The automatic detection of mental health dynamics represents a challenging and socially relevant task, requiring models to capture subtle psychological changes over time. Large language models have demonstrated remarkable capabilities in clinical and diagnostic contexts, yet their outputs often lack consistency and may diverge significantly across different model families and prompting strategies. This paper presents the USAI team’s submission to the CLPsych 2026 Shared Task, targeting Tasks 1.1, 1.2, 2, and 3.1. We propose an ensemble-based approach combining multiple open-source large language models, where the contribution of each model is weighted according to its alignment with clinically grounded human annotations on the training set. Our system achieves competitive results across the evaluated subtasks, with particularly strong performance on Tasks 1.2 and 2.
P2P - from Posts to Patterns: An LLM Ensemble Approach to Mental Health Dynamics Detection
Mira, Antonietta;
2026-01-01
Abstract
The automatic detection of mental health dynamics represents a challenging and socially relevant task, requiring models to capture subtle psychological changes over time. Large language models have demonstrated remarkable capabilities in clinical and diagnostic contexts, yet their outputs often lack consistency and may diverge significantly across different model families and prompting strategies. This paper presents the USAI team’s submission to the CLPsych 2026 Shared Task, targeting Tasks 1.1, 1.2, 2, and 3.1. We propose an ensemble-based approach combining multiple open-source large language models, where the contribution of each model is weighted according to its alignment with clinically grounded human annotations on the training set. Our system achieves competitive results across the evaluated subtasks, with particularly strong performance on Tasks 1.2 and 2.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
