New AI framework combines LLMs with multiple-instance learning for cognitive distortion detection

A novel artificial intelligence framework integrates Large Language Models with Multiple-Instance Learning architecture to improve detection of cognitive distortions, which are closely associated with mental health disorders. The system processes each utterance by breaking it down into Emotion, Logic, and Behavior components. LLMs analyze these ELB components to identify multiple distortion instances, each assigned a type, expression, and model-generated salience score. These instances are then combined using a Multi-View Gated Attention mechanism for final classification. Testing on Korean (KoACD) and English (Therapist QA) datasets showed that incorporating ELB components and LLM-inferred salience scores enhances classification performance, particularly for distortions with high interpretive ambiguity. The approach addresses challenges in automatic detection including contextual ambiguity, co-occurrence, and semantic overlap. By leveraging LLM reasoning capabilities within a MIL architecture, the framework aims to improve both interpretability and expression-level analysis. The research was documented in arXiv preprint 2509.17292v3 with announcement type replace-cross.