Title: v-HUB: A Benchmark for Video Humor Understanding from Vision and Sound

Abstract:

The ability of artificial intelligence systems to interpret humor offers significant practical applications, such as improving engagement levels in interactions between humans and machines. To assess and diagnose how well multimodal large language models (MLLMs) grasp humor, we present v-HUB, a new benchmark designed specifically for video humor comprehension. This benchmark features a carefully selected set of short, non-verbal videos that mirror real-world situations where humor is conveyed entirely through visual signals. Each video is accompanied by detailed annotations to facilitate various evaluation tasks and analytical studies, including an innovative examination of how environmental audio can amplify humorous content. To enhance the benchmark's utility, we have developed an open-ended question-and-answer task, allowing v-HUB to be easily incorporated into existing video understanding frameworks. We tested a wide range of MLLMs, spanning from dedicated Video-LLMs to versatile OmniLLMs capable of native audio processing, across both open-source and proprietary models. The results reveal the substantial challenges MLLMs encounter when trying to understand humor based solely on visual information. Furthermore, our analysis shows that adding audio cues significantly aids in comprehending video humor, underscoring the potential of integrating multiple modalities to tackle complex video understanding challenges.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC