Understanding Humorous Contradictions in Comics: A Multi-Agent Debate Approach

Abstract

Recent advancements in large language models (LLMs) have demonstrated impressive reasoning capabilities, yet their ability to understand and generate nuanced humor—particularly in contexts involving juxtaposition and contradictory narratives—remains a significant challenge. This thesis explores the application of advanced reasoning techniques, specifically multi-agent debate, to enhance LLMs’ performance in understanding and interpreting humor derived from juxtaposed comic panels. . Leveraging the YesBut benchmark—a dataset designed to evaluate AI models on tasks ranging from literal description generation to deep narrative reasoning, we explore how multi-agent debate can address key limitations in current models, such as visual misinterpretation, reasoning gaps, and hallucination. By simulating collaborative critique among multiple AI agents, our approach encourages more robust reasoning, enabling models to better resolve contradictions, infer underlying philosophies, and generate coherent, contextually appropriate humor. Our experiments demonstrate that multi-agent debate not only improves accuracy in humor comprehension tasks but also enhances the model’s ability to generalize across diverse creative expressions. This work advances the development of socially intelligent AI systems capable of nuanced, human-like understanding of humor and creative narratives. The findings highlight the potential of multi-agent reasoning frameworks to bridge the gap between AI and humanlevel humor interpretation, paving the way for more sophisticated applications in automated content generation, entertainment, and human-AI