Cognitive Conceptualization Diagrams (CCDs)

• CCDs are structured representations that map how client distress arises from enduring beliefs, triggers, thoughts, emotions, and behaviors in CBT.
• They bridge clinical frameworks with computational models by transforming traditional case formulations into machine-readable state representations.
• They enable high-fidelity synthetic CBT simulations and dynamic, multi-agent dialogue generation with accuracy in representing client cognition.

Cognitive Conceptualization Diagrams (CCDs) are structured representations of case formulation in Cognitive Behavioral Therapy (CBT) that organize how a client’s current distress is linked to enduring beliefs, situational triggers, cognitions, affect, and action. In contemporary computational work, the same construct has been adapted from a clinical formulation tool into a machine-usable representation for client simulation, dialogue generation, hidden evaluation state, and dynamically reconstructed therapist-side reasoning state in LLM-based counseling systems (Zhou et al., 3 Sep 2025, Liu et al., 8 Apr 2026).

1. Clinical basis and canonical components

In standard CBT, a cognitive conceptualization explains how a client’s current difficulties arise from the interaction of enduring beliefs, assumptions, triggers, thoughts, emotions, and actions. Recent computational papers explicitly tie CCDs to Judith Beck’s framework, describing the CCD as “a commonly used representation of a patient’s cognitive model in CBT,” “one of the most widely used tools in CBT,” and, in another formulation, “a clinical tool designed to understand and represent an individual’s cognitive patterns and psychological processes” for case formulation and intervention planning (Zhou et al., 3 Sep 2025, Liu et al., 8 Apr 2026).

The fullest CCD schema used across these papers contains eight components.

Component	Brief description
Relevant History	Important life experiences shaping current beliefs
Core Beliefs	Deep beliefs about self, others, world
Intermediate Beliefs	Rules, assumptions, and attitudes derived from core beliefs
Coping Strategies	Ways of managing distress
Situation	Triggering event or context
Automatic Thoughts	Immediate evaluative thoughts triggered by the situation
Emotions	Affective responses arising from those thoughts
Behaviors	Observable actions resulting from the thought-emotion pattern

The intended dependency structure is consistent across the CBT-oriented papers. Relevant history shapes core beliefs; core beliefs support intermediate beliefs; situations trigger automatic thoughts; automatic thoughts shape emotions and behaviors; and coping strategies mediate how distress is managed (Zhou et al., 3 Sep 2025, Qin et al., 3 Jun 2026). In resistance-aware counseling models, this same chain is treated not only as a therapist’s explanatory note but as an explicit latent state representation that conditions simulated client behavior (Qin et al., 3 Jun 2026).

2. Representation, schema design, and computational simplification

Although the full CCD comprises eight components, computational systems often operationalize a reduced form. DiaCBT states that it “select[s] six key components for formulating a patient’s cognitive model,” namely Core Beliefs, Intermediate Beliefs, Situation, Automatic Thoughts, Emotions, and Behaviors, and identifies these as the components “most relevant to understanding and simulating CBT-based interactions” (Zhou et al., 3 Sep 2025). This reduction preserves the core belief-to-thought-to-emotion/behavior chain while omitting Coping Strategies and Relevant History from the main working version, even though history is still used as context during generation.

In practice, these CCDs are usually represented as semi-structured natural-language records rather than as fully formal graphs. DiaCBT’s appendix gives a concrete six-field record with named slots for core beliefs, intermediate beliefs, situations, automatic thoughts, emotions, and behaviors, while CARS uses a more decomposed schema in which each belief entry contains a theme, an automatic thought with Situation, Cognition, and Reaction, and an intermediate belief with Attitude, Assumption, and Rule (Zhou et al., 3 Sep 2025, Qin et al., 3 Jun 2026). CCD-CBT likewise uses an eight-component formulation, but explicitly notes that it does not provide “a mathematical graph definition, adjacency matrix, ontology language, JSON schema, or serialized formal notation” for the CCD; its structure is procedural and component-based rather than given as an explicit graph data structure (Liu et al., 8 Apr 2026).

Two consequences follow from these design choices. First, the operational CCD is usually more text-slot-oriented than the richer diagrammatic formulations often used in clinical practice. Second, the computational papers treat CCDs less as static pictures than as structured state objects whose fields can be queried, conditioned on, or progressively reconstructed. A plausible implication is that the computational CCD has already become a hybrid object: clinically inspired in content, but engineered as a machine-readable latent schema rather than only a clinician-facing diagram.

3. CCDs in corpus construction and client simulation

DiaCBT makes CCDs the backbone of synthetic CBT dialogue generation. It assigns them five distinct roles: client profile representation, prompting scaffold for synthetic dialogue generation, mechanism for scenario diversity, latent plan for multi-session CBT progression, and simulation state for evaluation-time client role play (Zhou et al., 3 Sep 2025). In its data creation pipeline, human annotators first collect and label CBT sessions and strategies; LLMs then generate CCDs to enrich client profiles; full client-therapist chat sessions are built using CCDs, CBT sessions, and a scripted framework; and experts review and edit the resulting cases (Zhou et al., 3 Sep 2025).

The inputs for CCD generation are drawn from CBT-related datasets, specifically C2D2 and PatternReframe. DiaCBT states that it integrates situation, negative thought, cognitive pattern, and relevant history from these resources “as contexts for generating CCD-based cognitive models” (Zhou et al., 3 Sep 2025). The generated CCDs then guide multi-session dialogue generation: the client side is prompted to role-play “based on a CCD,” while therapist turns are shaped by few-shot CBT session structure and strategy labels. This arrangement is intended to produce client utterances that are psychologically coherent at the level of beliefs, situations, emotions, and behaviors rather than merely demographic or stylistic persona cues.

The corpus is explicitly longitudinal. DiaCBT contains 108 cases and 540 sessions, with five sessions per case, and the paper reports a session-wise strategy progression in which early sessions emphasize Information Gathering and Working with Automatic Thoughts, while later sessions emphasize Working with Intermediate and Core Beliefs (Zhou et al., 3 Sep 2025). The same CCD logic is also used at evaluation time: the test set consists of 140 distinct client CCDs, “only accessible to the AI client and not to the therapist agent” (Zhou et al., 3 Sep 2025). This hidden-state design makes the CCD function as ground-truth client cognition for simulation-based evaluation, not as an input directly observed by the therapist model.

The validation evidence is unusually direct. Three experts evaluated whether simulated clients matched their assigned CCDs, and DiaCBT reports that over 80% of simulated clients were rated very to extremely accurate for each cognitive-model component, with all six CCD components receiving average ratings ranging from very to extremely accurate (Zhou et al., 3 Sep 2025). At the same time, the paper notes a limitation: it does not report inter-annotator agreement specifically for CCD construction or CCD component annotation, even though it does report Fleiss’ Kappa of $\kappa = 0.685$ with $p<0.001$ for human evaluation of model outputs (Zhou et al., 3 Sep 2025).

4. Dynamic reconstruction, information asymmetry, and resistance-aware CCDs

A major subsequent shift is from static CCDs to dynamic, interaction-dependent CCDs. CCD-CBT explicitly contrasts its approach with prior systems that “predefine the global cognitive model prior to the session” and “keep it fixed throughout the interaction,” arguing instead for “a dynamically reconstructed Cognitive Conceptualization Diagram (CCD), updated by a dedicated Control Agent” under information asymmetry (Liu et al., 8 Apr 2026). In this framework, there are two distinct CCDs: the hidden ground-truth client-side CCD, $\mathrm{CCD}^{\text{client}_i}$, and the therapist-side inferred CCD, $\mathrm{CCD}^{\text{therapist}_i}$, maintained by the Control Agent (Liu et al., 8 Apr 2026).

The resulting multi-agent loop is explicit:

$$r_t = \mathcal{LLM}\big( \mathrm{CCD}^{\text{client}_i},\, a_i,\, \mathcal{H}_{t-1} \big)$$

$$s_t = \mathcal{LLM}\big( \mathrm{CCD}^{\text{therapist}_i},\, \pi_t,\, \mathcal{H}_{t-1} \big)$$

$$\pi_t = \mathcal{LLM}\big( \mathrm{CCD}^{\text{therapist}_i},\, \mathbf{z}_t,\, \mathcal{H}_{t-1} \big)$$

Here the client response depends on hidden client cognition, the therapist response depends on inferred therapist-side cognition plus current strategy, and the Control Agent updates strategy from the inferred CCD and a phase tracker $\mathbf z_t$ (Liu et al., 8 Apr 2026). The therapy process is organized into four phases—Identification, Assessment, Intervention, and Summary—and the Identification phase follows a standardized 8-step cognitive process that progressively fills in CCD components from client disclosures (Liu et al., 8 Apr 2026).

CCD-CBT operationalizes this at scale. It constructs 7,500 distinct CCDs from C2D2, generates dialogues under three client attitudes, filters candidates with an LLM proxy of CTRS, excludes dialogues with mean CTRS below 4, and retains 98% of candidates, yielding 4,500 curated sessions (Liu et al., 8 Apr 2026). The model trained on CCDChat outperforms strong baselines on CTRS dimensions and positive-affect enhancement, and an ablation directly comparing “CBT w/o CCD” to “CBT w/ CCD” shows gains on every reported counseling dimension, with the largest drops from removing CCD in Strategy and Guided Discovery (Liu et al., 8 Apr 2026). The inferred therapist-side CCD is also evaluated directly: over $N=450$, overall reconstruction fidelity is 2.77 on a 3-point scale, with Automatic Thoughts the hardest component at 2.48 (Liu et al., 8 Apr 2026).

A distinct extension appears in resistance-aware counseling. The CARS simulator defines a client as

$S = \langle b, p, q, s, t, M \rangle$

where $p<0.001$0 is the stable main CCD and $p<0.001$1 is a session-specific CCD (Qin et al., 3 Jun 2026). The main CCD contains enduring belief structure; the session-specific CCD contains topic-triggered intermediate beliefs, automatic thoughts, and associated emotional reactions. Resistance is then modeled as a belief-driven state transition: the system identifies the current topic, checks whether it triggers $p<0.001$2, and tests whether the counselor’s move is consistent with the client’s maladaptive assumptions or automatic thoughts. If it is, the negative expectation is confirmed, emotion worsens, and “the client’s goal shifts toward defense or avoidance”; if not, cooperation becomes more likely (Qin et al., 3 Jun 2026). In this formulation, CCDs no longer merely summarize a case: they become the mechanism that determines when resistance should occur, why it occurs, and how it should change across turns.

5. Broader diagrammatic analogues and extensions beyond psychotherapy

Outside psychotherapy, several systems study externalized conceptual structures with functions that closely parallel CCDs, even when they do not use the term. In GenAI-assisted hypothesis exploration, an ordered node-link tree serves as a shared representation for high-level conceptual reasoning. It is hierarchical, ordered by depth and horizontal branching, supports node selection, backtracking, and iterative expansion, and is described by participants as “guardrails” for exploration (Ding et al., 21 Mar 2025). In a design probe with 22 participants, the resulting diagrams contained 21.82 nodes on average, with average maximum depth 3.86 levels, average maximum breadth 8.09, and average backtracking 3.95 instances; 20 out of 22 participants gave positive feedback about future use (Ding et al., 21 Mar 2025). Although the paper studies data-analysis hypothesis formation rather than CBT, a plausible implication is that CCDs can also function as externalized control surfaces that scaffold breadth, depth, backtracking, and evidence-linked refinement rather than only serving as end-state documentation.

A related result appears in multimodal planning. A framework based on self-generated intermediate conceptual diagrams improves GPT-4o from 35.5% to 90.2% on Blocksworld, while removing diagrams drops performance to 58%, and replacing rendered diagrams with code without rendering drops performance to 24% (Borazjanizadeh et al., 14 Mar 2025). The diagrams are state representations integrated into graph-of-thought search, beam selection, and backtracking, not decorative visualizations. A plausible implication is that CCDs may benefit from similar separation between schema induction, rendered state representation, and search or deliberation control, especially when the reasoning task is relational and multi-step.

Cognitive Move Diagrams (CMDs) are another neighboring formalism. They do not model CBT case formulation, but they define Agent Cognitive State, Group Cognitive State, and Group Cognitive Move from measured outcome variables and visualize state transitions as a diagram of group cognition over changing conditions (Iorio et al., 2014). For CCD work concerned with longitudinal change, bias detection, or group-level reasoning, a cautious inference is that CMDs provide a formal template for state-transition semantics that ordinary static conceptual diagrams often lack.

Other theoretical work suggests richer semantics for future CCDs. CatCog argues for typed, compositional, shared-space diagrammatics in which structurally different compositions remain directly comparable in a common output space (Al-Mehairi et al., 2016). Abstract commonsense conceptualization introduces explicit instance-to-concept mappings, abstract event layers, and typicality-filtered inheritance (He et al., 2022). “Shapes of Cognition” treats frames, inheritance hierarchies, scripts, lexical construction schemas, episodic instances, and control schemas as recurring “shapes,” and explicitly invites thinking about models as “pictures, be they diagrams or templates” (McShane et al., 16 Sep 2025). These works do not define CCDs directly, but they suggest that a general CCD formalism could be more typed, multi-layered, and process-oriented than current CBT implementations.

6. Limitations, contested scope, and likely directions of development

Current CCD implementations are effective but partial. DiaCBT openly simplifies the full eight-component model to six fields for its main computational use, represents them as textual slots rather than richer diagrammatic or temporally evolving formulations, and does not make the therapist model explicitly predict or update CCD state during fine-tuning (Zhou et al., 3 Sep 2025). CCD-CBT addresses static-profile limitations by reconstructing therapist-side CCDs online, but it remains a single-session, text-only framework, and it notes that its dataset covers only a subset of possible CCDs and is primarily Chinese in source and application context (Liu et al., 8 Apr 2026). CARS is explicitly CBT-centric, uses a fixed CCD library, and acknowledges that some forms of resistance, such as passive withdrawal, may be underrepresented (Qin et al., 3 Jun 2026).

The representational limitations are equally clear. DiaCBT’s operational CCD omits coping strategies and relevant history from the main working form; CCD-CBT notes the absence of an explicit graph data structure; and the resistance-aware work uses structured prompting schemas rather than a typed graph with explicit edge ontologies (Zhou et al., 3 Sep 2025, Liu et al., 8 Apr 2026, Qin et al., 3 Jun 2026). Outside psychotherapy, tree-based conceptual work has shown strong support for branching exploration, but also exposes what trees do poorly: cross-links, convergence, contradiction, uncertainty, temporality, and many-to-many relations (Ding et al., 21 Mar 2025). That matters because real case formulations often require shared causes, feedback, contradictory evidence, and longitudinal or intervention-planning structure.

Taken together, these results suggest that CCD research is moving from static case notes toward explicit state representations for mixed-initiative reasoning, simulation, and evaluation, but has not yet converged on a general formalism. A plausible next step is a richer CCD language with typed relations such as refinement, support, contradiction, causation, moderation, and intervention, plus explicit encoding of evidence, confidence, temporality, and cross-session revision. A second likely direction is broader information asymmetry: therapist-side systems that must infer, revise, and test a CCD rather than receiving it as ground truth. A third is cross-domain generalization, in which CCDs are treated less as psychotherapy-specific diagrams and more as externalized conceptual workspaces for structured reasoning, planning, and explanation.