Can AI understand humor? It's a question that has puzzled researchers for years. While large language models (LLMs) excel at many language-based tasks, humor—with its nuances, ambiguities, and reliance on context—has remained a significant hurdle. A new study explores whether giving LLMs access to more than just text can help them finally get the joke. The research focuses on "multimodal prompting," providing the AI with both the text of a joke and an audio version generated by a text-to-speech system. The idea is that the audio adds another layer of information, capturing elements like phonetic ambiguity (think puns) that might be missed in text alone. The results, tested across several humor datasets, are promising. LLMs given both text and audio consistently generated better explanations of jokes than those relying solely on text. This suggests that multimodal input can significantly enhance an LLM's humor comprehension. The researchers investigated how the model processes information, discovering that it effectively captures phonetic ambiguities—the different possible interpretations of a word based on how it sounds—in its internal representations. For example, in the pun "Patience is a heavy weight," the model recognizes both the intended meaning of "weight" and the alternative meaning, "wait." However, the study also highlights challenges. LLMs are incredibly sensitive to the way prompts are phrased, meaning slight changes can significantly alter the output. Additionally, while audio helps with puns, it doesn't fully capture elements like timing and rhythm crucial to other forms of humor. The future of AI humor comprehension likely lies in even richer multimodal inputs. Imagine an AI processing video alongside text and audio, picking up on facial expressions and body language. This research is a step towards AI that not only understands language but also the subtle ways we use it to make each other laugh.