Attribute-Grounded Selective Reasoning for Artwork Emotion Understanding with Multimodal Large Language Models

Researchers introduced FAB-G, an attribute-grounded selective reasoning framework for artwork emotion understanding with multimodal large language models.

They also extended EmoArt with 1,400 new salience annotations from art-trained annotators to better link emotions to relevant visual attributes.

The approach outperformed prompting-based baselines on emotion, arousal, valence, and explanation quality.

The work addresses a common multimodal failure mode by making predictions more interpretable and improving cross-dataset transfer.