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

The paper introduces HumorGen, a framework that leverages a Mixture-of-Thought approach with six cognitive personas to synthesize high-quality humor data, demonstrating that cognitive-driven data curation is more critical than model scale or specific alignment algorithms for enabling a 7B-parameter model to outperform larger baselines in humor generation.

The Gist

Imagine you are trying to teach a very smart, very serious robot how to tell a joke.

The robot is like a perfectly polite librarian. Its entire life has been spent reading books and learning that the goal is to predict the most likely next word. If you say, "The sky is blue," the robot knows the next word is probably "and." It's great at logic, math, and writing helpful emails.

But here's the problem: Humor is the opposite of "most likely."

A joke works because it surprises you. It takes a logical path and then suddenly swerves into something weird, absurd, or unexpected. If the robot tries to tell a joke using its normal "predict the next word" brain, it just gives you a boring, safe explanation of why something is funny, or a joke that everyone has heard a thousand times. It's like a comedian who explains the punchline before they even say it.

This paper introduces a new way to fix this, called HumorGen. Here is how it works, broken down into simple steps:

1. The "Six Personalities" Trick (Cognitive Synergy)

Instead of asking the robot to "be funny" (which is too vague), the researchers gave the robot six distinct personalities to wear like costumes. Think of this as a comedy writing room where six different comedians are brainstorming ideas for the same news headline.

They created six "Cognitive Personas":

• The Absurdist: Thinks in dreams and nonsense (like a dream where your cat is driving a bus).
• The Cynic: The grumpy guy who points out how silly the world is.
• The Neurotic: The anxious over-thinker who worries about tiny details.
• The Wordsmith: Loves puns and playing with language.
• The Optimist: Finds the happy, silly side of bad situations.
• The Observer: Notices the weird, awkward things we all do but never talk about.

The system asks all six of these "personalities" to write a joke for the same prompt. This forces the robot to stop being a boring librarian and start exploring weird, creative, low-probability ideas where real humor lives.

2. The "Taste Test" (Data Curation)

Once the six personalities generate hundreds of jokes, the researchers don't just pick the first one. They use a super-smart "Judge" (another AI) to rate them.

Imagine a talent show. The six personalities perform, and the Judge scores them. The best jokes (the ones that actually make you laugh) are saved. The bad ones are thrown in the trash.

This creates a "Greatest Hits" album of high-quality jokes. This is the most important part of the paper: The quality of the jokes they fed the robot mattered more than the size of the robot.

3. The "Student" vs. The "Teacher"

The researchers took a smaller, cheaper robot (a 7-billion-parameter model, which is like a smart tablet) and taught it using this "Greatest Hits" album.

They tried two different teaching methods:

• Method A (SFT): "Here are the best jokes. Memorize them."
• Method B (DPO/GRPO): "Here is a good joke and a bad joke. Learn to pick the good one."

The Surprise Finding: Method A (just showing the good jokes) worked just as well as the complex Method B. In fact, the complex methods didn't help much. The paper concludes that if you feed a robot really good, diverse data, you don't need fancy math to make it smart. A small robot with great data beats a giant robot with average data.

4. The "Explainer Trap" (What NOT to do)

The researchers tried one more thing: They taught the robot to think out loud before telling the joke (like showing its work in math class).

They thought this would help. Instead, it made the robot worse.

• Without thinking: The robot tells a quick, punchy joke. Boom.
• With thinking: The robot starts explaining why the joke is funny before telling it. "Okay, I am going to make a joke about a cat. Cats are funny because..."

This is called the "Explainer Trap." It's like a magician explaining the secret of the trick before pulling the rabbit out of the hat. The magic is ruined. The paper found that for humor, less thinking is more.

The Big Takeaway

The main lesson of this paper is simple: To make a robot funny, don't just make it bigger or smarter. Give it a diverse set of "personalities" to play with, feed it the funniest examples you can find, and tell it to stop over-explaining.

They managed to train a small, open-source robot (HumorGen) that is funnier than much larger, expensive, proprietary robots used by big tech companies. They proved that good data is the secret sauce, not just raw computing power.

Technical Summary

1. Problem Statement

Humor generation remains a significant challenge for Large Language Models (LLMs). Standard LLM training objectives (minimizing perplexity/predicting the most likely next token) inherently conflict with the requirements of comedy, which rely on incongruity, surprise, and low-probability semantic shifts.

• The "Alignment Tax": Standard models tend to converge on the most probable (generic) continuations, resulting in safe, predictable, or boring outputs.
• Limitations of Existing Methods: Current approaches relying on Chain-of-Thought (CoT) or logical reasoning often fail because humor requires divergent thinking, not linear logic. Furthermore, existing instruction-tuning methods fail to capture the diverse cognitive styles humans use to construct jokes.
• Scale vs. Data: It is unclear whether increasing model size or applying complex preference alignment algorithms (like DPO) is more effective than curating high-quality, diverse training data for this subjective task.

2. Methodology: The Cognitive Synergy Framework

The authors propose a novel framework that operationalizes psychological humor theories into a Mixture-of-Thought (MoT) architecture to generate diverse, high-quality synthetic humor data.

A. Cognitive Personas (MoT)

Instead of a single reasoning path, the framework deploys six distinct cognitive personas, each grounded in a specific psychological theory of humor. These personas act as "latent experts" to force the model into low-probability, high-variance regions of the semantic space:

1. The Neurotic: Based on Relief Theory (tension release via anxiety/overthinking).
2. The Cynic: Based on Superiority Theory (social critique/sarcasm).
3. The Observer: Based on Incongruity (mundane social awkwardness).
4. The Wordsmith: Based on Linguistic Ambiguity (puns/double entendres).
5. The Optimist: Based on Benign Violation (wholesome misinterpretation).
6. The Absurdist: Based on Incongruity/Surrealism (non-sequiturs/dream logic).

B. Data Synthesis Pipeline

1. Generation: A teacher ensemble (Kimi-K2 and Qwen 2.5-32B) generates 24 candidates per prompt (4 per persona).
2. Evaluation: An LLM judge (Llama 3.3-70B) performs pairwise evaluations to compute Elo ratings for all candidates.
3. Dataset Construction:
   • SFT Data: Top 10 ranked candidates per prompt (to prevent mode collapse).
   • DPO Data: Pairs of top-5 vs. bottom-5 ranked jokes.
   • O-GRPO Data: All 24 candidates per prompt for group-relative optimization.

C. Training Strategies

The authors fine-tune a 7B-parameter student model (Qwen-7B) using three distinct approaches:

1. Supervised Fine-Tuning (SFT): Trained on the top-ranked "winner" jokes.
2. Direct Preference Optimization (DPO): Aligns the model using pairwise preferences (Chosen vs. Rejected).
3. Offline Group Relative Policy Optimization (O-GRPO): A novel variant of GRPO that uses a fixed group of 24 candidates to compute relative advantages without online sampling.

D. Cognitive Synergy Distillation (CSD)

A specific ablation study where the student model is trained on the teacher's reasoning traces (the "thought" process) alongside the final joke. This tests whether teaching the model how to think improves humor generation.

3. Key Contributions

1. Cognitive Synergy Framework: A methodology for generating diverse humor data by explicitly modeling divergent thinking through psychological personas, effectively mitigating mode collapse.
2. Empirical Findings on Alignment: The study demonstrates that for humor generation, data quality (cognitive curation) is far more critical than alignment algorithms or model scale.
   • DPO and O-GRPO did not outperform the SFT baseline.
   • The 7B model trained on high-quality SFT data outperformed larger baselines (Qwen-32B, GPT-OSS-120B).
3. The "Explainer Trap": The authors discovered that training models to output reasoning traces (CoT) for humor is detrimental. Models trained with reasoning traces ("Think" variants) fell into the trap of explaining the joke rather than delivering it, significantly lowering performance.
4. State-of-the-Art Open-Weight Performance: The resulting HumorGen-7B model achieves competitive performance with frontier proprietary models (GPT-5, Kimi-K2) despite being 10x-20x smaller.

4. Experimental Results

• Performance Rankings: Using Bradley-Terry ratings on a held-out SemEval 2026 test set:
  • HumorGen-SFT-7B: 1083.9 BT Rating (Top Open-Weight).
  • Qwen-32B: 964.3 BT Rating.
  • GPT-OSS-120B: 989.2 BT Rating.
  • Frontier Models (GPT-5, Kimi-K2): ~1200–1300 BT Rating.
• Alignment Ineffectiveness: DPO (1079.9) and O-GRPO (1034.5) performed similarly to or worse than SFT, confirming that preference optimization adds little value once high-quality SFT data is established.
• Persona Analysis: The Neurotic (63.4% win rate) and Absurdist (55.8%) personas were the most successful in generating winning jokes, while the Wordsmith (34.9%) underperformed, suggesting forced puns are less effective than absurdity or anxiety-based humor.
• Human Evaluation: Human evaluators agreed with the LLM judge on 58.3% of "Good vs. Good" comparisons, validating the automated evaluation pipeline.

5. Significance and Conclusion

The paper fundamentally shifts the paradigm for creative generation tasks:

• Data > Scale: For subjective tasks like humor, curating diverse, theoretically grounded data is more impactful than simply increasing model parameters or applying complex RLHF/DPO techniques.
• Divergent vs. Convergent Thinking: Standard reasoning (CoT) is counter-productive for humor. The "Cognitive Synergy" approach proves that simulating diverse human cognitive styles is the key to breaking the "safe" mode of LLMs.
• The "Explainer Trap": A critical insight for future creative AI: forcing models to "show their work" (reasoning traces) can degrade creative output by shifting the model's focus from delivery to explanation.

Future Work: The authors plan to explore multilingual humor, multimodal humor (memes/images), and extending the persona framework to other creative domains. Code and data will be released upon publication.