Researchers develop HumorGen framework using cognitive personas to teach smaller AI models to generate better comedy than larger instruction-tuned models

A 7-billion-parameter student model fine-tuned with HumorGen outperforms larger instruction-tuned baselines at humor generation

The Cognitive Synergy Framework deploys six distinct personas (including The Absurdist and The Cynic) using a Mixture-of-Thought approach to generate diverse comedic content

Researchers compare Direct Preference Optimization (DPO) with a novel Offline Group Relative Policy Optimization (O-GRPO) training method

The framework addresses LLMs' core challenge: standard next-word prediction training conflicts with the surprise and incongruity required for humor