AVMeme Exam: A Multimodal Multilingual Multicultural Benchmark for LLMs' Contextual and Cultural Knowledge and Thinking
Researchers created AVMeme Exam, a human-curated benchmark of over 1,000 iconic internet sounds and videos to test if multimodal AI can understand context and culture beyond surface content. The study found that current models struggle significantly with textless music and sound effects, revealing a major gap in their ability to reason about implicit meaning.