MINDS at SemEval-2026-Task 1: Enhancing Humor Generation through RAG and Synthetic DPO Alignment

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

Summary

The MINDS team's work for SemEval-2026-Task 1 (Subtask A) addresses the challenges of humor generation, which are complicated by subjectivity and limited automatic metrics. They evaluated three instruction-tuned models—Llama 3.1, Gemma 2, and Qwen 2.5—using a round-robin LLM judging framework. The study investigated the impact of Retrieval-Augmented Generation (RAG) and Direct Preference Optimization (DPO) on performance. Findings indicate Llama 3.1 as the strongest baseline model, and DPO consistently improved humor quality across various configurations. This research also confirms the efficacy of LLM-based judging as a practical training signal for optimizing subjective generation tasks.

Key takeaway

For NLP Engineers developing humor generation systems, consider integrating Direct Preference Optimization (DPO) into your training pipeline, as it consistently improves humor quality. You should also explore Llama 3.1 as a robust baseline model. Furthermore, adopt LLM-based judging frameworks to provide practical training signals for optimizing subjective generation tasks, overcoming the limitations of traditional automatic metrics and enhancing model performance.

Key insights

DPO and LLM-based judging enhance humor generation, with Llama 3.1 performing best among evaluated models.

Principles

• DPO consistently improves humor quality.
• LLM-based judging is an effective training signal.
• Subjective generation tasks benefit from LLM judging.

Method

Evaluated Llama 3.1, Gemma 2, and Qwen 2.5 using a round-robin LLM judging framework to assess RAG and DPO impact on humor generation for SemEval-2026-Task 1.

In practice

• Use Llama 3.1 as a strong humor generation baseline.
• Apply DPO for improved humor quality.
• Implement LLM-based judging for subjective tasks.

Topics

• Humor Generation
• Retrieval-Augmented Generation
• Direct Preference Optimization
• LLM Judging
• Llama 3.1
• SemEval-2026

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

Related on AIssential

• Direct Preference Optimization for Chatbot Fine-Tuning: An Empirical Study — DPO offers a computationally efficient and simplified pipeline for competitive LLM chatbot fine-tuning.
• Combining Retrieval-Augmented Text Generation with LLMs for Reading Content Recommendations — RAG significantly enhances LLM-generated personalized reading content, improving relevance and groundedness.
• Steerable Cultural Preference Optimization of Reward Models — SCPO enables LLMs to align with diverse cultural preferences, improving minority performance and data efficiency.
• Token-weighted Direct Preference Optimization with Attention — Using an LLM's intrinsic attention to weight tokens during preference optimization significantly improves alignment and performance.
• Recovering Diversity Without Losing Alignment: A DPO Recipe for Post-Trained LLMs — Post-training LLMs can regain output diversity without losing alignment by carefully constructing DPO preference data.
• I fine-tuned an LLM to be C-3PO to test which training data format works best for persona injection [P] — First-person statements are most effective for LLM persona injection, outperforming chat demos and synthetic docs.

Open in AIssential →

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