1 Introduction↩︎

Humor generation is a sophisticated creative task requiring mastery of context, nuance, and linguistic ambiguity [1], [2]. While Large Language Models (LLMs) excel at logical reasoning, reliable humor generation remains an open problem because standard training objectives, which minimize perplexity, conflict with the incongruity and surprise required for comedy. This “alignment tax” often results in models that are safe and helpful but produce predictable, boring jokes or tedious explanations of humor.

Recent efforts to improve LLM humor generation have focused on logical “thought leaps” [3] or multistep reasoning [4]. While these approaches improve performance on specific humor generation tasks, they often miss the diverse cognitive styles underlying human humor, and primarily rely on instruction tuning, limiting their ability to capture the variety of ways humans construct jokes.

To bridge this gap, we introduce the Cognitive Synergy Framework which advance beyond generic instruction tuning by operationalizing psychological humor theories within a Mixture-of-Thought (MoT) framework designed to encourage creative divergence. Traditional language modeling is highly susceptible to mode collapse in creative generation, converging toward the most probable (and therefore most generic) continuations. By instantiating six distinct “cognitive personas” (e.g., The Absurdist, The Cynic) as latent experts within the MoT framework, we consistently route the generation process into the low-probability, high-variance regions of the semantic space where humor naturally occurs. This ensemble approach mitigates mode collapse and yields a diverse, theoretically grounded corpus of synthetic data, enabling us to distill multi-faceted humor generation capabilities from a frontier teacher model into a highly efficient 7B-parameter student model.

Due to the highly subjective nature of humor, we investigate whether preference alignment (e.g., Direct Preference Optimization (DPO) and Offline Group Relative Policy Optimization (O-GRPO)) improves over supervised fine-tuning. Our experiments show that neither alignment method visibly improves the models over the SFT baseline as DPO achieves similar performance to SFT, while O-GRPO is less impressive. Thus, under our setup, the alignment exercises did not improve the models, and the quality of the underlying cognitive data (Cognitive Synergy Framework) is the primary driver of generation performance.

Our contributions are:

• We introduce the Cognitive Synergy Framework (CSF), a methodology that operationalizes psychological humor theories into six cognitive personas to generate diverse, theory-grounded humor data via Mixture-of-Thought.

• We investigate whether preference alignment (DPO, O-GRPO) improves over SFT for humor generation, finding that neither method yields consistent gains beyond a well-curated SFT baseline in this subjective humor generation domain.

• We introduce Humor Transfer Bench (HTB), a 400-prompt benchmark spanning eight input domains to test generalization in textual humor generation.

• We show that our 7B student models, HumorGen variants, achieves top-tier open-weight performance on HTB and SemEval MWAHAHA humor-generation benchmark and remains competitive with much larger proprietary systems, showing that high-quality data can outweigh model size for humor.

2 Related Work↩︎

2.1 Computational Humor Generation↩︎

Computational humor research has largely focused on detection and recognition tasks [5], [6], while humor generation remains comparatively underexplored. As a result, existing work on humor generation is often fragmented, targeting specific forms such as puns [7], domain-based humor [8], punchline generation [9], or language-based humor generation [3], [10]. However, while classical humor theories [11]–[14] offer valuable linguistic and semantic interpretations of humor, they do not directly translate into reliable generative mechanisms capable of consistently producing effective or high-quality humor.

2.2 Reasoning-Enhanced Humor Generation↩︎

With the rise of Large Language Models (LLMs), recent work has explored prompting and reasoning strategies for humor generation. Prior studies show that Chain-of-Thought (CoT) reasoning [15], while effective for logical tasks, is often misaligned with humor generation, which depends on divergence, incongruity, and non-linear associations [3], [4], [16]. As a result, CoT-based prompting often produces coherent but non-comedic outputs.

To address this, Creative Leap of Thought (CLoT) [3] introduces leap-of-thought reasoning to encourage non-obvious associations during humor generation. The LoL framework [4] builds on CLoT by incorporation of external knowledge to support multi-hop creative reasoning, while related work explores structured multistep reasoning for humor generation [16]. In contrast, naive prompting with ChatGPT [5] shows strong repetition patterns, with 90% of 1,008 generated jokes repeating only 25 templates. These findings suggest that standard reasoning paradigms remain limited for humor generation.

2.3 Preference Optimization in Creative Generation↩︎

Preference optimization methods have been widely studied for aligning LLMs with human preferences in subjective generation tasks [17]–[19]. With methods such as Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO) having demonstrated effectiveness across domains such as code and image generation [20], [21].

In humor generation, prior work integrates Supervised Fine-Tuning (SFT) with DPO-based alignment [4]. However, many preference optimization methods rely on online sampling or pairwise comparisons, which can be unstable and computationally expensive in highly subjective settings. We build on these ideas by additionally exploring an offline group-relative formulation (O-GRPO) using fixed preference datasets without online sampling.

3 The Cognitive Synergy Framework↩︎

Humor generation needs both coherence and surprise: a joke must first establish a plausible setup, then introduce an unexpected shift. Standard LLM decoding tends to favor high-probability next-token predictions, which often reduces surprise and yields generic outputs.

To address this, we introduce the Cognitive Synergy Framework (CSF), which leverages Mixture-of-Thought (MoT) [22] for humor generation by running multiple persona-guided reasoning paths in parallel, where each Cognitive Persona (see Section 3.2) encodes a different humor mechanism (e.g., absurdity, social critique, wordplay), enabling the model to explore diverse comedic directions instead of converging to a single safe continuation. The resulting candidate set is therefore both coherent and diverse, improving downstream selection and distillation for humor-focused training.

3.1 Divergent Reasoning via MoT↩︎

Unlike standard Chain-of-Thought (CoT) prompting, which optimizes for a single logical path, our framework generates \(K\) distinct reasoning traces in parallel. Given an input premise \(x\), we sample a set of diverse reasoning paths \(\{z_1, z_2, \dots, z_K\}\) seeded by different cognitive priors, yielding a pool of \(K\) candidates that is retained in full. All candidates are subsequently ranked and used for training, so the model learns from the full distribution of generated joke candidates rather than a single top-ranked joke.

3.2 Cognitive Personas↩︎

To guide this diversity, we define six Cognitive Personas, each grounded in a specific psychological theory of humor (Table 1; see Appendix 15 for a full overview of the underlying theories). These personas act as soft constraints on the reasoning process, ensuring that our candidate pool covers a wide spectrum of comedic mechanisms.

By using these personas, we created a “synergy” between different styles of thought. This structural diversity proved critical for our subsequent alignment stage, as it provided a rich variety of distinct candidates for the model to learn from during preference optimization.

4 Methodology↩︎

We frame humor generation as a conditional language modeling task, where the goal is to generate a humorous response \(y\) given a context \(x\). We first build a strong initialization model via supervised fine-tuning (SFT) using MoT-generated candidates and silver LLM supervision. Starting from this model, we explore two preference alignment strategies based on LLM-judged pairwise evaluations: Direct Preference Optimization (DPO) [23] and an offline group-relative variant, O-GRPO, both of which aim to better align outputs with learned preference signals.

4.1 Supervised Fine-Tuning (SFT)↩︎

This initial stage establishes baseline humor capabilities and internalizes the various cognitive personas. We construct a dataset \(\mathcal{D}_{SFT}\) using a “Silver Teacher” protocol. Given the candidate pool \(\mathcal{C}_{total}\) generated by our Mixture-of-Thought (MoT) ensemble, we employ a pairwise LLM evaluation system to compute Elo ratings for all candidates. We select the top-ranked candidates for each prompt based on these Elo ratings: \[y^* = \text{argmax}_{y \in \mathcal{C}_{total}} \text{Score}_{LLM}(y|x)\] We fine-tune a base Qwen-7B model using standard cross-entropy loss to maximize the likelihood of these “winner” responses: \[\mathcal{L}_{SFT}(\theta) = -\mathbb{E}_{(x, y^*) \sim \mathcal{D}_{SFT}} [ \log \pi_\theta(y^* | x) ]\] This stage effectively distills the creative diversity of the larger teacher model into the student model.

4.2 Direct Preference Optimization (DPO)↩︎

To further align the model with humor preferences, we employ DPO using a dataset \(\mathcal{D}_{DPO}\) of high-quality pairwise preferences derived from the LLM-judged Elo rankings. Each pair \((y_w, y_l)\) consists of a high-ranking joke \(y_w\) and a low-ranking candidate \(y_l\) for the same prompt, selected based on their Elo gap. We optimize the policy \(\pi_\theta\) directly without a reward model:

\[\begin{gather} \mathcal{L}_{DPO}(\pi_\theta; \pi_{ref}) = - \mathbb{E}_{(x, y_w, y_l) \sim \mathcal{D}_{DPO}} \Bigg[ \log \sigma \Bigg( \\ \beta \log \frac{\pi_\theta(y_w|x)}{\pi_{ref}(y_w|x)} - \beta \log \frac{\pi_\theta(y_l|x)}{\pi_{ref}(y_l|x)} \Bigg) \Bigg] \end{gather}\]

4.3 Offline Group Relative Policy Optimization (O-GRPO)↩︎

Beyond pairwise preference alignment, we experiment with an offline variant of Group Relative Policy Optimization, O-GRPO, which adapts GRPO [24] to a fixed preference dataset. Rather than sampling candidates and computing rewards online during training, O-GRPO operates on the pre-constructed candidate pool: all 24 candidates per prompt are ranked once via a complete Bradley-Terry tournament, and group-relative advantages are derived from those static scores. Specifically, for each joke \(y_i\), we compute its advantage as \[A_i = \frac{r_i - \mu_{\mathrm{group}}}{\sigma_{\mathrm{group}} + \epsilon} \label{eq:ogrpo95main}\tag{1}\] where \(r_i\) is its pre-computed ranking score. We then apply exponentially weighted SFT, where each candidate contributes proportionally to the weights defined in Appendix 16. This upweights relatively stronger responses and downweights weaker ones using fixed offline rankings, without live reward computation.

4.4 Cognitive Synergy Distillation (CSD)↩︎

In the SFT stage of the training pipeline, the teacher’s persona-specific reasoning traces are not used, and the student model observes only the final jokes. In contrast, CSD trains the student on both the reasoning traces and the corresponding joke:

<think> persona-specific brainstorming </think> joke

This is process distillation: the student learns not just what to generate but how the teacher planned it. For DPO, both chosen and rejected responses include reasoning traces (symmetric format), so the model cannot shortcut by learning that the mere presence of reasoning correlates with winning; it must learn which content leads to better jokes.

At inference, the model generates reasoning followed by the joke. The reasoning is stripped for evaluation (ensuring fair comparison with non-CSD models) but retained for interpretability. Unlike generic CoT, ineffective for humor [3], [16], CSD’s reasoning is grounded in specific humor theories through the cognitive personas, making it a form of theory-grounded creative distillation.

5 Experimental Setup↩︎

5.1 Datasets and Data Synthesis↩︎

We utilize the official SemEval 2026 Task 1 (MWAHAHA) experimental set [25], comprising 1,200 news headlines and word-pair prompts as inputs to our generation pipeline. Using the Cognitive Synergy Framework, we generate 24 candidates per prompt (4 per persona \(\times\) 6 personas) from a teacher ensemble of Kimi-K2 and Qwen 2.5-32B-Instruct, yielding a raw pool of \(\sim\)​28,800 candidates. These candidates are scored and ranked via a pairwise LLM evaluation system using Llama 3.3-70B-Instruct as the judge, producing per-prompt Elo ratings for all 24 candidates. We construct three training subsets from these rankings:

• SFT Data (\(\mathcal{D}_{SFT}\), \(N=12,000\)): For each of the 1,200 prompts, we select the top 10 Elo-ranked candidates (rather than only the single best). Using multiple top-ranked candidates per prompt avoids mode collapse: the student learns a diverse range of humor styles (e.g., wordplay, absurdity, sarcasm) instead of collapsing toward one dominant style.

• DPO Data (\(\mathcal{D}_{DPO}\), \(N=6,000\)): For each prompt, we construct 5 preference pairs by randomly pairing candidates from the top-5 Elo-ranked jokes (chosen, \(y_w\)) with candidates from the bottom-5 Elo-ranked jokes (rejected, \(y_l\)). This yields 5 pairs \(\times\) 1,200 prompts = 6,000 preference pairs, with a sharp quality gap between chosen and rejected responses.

• O-GRPO Data (\(\mathcal{D}_{GRPO}\)): We use all 24 candidates per prompt across the 1,200 prompts, computing normalized group-relative Elo advantages per group (\(G=24\)). This exposes the model to the full quality spectrum within each prompt group.

Appendix 14 reports the persona composition of each alignment corpus, confirming that all six cognitive personas contribute to SFT, DPO, and O-GRPO training data.

For evaluation, we use two benchmarks. (1) SemEval-MWAHAHA: from the official 300-prompt test set, we evaluate on the first 50 headlines for controlled leaderboard comparison. (2) Humor Transfer Bench (HTB): our new benchmark with 400 prompts (8 domains \(\times\) 50 prompts) designed to test transfer beyond headline-style inputs.

5.2 Baselines↩︎

We evaluate 15 total models: frontier systems (GPT-5, Kimi-K2, Gemini-2.5-Pro), open-weight general models (Qwen3-32B, GPT-OSS-120B, Base Qwen-7B), HumorGen variants (SFT, DPO, O-GRPO and their Think versions, plus HumorGen-Com-7B), and humor-specialized baselines (phi2-Humor [26], JokeGPT [27]).

5.3 Implementation Details↩︎

All models were trained on NVIDIA H100 (80GB) GPUs using LoRA [28] (\(r\)=16) with the Unsloth library. SFT ran for 3 epochs; DPO and O-GRPO for 5 epochs, both with early stopping (patience=2). Candidate ranking for the full pool consumed \(\sim\)​132 H100 node-hours. For O-GRPO, groups of \(G=24\) candidates per prompt maximize the advantage-weighted learning signal.

5.4 Evaluation Protocols↩︎

5.4.0.1 Ranking methodology:

For each pair of jokes (A, B), the LLM judge selects the funnier one (or tie), with presentation order randomized per match to mitigate position bias. We aggregate all outcomes into a contest matrix, fit a Bradley-Terry (BT) model [29], [30] via the MM algorithm to estimate latent ratings, and report Elo-scale ratings with 95% bootstrap confidence intervals (100 samples). For cross-judge Kendall comparisons, exact BT ties are broken deterministically by model-name order before rank correlation is computed; key model comparisons have non-overlapping confidence intervals and are statistically significant. We do not report reference-based metrics (e.g., BLEU, ROUGE) because humor generation is inherently open-ended and subjective, motivating pairwise preference-based evaluation.

1. Automated Pairwise Evaluation: We report Llama-3.3-70B and Qwen 2.5 72B judged results for both benchmarks in the main paper: SemEval leaderboard (15 models, 5,250 pairwise comparisons) and HTB full benchmark (400 prompts, 15 models, 42,000 pairwise comparisons).

2. Human Validation: Human evaluators judge 60 curated pairwise comparisons across 12 ablation categories; they are blinded to model identity and presentation order is randomized to mitigate position bias. We report inter-annotator agreement, LLM–human consensus, and correlation with automated BT ratings.

5.5 Humor Transfer Bench (HTB)↩︎

A key challenge in evaluating textual humor generation models is the absence of diverse benchmark datasets. The only existing benchmark for this task is the SemEval 2026 MWAHAHA test set, which uses news headlines as prompts, a single, narrow input distribution that limits assessment of model generalizability. To address this gap, we introduce the Humor Transfer Bench (HTB), a new benchmark of 400 prompts spanning eight structurally distinct input domains (50 each): Neutral Facts, Everyday Life, Abstract Concepts, Dialogic Quotations, Scenario Inputs, Analogical Prompts, Direct Instructional, and News Headlines. No model is fine-tuned on HTB data; any performance gain on out-of-distribution domains therefore reflects genuine humor transfer. Full domain descriptions and design rationale are in Appendix 12.

6 Results↩︎

6.1 Model Performance↩︎

We evaluate all 15 models on the Humor Transfer Bench under the Llama 3.3-70B judge (42,000 pairwise comparisons). Figure 3 shows the Bradley-Terry leaderboard; Figure 4 shows the head-to-head win-rate heatmap across all model pairs. HumorGen-SFT-7B ranks 3rd (BT 1128.14) and HumorGen-DPO-7B ranks 4th (BT 1123.72), ahead of Gemini-2.5-Pro (BT 1059.07), GPT-OSS-120B (BT 1048.19), and Qwen3-32B (BT 990.44), all of which are 4–18\(\times\) larger. Humor-finetuned baselines phi2-Humor and JokeGPT rank 12th and 15th, confirming that domain specialization alone, without cognitive-driven data curation, does not generalize. To verify rating stability across evaluators, we run Qwen 2.5-72B as a second independent judge; the cross-judge Kendall \(\tau = 0.8667\) (\(p=1.54\times10^{-7}\)) confirms that model ordering is robust across judges. We additionally evaluate both judges on the SemEval news-headline test set (15 models, 50 headlines), where cross-judge agreement reaches Kendall \(\tau = 0.8286\) (\(p=9.95\times10^{-7}\)). Full leaderboard tables with 95% CI and all Qwen-judge results are reported in Appendix 11. Overall, HumorGen 7B variants consistently outperform significantly larger models across unseen prompt domains, validating the generalization capacity of the Cognitive Synergy Framework.

6.2 Preference Alignment↩︎

We investigate whether preference alignment (DPO, O-GRPO) improves over SFT. Across both benchmarks and judges, SFT and DPO perform similarly, with negligible and inconsistent differences: on HTB (Llama judge), SFT (BT 1128.14) slightly outperforms DPO (BT 1123.72), while under the Qwen judge, DPO (BT 1138.05) narrowly exceeds SFT (BT 1132.41). In both cases, the 95% confidence intervals overlap, indicating no statistically significant difference. O-GRPO consistently underperforms both SFT and DPO (e.g., BT 1071.13 on HTB under the Llama judge, rank 7th). This gap is attributable to a skewed advantage distribution in the O-GRPO training data (median advantage \(= -0.479\)), where a small number of high-quality samples dominate the group mean, causing most candidates to receive negative relative advantages. As a result, the optimization signal primarily penalizes weaker samples rather than reinforcing strong ones, leading to stable underperformance. These findings suggest a data quality ceiling: when SFT data is already well-curated via CSF, additional preference optimization yields limited gains in this subjective humor generation setting.

6.3 CSD and the Explainer Trap↩︎

The “explainer trap” emerges when we train the 7B HumorGen variants to think by applying Cognitive Synergy Distillation (CSD) so the student is trained on the teacher’s <think> reasoning traces alongside the joke (see § Methodology). On automatic leaderboards, Think variants usually rank below their matched non-thinking counterparts (e.g., SFT-Think 6th vs.SFT 3rd on HTB under Llama; DPO-Think 5th vs.DPO 4th), but the effect is not uniform: DPO-Think ranks 3rd under the Qwen judge on HTB, ahead of DPO (4th). Distilling reasoning traces appears to bias outputs toward explaining the joke rather than delivering it (Appendix 20), though we did not test whether teacher models themselves over-explain. This extends prior work showing limited benefit of CoT for humor [3], [16]: reasoning-augmented training does not reliably improve judged funniness in our setup.

6.4 Comedian Adaptation↩︎

To further probe the data quality thesis, we fine-tuned HumorGen-SFT on 998 stand-up jokes (Shaun Eli [31]). Results show that performance regressed sharply (BT: 1083.9 \(\to\) 653.1), as performance-native stand-up is optimized for live delivery rather than written punchlines, confirming that data origin matters as much as data quantity. Full analysis of the comedian adaption experiment is provided in Appendix 24.

6.5 Human Evaluation↩︎

Three evaluators gave 170 blind pairwise judgments on 60 curated pairs (12 categories, 5 each) over 50 held-out headlines. Inter-annotator agreement was 31.7% (one-third of pairs), reflecting humor’s subjectivity. To formalize this, we computed Krippendorff’s alpha for the multi-annotator subset, achieving \(\alpha = 0.425\) across three evaluators and \(\alpha = 0.412\) across two, which is consistent with the highly subjective nature of humor evaluation. The LLM judge matched human consensus on 58.3% of pairs (Gold) and individual votes at 52.4% (Micro-Avg). In this “Good vs.Good” regime (high-quality outputs, no objectively worse option), 58.3% indicates the judge captures shared preferences well above chance. More details of the human evaluation experiments are provided in Appendix 21.

7 Analysis↩︎

7.1 What Makes Jokes Win?↩︎

For each HTB pairwise comparison, the Llama judge assigns structured tags to the winner (humor mechanisms and delivery) and to the loser (flaw tags). Table [tab:feature95contribution] aggregates all canonical tags across 41,976 non-tie comparisons from the full HTB Llama evaluation. Percentages denote the share of comparisons in which a tag was assigned (multiple tags per comparison are allowed).

Incongruity and absurdity dominate winning mechanisms (\(>\)​84%); punchline positioning (96.1%) and conciseness (72.0%) dominate delivery. Among losers, weak punchlines (69.0%) and cliché (64.9%) are most frequent; overexplained (22.3%) and buried punchlines (12.9%) align with the explainer trap (see section 6.3). Two recurring failure patterns emerge: (1) generic punchlines defaulting to safe, high-probability completions, and (2) overextended setups burying the joke. See Appendix 26 for qualitative examples.

8 Conclusion↩︎

We introduce the Cognitive Synergy Framework, which operationalizes psychological humor theories into six cognitive personas to generate diverse, high-quality humor data via Mixture-of-Thought. HumorGen model variants achieve strong performance among open-weight models and are competitive with frontier systems, outperforming significantly larger models, suggesting that targeted CSF data curation matters more than model scale for humor generation. Our central finding is a data quality ceiling: when SFT data is diverse and well-curated, preference optimization (DPO, O-GRPO) yields no significant gains. We also observe that reasoning-augmented training can reduce judged funniness, while CSF-curated synthetic data outperforms human-written jokes when used in training. Human evaluation further supports that the LLM judge captures subtle preference differences even in subjective “Good vs.Good” pairwise comparisons among high-quality jokes. Overall, our results indicate that in modeling humor generation, performance gains are driven less by optimization complexity and more by the quality and structure of the underlying training data.

9 Limitations↩︎

This study defines a focused empirical scope that future work should broaden. Our analysis is limited to English, text-only humor and two benchmarks (SemEval-MWAHAHA and Humor Transfer Bench), motivating evaluation in additional languages, domains, and multimodal settings. Human evaluation is constrained by the inherent subjectivity of humor and the presence of multiple valid outputs per prompt; we therefore use a small annotator pool (\(N=3\)) over 60 sampled pairwise comparisons, and larger evaluations would provide higher statistical coverage. Model comparisons are restricted to 15 selected models (including models from Qwen, Llama, GPT-model families, and specialized humor models) due to compute constraints; however, this can be expanded to examine more models in the future. Finally, given the subjective nature of humor, there is a tendency for models to produce jokes that are offensive (see Appendix 25) to some listeners or be culturally insensitive.

10 Ethics Statement↩︎

Humor generation risks producing offensive content. Our framework encourages creative mechanisms (e.g., wordplay, absurdity) over denigration or prejudice. Training data derives from the SEMEVAL 2026 MWAHAHA headlines and our synthetic candidates generated from those prompts; evaluation additionally uses HTB prompts for transfer assessment. Human evaluators were volunteers recruited by invitation; no payment was provided. All content in this work was written and verified by the authors. We used an LLM-based tool solely for grammar correction, language polishing, and LaTeX formatting assistance. The authors take full responsibility for the final content.

Appendix - Table of Contents↩︎

11 Benchmark Leaderboards
11.2 HTB Full Results (Llama 3.3-70B Judge)
11.3 HTB Full Results (Qwen 2.5-72B Judge)
11.4 SemEval Full Results (Llama 3.3-70B Judge)
11.5 SemEval Full Results (Qwen 2.5-72B Judge)
12 HTB Design and Domain Descriptions
13 Per-Persona Analysis
14 Training Dataset Persona Distribution
15 Psychological Theories of Humor
17 Training Details and Hyperparameters
16 O-GRPO Objective
17.1 Hyperparameter Configurations
17.2 Training Dynamics and Results
17.3 Evaluation Loss Trajectories
17.4 Comedian Adaptation Hyperparameters
18 Full Persona Prompts
19 Immersive Persona Comparison
20 Think vs.Non-Think
21 Human Evaluation Details
22 Evaluation UI
24 Comedian Adaptation Analysis
25 Qualitative Output Examples: Out-of-Distribution Headlines
26 Failure Mode Examples

11 Benchmark Leaderboards↩︎

Full leaderboard tables (BT ratings with 95% bootstrap confidence intervals) and plots for all four evaluation conditions: HTB and SemEval, under both the Llama 3.3-70B and Qwen 2.5-72B judges.

11.1 Bootstrap and CI Estimation Details↩︎

For each evaluation setting (HTB/SemEval \(\times\) Llama/Qwen), we build a complete model–model outcome matrix from pairwise comparisons and fit Bradley–Terry ratings via the MM algorithm. We estimate uncertainty with non-parametric bootstrap resampling (\(B=100\)): each bootstrap replicate re-samples the evaluation units with replacement, reconstructs the full contest matrix, and re-fits the BT model. We report 95% confidence intervals as percentile bounds (2.5th, 97.5th) over the bootstrap rating distribution for each model. This procedure is applied consistently to all leaderboard tables in this appendix.

11.2 HTB Full Results (Llama 3.3-70B Judge)↩︎

Table 2 reports full BT ratings with 95% bootstrap confidence intervals for all 15 models on HTB under the Llama 3.3-70B judge. Figure 5 visualises the leaderboard; Figure 6 shows the head-to-head win-rate matrix. HumorGen-SFT-7B and HumorGen-DPO-7B rank 3rd and 4th with non-overlapping CIs from all models below rank 6, confirming statistically significant separation from Gemini-2.5-Pro, GPT-OSS-120B, and all smaller baselines.

11.3 HTB Full Results (Qwen 2.5-72B Judge)↩︎

Table 3 and Figures 7–8 report HTB results under the Qwen 2.5-72B judge. The rank ordering is highly consistent with the Llama judge (Kendall \(\tau = 0.8667\), \(p=1.54\times10^{-7}\)), confirming that the leaderboard is robust to judge choice. Notable difference: HumorGen-DPO-Think-7B rises to 3rd under Qwen, suggesting the Qwen judge is slightly more receptive to reasoning-augmented outputs.

11.4 SemEval Full Results (Llama 3.3-70B Judge)↩︎

Table 4 and Figures 9–10 report results on the SemEval 2026 MWAHAHA test set (first 50 headlines, 15 models, 5,250 pairwise comparisons) under the Llama 3.3-70B judge. HumorGen-SFT-7B and HumorGen-DPO-7B rank 4th and 5th, behind the three frontier systems (GPT-5, Kimi-K2, Gemini-2.5-Pro) and ahead of all open-weight general models. The gap between SFT and DPO (4.88 BT points) is within overlapping confidence intervals, confirming no statistically meaningful difference.

11.5 SemEval Full Results (Qwen 2.5-72B Judge)↩︎

Table 5 and Figures 11–12 report SemEval results under the Qwen 2.5-72B judge. Rank ordering is consistent with the Llama judge at Kendall \(\tau = 0.8286\) (\(p=9.95\times10^{-7}\)). Under Qwen, HumorGen-DPO-7B rises to 3rd (BT 1202.76) and HumorGen-SFT-7B to 4th (BT 1193.81), both ahead of Gemini-2.5-Pro, consistent with the HTB Qwen pattern of the Qwen judge rating DPO-aligned outputs slightly higher.

11.5.0.1 Cross-benchmark observations.

Three consistent patterns emerge across all four evaluation conditions (HTB and SemEval, Llama and Qwen judges). First, HumorGen-SFT-7B and HumorGen-DPO-7B cluster tightly at ranks 3–5, with BT differences between them within confidence interval overlap in every condition, consistent with the data quality ceiling finding in the main paper. Second, Think variants most often trail their matched non-thinking counterparts on automatic rankings (e.g., all three pairs on HTB–Llama and SemEval–Llama), though DPO-Think outranks DPO under the Qwen judge on HTB. Third, humor-specialized baselines phi2-Humor and JokeGPT consistently rank 12th–15th across both benchmarks and both judges, confirming that fine-tuning on humor data alone, without cognitive-driven curation, does not yield competitive performance. The high cross-judge agreement (\(\tau \geq 0.83\) on both benchmarks) indicates these orderings are stable and not an artifact of judge choice.

12 Humor Transfer Bench (HTB) Design and Domain Descriptions↩︎

The Humor Transfer Bench (HTB) tests whether humor generation models generalize beyond the single prompt distribution present in existing benchmarks. HTB holds the task constant (Generate a funny joke given this prompt) while varying the input domain. Each domain is defined by a distinct syntactic and pragmatic structure, making domain membership an observable property rather than a subjective label. The eight domains are described in Table 6.

12.1 Humor Transfer Bench Data Card↩︎

HTB is structured as eight equal-size prompt categories so that models can be compared consistently across varied humor-input styles.

13 Per-Persona Analysis↩︎

A key design property of the Cognitive Synergy Framework is that no single persona is hand-picked as the “best”. All six personas generate candidates, the judge ranks them, and the Elo-based curation process decides which jokes enter training. Table 8 reports three independent measures of curation-stage performance, all derived from the HumorRank pairwise tournament over 28,789 candidates across 1,200 prompts.

• Pairwise win rate: proportion of individual 1v1 match-ups a persona’s jokes won (source: full match history).

• Mean rank (of 24): average within-prompt rank across all 1,200 prompts (rank 1 = best in group; source: per-prompt rankings).

• Top-3 rate: proportion of prompts in which the persona produced a top-3 ranked joke.

All three metrics tell the same story and reinforce each other: Neurotic leads across every measure, while Wordsmith consistently trails. Crucially, all six personas remain competitive across different prompts. Wordsmith, despite ranking lowest, still achieves a top-3 placement on 3.1% of prompts, and all personas contribute to the final training corpora (see Table 9).

The ordering reflects the humor mechanisms each persona draws on. Neurotic and Absurdist generate high incongruity and surprise through escalating anxiety and surreal logic respectively, consistent with the dominant winning tags in Table [tab:feature95contribution]. Cynic relies on satire and superiority framing, which scores similarly to Absurdist. Observer occupies the boundary just below 50%, producing reliably solid jokes that rarely dominate. Optimist and Wordsmith trail because wholesome reframings lack edge, and forced wordplay is penalised by the judge. The consistency of the ordering across all three metrics confirms it is not an artefact of any single evaluation perspective.

14 Training Dataset Persona Distribution↩︎

We trace each example in the SFT, DPO, and O-GRPO corpora back to its originating cognitive persona in the generation pool. Persona labels are taken from the raw CSF candidate pool and joined to the final alignment files by exact joke-text match (100% coverage on the v4 corpora). Table 9 summarizes the resulting composition.

The raw candidate pool is balanced across all six personas (\(\sim\)​16.7% each). Elo-based selection for SFT (top 10 per prompt) and DPO (top 5 vs.bottom 5) shifts mass toward Neurotic, Absurdist, and Cynic on the preferred side, which directly reflects their higher mean ranks and top-3 rates in Table 8, while Wordsmith and Optimist are more frequent among rejected DPO examples. O-GRPO retains all 24 candidates per prompt, preserving uniform persona coverage before advantage weighting. Taken together, Tables 8 and 9 show that CSF does not select a single dominant persona: the student model is trained on a distribution spanning all six cognitive styles, weighted by judge-assessed quality rather than by design.

15 Psychological Theories of Humor↩︎

Each cognitive persona in the Cognitive Synergy Framework is grounded in a classical psychological theory of humor. Table 10 summarizes the five theories and their corresponding personas. These theories do not provide a generative recipe for joke construction but characterize the semantic and linguistic structures that make language humorous, which informed our persona design.

The empirical win rates in Table 8 are consistent with these theories: Relief (Neurotic, 63.4%) and Incongruity (Absurdist, 55.8%) generate the most consistently preferred jokes, echoing the dominance of incongruity and absurdity in the judge-assigned mechanism tags (Table [tab:feature95contribution]). Linguistic humor (Wordsmith, 34.9%) scores lowest, suggesting that forced wordplay is penalized more than tension-release or expectation-violation mechanisms.

16 O-GRPO Objective↩︎

O-GRPO is an offline variant of Group Relative Policy Optimization (GRPO) [24] adapted to our humor distillation pipeline. Standard GRPO samples fresh completions at each training step and scores them on the fly; for humor, each score requires pairwise LLM judging, which is too expensive to embed in the training loop. O-GRPO instead uses the fixed CSF candidate pool: all \(G{=}24\) jokes per prompt are ranked once before alignment begins, and training reads from pre-computed scores only.

16.0.0.1 Training data.

For each of the 1,200 SemEval prompts, the MoT ensemble produces 24 candidates (4 per persona \(\times\) 6 personas). Llama-3.3-70B judges a complete pairwise tournament per prompt (Figure 2, step B), yielding an Elo rating \(r_i\) for each candidate \(y_i\). This gives \(\mathcal{D}_{GRPO}\) with 28,800 examples. By contrast, DPO uses five binary pairs per prompt (top-5 vs.bottom-5 Elo jokes; \(N{=}6{,}000\)). O-GRPO retains the full group so that every candidate is weighted relative to its peers on the same prompt, not only the extremes.

16.0.0.2 Advantages and loss.

Raw Elo ratings are comparable only within a prompt (typical spread \(\sim\)​107 points), so we normalize per group before training. For each prompt \(x\), the advantage of candidate \(y_i\) is the group z-score (Equation 1 ): \[A_i = \frac{r_i - \mu_{\mathrm{group}}}{\sigma_{\mathrm{group}} + \epsilon}, \quad \epsilon = 10^{-6}.\] Values are computed once at data preparation and stored with each example. At training time, advantages map to softmax weights over the group, \[\begin{align} \label{eq:ogrpo95weights} w_i &= \frac{\exp(A_i / T)}{\sum_{j=1}^{G}\exp(A_j / T)}, \end{align}\tag{2}\] with temperature \(T{=}1.0\) (Table 11). The objective is advantage-weighted SFT: each candidate contributes cross-entropy loss \(\ell_i(\theta) = -\log \pi_\theta(y_i \mid x)\), computed on assistant-response tokens only, \[\begin{align} \label{eq:ogrpo95loss} \mathcal{L}_{O\text{-GRPO}}(\theta) &= \mathbb{E}_{x}\!\left[\sum_{i=1}^{G} w_i \,\ell_i(\theta)\right]. \end{align}\tag{3}\] Higher-ranked jokes receive larger gradient weight; lower-ranked jokes are down-weighted but still present in the loss. There is no PPO-style clipping and no live reward model during training.

16.0.0.3 Training and observed dynamics.

O-GRPO initializes from the HumorGen-SFT checkpoint and runs for up to 5 epochs with early stopping (patience \(=2\)); full hyperparameters are in Table 11. In practice, the advantage distribution across \(\mathcal{D}_{GRPO}\) is strongly right-skewed (median \(A = -0.479\)): a few high-Elo candidates per prompt pull up \(\mu_{\mathrm{group}}\), leaving most jokes with negative advantages. The optimizer therefore spends much of its signal suppressing weak outputs rather than reinforcing strong ones, which aligns with O-GRPO’s underperformance relative to SFT and DPO (Results: Preference Alignment).

17 Training Details and Hyperparameters↩︎

This section provides a comprehensive record of the training configurations and experimental results for the HumorGen model suite. All models were fine-tuned using the Qwen 2.5-7B-Instruct base architecture on NVIDIA H100-80GB GPUs.

17.1 Hyperparameter Configurations↩︎

Table 11 consolidates the core hyperparameters used across the three major training phases: Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO).

17.2 Training Dynamics and Results↩︎

Table 12 summarizes the convergence behavior and final metrics for the primary alignment experiments.

17.3 Evaluation Loss Trajectories↩︎

Tables 13–15 provide evaluation loss trends for the GRPO and SFT-Think experiments, illustrating the convergence patterns that informed our early stopping decisions.

17.4 Comedian Adaptation Hyperparameters↩︎

Table 16 specifies the unique settings required to mitigate catastrophic forgetting during the human stand-up comedian adaptation phase.

18 Full Persona Prompts↩︎

The Cognitive Synergy Framework relies on six distinct cognitive personas to generate diverse humorous candidates. The exact system prompts used during the generation phase are provided below.

Table [tab:personas95appendix]: The exact system instructions for the six Cognitive Personas. Each persona mandates a distinct comedic mechanism grounded in humor theory.

19 Immersive Persona Comparison↩︎

To illustrate how each Cognitive Persona interprets and subverts the same input premise, we present a side-by-side comparison of six candidates generated from a single SemEval headline.

20 Think vs.Non-Think↩︎

This section illustrates the “Explainer Trap” failure mode. In paired examples, non-thinking models tend to deliver punchier punchlines, while Think variants more often produce verbose, analytical outputs that explain the humor rather than deliver it.

21 Human Evaluation Details↩︎

21.0.0.1 Instructions to participants.

Evaluators signed in via Google and were shown the instructions in Figures 13 and 14 before starting: view a headline and two anonymous jokes, vote for the funnier one, and complete 60 pairs. Model identities were not disclosed.

21.0.0.2 Evaluator demographics.

Three annotators were selected from an initial qualification pool of 12 candidates, all Master’s-level students in Computer Science or Artificial Intelligence. Each candidate in the pool completed a qualification task comprising humor identification and pairwise joke ranking on a shared example set. The three annotators retained for the main evaluation were chosen based on annotation reliability, specifically inter-annotator agreement and response consistency on the qualification set. Participation was voluntary; no compensation was provided.

21.0.0.3 Metrics and recruitment.

180 votes (3 evaluators, 60 pairs each). Human agreement: 31.7%; LLM vs.consensus (Gold Standard): 58.3%; micro-avg: 52.4%. Position bias mitigated via random A/B ordering. We additionally report Krippendorff’s \(\alpha = 0.425\) (3 evaluators) and \(\alpha = 0.412\) (2 evaluators), indicating moderate agreement consistent with the inherent subjectivity of humor.

21.0.0.4 Agreement definitions.

Human agreement: proportion of pairs with unanimous annotator choice. Gold Standard (LLM–consensus): proportion of pairs where the LLM judge matches the human majority vote. Micro-average accuracy: proportion of all individual votes (across all evaluators and pairs) that agree with the LLM’s choice.

21.0.0.5 Category design.

60 pairs, 12 categories (5 each). Table 17.

22 Evaluation UI↩︎

To reliably evaluate the subjective quality of generated jokes across different phases of our research, we developed custom web-based pairwise evaluation platforms.

23 Humor Ranking Granularity – Sample from our Jokes Ranking for creating the Training Dataset↩︎

24 Comedian Adaptation Analysis↩︎

25 Qualitative Output Examples: Out-of-Distribution Headlines↩︎

26 Failure Mode Examples↩︎

Beyond the Explainer Trap (discussed in §4.3), we document two additional failure patterns observed across model variants. The examples below are drawn from held-out evaluation outputs.

References↩︎