Multi-Agent LLM Pipeline Detects Delusion in Audio Diaries
Researchers have developed an automated multi-agent large language model (LLM) pipeline for detecting and classifying delusion-related content in naturalistic audio diaries. The system analyzes transcripts from individuals with moderate persecutory ideation, extracting language suggestive of delusional beliefs along with associated affective and behavioral responses. Using an ensemble of three foundation models, the study found that detailed diagnostic prompts reduce false positives for delusional theme classification but constrain interpretation of affective or behavioral responses. The work highlights both the potential and limitations of LLMs in automating mental health symptom detection from speech data.
Key facts
- Pipeline uses multi-agent LLMs for fine-grained multi-label extraction
- Evaluates ensemble of three foundation models
- Detailed diagnostic prompts reduce false positives for delusional themes
- Prompts constrain interpretation of affective/behavioral responses
- Data from naturalistic audio diaries of people with persecutory ideation
- LLMs require annotated data primarily for evaluation, not training
- Published on arXiv with ID 2605.24755
- Focus on automated detection of symptom exacerbation
Entities
Institutions
- arXiv