CCD-CBT: Multi-Agent Therapeutic Interaction for CBT Guided by Cognitive Conceptualization Diagram

Chang Liu; Changsheng Ma; Yongfeng Tao; Bin Hu; Minqiang Yang

cs.CLApr 8, 2026

CCD-CBT: Multi-Agent Therapeutic Interaction for CBT Guided by Cognitive Conceptualization Diagram

Authors: Chang Liu, Changsheng Ma, Yongfeng Tao, Bin Hu, Minqiang Yang

Organizations: School of Information Science and Engineering, Lanzhou University

Abstract

Large language models show potential for scalable mental-health support by simulating Cognitive Behavioral Therapy (CBT) counselors. However, existing methods often rely on static cognitive profiles and omniscient single-agent simulation, failing to capture the dynamic, information-asymmetric nature of real therapy. We introduce CCD-CBT, a multi-agent framework that shifts CBT simulation along two axes: 1) from a static to a dynamically reconstructed Cognitive Conceptualization Diagram (CCD), updated by a dedicated Control Agent, and 2) from omniscient to information-asymmetric interaction, where the Therapist Agent must reason from inferred client states. We release CCDCHAT, a synthetic multi-turn CBT dataset generated under this framework. Evaluations with clinical scales and expert therapists show that models fine-tuned on CCDCHAT outperform strong baselines in both counseling fidelity and positive-affect enhancement, with ablations confirming the necessity of dynamic CCD guidance and asymmetric agent design. Our work offers a new paradigm for building theory-grounded, clinically-plausible conversational agents.

Figures & tables

Figure 1: Comparison of CBT simulation paradigms: early prompt-based methods (upper), recent static cognitive modeling (middle), and our dynamic multi-agent framework (lower).

Figure 2: Overview of the CCD-CBT framework. The system consists of a Client Agent, a Therapist Agent and a Control Agent that interact based on the Specific Cognitive Conceptualization Diagram (CCD), guiding the dialogue through four therapeutic phases: Identification, Assessment, Intervention, and Summary. * denotes key states passed to the subsequent phase.

Datasets	Language	Available	Strategy	CBT-based	CCD-guided	Avg. turns
Psych8k ( Liu et al., 2023 )	English	✓	✗	✗	✗	1.0
SmileChat ( Qiu et al., 2024a )	Chinese	✓	✗	✗	✗	10.4
HealMe ( Xiao et al., 2024 )	English	✗	✓	✓	✗	3.0
CBT-LLM ( Na, 2024 )	Chinese	✓	✓	✓	✗	1.0
CACTUS ( Lee et al., 2024 )	English	✓	✓	✓	✗	16.6
PsyDTCorpus ( Xie et al., 2025 )	Chinese	✓	✓	✓	✗	18.0

Table 1: Comparison of existing mental health dialogue datasets. ✓ indicates support or availability; ✗ indicates not supported or unavailable.

Figure 3: Distribution of situational contexts in CCD-CBT. The dataset covers seven life domains, with Family (25%) and Work (20%) being the most frequent triggers for psychological distress.

Models	General Counseling Skills			CBT-specific Skills
Models	Understanding	Interpersonal Eff.	Collaboration	Guided Discovery	Focus	Strategy
PsyChat	$4.50$	$4.36$	$4.48$	$4.20$	$4.19$	$3.77$
SoulChat	$4.32$	$4.35$	$4.42$	$3.92$	$4.13$	$3.91$
MeChat	$4.52$	$4.44$	$4.48$	$3.98$	$4.24$	$4.06$
Camel	$4.78$	$4.52$	$4.68$	$4.50$	$4.39$	$4.16$
SoulChat2.0	$4.88$	$4.68$	$4.83$	$4.61$	4.57	$4.22$

Table 2: Comparison of general counseling and CBT-specific skills across models. Higher scores indicate better performance.

Models	Positive Attitude		Neutral Attitude		Negative Attitude
Models	positive $\uparrow$	negative $\downarrow$	positive $\uparrow$	negative $\downarrow$	positive $\uparrow$	negative $\downarrow$
PsyChat	+1.90	-1.38	+1.88	-1.82	+1.94	-1.84
SoulChat	+1.94	-1.74	+0.88	-1.34	+0.06	-0.24
MeChat	+2.66	-1.56	+1.62	-0.76	+0.12	-0.56
Camel	+2.22	-0.46	+0.40	-1.90	+0.18	-1.52
SoulChat2.0	+4.46	-1.18	+3.80	-1.50	+1.28	-0.80

Table 3: Results of PANAS across various models. This table illustrates the shifts in positive and negative emotions before and after counseling sessions, categorized by the client’s attitude. Higher positive values and lower negative values indicate better performance.

Models	General Counseling Skills			CBT-specific Skills
Models	Understanding	Interpersonal Eff.	Collaboration	Guided Discovery	Focus	Strategy
w/o CBT	$4.42$	$4.35$	$4.54$	$4.34$	$4.34$	$4.11$
CBT w/o CCD	$4.71$	$4.51$	$4.71$	$4.63$	$4.47$	$4.23$
CBT w/ CCD	4.90	4.69	4.87	4.74	4.57	4.36

Table 4: Ablation study evaluating the contribution of CBT integration and the control module.

Panel of Figure 4 — Figure 4: Human evaluation results comparing CCDChat with CACTUS and PsyDTCorpus across four dimensions. In 100 pairwise comparisons per dimension, CCDChat is consistently preferred over both baselines. All differences are statistically significant ( $p<0.05$ ) according to one-sided binomial tests.

Component	Not	Slightly	Very	Avg.
Overall	$0$	$104$	$346$	$2.77$
Situation	$1$	$135$	$314$	$2.70$
Emotions	$3$	$152$	$295$	$2.65$
Behaviors	$1$	$108$	$341$	$2.76$
Core Beliefs	$1$	$129$	$320$	$2.71$
Relevant History	$2$	$169$	$279$	$2.62$

Table 5: CCD reconstruction similarity ratings across components (total

N=450

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: The components of a Cognitive Conceptualization Diagram (CCD) and the relationships between them.

Category	General Counseling Skills			CBT-specific Skills
Category	Understanding	Interpersonal Eff.	Collaboration	Guided Discovery	Focus	Strategy
AVG. NEG	$4.92$	$4.75$	$5.07$	$4.85$	$4.73$	$4.39$
AVG. NEU	$5.07$	$4.76$	5.17	4.97	4.90	4.41
AVG. POS	5.11	4.82	$5.16$	$4.85$	$4.88$	$4.36$

Appendix

Table 6: Impact of client attitude on simulator performance.

Category	General Counseling Skills			CBT-specific Skills
Category	Understanding	Interpersonal Eff.	Collaboration	Guided Discovery	Focus	Strategy
Helpless	$5.03$	$4.73$	$5.13$	4.93	$4.83$	4.43
Unlovable	5.04	$4.75$	$5.11$	$4.81$	$4.83$	$4.35$
Worthless	$5.03$	4.86	5.15	$4.92$	4.85	$4.37$

Appendix

Table 7: Impact of client’s core belief on simulator performance.

Category	Total	Client	Therapist
Cases	$4500$	-	-
Characters per case	$4570.91$	$2901.64$	$1669.27$
Characters per utterance	$281.27$	$178.55$	$102.72$

Appendix

Table 8: Statistics of the CCDChat dataset.

Datasets	General Counseling Skills			CBT-specific Skills
Datasets	Understanding	Interpersonal Eff.	Collaboration	Guided Discovery	Focus	Strategy
CACTUS	$4.84$	$4.64$	$5.06$	$4.68$	$4.44$	$4.32$
PsyDTCorpus	$5.06$	$4.80$	$5.30$	$4.92$	4.92	$4.24$
CCDChat	5.20	4.94	5.32	4.96	4.92	4.50

Appendix

Table 9: Automated CTRS evaluation of 150 randomly sampled cases across different datasets using GPT-4o-mini.

Model	CBT-specific Skills
Model	Guided Discovery	Focus	Strategy
Camel	$4.78$	$5.03$	$4.90$
SoulChat2.0	$4.38$	$4.82$	$4.48$
Ours	4.85	5.26	5.34

Appendix

Table 10: Cross-model evaluation results on 150 randomly sampled instances using Gemini-3-Flash as the judge.

Explore similar work

DMT-CBT: Longitudinal Therapeutic State Modeling for CBT Counseling
Jun 2, 2026Chang Liu, Shuyi Zhang, Changsheng Ma +3Mental Health Counseling Multimodal Large Language Models
School of Information Science and Engineering, Lanzhou University
PSY-STEP: Structuring Therapeutic Targets and Action Sequences for Proactive Counseling Dialogue Systems
Apr 6, 2026Jihyun Lee, Yejin Min, Yejin Jeon +3Conversational Agents Cognitive Behavioral Therapy
Graduate School of Artificial Intelligence, POSTECH · MILA · McGill University +1
Steering the Compass: Aligning Dynamic Psychological Counseling Conversations with Cognitive Behavioral Therapy Strategies
Sep 17, 2026Zimu Wang, Yiwen Jiang, Xiangyu Zhao +10Mental Health Counseling Synthetic Data Generation for Language Models
AIM for Health Lab, Monash University · University of Liverpool · Massachusetts Institute of Technology +2

Jun 2, 2026J/K move · Enter open · S save

DMT-CBT: Longitudinal Therapeutic State Modeling for CBT Counseling

Chang Liu, Shuyi Zhang, Changsheng Ma, Yongfeng Tao+2

School of Information Science and Engineering, Lanzhou University

Mental Health Counseling Multimodal Large Language Models Cognitive Behavioral Therapy

Open paper AskPDF arXiv

Large language models (LLMs) have shown growing potential for Cognitive Behavioral Therapy (CBT) counseling. However, most existing approaches still formulate counseling as a local response generation problem, focusing on empathetic replies within short, text-only, or single-session interactions. We argue that this formulation fundamentally mismatches the nature of real psychotherapy. In clinical CBT, therapy is a longitudinal process in which therapists continuously infer, update, and intervene on evolving therapeutic states across sessions. Realistic CBT further involves multimodal inference and delayed cross-session intervention effects, requiring models to capture longitudinal therapeutic state evolution under partial observability. We propose DMT-CBT, a framework for Dynamic Modeling of evolving Therapeutic states in CBT counseling. DMT-CBT maintains structured therapeutic states across sessions while incorporating multimodal behavioral grounding and tool-augmented intervention to support adaptive therapeutic reasoning. Based on this framework, we construct DMTCorpus, a synthetic multi-session multimodal CBT counseling dataset featuring evolving therapeutic states, image-grounded client behaviors, and cross-session intervention continuity. Experimental results show that DMT-CBT improves counseling fidelity and therapeutic alliance, produces more favorable longitudinal affective trajectories, and preserves therapeutic states more faithfully than post-hoc extraction approaches.

Abstract

Explore similar work

DMT-CBT: Longitudinal Therapeutic State Modeling for CBT Counseling

PSY-STEP: Structuring Therapeutic Targets and Action Sequences for Proactive Counseling Dialogue Systems

Steering the Compass: Aligning Dynamic Psychological Counseling Conversations with Cognitive Behavioral Therapy Strategies