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

arXiv cs.CL / 4/14/2026

📰 NewsSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

Humor generation is difficult for standard LLM training because next-token prediction discourages the surprise and incongruity that comedy relies on.
The paper proposes the “Cognitive Synergy Framework,” using a Mixture-of-Thought setup with six persona-based cognitive perspectives to synthesize diverse humor data from psychological theories.
A theoretically grounded dataset produced by these personas is used to fine-tune a 7B-parameter student model.
The authors compare training methods, finding that their 7B model strongly outperforms larger instruction-tuned baselines and performs competitively with state-of-the-art proprietary models.
The work concludes that persona-driven cognitive data curation is more important than alignment algorithms or sheer model scale for achieving strong humor generation quality.

Abstract

Humor generation poses a significant challenge for Large Language Models (LLMs), because their standard training objective - predicting the most likely next word - inherently conflicts with the surprise and incongruity needed for comedy. To bridge this gap, we introduce the Cognitive Synergy Framework, a theoretically grounded methodology for generating high-quality humor data inspired by psychological theories of humor. Utilizing a Mixture-of-Thought (MoT) approach, we deploy six cognitive personas (e.g., The Absurdist, The Cynic) to synthesize diverse comedic perspectives for a given prompt. This framework creates a theoretically grounded dataset, which we use to fine-tune a 7B-parameter student model. We compare Direct Preference Optimization (DPO) and a novel Offline Group Relative Policy Optimization (O-GRPO); our 7B model significantly outperforms larger instruction-tuned baselines and achieves performance competitive with state-of-the-art proprietary models. We find that cognitive-driven data curation is far more critical than alignment algorithms or model scale for humor generation. Code and data will be available upon publication.