pith. sign in

arxiv: 2604.06551 · v1 · submitted 2026-04-08 · 💻 cs.CL

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

Chang Liu , Changsheng Ma , Yongfeng Tao , Bin Hu , Minqiang Yang This is my paper

Pith reviewed 2026-05-10 18:42 UTC · model grok-4.3

classification 💻 cs.CL
keywords cognitive behavioral therapymulti-agent systemslarge language modelscognitive conceptualization diagrammental health supporttherapeutic simulationinformation asymmetrysynthetic dataset
0
0 comments X

The pith

A multi-agent framework with dynamic cognitive diagrams and asymmetric roles improves LLM-based CBT simulations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Existing approaches to simulating Cognitive Behavioral Therapy with large language models rely on static client profiles and agents that possess complete knowledge of the situation, which overlooks the evolving understandings and partial information that characterize actual therapy sessions. CCD-CBT addresses this by introducing a Control Agent that continuously updates a Cognitive Conceptualization Diagram to guide the interaction while requiring the Therapist Agent to reason solely from inferred client states rather than full access. The resulting synthetic dataset, CCDCHAT, is used to fine-tune models that achieve stronger performance on clinical fidelity measures and greater gains in client positive affect compared to strong baselines. Ablation experiments demonstrate that both the ongoing diagram updates and the information asymmetry are essential to these improvements.

Core claim

By shifting CBT simulation from static profiles and omniscient single-agent designs to a dynamically reconstructed Cognitive Conceptualization Diagram maintained by a Control Agent together with enforced information asymmetry between Therapist and Client Agents, the CCD-CBT framework generates more clinically plausible therapeutic dialogues in language models.

What carries the argument

The CCD-CBT multi-agent framework, in which a Control Agent maintains and updates a Cognitive Conceptualization Diagram to structure information-asymmetric exchanges between a Therapist Agent and a Client Agent.

If this is right

  • Fine-tuned models on the CCDCHAT dataset outperform baselines in counseling fidelity as assessed by clinical scales.
  • These models also produce larger improvements in positive affect for simulated clients.
  • Removing dynamic CCD updates or the asymmetric agent roles leads to measurable drops in performance, confirming both components are required.
  • The approach supplies a concrete method for constructing theory-grounded conversational agents that better align with established CBT principles.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same structure of live diagram updates plus information asymmetry could be tested in simulations of other therapy modalities to check whether the gains generalize.
  • Deploying the trained models in pilot apps with consenting users would allow direct comparison of engagement and reported helpfulness against non-asymmetric systems.
  • Adding mechanisms for the diagram to incorporate real-time user feedback might further reduce the gap between synthetic training and live sessions.

Load-bearing premise

That a synthetic multi-turn dataset generated by the framework, together with ratings from clinical scales and expert therapists, sufficiently represents the dynamic and asymmetric character of real human therapy interactions.

What would settle it

A controlled study measuring actual symptom reduction in real clients who interact over multiple sessions with a model fine-tuned on CCDCHAT versus clients using baseline models or standard care.

Figures

Figures reproduced from arXiv: 2604.06551 by Bin Hu, Chang Liu, Changsheng Ma, Minqiang Yang, Yongfeng Tao.

Figure 1
Figure 1. Figure 1: Comparison of CBT simulation paradigms: early prompt-based methods (upper), recent static cog￾nitive modeling (middle), and our dynamic multi-agent framework (lower). illustrated in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the CCD-CBT framework. The system consists of a Client Agent, a Therapist Agent and a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of situational contexts in CCD [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Human evaluation results comparing CCDCHAT with CACTUS and PsyDTCorpus across four dimen￾sions. 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. pairwise comparisons on 50 randomly sampled di￾alogues from each of CCDCHAT, CACTUS, and PsyDTCorpus. Each dialogue was… view at source ↗
Figure 5
Figure 5. Figure 5: The components of a Cognitive Conceptual [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: CCD generation prompt. • Neutral attitude: The client shows occa￾sional willingness to follow guidance and co￾operation, but may also display skepticism, hesitation, or mild defensiveness depending on the context. • Positive attitude: The client demonstrates high levels of engagement, openness, and co￾operation throughout the session, actively par￾ticipating in therapeutic tasks and discussions. The detail… view at source ↗
Figure 8
Figure 8. Figure 8: Therapist Agent prompt. constructing and updating a structured CCD based on the dialogue history. Following a standardized 8-step cognitive process, it generates the current CCD and determines the next therapeutic phase, ensuring the intervention remains aligned with core CBT principles. Assessment Agent. The Assessment Agent functions as a CBT evaluation assistant, employing a Scale Assessment Method (Bec… view at source ↗
Figure 10
Figure 10. Figure 10: Assessment Agent prompt [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: Identification Agent prompt. the current cognitive-affective state in a structured format, which directly guides the subsequent therapeutic intervention in alignment with CBT principles. Intervention Agent. The Intervention Agent serves as a CBT cognitive intervention assistant, responsible for guiding clients through structured thought restructuring and behavioral change. As illustrated in [PITH_FULL_IMA… view at source ↗
Figure 11
Figure 11. Figure 11: Intervention Agent prompt. As shown in [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 14
Figure 14. Figure 14: CTRS Evaluation Prompt [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗
Figure 13
Figure 13. Figure 13: Human Evaluation Prompt. C.1 Human Evaluation To ensure clinical validity and therapeutic quality, two licensed CBT practitioners evaluated dialogues using a structured 4-point rubric. The assessment covered four dimensions: Helpfulness, Empathy, Logical Coherence, and Guidance. The detailed scoring criteria for Empathy, Logical Coherence, and Guidance follow the definitions in Lee et al. (2024); the Help… view at source ↗
Figure 15
Figure 15. Figure 15: CTRS Evaluation Prompt. on key cognitions or behaviors, change strategies, technique application, and homework). For our evaluation, we selected a streamlined subset of six criteria that are reliably observable in text-based di￾alogues: General skills: Understanding, Interper￾sonal Effectiveness, Collaboration;CBT-specific skills: Guided Discovery, Focus, Strategy. The prompt used to guide CTRS scoring by… view at source ↗
Figure 16
Figure 16. Figure 16: Ablation Experiment Prompt. Strong, Enthusiastic, Proud, Alert, Inspired, De￾termined, Attentive, Active, Interest) and one for negative emotions (e.g., Distressed, Upset, Guilty, Scared, Hostile, Irritable, Ashamed, Nervous, Jit￾tery, Afraid). Respondents rate each item on a 5-point scale from 1 (very slightly or not at all) to 5 (extremely), indicating how much they have experienced each emotion. In thi… view at source ↗
Figure 17
Figure 17. Figure 17: Example of Cognitive Conceptualization Diagram (CCD) for the client. [PITH_FULL_IMAGE:figures/full_fig_p018_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Example of identification phase [PITH_FULL_IMAGE:figures/full_fig_p018_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Example of assessment phase. Strategy 1: Guide client to provide evidence supporting and opposing automatic thoughts. Counselor: Well, I understand your concerns; these thoughts can be hard to shake off. However, resting when you’re sick is actually part of enhancing your work performance. Is it possible that, to some extent, you can only be more efficient at work after a good rest? Have you ever felt mor… view at source ↗
Figure 20
Figure 20. Figure 20: Example of intervention phase [PITH_FULL_IMAGE:figures/full_fig_p019_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Example of summary phase [PITH_FULL_IMAGE:figures/full_fig_p020_21.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces CCD-CBT, a multi-agent framework for CBT simulation that replaces static cognitive profiles and omniscient single-agent setups with a Control Agent that dynamically reconstructs and updates a Cognitive Conceptualization Diagram (CCD) while enforcing information asymmetry between Therapist and Client agents. It generates and releases the synthetic CCDCHAT multi-turn dataset under this framework, then reports that LLMs fine-tuned on CCDCHAT outperform baselines on clinical scales and expert therapist ratings for counseling fidelity and positive-affect enhancement, with ablations supporting the necessity of dynamic CCD guidance and asymmetric roles.

Significance. If the central claims hold after addressing evaluation independence, the work could advance theory-grounded mental-health agents by better approximating real therapy's dynamic and asymmetric information flow; the dataset release and explicit use of CBT's CCD construct are positive contributions that could support reproducible follow-on research.

major comments (3)
  1. [§3 and §4.1] §3 (CCD-CBT Framework) and §4.1 (CCDCHAT Construction): The entire CCDCHAT dataset is generated inside the proposed multi-agent pipeline (Control Agent updating CCD, Therapist reasoning from partial states). This makes training and test distributions dependent on the exact mechanisms being evaluated, so reported gains in fidelity and affect may reflect distribution match rather than genuine capture of real therapy asymmetry.
  2. [§4.3] §4.3 (Evaluation Protocol): The manuscript provides no details on whether expert ratings or clinical-scale assessments used held-out dialogues generated by a different process, real therapist-client transcripts, or external benchmarks. Without such separation, the outperformance claim and the ablation results confirming dynamic CCD and asymmetry cannot be distinguished from artifacts of the synthetic generation process.
  3. [Ablation Studies] Ablation Studies (Table 2 or equivalent): The 'no dynamic CCD' and 'symmetric agent' conditions must be shown to have been created without reusing the same Control Agent or generation pipeline; otherwise the ablations do not isolate the claimed contributions and the necessity argument is weakened.
minor comments (2)
  1. [Abstract and §4] The abstract and §4 could state the exact number of dialogues in CCDCHAT, the number of expert raters, and inter-rater agreement statistics for transparency.
  2. [§3] Notation for the CCD update rule and the asymmetric state representations could be formalized with a small diagram or pseudocode to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of our evaluation methodology that we will clarify in the revision. We address each major comment below.

read point-by-point responses

  1. Referee: [§3 and §4.1] §3 (CCD-CBT Framework) and §4.1 (CCDCHAT Construction): The entire CCDCHAT dataset is generated inside the proposed multi-agent pipeline (Control Agent updating CCD, Therapist reasoning from partial states). This makes training and test distributions dependent on the exact mechanisms being evaluated, so reported gains in fidelity and affect may reflect distribution match rather than genuine capture of real therapy asymmetry.

    Authors: We acknowledge that CCDCHAT is entirely generated via the CCD-CBT multi-agent pipeline, as this controlled synthetic generation is the core of our contribution for simulating dynamic, asymmetric CBT interactions. The train/test split uses held-out client profiles and distinct generation seeds not seen in training to create distributional separation within the framework. This design allows us to isolate the effects of our proposed mechanisms rather than claiming direct equivalence to real therapy data. We will revise §4.1 to explicitly describe the splitting procedure and add a Limitations section discussing the synthetic nature of the data and its implications for generalizability to real-world transcripts. revision: yes

  2. Referee: [§4.3] §4.3 (Evaluation Protocol): The manuscript provides no details on whether expert ratings or clinical-scale assessments used held-out dialogues generated by a different process, real therapist-client transcripts, or external benchmarks. Without such separation, the outperformance claim and the ablation results confirming dynamic CCD and asymmetry cannot be distinguished from artifacts of the synthetic generation process.

    Authors: All clinical-scale assessments and expert ratings were performed on held-out dialogues from CCDCHAT, generated with the same pipeline but using previously unseen client profiles and conversation seeds to ensure they were not part of the training distribution. No real therapist-client transcripts or external benchmarks were used, given ethical and privacy constraints on accessing such data at scale. We will update §4.3 with a full description of this held-out protocol, including how dialogues were sampled for expert review and the exact clinical scales applied, to eliminate ambiguity. revision: yes

  3. Referee: [Ablation Studies] Ablation Studies (Table 2 or equivalent): The 'no dynamic CCD' and 'symmetric agent' conditions must be shown to have been created without reusing the same Control Agent or generation pipeline; otherwise the ablations do not isolate the claimed contributions and the necessity argument is weakened.

    Authors: The ablation variants were produced through independent generation runs in which the Control Agent's CCD update logic was disabled for the 'no dynamic CCD' condition and full state sharing was enabled for the 'symmetric agent' condition. These runs used the same agent code base but with the targeted modifications and separate random seeds, ensuring the performance differences arise from the ablated components. We will expand the ablation section to document the exact generation configurations and confirm the independence of each condition from the main dataset. revision: yes

Circularity Check

0 steps flagged

No significant circularity: novel multi-agent components and external expert validation keep derivation self-contained

full rationale

The paper introduces a multi-agent framework with a Control Agent for dynamic CCD updates and asymmetric Therapist/Client roles, generates the CCDCHAT synthetic dataset under this framework, and reports that fine-tuned models outperform baselines on counseling fidelity and positive-affect metrics per clinical scales and expert therapist evaluations. No load-bearing step reduces to a self-definition, a fitted parameter renamed as prediction, or a self-citation chain; the ablations test the added components within the generated data, yet the human-expert and scale-based evaluations supply independent external grounding. The central claims therefore rest on empirical comparisons rather than any construction that equates outputs to inputs by definition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the assumption that LLMs can role-play therapeutic interactions when provided with appropriate guiding structures; the Control Agent and asymmetric design are new constructs introduced here without independent external validation in the abstract.

axioms (1)
  • domain assumption Large language models can simulate realistic therapeutic interactions when guided by structured cognitive models such as the CCD.
    The entire simulation pipeline depends on LLMs' capacity to maintain consistent client and therapist personas under the described constraints.
invented entities (1)
  • Control Agent no independent evidence
    purpose: Dynamically reconstruct and update the Cognitive Conceptualization Diagram during multi-turn interactions.
    This agent is introduced as a new component to handle CCD updates separate from the Therapist and Client agents.

pith-pipeline@v0.9.0 · 5478 in / 1406 out tokens · 41115 ms · 2026-05-10T18:42:59.076029+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    GPT-4o System Card

    Psychatbot: a psychological counseling agent towards depressed chinese population based on cog- nitive behavioural therapy.ACM Transactions on Asian and Low-Resource Language Information Pro- cessing. Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu. 2023. SoulChat: Improving LLMs’ empathy, listening, and comfort ...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Diacbt: A long-periodic dialogue corpus guided by cognitive conceptualization diagram for cbt-based psychological counseling.arXiv preprint arXiv:2509.02999, 2025

    Development and validation of brief measures of positive and negative affect: the panas scales.Jour- nal of personality and social psychology, 54(6):1063. CONSTITUTION OF WHO. 2020. World health orga- nization.Air Quality Guidelines for Europe, (91). Mengxi Xiao, Qianqian Xie, Ziyan Kuang, Zhicheng Liu, Kailai Yang, Min Peng, Weiguang Han, and Jimin Huang...

    work page arXiv 2020