to “think through” problems using psychological theories. Our Contributions. We introduce coTherapist , a lightweight agent emulating expert therapeutic behavior ( 24 ) (Figure 1 ). Designed strictly as a clinician’s assistant, not a replacement, it is evaluated on simulated queries. Unlike computationally intensive LLMs ( 46 ) , we focus on aligning SLMs with clinical standards ( 22 ; 14 ) , enabling deployment in resource-constrained settings ( 65 ; 67 ) . Prioritizing behavioral fidelity over raw performance, our contributions are: (1) Domain-Specific Psychotherapy Knowledge Dataset. We curated a dataset of 800 million+ tokens, comprising undergraduate and postgraduate textbooks, lecture notes and videos, practice materials, etc. This corpus grounds the model in evidence-based therapeutic language, conceptual reasoning, and technical accuracy. (2) The coTherapist Framework. We develop a multi-stage pipeline for a 1 Billion parameter SLM, integrating continued pretraining for domain adaptation, LoRA fine-tuning for dialogue alignment ( 11 ) , and Retrieval-Augmented Generation (RAG) for evidentiary support ( 43 ) . Additionally, we implement an agentic reasoning structure to facilitate step-by-step clinical formulation. (3) Therapist Behavior Rating Scale. We propose T-BARS , a novel framework assessing four key pillars: Behavioral Style, Conceptual Reasoning, Relational Competence, and Technique Execution, across 20 sub-skills. Using LLM-based judges and psychometric profiling ( 27 ) , we find that coTherapist achieves a 3 improvement in empathy and relational clarity over baselines, despite modest gains in standard metrics. Human evaluation further confirms expert preference for coTherapist ’s safety and helpfulness. 1 1 1 We are committed to releasing the source codes and dataset upon acceptance of the paper. 2. Related Work The application of NLP in mental health has evolved from simple rule-based chatbots to complex generative systems ( 54 ) . Recent advancements demonstrate that LLMs can be aligned with professional therapeutic standards ( 51 ) . For instance, models fine-tuned for CBT have shown significantly higher adherence to counseling strategies compared to base models ( 24 ) . Beyond behavioral adherence, research indicates that LLMs can infer latent psychological structures, accurately reasoning about personality patterns based on Big Five traits ( 31 ) . This capability is essential for developing a consistent “therapist persona,” shifting the field’s focus from client-facing chatbots to AI assistants that support clinical decision-making, documentation, and retrieval of evidence-based criteria. Despite progress in generative capabilities, there remains limited research on designing AI systems specifically to reduce a practitioner’s need to constantly consult external manuals or textbooks during practice ( 40 ) . Generic models often struggle with this due to hallucinations. To mitigate this, RAG has emerged as a standard for grounding LLMs ( 12 ) in external knowledge bases ( 26 ) , a technique we adapt specifically for psychotherapy manuals ( 49 ) . Turning a general SLM into an effective AI requires the model to adopt core counseling skills and clinical reasoning patterns. Generic SLMs often fall short: prior research

Read the complete text