AICA-Bench: Holistically Examining the Capabilities of VLMs in Affective Image Content Analysis
Jul 2, 2026·,,,,,
Dong She
Xianrong Yao
Liqun Chen
Jinghe Yu
Yang Gao
Zhanpeng Jin
Figure 1 from the paper.Abstract
AICA-Bench evaluates affective image understanding, emotion reasoning, and emotion-guided content generation. An evaluation of 23 vision-language models reveals weaknesses in emotion-intensity calibration and open-ended descriptions. Grounded Affective Tree prompting combines visual scaffolding with hierarchical reasoning to address these gaps without model training.
Type
Publication
Findings of the Association for Computational Linguistics: ACL 2026 (ACL 2026 Findings), pp. 13501–13528