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

Edward Ajayi; Prasenjit Mitra

cs.CLMar 19, 2026

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

Authors: Edward Ajayi, Prasenjit Mitra

Organizations: Carnegie Mellon University Africa

Abstract

Humor generation poses a significant challenge for Large Language Models (LLMs), because their standard training objective (next-token prediction) inherently conflicts with the surprise and incongruity required for comedy. To bridge this gap, we introduce the Cognitive Synergy Framework, a methodology for generating highquality humor data inspired by psychological theories of humor. Utilizing a Mixtureof-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 produces a theory-grounded dataset, which we use to fine-tune a 7B-parameter student model. We further evaluate two alignment strategies, Direct Preference Optimization (DPO) and an offline group-relative variant O-GRPO, finding that neither improves over SFT. However, our 7B HumorGen model variants significantly outperform larger instruction-tuned baselines and achieve top-tier open-weight performance while remaining competitive with frontier proprietary systems. These results suggest that cognitively driven data curation is more critical than alignment algorithms or model scale for humor generation.

Figures & tables

Panel of Figure 1 — Figure 1: Example of an LLM-generated joke based on a news headline prompt, synthesized using the Cognitive Synergy Framework.

Figure 2: The HumorGen training pipeline. Input headlines are processed by the Cognitive Synergy module (MoT), generating diverse candidates from 6 distinct personas. Candidates are ranked via pairwise LLM judging (Llama 3.3-70B) to compute Elo ratings. The base policy is fine-tuned on the top-ranked candidates. The model is further optimized via two parallel experimental branches: Pairwise DPO (top) or Group-Relative O-GRPO (bottom) driven by Elo-based preference data.

Persona	Humor Theory	Mechanism	Cognitive Focus
Neurotic	Relief Theory	Tension Release	Internal anxiety, overthinking, and social insecurity.
Cynic	Superiority Theory	Social Critique	Hypocrisy, biting sarcasm, and moral contradictions.
Observer	Incongruity	Social Mapping	Mundane minutiae and unwritten awkward social norms.
Wordsmith	Linguistic	Ambiguity	Puns, double entendres, and phonological play.
Optimist	Benign Violation	Recontextualization	Wholesome misinterpretations of potentially negative traits.
Absurdist	Incongruity	Surrealism	Non-sequiturs, dream logic, and fractured causality.

Table 1: The six Cognitive Personas used in our framework. We map each persona to a foundational humor theory and a specific cognitive focus to ensure divergent candidate generation.

Figure 3: HTB BT leaderboard (Llama judge, 400 prompts, 15 models). HumorGen-SFT and DPO-7B rank 3rd/4th, outperforming models 4–18 $\times$ their size.

Figure 4: HTB pairwise win-rate heatmap (Llama judge). Each cell shows the row model’s win rate against the column model.

Feature	%
Humor Mechanisms (winner)
Incongruity	85.2
Absurdity	84.9
Surprise	57.2
Wordplay	42.0
Irony	22.5

Table 2: Judge-assigned tags on HTB (Llama; 41,976 non-tie comparisons). % of comparisons tagged.

Appendix figures & tables32 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Rationale for Evaluation	Qualitative Observation
DeepSeek-67B	Top-tier open-source model available during initial development.	Generated jokes with poor comedic timing; subjectively less funny relative to Qwen-2.5-32B.
Qwen-2.5-32B	Selected as a robust baseline due to its strong performance-to-size ratio and high instruction-following capabilities, which proved highly effective for structurally complex joke generation.	Generated structurally sound and elaborate jokes, though often lacked concise punchiness.
Kimi-K2	Explored specifically to introduce stylistic diversity alongside Qwen.	Generated consistently punchy, concise jokes with excellent comedic timing.

Appendix

Table 3: Observations during the teacher model selection phase.

Feature	Qwen-2.5-32B	Kimi-K2
Average Word Count	$\sim$ 54.7 words	$\sim$ 37.5 words
Structural Style	Elaborate setups, detailed narratives	Concise setups, rapid punchlines
Cognitive Tone	Analytical and descriptive	Punchy, sharp, and conversational

Appendix

Table 4: Stylistic and cognitive differences between the two teacher models.

Dataset Partition	Qwen-2.5-32B Contribution	Kimi-K2 Contribution
SFT Training Corpus	44.8%	55.2%
DPO Chosen Pool	40.9%	59.1%
DPO Rejected Pool	59.6%	40.4%
O-GRPO Full Pool	50.1%	49.9%

Appendix

Table 5: Training data distribution contribution between the two teacher models across all alignment phases.

Post-ranking statistic	Kimi-K2	Qwen-2.5-32B
Mean Elo	1003.9	996.2
Mean finish rank ( $\downarrow$ )	11.5	13.4
Rank-1 wins	603 (54.8%)	497 (45.2%)
Share of top-3 positions	63.6%	36.4%
Share of bottom-half positions (#11–24)	46.2%	53.8%

Appendix

Table 6: Post-ranking candidate quality by teacher across 1,100 prompts. Rankings are computed within each 24-candidate prompt group before constructing the SFT and preference-alignment corpora.

Judge model	Observation	Selected
Llama-3.3-70B-Instruct	Produced the most stable pairwise preferences and showed the strongest correspondence with human judgments in our blind calibration study.	Yes (primary)
Qwen-2.5-72B	Produced rankings broadly consistent with the primary judge, while providing an independent model-family perspective for cross-judge validation.	Yes (secondary)
Qwen3-32B	Produced less consistent ordering than the larger Qwen-2.5-72B judge in our preliminary comparisons.	No
GPT-OSS-120B	Produced more variable preferences and was less consistent with the primary and secondary judges.	No
Gemini-2.5-Pro	Provided generally coherent judgments, but its ordering was less consistent with the selected judges; its closed access also limited reproducibility.	No

Appendix

Table 7: Qualitative comparison of candidate pairwise judges. Llama-3.3-70B-Instruct was selected as the primary curation judge, and Qwen-2.5-72B was selected as the secondary validation judge.

Rank	Model	BT Rating	95% CI	Wins / Losses
1	GPT-5	1336.18	[1323.3, 1348.3]	4738 / 862
2	Kimi-K2	1259.98	[1249.7, 1268.5]	4362 / 1237
3	HumorGen SFT-7B	1128.14	[1118.3, 1138.1]	3565 / 2035
4	HumorGen DPO-7B	1123.72	[1115.7, 1134.9]	3536 / 2064
5	HumorGen DPO-Think-7B	1116.65	[1107.9, 1127.1]	3489 / 2110
6	HumorGen SFT-Think-7B	1085.31	[1075.8, 1096.5]	3281 / 2319

Appendix

Table 8: HTB leaderboard, Llama judge (400 prompts, 42,000 comparisons).

Figure 5: HTB BT leaderboard, Llama judge (400 prompts, 15 models).

Figure 6: HTB win-rate heatmap, Llama judge (row beats column %).

Rank	Model	BT Rating	95% CI	Wins / Losses
1	GPT-5	1268.25	[1256.5, 1281.1]	4418 / 1181
2	Kimi-K2	1228.24	[1217.3, 1239.4]	4192 / 1408
3	HumorGen DPO-Think-7B	1147.52	[1136.8, 1158.8]	3686 / 1914
4	HumorGen DPO-7B	1138.05	[1130.3, 1147.2]	3623 / 1976
5	HumorGen SFT-7B	1132.41	[1121.0, 1141.2]	3586 / 2014
6	HumorGen SFT-Think-7B	1105.57	[1094.0, 1115.2]	3406 / 2194

Appendix

Table 9: HTB leaderboard, Qwen judge (400 prompts, 42,000 comparisons). Kendall

\tau=0.8667

vs. Llama judge.

Figure 7: HTB BT leaderboard, Qwen judge (400 prompts, 15 models).

Figure 8: HTB win-rate heatmap, Qwen judge (row beats column %).

Rank	Model	BT Rating	95% CI	Wins / Losses
1	GPT-5	1378.73	[1346.1, 1421.5]	605 / 94
2	Kimi-K2	1279.63	[1245.0, 1322.1]	549 / 151
3	Gemini-2.5-Pro	1247.80	[1212.4, 1279.7]	528 / 172
4	HumorGen SFT-7B	1140.37	[1107.9, 1173.2]	449 / 251
5	HumorGen DPO-7B	1135.25	[1101.9, 1160.2]	445 / 255
6	HumorGen GRPO-7B	1089.84	[1060.7, 1114.1]	409 / 291

Appendix

Table 10: SemEval leaderboard, Llama judge (50 headlines, 15 models, 5,250 comparisons).

Figure 9: SemEval BT leaderboard, Llama judge (50 headlines, 15 models).

Figure 10: SemEval win-rate heatmap, Llama judge (row beats column %).

Rank	Model	BT Rating	95% CI
1	GPT-5	1285.15	[1258.2, 1320.8]
2	Kimi-K2	1210.51	[1188.5, 1233.9]
3	HumorGen DPO-7B	1202.76	[1175.5, 1227.0]
4	HumorGen SFT-7B	1193.81	[1162.0, 1228.2]
5	Gemini-2.5-Pro	1144.23	[1114.3, 1171.8]
6	HumorGen DPO-Think-7B	1132.13	[1103.0, 1170.3]

Appendix

Table 11: SemEval leaderboard, Qwen judge (50 headlines, 15 models). Kendall

\tau=0.8286

vs. Llama judge.

Figure 11: SemEval BT leaderboard, Qwen judge (50 headlines, 15 models).

Figure 12: SemEval win-rate heatmap, Qwen judge (row beats column %).

Dom.	Name	Status	Description and Example
A	Neutral Facts	OOD	Surprising declarative facts. “Bananas are technically berries.”
B	Everyday Life	OOD	First-person observational statements. “My printer hates me.”
C	Abstract Concepts	OOD	Bare noun-phrase prompts. “Bureaucratic nostalgia.”
D	Dialogic Quotations	OOD	Attributed utterances. “The fridge said: I’m judging you.”
E	Scenario Inputs	OOD	Situational irony setups. “A vampire is working in HR.”
F	Analogical Prompts	OOD	Cross-domain relational mappings. “Life is a subscription service you never read the terms for.”

Appendix

Table 12: HTB domain descriptions used in evaluation.

Field	Value
Dataset name/version	Humor Transfer Bench (HTB), v1
Primary objective	Measure humor transfer under prompt-domain shift while holding the generation task fixed
Task/instruction	Text-to-text humor generation ( Generate a funny joke given this prompt )
File format/encoding	TSV, UTF-8, Unix line endings
Schema (columns)	id , domain , prompt
Identifier format	HTB_{DOMAIN}_{NNN} (e.g., HTB_A_001 )

Appendix

Table 13: HTB data card aligned with the benchmark design specification.

Persona	Pairwise win rate	Mean rank (of 24)	Top-3 rate
Neurotic	63.4%	9.3	26.6%
Absurdist	55.8%	11.3	17.6%
Cynic	55.2%	11.4	15.9%
Observer	49.1%	12.5	6.2%
Optimist	41.1%	14.4	5.7%
Wordsmith	34.9%	16.1	3.1%

Appendix

Table 14: Per-persona curation-stage performance across three independent metrics. Pairwise win rate : fraction of 1v1 judge comparisons won. Mean rank : average within-prompt rank out of 24 candidates (lower = better). Top-3 rate : fraction of prompts where the persona placed in the top 3. All three metrics converge on the same ordering.

Persona	O-GRPO pool	SFT (top-10)	DPO chosen	DPO rejected
	N=28,789	N=12,000	N=6,000	N=6,000
Neurotic	4,800 (16.7%)	2,899 (24.2%)	1,650 (27.5%)	328 (5.5%)
Absurdist	4,796 (16.7%)	2,364 (19.7%)	1,316 (21.9%)	783 (13.1%)
Cynic	4,796 (16.7%)	2,269 (18.9%)	1,257 (20.9%)	679 (11.3%)
Observer	4,798 (16.7%)	2,015 (16.8%)	934 (15.6%)	1,007 (16.8%)
Optimist	4,799 (16.7%)	1,375 (11.5%)	542 (9.0%)	1,334 (22.2%)

Appendix

Table 15: Persona composition of the alignment training corpora. Counts are followed by row percentages within each column (each column sums to 100%). All six personas appear in every corpus.

Theory	Persona	Summary
Relief Theory	Neurotic	Humor functions as a release mechanism for suppressed tension, anxiety, or socially restricted thoughts. Jokes provide a culturally sanctioned outlet for uncomfortable truths, transforming nervous energy into laughter through cathartic release.
Superiority Theory	Cynic	Dating back to Plato and developed by Hobbes (1650), humor arises from feelings of superiority over others’ failures, social hypocrisy, or misfortune. Laughter emerges through recognizing incompetence or pretension, reinforcing the observer’s relative standing via critique or sarcasm.
Incongruity Theory	Observer, Absurdist	Kant (1790) and Schopenhauer (1819): humor arises from violated expectations through unexpected juxtapositions or logical impossibilities. The Observer persona reflects everyday social mismatches; the Absurdist extends this to surreal and illogical extremes.
Linguistic Theories	Wordsmith	Raskin’s (1985) Semantic Script Theory of Humor and Attardo’s extensions: humor arises from language itself via ambiguity, phonological similarity, double meanings, and wordplay. Humor emerges from the gap between literal meaning and intended interpretation.
Benign Violation Theory	Optimist	McGraw and Warren (2010): humor occurs when a norm violation is simultaneously perceived as benign. The Optimist persona operationalizes this through positive reinterpretation of otherwise negative or threatening situations.

Appendix

Table 16: Psychological humor theories grounding each CSF cognitive persona Lintott (2016) ; Scheel (2025) ; McGraw and Warren (2010) .

Normalization Method	Mean	Median	Std Dev	Fraction $<0$
Standard Z-Score	0.000	-0.479	1.000	50.5%
Winsorized Z-Score ( $k=2$ )	-	-0.479	-	50.5%
Rank-Normalized (RankNorm)	0.000	0.000	0.602	49.9%

Appendix

Table 17: Advantage distribution properties for different normalization strategies in O-GRPO. RankNorm effectively centers the median and creates a balanced learning signal.

Rank	Model	BT Rating	95% CI
1	GPT-5	1379.04	[1348.3, 1418.4]
2	Kimi-K2	1285.59	[1259.8, 1327.1]
3	Gemini-2.5-Pro	1254.02	[1221.8, 1279.5]
4	HumorGen SFT-7B	1145.16	[1120.4, 1177.0]
5	HumorGen DPO-7B	1137.32	[1110.1, 1165.1]
6	HumorGen GRPO-7B	1093.50	[1061.2, 1121.6]

Appendix

Table 18: Full Leaderboard (50 Headlines) on the SemEval benchmark comparing O-GRPO models trained with standard vs. RankNorm advantage estimation. The results show that RankNorm variants consistently underperform their standard O-GRPO counterparts, demonstrating that forcing a symmetric advantage distribution degrades the learning signal.

Parameter	SFT / SFT-Think	DPO / DPO-Think	GRPO / GRPO-Think
Learning Rate	$2\times 10^{-4}$ (Linear)	$5\times 10^{-7}$ (Constant)	$1\times 10^{-6}$ (Constant)
Batch Size (Global)	16	16	16
Epochs (Configured)	3	5	5
Max Sequence Length	1024	1024 (512 prompt)	1024
Optimizer	AdamW (8-bit)	AdamW (8-bit)	AdamW
LoRA Rank ( $r$ )	16	16	16

Appendix

Table 19: Consolidated hyperparameters for the HumorGen training pipeline. SFT-Think, DPO-Think, and GRPO-Think used identical settings to their base counterparts.

Model Variant	Steps	Final Epoch	Final Loss	Eval Loss (Min)	Runtime
HumorGen-SFT-7B	900	1.26*	1.258	1.342	7.2m
HumorGen-SFT-Think	900	1.26*	1.768	1.908	7.8m
HumorGen-DPO-7B	1,550	4.34*	0.512	0.742	18.5m
HumorGen-DPO-Think	1,550	4.34*	0.528	0.756	21.2m
HumorGen-GRPO-7B	6,050	3.66*	0.456	1.593	23.7m
HumorGen-GRPO-Think	6,850	4.02*	5.901	1.461	6.0h

Appendix

Table 20: Training metrics across all HumorGen variants. (*) Training was terminated by early stopping at the best recorded eval loss.

Parameter	Value
Learning Rate	$5\times 10^{-5}$ (Cosine)
Batch Size	16
Epochs	2
Warmup Ratio	0.05
Optimizer	AdamW (8-bit)
Data Volume ( $N$ )	998 jokes

Appendix

Table 22: Hyperparameters for the Comedian SFT (Ablation-C) model.

Figure 15: Instructions screen (HumorGen Blind Eval): As shown to participants before voting. Full text: “You’ll see a headline and two anonymous jokes. Vote for the one you find funnier (no labels!) There are 60 pairs to rate.”

Figure 16: Sign-in / instructions (alternative view): Google sign-in step before the pairwise task.

Class	Sub-Category	N	Research Question
Think Tax	1a. SFT vs. SFT-Think	5	Do humans penalize CoT over-reasoning in SFT?
	1b. DPO vs. DPO-Think	5	Do humans penalize CoT over-reasoning in DPO?
	1c. GRPO vs. GRPO-Think	5	Do humans penalize CoT over-reasoning in GRPO?
SFT-7B	2a. SFT-7B vs. GPT-4o	5	Can a 7B model hold its own against ${\sim}$ 1.5T weights?
	2b. SFT-7B vs. Gemini	5	Can a 7B model challenge a 1T+ frontier API?
	2c. SFT-7B vs. Kimi	5	Can the student beat the teacher that generated its data?

Appendix

Table 24: Human evaluation category design (60 pairs, 12 sub-categories, 5 each).

Figure 17: Preliminary Evaluation Interface: Used internally during early experimentation to confirm our core hypothesis regarding Cognitive Synergy. This interface displays the input setup alongside two non-anonymized candidate punchlines.

Figure 18: Blind Human Evaluation Interface: Deployed to volunteer annotators for unbiased A/B testing. Model identities are hidden and candidate order is randomized.

Humor Explanation Generation

See all 8 papers

HumorRank: A Tournament-Based Leaderboard for Evaluating Humor Generation in Large Language Models
Mar 31, 2026Edward Ajayi, Prasenjit MitraHumor Explanation Generation LLM Evaluation
Carnegie Mellon University Africa Kigali, Rwanda
Re-defining Humor Data Objects for AI Humor Research
May 24, 2026Anna Arnett, Bang Nguyen, Meng JiangLLM Prompting Humor Explanation Generation
Department of Computer Science and Engineering, University of Notre Dame
lmfaoooo at SemEval-2026 Task 1: Humor Is an Audience. Preference Modeling for Constrained Humor Generation
Apr 14, 2026Alexey Tikhonov, Alexey IvanovHumor Explanation Generation Text Generation
Inworld.AI Berlin, Germany · OpenAI Mountain View, CA

Mar 31, 2026J/K move · Enter open · S save

HumorRank: A Tournament-Based Leaderboard for Evaluating Humor Generation in Large Language Models

Edward Ajayi, Prasenjit Mitra

Carnegie Mellon University Africa Kigali, Rwanda

Humor Explanation Generation LLM Evaluation Generative AI Evaluation Text Generation LLM-as-a-Judge

Open paper AskPDF arXiv

Evaluating humor in large language models (LLMs) is an open challenge because existing approaches yield isolated, incomparable metrics rather than unified model rankings, making it difficult to track progress across systems. We introduce HumorRank, a tournament-based evaluation framework and leaderboard for textual humor generation. On two public benchmarks (SemEval-2026 MWAHAHA and Humor Transfer Bench), we conduct extensive automated pairwise evaluation across nine models spanning proprietary, open-weight, and specialized systems. Pairwise judgments are produced by LLM judges grounded in the General Theory of Verbal Humor (GTVH): each judge integrates structured comedic analysis into adjudication, jointly yielding a preference decision, an interpretable rationale, and mechanism, delivery, and failure tags rather than a black-box funniness score. Judgments are aggregated via an Adaptive Swiss tournament, with Bradley-Terry Maximum Likelihood Estimation (MLE) producing globally consistent humor generation capability rankings. Rankings are cross-judge stable: independent LLM judges (Llama 3.3 70B and Qwen 2.5 72B) yield Kendall tau = 0.889 on both benchmarks, and a human calibration study shows human-LLM agreement tracking human-human agreement on hard funny-versus-funny pairs. Our results demonstrate that HumorRank yields statistically grounded model stratifications, showing that humor quality is associated with mastery of comedic mechanisms such as incongruity, conciseness, escalation, and absurdity rather than model scale alone, with specialized fine-tuned models reaching parity with far larger systems. HumorRank thus provides a scalable, interpretable, and reproducible methodology for benchmarking and understanding LLM-generated humor.

Abstract

Humor Explanation Generation

Explore similar work

HumorRank: A Tournament-Based Leaderboard for Evaluating Humor Generation in Large Language Models

Re-defining Humor Data Objects for AI Humor Research

lmfaoooo at SemEval-2026 Task 1: Humor Is an Audience. Preference Modeling for Constrained Humor Generation