PRISM framework helps LLMs pinpoint true drivers of negative thoughts

Large Language Models are increasingly used for emotional support, including negative thought reframing—a task that depends on modifying cognitive appraisals, the subjective interpretations that drive negative emotions. Current LLM frameworks attempt to model appraisal by exhaustively evaluating all possible dimensions, treating each as equally important. But in real conversations, some appraisals matter far more than others depending on context. A new arXiv paper from researchers including Hainiu Xu and Yulan He asks the overlooked question: can LLMs infer which appraisal dimensions are actually salient in a given emotional support conversation?

To address this, the team built AppraiSal, a benchmark containing 996 emotional support conversations with human-annotated mental states and salient cognitive appraisal dimensions. They also proposed PRISM, a multi-agent probabilistic framework built on Bayesian Inverse Planning, which models how people reason about others' goals and beliefs. Experiments show PRISM delivers consistent improvements across LLMs of different sizes, particularly on the most salient appraisal dimensions. The work suggests that treating appraisal dimensions as context-dependent—rather than uniform—can meaningfully improve how AI understands and reframes negative thoughts in mental health applications.