A fully unsupervised framework (aRAG) that reframes LLM mental-health screening as completing standardized clinical questionnaires (e.g., Beck Depression Inventory-II, Self-Harm Inventory, SCOFF). For each questionnaire item, an intrinsic-dimension-based adaptive neighborhood (ABIDE-ZS) retrieves an item-specific optimal number k* of the most semantically relevant user posts, which an LLM then scores zero-shot; item scores are summed into a severity classification. This yields interpretable, training-data-free assessment that matches or outperforms supervised state-of-the-art on eRisk depression-severity benchmarks and generalizes to self-harm, anorexia, and pathological-gambling screening.
This paper introduces a questionnaire-guided mental-health screening framework that uses adaptive Retrieval-Augmented Generation (aRAG) to let large language models complete standardized clinical psy…