A team of Chinese scientists, in collaboration with international peers, is tackling the challenge of humor in artificial intelligence. They aim to develop a model capable of creating its own jokes and providing witty comments on encountered images, similar to how internet users do. Humor, being highly subjective and context-dependent, presents a difficult task for translation and AI. The researchers from Sun Yat-sen University, in partnership with Singapore Management University and Harvard University, hope to enhance creativity by creating a model that can generate jokes that resonate with different individuals. In a survey involving 154 internet users, the AI's jokes in English, Japanese, and Mandarin Chinese were found funnier than comparable models such as Microsoft Research's LLaVA-1. 5 and OpenAI's GPT-4v. The team addressed the challenge of reasoning-based language models by developing a "Leap-of-Thought" (LoT) capability that allows the model to make creative connections between different concepts.

By fostering innovation and intellectual leaps, this capability has the potential to advance creative applications. The team drew inspiration from the Japanese game Oogiri, which encourages creative and humorous responses to images. They created a data set called Oogiri-GO, comprising over 130, 000 samples in Chinese, Japanese, and English, which featured humorous responses obtained from the internet. By using Associable Instruction Tuning and self-refinement techniques, the researchers trained Alibaba's Qwen model to respond to images and texts with human-like leap-of-thought thinking. The model has shown promise in generating entertaining and interactive content, and the team believes it could play a role in developing humorous sketches and creative scripts. This research represents an exploration into the leap-of-thought capability in large language models, blurring the boundaries between humans and AI and fostering creative exploration and discovery.