HumorGen: Cognitive Synergy for Humor Generation in Large Language Models via Persona-Based Distillation

A team of researchers led by Edward Ajayi and Prasenjit Mitra has introduced HumorGen, a novel framework designed to solve a core weakness in Large Language Models: their inability to reliably generate humor. The standard next-word prediction objective of LLMs inherently conflicts with the surprise and incongruity essential to comedy. To bridge this gap, the team developed the theoretically grounded Cognitive Synergy Framework, which employs a Mixture-of-Thought (MoT) approach.

This framework deploys six distinct cognitive personas—including The Absurdist, The Cynic, and The Satirist—to synthesize a wide range of comedic perspectives for any given prompt. This method creates a high-quality, psychologically-informed dataset, which was then used to fine-tune a relatively small 7-billion-parameter student model. The researchers tested alignment techniques like Direct Preference Optimization (DPO) and a novel Offline Group Relative Policy Optimization (O-GRPO).

The results were striking: their compact 7B model significantly outperformed larger, general-purpose instruction-tuned models and achieved performance competitive with state-of-the-art proprietary models like GPT-4 in humor generation tasks. The key finding is that cognitive-driven data curation is far more critical for this capability than the choice of alignment algorithm or raw model scale. The code and data will be made available upon publication, offering a new blueprint for instilling specific, complex cognitive traits into AI.