PaperlayerView exact revision on arXiv

Source-linked AI summary

MACBT: A Multi-Agent Cognitive Behavioral Therapy Decision Support System with Longitudinal Memory

De Jiang, Shuo Zhang, Weiwei Liao, Jianying Zhang, Chuanhui Yu, Hongen Liao, Kehong Yuan

arXiv:2609.30939v1cs.AI

TL;DR

Clinicians face substantial preparation, documentation, and longitudinal tracking demands in depression CBT, while existing systems do not combine clinician-facing CBT treatment logic with CBT-specific cross-session memory. The paper presents MACBT with CD Memory and reports stronger dialogue ratings and improved session quality with memory augmentation.

  • Problem

    Clinicians must review prior notes, document sessions, and track changing cognitive pathology, while existing systems do not combine CBT treatment logic, multi-agent reasoning, and cross-session cognitive-distortion memory in a clinician-facing tool.

    Source: §I · p. 1

  • Method

    MACBT assigns CBT’s five treatment stages to collaborative agents and pairs them with CD Memory, which tracks distortion characteristics and restructuring efficacy to generate intervention priorities and pre-session reports.

    Source: §I · p. 2, C. Longitudinal Memory System · p. 5

  • Results

    MACBT achieves the highest professionalism (2.62) and clinical authenticity (2.25) among compared systems; CD Memory raises session quality by 12.6% and achieves a longitudinal mean of 2.29.

    Source: §I · p. 2

  • Takeaways & Limitations

    The system provides clinicians with CBT-stage decision support and distortion-based pre-session reports intended to support planning across sessions.

    Source: §I · p. 2

    The longitudinal test covers only three sessions, so the authors cannot claim the memory design scales to distortions and techniques accumulated over months of therapy.

    Source: B. Limitations · p. 7

Abstract

from arXiv · show

Cognitive behavioral therapy (CBT) is an evidence-based first-line treatment for depression, yet its scale is constrained by the time clinicians spend on pre-session preparation, post-session documentation, and longitudinal cognitive-pathology tracking. We present a clinician-facing AI decision-support system that combines a multi-agent CBT framework (MACBT) with a CBT-specific longitudinal memory module (CD Memory). MACBT encodes the five-stage CBT workflow (assessment, Socratic questioning, cognitive restructuring, behavioral experiments, and treatment monitoring) into five collaborative agents. CD Memory tracks cognitive-distortion type, frequency, severity, and restructuring efficacy across sessions to generate pre-session pathology reports and intervention-priority recommendations. We construct a Chinese CBT dialogue corpus via dual-role large language model simulation and train a Qwen3-14B backbone with supervised fine-tuning and direct preference optimization. Evaluation with GPT-4 judges shows MACBT outperforms MeChat, SoulChat, PsyChat, and CPsyCounX in professionalism (2.62) and clinical authenticity (2.25). The full memory-augmented system further improves session quality by 12.6% and achieves a longitudinal mean of 2.29 on cross-session continuity, intervention progression, and personalization.

I. INTRODUCTION

The paper targets clinician time and continuity burdens in depression CBT, proposing MACBT and CD Memory as clinician-facing decision support. It reports stronger dialogue-system ratings and improved session quality with memory augmentation.

I. INTRODUCTION

  • Clinicians face pre-session review, post-session documentation, and longitudinal tracking burdens that can consume 30–45 minutes per patient for preparation and documentation alone.These demands compress direct therapeutic contact and may increase information loss between sessions.Source: §I · p. 1
  • The paper identifies a gap in clinician-facing tools combining CBT treatment logic, multi-agent reasoning, and cross-session cognitive-distortion memory.Prior systems are described as patient-facing or single-session, while generic memory lacks CBT-specific constructs.Source: §I · p. 1
  • MACBT encodes CBT’s five-stage treatment logic in collaborative agents, while CD Memory tracks cognitive distortions across sessions and generates intervention priorities and pre-session reports.The five agents address assessment, Socratic questioning, cognitive restructuring, behavioral experiments, and treatment monitoring.Source: §I · pp. 1–2
  • MACBT achieves the highest professionalism (2.62) and clinical authenticity (2.25) among the compared systems; adding CD Memory raises session quality by 12.6%.The memory-augmented system achieves a longitudinal mean of 2.29 across continuity, intervention progression, and personalization.Source: §I · p. 2

II. RELATED WORK

Prior digital CBT and mental-health systems provide dialogue, specialized agents, or long-term memory, but the paper identifies a lack of clinician-facing systems that integrate sequential CBT treatment stages with structured cognitive-pathology memory. Its corpus construction uses structured personas and dual-role simulation, and the corpus leads reported CBT technique coverage comparisons.

II. RELATED WORK

  • Prior systems separately provide rule-based patient-facing CBT, LLM counseling, psychiatric assessment agents, or generic and session-level memory, leaving CBT-stage reasoning and structured cross-session distortion tracking uncombined.The paper frames this as a gap in clinician-facing tools that expose intermediate reasoning for supervision.Source: §II · p. 2

A. CBT Dialogue Corpus Construction

  • The corpus begins with 3,134 anonymized counseling reports converted into structured personas, then uses separate patient and counselor agents for iterative dialogue.The counselor follows the five-stage CBT workflow, while the patient reveals information gradually; the patient agent’s limited dialogue history is described as producing emotional exploration rather than a flat, overly cooperative tone.Source: A. CBT Dialogue Corpus Construction · pp. 2–3, Figure 3 · p. 3
  • The corpus’s WAI scores are 4.51 on bond, 3.89 on goal, and 3.73 on task, outperforming SMILE, CACTUS, and SimPsyDial.The figure reports the highest bond and goal scores for the corpus.Source: Figure 1 · p. 3, A. CBT Dialogue Corpus Construction · p. 3
  • Filtering yields 2,000 multi-turn dialogues and roughly 68,000 supervised fine-tuning samples, after applying counselor-response constraints and dialogue stopping conditions.Responses are limited to 40 Chinese characters and at most one question; dialogues otherwise stop at 50 turns.Source: A. CBT Dialogue Corpus Construction · pp. 2–3
  • The corpus scores 0.287 on emotional variability, 0.171 on emotional-word density, and 0.421 on contextual responsiveness, approaching RealCBT’s respective scores of 0.312, 0.184, and 0.463.It substantially exceeds the simulated baselines on these measures.Source: A. CBT Dialogue Corpus Construction · p. 3
  • The largest reported improvement is contextual responsiveness: 0.421 versus 0.289 for SMILE and 0.274 for SimPsyDial, which the paper links to adapting questions to immediate emotional signals.The emotional-variability gap from RealCBT is 0.025, compared with gaps exceeding 0.11 for SMILE and SimPsyDial.Source: A. CBT Dialogue Corpus Construction · p. 3
  • 35.6% overall CBT technique coverage exceeds the strongest baseline’s 31.9%, with the corpus leading on each of four techniques.Coverage is reported as 45.8% for Socratic questioning, 43.2% for cognitive restructuring, 22.8% for behavioral experiments, and 30.6% for consolidation.Source: Figure 2 · p. 3, A. CBT Dialogue Corpus Construction · p. 3

B. MACBT Framework

MACBT organizes the CBT workflow through a navigator and four intervention agents sharing working memory. The section also describes Qwen3-14B training with SFT and DPO.

B. MACBT Framework

  • Qwen3-14B is fully fine-tuned and then aligned with DPO using 1,347 preference pairs selected from four candidate responses per context.Responses are scored for comprehensiveness, professionalism, authenticity, and safety; the highest- and lowest-scoring are selected as preference pairs.Source: B. MACBT Framework · pp. 3–4
  • Over 4,000 SFT steps, final training and validation cross-entropy losses are 0.536 and 0.880, respectively.DPO alignment runs for 200 steps; its reward margin reaches 2.07 and its loss converges to 0.456.Source: Figure 4 · p. 4, Figure 5 · p. 4
  • The navigator maintains CBT stage structure and transitions; specialized agents question, restructure, and design behavioral experiments.The restructuring agent identifies ten distortion categories and uses corresponding techniques; the behavioral experiment agent designs SMART experiments and tracks execution.Source: B. MACBT Framework · p. 4
  • The assessment navigator schedules four intervention agents, which share working memory, while treatment monitoring produces a structured clinician report.The shared memory stores the current stage, identified distortions, and intervention history; the monitoring agent's report is for the clinician.Source: Figure 6 · p. 4, B. MACBT Framework · p. 4
  • A representative session moves from identifying an all-or-nothing distortion through evidence probing, continuum rating, a behavioral task, and efficacy recording.Treatment monitoring records the distortion and intervention efficacy for the next session.Source: B. MACBT Framework · p. 4

C. Longitudinal Memory System

The longitudinal memory system combines within-session memory with after-session evaluation and updates. CD Memory tracks cognitive distortions and intervention efficacy to prioritize treatment targets and inform later sessions.

C. Longitudinal Memory System

  • CD Memory records each detected distortion's type, triggering utterance, 1–5 severity, and restructuring-efficacy history.Efficacy reflects the patient's immediate attitude shift, emotional improvement, and willingness to try a behavioral experiment.Source: C. Longitudinal Memory System · p. 5
  • At each session's start, CD Memory combines normalized recency, frequency, and severity into an intervention priority score.The distortion with the highest score becomes the primary intervention target.Source: C. Longitudinal Memory System · p. 5
  • The system retrieves historically effective and ineffective techniques for the target distortion, building a patient-specific technique-response map over time.For a patient's all-or-nothing thinking, the example recommends continuum techniques when effective and avoids probability estimation when it performs poorly.Source: C. Longitudinal Memory System · p. 5
  • Dynamic prompts combine the priority distortion and background, recommended or avoided techniques, and the current CBT stage.The static prompt defines the CBT role and safety constraints.Source: C. Longitudinal Memory System · p. 5
  • After each session, treatment monitoring evaluates distortion improvement, technique efficacy, and experiment completion, then updates CD Memory and the next-session treatment plan.The plan is injected into the next session's dynamic prompt.Source: C. Longitudinal Memory System · p. 5, Figure 8 · p. 5

A. Evaluation Setup

The evaluation uses GPT-4 as a judge to assess session quality and longitudinal quality across separate test sets.

A. Evaluation Setup

  • GPT-4 judges session quality on four 0–3 Likert dimensions: comprehensiveness, professionalism, authenticity, and safety.The single-session evaluation uses 100 virtual patient profiles.Source: A. Evaluation Setup · p. 5
  • Longitudinal quality is scored on cross-session continuity, intervention progression, and personalization, each on a 0–3 scale.The longitudinal evaluation uses 30 three-session test units.Source: A. Evaluation Setup · p. 5
  • The setup separates single-session testing with 100 virtual patient profiles from longitudinal testing with 30 three-session units.Both evaluations use GPT-4 as judge.Source: A. Evaluation Setup · p. 5

B. Main Results

MACBT achieves the best overall mean among the compared systems, with leading professionalism and authenticity scores. The section also reports gains from CBT-specific training data and DPO, and favors full fine-tuning over the tested LoRA settings.

B. Main Results

  • MACBT records the highest professionalism (2.62) and authenticity (2.25), and the best overall mean (1.83) against MeChat, SoulChat, PsyChat, and CPsyCounX.Its comprehensiveness is slightly lower than CPsyCounX because the navigator focuses depth-first on the current CBT stage rather than listing multiple topics.Source: B. Main Results · pp. 5–6
  • The authors attribute MACBT's professionalism lead over CPsyCounX (2.62 vs. 2.48) to CBT-specific fine-tuning and hard stage constraints preventing treatment-phase skipping.They attribute the authenticity lead (2.25 vs. 2.11) to dual-role data simulating patient-counselor information asymmetry.Source: B. Main Results · p. 6
  • CBT-specific data raises professionalism to 2.48 from 2.08 with generic mental-health data; DPO adds 0.14 professionalism and 0.27 authenticity.The progression also reports 1.74 professionalism for non-fine-tuned Qwen3-14B with the MACBT prompt.Source: B. Main Results · p. 6, Table IV · p. 6
  • Full fine-tuning is selected over LoRA: rank-16 LoRA scores 2.21 professionalism, while rank-64 scores 2.38 but still trails on professionalism and mean score.The comparison is reported in Table V.Source: B. Main Results · p. 6, Table V · p. 6
  • As training dialogues increase from 500 to 2,000, professionalism rises from 1.89 to 2.48; comprehensiveness and authenticity saturate after 1,500 dialogues.The 1,500-to-2,000 increase still yields a 0.15 professionalism gain, and DPO lifts final professionalism to 2.62.Source: B. Main Results · p. 6

C. Longitudinal Results

The full memory-augmented system achieves the highest reported session quality and longitudinal mean, while ablations show that persistent, structured memory supports continuity, progression, and personalization.

C. Longitudinal Results

  • 2.29 longitudinal mean and 2.06 session quality are the full memory-augmented system’s highest scores among the compared variants.The comparison covers MACBT without memory, generic cross-session summaries, and single-session CD Memory.Source: C. Longitudinal Results · p. 6
  • Professionalism rises by 0.73 from 500 to 2,000 dialogues, making it the most responsive dimension to training-data scale.The comparison is reported in Fig. 9.Source: Figure 9 · p. 6
  • Generic cross-session summaries improve continuity from 0.87 to 1.48 but fall short of structured CD Memory, which retains distortion-level planning granularity.The full system’s longitudinal mean is 2.29, compared with 1.72 for single-session CD Memory.Source: C. Longitudinal Results · p. 6
  • The longitudinal mean of 2.29 comprises continuity (2.38), intervention progression (2.27), and personalization (2.21).The reported dimensions correspond to avoiding repeated rediscovery, adjusting treatment focus, and tailoring techniques to accumulated efficacy records.Source: C. Longitudinal Results · p. 6
  • Removing restructuring-efficacy records primarily lowers personalization from 2.21 to 1.81, while removing inter-session planning lowers intervention progression from 2.27 to 1.74.Removing cross-session persistence degrades all three longitudinal metrics.Source: C. Longitudinal Results · p. 6

V. DISCUSSION AND CONCLUSION

The paper presents MACBT and CD Memory as clinician-facing support for depression CBT, addressing preparation, in-session decisions, and longitudinal tracking. Its conclusion highlights structured multi-agent reasoning and cross-session cognitive-distortion memory.

V. DISCUSSION AND CONCLUSION

  • MACBT and CD Memory are presented as a decision-support system and CBT-specific longitudinal memory module for depression CBT.The conclusion frames them as addressing clinician preparation, in-session decision support, and longitudinal tracking.Source: §V · pp. 6–7
  • The longitudinal comparison covers three sessions, and the component ablation examines longitudinal memory components.The supplied table captions identify these respective scopes.Source: Table VII · p. 7, Table VIII · p. 7
  • Structured multi-agent reasoning supports in-session decision support, while cross-session cognitive-distortion memory supports longitudinal tracking.The system’s five collaborative agents and distortion-level tracking are described as complementary contributions.Source: §V · pp. 6–7

A. Key Findings

The discussion reports gains from explicit CBT stage logic, preference alignment, and longitudinal memory, while noting evaluation and data-construction limits. The evidence is bounded by simulated patients, GPT-4 judging, and a short longitudinal test.

A. Key Findings

  • Explicit CBT stage logic in specialized agents improves professionalism and clinical authenticity, while DPO adds 0.14 professionalism and 0.27 authenticity over generic mental-health SFT.The professionalism comparison is 2.08 versus 2.48, and the reported authenticity improvement is the largest.Source: A. Key Findings · p. 7
  • Full fine-tuning reaches 2.48 professionalism versus 2.38 for rank-64 LoRA, whose mean score still trails.The proposed explanation—that LoRA may not fully adapt latent CBT stage boundaries—is qualified as possible.Source: A. Key Findings · p. 7
  • The longitudinal mean rises from 0.87 with no memory to 1.72 with single-session CD Memory and 2.29 with full persistence.The authors identify longitudinal memory as the largest single contributor to cross-session continuity and describe pre-session reports as structured distortion-level summaries with intervention recommendations.Source: A. Key Findings · p. 7

B. Limitations

  • The evaluation uses GPT-4 judges and simulated patients rather than licensed clinicians or real patients.The authors caution that GPT-4 may overweight fluent surface form and underweight clinical nuance; longitudinal testing covers three sessions, not a standard 8–20-session course.Source: B. Limitations · p. 7
  • The 2,000-dialogue corpus uses anonymized counseling reports rather than verbatim CBT transcripts, and its patient personas remain static.The authors also describe the corpus as modest by modern LLM training standards and say safety guardrails, crisis detection, and regulatory compliance need further work before deployment.Source: B. Limitations · p. 7

C. Ethical Considerations and Safety

MACBT uses layered protections and clinician oversight, while the authors identify unresolved risks and propose validation and memory extensions. Future directions include testing across full treatment courses and adapting the design to other structured therapies.

C. Ethical Considerations and Safety

  • Safety protections operate at content, workflow, and human-oversight layers, with outputs framed as clinician draft suggestions rather than direct patient advice.The content layer forbids diagnosis and medication advice and routes urgent language to an escalation message; the workflow layer can terminate or hand off out-of-scope dialogue.Source: C. Ethical Considerations and Safety · p. 7
  • Unresolved risks include judge bias toward fluent responses, uncertain transfer beyond Chinese contexts, report-derived bias, and privacy needs for sensitive longitudinal records.The authors call for clinician validation and strict access controls, encryption, and retention limits for memory records.Source: C. Ethical Considerations and Safety · p. 8

D. Future Work

  • Future work includes clinician-in-the-loop validation across 8–20 sessions, memory compression, multimodal assessment cues, and clinician-feedback learning of priority weights.The design could also be adapted to DBT or ACT.Source: D. Future Work · p. 8
Loading 2609.30939v1…