@inproceedings{magzym-etal-2026-dual,
title = "Dual Controller Architecture to Enable Adherence of {LLM}-based Counselors to Session- and Discourse-level Aspects of Motivational Interviewing Theory",
author = "Magzym, Yermakhan and
Carcone, April Idalski and
Towner, Elizabeth and
M.D., Christopher Mann and
Badr, M. Safwan and
Kotov, Alexander",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2603/",
pages = "59120--59163",
ISBN = "979-8-89176-459-0",
abstract = "Motivational Interviewing (MI) theory posits that effective health behavior counseling requires adherence to a session structure that sequentially progresses through four fundamental processes of MI: engaging, focusing, evoking, and planning. While previous research has demonstrated remarkable capabilities of Large Language Model (LLM)-based conversational agents for MI counseling at mimicking individual utterances of human counselors, methods to ensure adherence of such agents to session- and discourse-level aspects of MI theory have yet to be explored. To address this limitation, we propose the Dual Controller Architecture (DCA), a neuro-symbolic approach to modeling counseling session discourse and structure for LLM-based MI agents with local and global controllers. The global controller ensures timely and appropriate transitions of MI sessions between the stages focused on the four processes. The local controller leverages expert-curated, process-specific distributions of MI counselor communication behaviors to align the agent{'}s responses with the objectives of each process and ensures the agent{'}s utilization of effective counselor communication patterns. Standard approaches to domain adaptation of LLMs, such as prompting, fine-tuning, and RAG, as well as the DCA, have been implemented as five versions of the Neural Agent for Obesity MI (NAOMI) and evaluated in a user study involving 49 participants with clinically documented obesity. The study revealed that the DCA demonstrates comparable or superior performance, with particularly strong gains in participants' perceived readiness, importance, and confidence for weight loss."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="magzym-etal-2026-dual">
<titleInfo>
<title>Dual Controller Architecture to Enable Adherence of LLM-based Counselors to Session- and Discourse-level Aspects of Motivational Interviewing Theory</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yermakhan</namePart>
<namePart type="family">Magzym</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">April</namePart>
<namePart type="given">Idalski</namePart>
<namePart type="family">Carcone</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Elizabeth</namePart>
<namePart type="family">Towner</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christopher</namePart>
<namePart type="given">Mann</namePart>
<namePart type="family">M.D.</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">M</namePart>
<namePart type="given">Safwan</namePart>
<namePart type="family">Badr</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Kotov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Motivational Interviewing (MI) theory posits that effective health behavior counseling requires adherence to a session structure that sequentially progresses through four fundamental processes of MI: engaging, focusing, evoking, and planning. While previous research has demonstrated remarkable capabilities of Large Language Model (LLM)-based conversational agents for MI counseling at mimicking individual utterances of human counselors, methods to ensure adherence of such agents to session- and discourse-level aspects of MI theory have yet to be explored. To address this limitation, we propose the Dual Controller Architecture (DCA), a neuro-symbolic approach to modeling counseling session discourse and structure for LLM-based MI agents with local and global controllers. The global controller ensures timely and appropriate transitions of MI sessions between the stages focused on the four processes. The local controller leverages expert-curated, process-specific distributions of MI counselor communication behaviors to align the agent’s responses with the objectives of each process and ensures the agent’s utilization of effective counselor communication patterns. Standard approaches to domain adaptation of LLMs, such as prompting, fine-tuning, and RAG, as well as the DCA, have been implemented as five versions of the Neural Agent for Obesity MI (NAOMI) and evaluated in a user study involving 49 participants with clinically documented obesity. The study revealed that the DCA demonstrates comparable or superior performance, with particularly strong gains in participants’ perceived readiness, importance, and confidence for weight loss.</abstract>
<identifier type="citekey">magzym-etal-2026-dual</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2603/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>59120</start>
<end>59163</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Dual Controller Architecture to Enable Adherence of LLM-based Counselors to Session- and Discourse-level Aspects of Motivational Interviewing Theory
%A Magzym, Yermakhan
%A Carcone, April Idalski
%A Towner, Elizabeth
%A M.D., Christopher Mann
%A Badr, M. Safwan
%A Kotov, Alexander
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F magzym-etal-2026-dual
%X Motivational Interviewing (MI) theory posits that effective health behavior counseling requires adherence to a session structure that sequentially progresses through four fundamental processes of MI: engaging, focusing, evoking, and planning. While previous research has demonstrated remarkable capabilities of Large Language Model (LLM)-based conversational agents for MI counseling at mimicking individual utterances of human counselors, methods to ensure adherence of such agents to session- and discourse-level aspects of MI theory have yet to be explored. To address this limitation, we propose the Dual Controller Architecture (DCA), a neuro-symbolic approach to modeling counseling session discourse and structure for LLM-based MI agents with local and global controllers. The global controller ensures timely and appropriate transitions of MI sessions between the stages focused on the four processes. The local controller leverages expert-curated, process-specific distributions of MI counselor communication behaviors to align the agent’s responses with the objectives of each process and ensures the agent’s utilization of effective counselor communication patterns. Standard approaches to domain adaptation of LLMs, such as prompting, fine-tuning, and RAG, as well as the DCA, have been implemented as five versions of the Neural Agent for Obesity MI (NAOMI) and evaluated in a user study involving 49 participants with clinically documented obesity. The study revealed that the DCA demonstrates comparable or superior performance, with particularly strong gains in participants’ perceived readiness, importance, and confidence for weight loss.
%U https://aclanthology.org/2026.emnlp-main.2603/
%P 59120-59163
Markdown (Informal)
[Dual Controller Architecture to Enable Adherence of LLM-based Counselors to Session- and Discourse-level Aspects of Motivational Interviewing Theory](https://aclanthology.org/2026.emnlp-main.2603/) (Magzym et al., EMNLP 2026)
ACL