Tag
This study investigates how LLMs ground abstract concepts compared to humans, finding a significant 'grounding gap' where models rely heavily on word associations rather than emotional or internal states. Using sparse autoencoders, the authors identify internal features related to grounding dimensions, suggesting LLMs possess this information but do not recruit it naturally during generation.
A comprehensive survey examining image classification into high-level and abstract categories, clarifying the tacit understanding of high-level semantics in computer vision through multidisciplinary analysis of commonsense, emotional, aesthetic, and interpretative semantics. The paper identifies persistent challenges in abstract concept image classification and emphasizes the importance of hybrid AI systems for addressing complex visual reasoning tasks.