Researchers crack AI humor generation with SemEval-2026
AI finally gets jokes—system outperforms humans in constrained humor tasks
Researchers Alexey Tikhonov and Alexey Ivanov have cracked a long-standing challenge in AI: generating humor that resonates with specific audiences. Their paper, submitted to SemEval-2026 Task 1 (MWAHAHA), introduces a two-stage system that first generates a wide pool of joke candidates using multi-step prompting, model ensembling, and diversity-encouraging decoding techniques.
The second stage is where it gets interesting: a preference model trained on 2,500 human pairwise judgments (released as part of their work) ranks candidates to approximate what a “reader” would find funniest. This approach sidesteps the ambiguity of absolute humor ratings by learning from relative preferences. Their system ranked 1st in English and Chinese subtasks and 2nd in Spanish, demonstrating robust cross-lingual and cross-domain performance.
- System ranked 1st in English/Chinese and 2nd in Spanish at SemEval-2026 Task 1 (MWAHAHA) for constrained humor generation
- Used a 'generate-many -> select-best' pipeline with multi-model ensembling and a preference model trained on 2.5K human pairwise judgments
- Preference model outperformed baselines across three datasets and showed strong cross-domain transfer capabilities
Why It Matters
AI humor generation could revolutionize marketing, entertainment, and conversational AI by tailoring jokes to specific audiences and cultures.