CCDWIKI: Decoupling Case Formulation from Response Generation in Longitudinal CBT Dialogue

Authors: Mingwen Zhang, Ruiyang Song, Qinying Wang, Yuchen Wang, Changsheng Ma, Minqiang Yang

Subject: Computation and Language (cs.CL)

arXiv: 2611.05383 [cs.CL] · Submitted 31 Jul 2026

DOI: https://doi.org/10.48550/arXiv.2611.05383

Overview

Cognitive Behavioral Therapy (CBT) requires more than empathic validation. It calls for a structured formulation of how situations, automatic thoughts, emotions, and behaviors interact. Therapists typically formalize this interplay through a Cognitive Conceptualization Diagram, and maintaining longitudinal coherence demands ongoing evidential adjudication—separating tentative hypotheses from established findings as therapy progresses.

Existing dialogue systems track conversational history through retrieval or static user profiles, but they lack the representational capacity to encode epistemic status. As a result, provisional interpretations are often prematurely treated as established facts, which can bias downstream interventions.

The CCDWiki Approach

To address this limitation, the authors introduce CCDWiki, a plug-and-play advisor that frames longitudinal therapy as evidence-grounded hypothesis adjudication.

• Mutable hypothesis entries: Each clinical hypothesis is encoded as a mutable entry bound to both corroborating and contradicting evidence.
• Cross-session updating: Across sessions, accumulating data strengthens well-supported entries, preserves provisional ones, and retracts assumptions that are contradicted.
• Concise guidance output: The resulting formulation is distilled into concise guidance for unmodified therapeutic language models—no backbone retraining required.

Results

Experiments show substantial improvements in both conceptualization and intervention quality without retraining the underlying model. This positions CCDWiki as a lightweight, model-agnostic layer for improving longitudinal CBT dialogue systems.

Why It Matters in 2026

As therapeutic language models move from single-session chat toward longitudinal care, preserving the distinction between hypothesis and fact becomes a core safety and quality requirement. CCDWiki's decoupled design—separating structured case formulation from response generation—offers a practical template for evidence-grounded reasoning that can be layered onto existing clinical dialogue systems.

via ArXiv CL+LG