Do Machines Think Emotionally? A Cognitive Appraisal Analysis of Large Language Models

Emotional intelligence is central to human life, and equipping LLMs with coherent emotional reasoning is critical for their use as conversational agents, simulators in human-subject research, and models of cognition. Yet, most existing approaches remain limited to surface-level, categorical prediction of emotions from stimuli, which often fail to generalize to novel, ambiguous, or culturally nuanced scenarios. To build systems that engage in true emotional reasoning, we focus on using the cognitive appraisal theory. Prior work leveraging appraisal theory has mostly studied other-appraisal, often restricted to a few coarse cognitive. Other-appraisal has the caveat of confounding assumptions about external agents’ demographics, not reflecting true emotional reasoning. Thus, a lack of systematic analysis remains regarding how LLMs internally represent emotions through their own cognitive structures. We introduce CoRE (Cognitive Reasoning for Emotions), the first large-scale benchmark of self-appraisals for emotions. CoRE spans 15 emotion categories, 16 appraisal dimensions, and an analysis across 7 models.