Organizations: Tsinghua University, China · Hangzhou Municipal Health Commission, China · Department of Child Health Care, Hangzhou Women’s Hospital (Hangzhou Maternity and Child Health Care Hospital), China · Hangzhou Fushu Health Technology Co., Ltd, China
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.
Figures & tables
Dataset
Variability
Word density
Responsiveness
RealCBT
0.312
0.184
0.463
SMILE
0.198
0.132
0.289
CACTUS
0.241
0.149
0.347
SimPsyDial
0.183
0.128
0.274
Ours
0.287
0.171
0.421
TABLE I: Emotional-arc comparison across datasets.
Fig. 1: Working Alliance Inventory scores across datasets. Our corpus achieves the highest bond (4.51) and goal (3.89) scores.
Fig. 2: CBT technique coverage comparison across SMILE, CACTUS, SimPsyDial, and our corpus. Our corpus leads on all four CBT techniques and achieves the highest overall coverage.
Fig. 3: Two-stage CBT dialogue corpus construction. Stage 1 extracts structured patient personas from anonymized counseling reports; Stage 2 simulates multi-turn dialogues between a profile-aware patient agent and a CBT-principled counselor agent.
Fig. 4: SFT training and validation cross-entropy loss over 4,000 steps. The final training loss is 0.536 and the final validation loss is 0.880.
Fig. 5: DPO preference alignment over 200 steps. The reward gap between chosen and rejected responses widens, the reward margin reaches 2.07, and the DPO loss converges to 0.456.
Hyperparameter
Setting
Backbone
Qwen3-14B
GPU
4 × NVIDIA A100 80G
Per-device batch size
1
Gradient accumulation steps
2
Learning rate
10−5
LR scheduler
cosine annealing
TABLE II: Training hyperparameters.
Fig. 6: Overview of MACBT. The assessment navigator schedules one of four intervention agents; all agents read from and write to shared working memory, and the treatment monitoring agent outputs a structured clinician report.
Fig. 8: Dual-loop longitudinal memory architecture. The intra-session loop uses Basic Memory and CD Memory to guide live dialogue; the inter-session loop evaluates the session and updates CD Memory and the treatment plan.
System
Comp.
Prof.
Auth.
Mean
MeChat [ 9 ]
1.28
1.95
1.88
1.46
SoulChat [ 10 ]
1.35
2.01
2.12
1.54
PsyChat [ 20 ]
1.42
2.23
2.08
1.62
CPsyCoun [ 11 ]
1.56
2.48
2.11
1.71
MACBT
1.44
2.62
2.25
1.83
TABLE III: Session-quality comparison (0–3 scale, higher is better).
Setting
Prof.
Auth.
Mean
No fine-tuning
1.74
1.98
1.55
Generic mental-health SFT
2.08
2.11
1.68
CBT-specific SFT
2.48
2.12
1.92
CBT SFT + DPO
2.62
2.25
2.10
TABLE IV: Ablation on training strategy (0–3 scale).
Strategy
Comp.
Prof.
Auth.
Mean
LoRA (rank=16)
1.18
2.21
1.97
1.79
LoRA (rank=64)
1.31
2.38
2.09
1.93
Full fine-tuning
1.31
2.48
1.98
1.92
TABLE V: Comparison of fine-tuning strategies (0–3 scale).
# Dialogues
Comp.
Prof.
Auth.
Mean
500
0.97
1.89
1.72
1.53
1,000
1.18
2.24
2.01
1.81
1,500
1.35
2.47
2.16
1.99
2,000
1.31
2.48
1.98
1.92
TABLE VI: Ablation on training data scale (0–3 scale).
Fig. 9: Performance versus training data scale. Professionalism is the most responsive dimension, rising by 0.73 from 500 to 2,000 dialogues.
System
Session quality
Cont.
Prog.
Long. mean
MACBT (no memory)
1.83
0.91
0.89
0.87
+ Generic summary
1.89
1.52
1.47
1.48
+ Single-session CD
1.96
1.78
1.71
1.72
Full CD Memory
2.06
2.38
2.27
2.29
TABLE VII: Longitudinal comparison over three sessions (0–3 scale).
System variant
Cont.
Prog.
Pers.
Mean
Full CD Memory
2.38
2.27
2.21
2.29
– Efficacy records
2.35
2.24
1.81
2.13
– Inter-session planning
1.97
1.74
2.09
1.93
– Cross-session persistence
1.78
1.64
1.70
1.71
TABLE VIII: Ablation of longitudinal memory components (0–3 scale).
Fig. 10: Heatmap of longitudinal ablation results. Removing cross-session persistence causes the largest drop across all three longitudinal dimensions.