PRISM: A Probabilistic Framework for Inferring Salient Cognitive Appraisals in Emotional Support Conversations
A recent paper published on arXiv (2607.28648) presents PRISM, a multi-agent framework based on Bayesian inference aimed at pinpointing key cognitive appraisal aspects in discussions about emotional support. This research fills a void in existing LLM-based models for reframing negative thoughts, which usually assess all potential dimensions without recognizing their varying importance in different contexts. The authors have created the AppraiSal benchmark, featuring 996 emotional support dialogues with human-annotated mental states, including key cognitive appraisal aspects. The investigation explores whether LLMs can detect these significant appraisal dimensions from conversations, a crucial element for enhancing emotional support efforts. The study is categorized as a cross-type announcement, indicating its relevance across various domains.
Key facts
- Paper arXiv:2607.28648 introduces PRISM, a multi-agent probabilistic framework.
- PRISM is grounded in Bayesian inference.
- The AppraiSal benchmark contains 996 emotional support conversations.
- Conversations are annotated with human-annotated mental states.
- The benchmark includes salient cognitive appraisal dimensions.
- Current LLM frameworks model cognitive appraisal by exhaustively evaluating all dimensions.
- The research questions whether LLMs can infer salient appraisal dimensions.
- The work focuses on negative thought reframing in emotional support.
Entities
Institutions
- arXiv