FACET benchmark reveals fragmented emotional intelligence in LLMs
A new framework called FACET (Functional Affective Competence and Empathy Test) has been developed by researchers to assess emotional intelligence in large language models. This psychometrically validated tool consists of 480 items designed by experts and is based on the Mayer-Salovey-Caruso four-branch ability model, which defines EI through the perception, facilitation, understanding, and management of emotions. An analysis of nine leading models, including GPT-5 and Claude-Sonnet-4, revealed that emotional intelligence is not a singular trait but rather consists of various cognitive and interactive aspects. Existing benchmarks often confuse basic politeness with profound emotional reasoning, neglecting the distinction between perceptual accuracy and interactive effectiveness. The research underscores the importance of maintaining the structural integrity of EI as LLMs are deployed in emotionally sensitive areas, emphasizing safety and alignment as essential challenges.
Key facts
- FACET framework comprises 480 expert-crafted items
- Based on Mayer-Salovey-Caruso four-branch ability model
- Evaluated nine frontier models including GPT-5 and Claude-Sonnet-4
- Emotional intelligence is fragmented across perception, cognition, and interaction
- Current benchmarks conflate politeness with affective reasoning
- LLMs are increasingly used in emotionally sensitive domains
- Study published on arXiv with ID 2605.24686
Entities
Institutions
- arXiv