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

Abstract

Generating humor is a complex challenge, not merely due to the difficulty of crafting fluent and novel jokes, but because the perception of "funny" is inherently audience-specific. Furthermore, the supervision data is often noisy; preferences fluctuate based on cultural context, specific situations, and the audience itself, leading to low agreement among annotators. This paper presents our approach for SemEval-2026 Task-1 (MWAHAHA), a competition centered on generating humor within explicit constraints. The task assesses systems through human preference judgments conducted via 1-on-1 arena-style comparisons.

Our methodology employs a "generate-many -> select-best" framework. Initially, we create a diverse array of candidate outputs for each instance by leveraging model ensembling, diversity-oriented decoding, and multi-step prompting. Subsequently, we utilize a preference model to identify the optimal output. This model simulates a "reader" by learning from human pairwise comparisons rather than relying on absolute funniness ratings. To facilitate this, we have made available 2,500 human pairwise judgments gathered through the Humor Arena prototype. Additionally, we introduce an interpretable pipeline designed to transform labeled comparisons into a robust preference model.

Evaluations across three distinct preference datasets demonstrate that our models consistently surpass baseline methods and exhibit superior cross-domain transfer capabilities. We applied this learned preference model to rank candidates for the MWAHAHA setting and have released intermediate artifacts, including candidate pools and their rankings, to support further research. Our system achieved first place in both the English and Chinese subtasks of MWAHAHA, and secured second place in the Spanish subtask.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC