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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The "lmfaoooo at SemEval-2026 Task 1" system addresses the challenge of humor generation under explicit constraints, where "funny" is audience-dependent and supervision is noisy. This system, which ranked 1st in the English and Chinese subtasks and 2nd in the Spanish subtask of MWAHAHA, employs a "generate-many - select-best" strategy. It first creates a diverse pool of candidate jokes using multi-step prompting, model ensembling, and diversity-oriented decoding. Subsequently, a preference model selects the best outputs by learning from 2.5K human pairwise judgments collected via the Humor Arena prototype, rather than relying on absolute funniness scores. This approach consistently outperforms baselines and demonstrates stronger cross-domain transfer across three preference datasets.

Key takeaway

For NLP engineers developing generative models for subjective content like humor, you should prioritize building robust preference models based on human comparisons. This approach, demonstrated by the top-ranking "lmfaoooo" system, effectively navigates audience-dependent notions of "funny" and noisy supervision. Consider implementing a "generate-many - select-best" pipeline, leveraging pairwise human judgments to refine your model's output and achieve superior cross-domain transfer.

Key insights

Humor generation significantly improves by modeling audience preferences through human comparisons.

Principles

• "Funny" is audience-dependent and context-sensitive.
• Human pairwise judgments offer robust preference signals.

Method

A "generate-many - select-best" pipeline involves multi-step prompting, model ensembling, and diversity-oriented decoding for candidate generation, followed by selection using a preference model trained on human pairwise comparisons.

In practice

• Collect human pairwise judgments to train preference models for subjective content generation tasks.
• Release intermediate artifacts like candidate pools and rankings to foster collaborative research.

Topics

• Humor Generation
• Preference Modeling
• SemEval-2026
• Natural Language Generation
• Human-in-the-Loop AI
• Model Ensembling

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Memes-as-Replies: Can Models Select Humorous Manga Panel Responses? — LLMs show promise in understanding social cues for humor, but struggle with visual information and subtle wit in meme reply selection.
• Humor Style Drives Laughter, Topic Shapes Acceptability: Evaluating Bilingual Personal and Political Robot-Delivered AI Jokes — Humor style dictates funniness, while joke topic determines acceptability for robot-delivered AI jokes.
• X+Slides: Benchmarking Audience-Conditioned Slide Generation — Benchmarking LLM-generated slides requires audience-specific utility and source grounding, not just completeness or visual quality.
• Re-defining Humor Data Objects for AI Humor Research — AI humor research improves by defining humor as a social interaction with context and explanations, not just a binary state.
• I built a public voting benchmark where models have to make memes out of daily news — Memebench evaluates LLM humor and creativity by generating memes from daily news for public A/B voting.
• RLHF: Aligning Language Models with Human Feedback — RLHF aligns language models by leveraging human judgments, recognizing that judging quality is easier than producing it.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.