CAMI: A Counselor Agent Supporting Motivational Interviewing through State Inference and Topic Exploration
Yizhe Yang, Palakorn Achananuparp, Heyan Huang, Jing Jiang, Kit Phey Leng, Nicholas Gabriel Lim, Cameron Tan Shi Ern, Ee-peng Lim · 2025
Read CAMI: A Counselor Agent Supporting Motivational Interviewing through State Inference and Topic Exploration by Yizhe Yang, Palakorn Achananuparp, Heyan Huang, Jing Jiang, Kit Phey Leng, Nicholas Gabriel Lim, Cameron Tan Shi Ern, Ee-peng Lim (2025) 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-2502-02807.txt.
Introduction
Yizhe Yang and co-authors present a 2025 computing paper with a counseling chatbot built on motivational interviewing. That therapy style helps clients resolve their own mixed feelings and talk themselves toward change. Automated counselors could widen access to such support if they can do the method justice.
Their agent, CAMI, runs three modules in a loop. It infers the client's inner state, explores topics tied to motivation, and then generates its reply. Simulated clients let the team grade the bot's interviewing skill, state reading, and topic handling against rival systems. CAMI beat the state of the art and sounded more like a real counselor. Removing either the state or topic module hurt performance in ablation tests. The paper argues structure, not just scale, makes the counselor.
Catalog record
| CREATOR | Yizhe Yang, Palakorn Achananuparp, Heyan Huang, Jing Jiang, Kit Phey Leng, Nicholas Gabriel Lim, Cameron Tan Shi Ern, Ee-peng Lim |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 111 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-02807
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
CAMI: A Counselor Agent Supporting Motivational Interviewing through State Inference and Topic Exploration Yizhe Yang, Palakorn Achananuparp, Heyan Huang, Jing Jiang, Kit Phey Leng, Nicholas Gabriel Lim, Cameron Tan Shi Ern, Ee-peng Lim Abstract Conversational counselor agents have become essential tools for addressing the rising demand for scalable and accessible mental health support. This paper introduces CAMI, a novel automated counselor agent grounded in Motivational Interviewing (MI) – a client-centered counseling approach designed to address ambivalence and facilitate behavior change. CAMI employs a novel STAR framework, consisting of client’s state inference, motivation topic exploration, and response generation modules, leveraging large language models (LLMs). These components work together to evoke change talk, aligning with MI principles and improving counseling outcomes for clients from diverse backgrounds. We evaluate CAMI’s performance through both automated and manual evaluations, utilizing simulated clients to assess MI skill competency, client’s state inference accuracy, topic exploration proficiency, and overall counseling success. Results show that CAMI not only outperforms several state-of-the-art methods but also shows more realistic counselor-like behavior. Additionally, our ablation study underscores the critical roles of state inference and topic exploration in achieving this performance. 1 Introduction Motivation. Motivational Interviewing (MI) is a client-centered counseling technique aimed at addressing ambivalence and facilitating behavior change in clients, particularly for issues like binge drinking and substance use ( Miller and Rollnick, 2002 ) . MI promotes intrinsic motivation and fosters a collaborative relationship between clients and counselors, ultimately improving clients’ commitment to intervention programs ( Martins and McNeil, 2009 ) . Although recent works building LLM-based counselor agents have shown some promising results, we observe at least three limitations with existing works. First, they do not follow the collaborative counseling style in MI. For example, Steenstra et al. (2024) describes MI in a prompt which may not provide LLM adequate guidance to perform a complex multi-turn conversational task such as MI counseling. Second, those works that use strategy planning overlook the client’s state and the need to explore a diverse range of topics that can evoke change talk based on the client’s unique motivations ( Hsu et al., 2023 ; Sun et al., 2024 ; Xie et al., 2024 ) . As a result, they fall short in providing client-centered counseling ( Steenstra et al., 2024 ) and systematic change talk exploration, especially given the inherent preference bias of LLMs ( Kang et al., 2024 ) . Finally, the evaluation of counselor behavior in most works focus on comparing the agent generated responses with the predefined “ground truth” responses at the turn level ( Sun et al., 2024 ) . Such an evaluation approach rules out other possible high quality responses. Meanwhile, human evaluation of counselor agents remains costly ( Xie et al., 2024 ; Hsu et al., 2023 ; Steenstra et al., 2024 ) . Objective. To tackle these challenges, we propose counselor agents that organically integrate psychotherapy skills, incorporating three novel ideas. First, we design a client’s state inference module that models the client’s state of mind to provide personalized counseling tailored to different clients. This module draws inspiration from the transtheoretical model of health behavior change ( Prochaska and DiClemente, 2005 ; Prochaska et al., 2008 ) , which defines five stages of change and suggests that the counselor should adjust their goals and strategies according to the client’s current stage. Second, we introduce a topic tree based on a real counseling dataset and a knowledge base (Wikipedia) that systematically models the topics of change talk exploration in MI counseling. With the help of the topic tree, the counselor can engage with a greater diversity of topics and focus more effectively on capturing client feedback to evoke change talk from clients with various inherent motivations. To the best of our knowledge, this is the first attempt to build a topic exploration method for a counselor agent. Third, we design a ranking-based approach for MI strategies guided response generation. Unlike previous approaches, which generate responses based on all selected strategies, we use a ranking approach for generated candidate responses for each strategy, effectively reducing the overuse of preferred strategies and mitigating bias in strategy selection. Finally, we develop a comprehensive automated evaluation framework using simulated clients that are carefully designed to exhibit realistic client behaviors. The above ideas have been incorporated into our proposed framework known as the STAR (State, Topic, Action, and Response) framework as shown in Figure 1 . Figure 1: The STAR Framework illustrates how an LLM-based counselor agent and a client simulator can be created in an MI-based counseling session. The counselor agent infers the client’s state, explores topics that motivate change talk, and utilizes MI techniques to generate appropriate responses. Check marks represent the selected actions, strategies, or states. The client agent may be simulated by adapting the STAR framework, incorporating modules for state transition, dynamic engagement, action selection, and response generation, allowing the client agent to closely align with the client’s profile and adapt its engagement level with the counselor accordingly. The “Precon.” and “Cont.” indicate Precontemplation and Contemplation, respectively. Contributions. Our work contributes to the field of counselor agent research as follows: (1) Based on the STAR framework, we develop CAMI, a counselor agent that integrates state inference, topic exploration, and strategy selection modules, all aligned with established psychotherapy skills. (2) Through ablation study, we show that the ability of CAMI to infer client’s state significantly improves its performance in MI-based counseling. (3) Our experiment also shows that topic tree and topic exploration further enhances CAMI’s effectiveness in evoking change talk. (4) Our evaluation, combining both automated methods and human expert assessments, demonstrates that CAMI achieves superior performance, outperforming several strong baseline models. 2 Related Work Developing LLM-based agents to support mental health counseling is an emerging research topic. Steenstra et al. (2024) developed LLM-based agents for addressing the alcohol use behavior only. The counselor agent in DIIR ( Xie et al., 2024 ) uses high-quality counseling sessions in the AnnoMI dataset ( Wu et al., 2022 ; Wu et al., 2023a ) to learn a set of strategies in the form of natural language inductive rules. During counseling, DIIR uses the current session context to retrieve the best-matching strategy which prompts LLM to generate the next response. Sun et al. (2024) proposed Chain-of-Strategy (CoS), a prompting method, to align counselor response generation with MI strategies. CARE ( Hsu et al., 2023 ) determines the specific counseling strategy most suitable for a given context and provides example responses for peer counselors. However, the above works focus on modeling MI strategies only and overlook the need to model the client’s state and to evoke change talk based on topic(s) that could match the client’s underlying motivation. We thus propose to perform client’s state inference and topic exploration on a topic tree specially constructed for MI counseling. To compare the performance of counseling agents, the above works rely on evaluation approaches that compare the generated responses against the predefined “ground truth” responses. One such approach computes word overlap or semantic similarity between the generated and ground truth responses. Another approach requires human judgement comparing the generated responses with the ground truth ones at the turn-level when the counseling agent is given all