Back to feed
arXiv cs.CL
arXiv cs.CL
7/16/2026
RAGthoven at SemEval-2026: Multi-Stage Humor Generation Pipeline Ties Frontier Model Baseline

RAGthoven at SemEval-2026: Multi-Stage Humor Generation Pipeline Ties Frontier Model Baseline

Original: RAGthoven at SemEval-2026 Task 1: A Multi-Stage Pipeline Walks Into a Benchmark and Barely Clears the Bar

Short summary

RAGthoven is a multi-stage LLM pipeline for multilingual humor generation (Planner, Best-of-N Writer, Reflector, Judge) grounded in computational humor theory, evaluated at SemEval-2026 Task 1. It ties for Rank 1 with the Gemini 2.5 Flash baseline across English, Spanish, and Chinese, leading in Spanish by 42 Elo points but trailing in English and Chinese within statistical ties. Agentic variants (ReAct-style and autonomous orchestration) did not outperform the simpler non-agentic pipeline despite higher tool-call budgets, suggesting diminishing returns from elaborate scaffolding with strong frontier models.

  • •RAGthoven ties Gemini 2.5 Flash baseline for Rank 1 in multilingual humor generation across three languages
  • •Agentic variants with higher tool-call budgets did not outperform the simpler non-agentic pipeline
  • •Results suggest language-dependent diminishing returns from multi-stage prompt engineering with strong frontier models

Generated with AI, which can make mistakes.

#research-breakthrough#ai-agents#ai-tools
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Operationalising Multi-Dimensional Evaluation for Conversational Agents: A Scalable, Governed Pipeline with Selective Re-evaluation and Model Benchmarking
arXiv CS.AI
Beyond a Joke: Multi-Angle Reasoning for Detecting and Explaining Harmful Humor in Memes
arXiv CS.AI
The original title is "NOWJ@COLIEE 2026: Adaptive Pipelines for Legal Retrieval and Reasoning"
arXiv cs.CL
HG-RAG: Hierarchy-Guided Retrieval-Augmented Generation for Structured Knowledge Graphs
arXiv CS.AI