This paper demonstrates how fine-tuning small 7-8B parameter LLMs can effectively deliver Cognitive Behavioral Therapy (CBT) for depression, achieving significant improvements in therapeutic competencies compared to base models.
-----
https://arxiv.org/abs/2412.00251
🤖 Original Problem:
Depression affects 20.6% of Americans, but many can't access CBT therapy due to cost, stigma, and therapist scarcity. AI could help, but current solutions using prompt engineering are limited and brittle.
-----
🔧 Solution in this Paper:
→ Researchers fine-tuned three small LLMs (Mistral 7B, Qwen 7B, Llama 8B) on 58 sets of synthetic CBT transcripts.
→ Each transcript set contained 20 therapy sessions, designed with proper phase progression and CBT techniques.
→ Models were evaluated through simulated patient interactions using DeepSeek-V2.5 as the patient.
→ Performance was measured using a modified Cognitive Therapy Rating Scale (CTRS).
-----
💡 Key Insights:
→ Small LLMs can effectively learn therapeutic competencies through targeted fine-tuning
→ CBT-tuned models maintained better session structure and technique implementation
→ Models showed strong empathy but struggled with agenda adherence and deep exploration
→ Performance degraded as context length approached 4000 tokens
-----
📊 Results:
→ CBT-tuned models outperformed base versions by 11.33 points on CTRS (p<0.001)
→ Llama 8B performed best with mean CTRS score of 67.86 ±7.24
→ All models showed significant improvements in therapeutic competencies
→ Strongest gains in Agenda Setting, Guided Discovery, and CBT technique application
