This research explores how Large Language Models (LLMs) can interpret facial expressions through numerical Valence-Arousal values instead of requiring visual input.

• LLMs show surprising ability to infer emotions from dimensional values alone
• This approach is more resource-efficient than Vision-Language Models
• Model performs best when provided with both context and dimensional values
• Simple numerical inputs enable LLMs to access non-verbal emotional cues

Medical Impact: This research opens new pathways for mental health assessment and patient emotional monitoring without requiring complex visual processing, potentially enabling more accessible and lightweight emotion recognition systems in healthcare settings.

Beyond Vision: How Large Language Models Interpret Facial Expressions from Valence-Arousal Values