Lattice at SemEval-2026 Task 1: Why did the prompt engineer break up with their LLM? Because zero-shot was zero-fun.

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

Summary

The Lattice Team contributed to the SemEval-2026 MWA-HAHA humor generation shared task (English dataset, subtask A) with two distinct LLM-based strategies. The winning approach involved a two-phase process: first, Deepseek-R1 32B generated multiple jokes in a few-shot framework from headlines and word pairs. Second, Llama-3.1 8B ranked these jokes using a voting protocol to identify the funniest. An alternative strategy generated ten times more jokes zero-shot with lighter, faster models, followed by a knock-out tournament for ranking. The Deepseek-R1 based model achieved a first-place tie among nine systems out of 32 valid submissions on the English dataset, presented at the 20th International Workshop on Semantic Evaluation (2026).

Key takeaway

For prompt engineers developing humor generation systems, you should prioritize few-shot prompting with larger models like Deepseek-R1 32B for initial content creation. Subsequently, employ a smaller, efficient model such as Llama-3.1 8B with a voting protocol for robust ranking and selection of the best outputs. This two-stage approach significantly outperforms zero-shot methods, ensuring higher quality results in competitive tasks like SemEval-2026.

Key insights

Few-shot generation with a larger LLM followed by a smaller LLM for ranking outperforms zero-shot for humor generation.

Principles

• Few-shot prompting enhances joke generation quality.
• LLM-based voting protocols effectively rank humor.
• Zero-shot generation yields lower quality, even with more outputs.

Method

Generate jokes using a few-shot Deepseek-R1 32B, then rank them with a Llama-3.1 8B voting protocol to select the best.

In practice

• Employ few-shot prompting for creative text generation tasks.
• Use smaller LLMs for efficient ranking of generated content.
• Avoid zero-shot for humor generation tasks.

Topics

• Humor Generation
• Large Language Models
• Few-shot Learning
• Zero-shot Learning
• LLM Ranking
• SemEval-2026

Best for: AI Engineer, Machine Learning Engineer, Research Scientist, AI Scientist, NLP Engineer, Prompt Engineer

Related on AIssential

• The easiest way to improve prompts — Zero-shot and few-shot prompting are core techniques for controlling language model output based on guidance provided.
• True Zero-shot MT — LLM evaluation faces a benchmark crisis, while specific models like Command R+ excel in RAG and multilingual tasks.
• Automatic Generation of Highlights for Academic Paper Via Prompt-based Learning — Prompt-based learning with LLMs can generate academic paper highlights effectively without extensive labeled data.
• Less Is More: Cognitive Load and the Single-Prompt Ceiling in LLM Mathematical Reasoning — LLM mathematical reasoning accuracy plateaus despite extensive prompt engineering, limited by undecidability and prompt complexity.
• Scikit-LLM vs. Traditional Text Classifiers: When Should You Use an LLM? — LLMs offer superior zero-shot text classification accuracy and competitive latency, especially with limited data.
• Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification — In-context learning for Turkish LVCs is highly prompt-sensitive, with few-shot demonstrations improving LLM calibration to match supervised baselines.

Open in AIssential →

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