@inproceedings{suppa-etal-2026-ragthoven,
    title = "{RAG}thoven at {S}em{E}val-2026 Task 1: A Multi-Stage Pipeline Walks Into a Benchmark and Barely Clears the Bar",
    author = "Suppa, Marek  and
      Ondrejov{\'a}, Vikt{\'o}ria  and
      Ganajov{\'a}, Lucia  and
      Karetka, Gregor  and
      Skala, Daniel",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.416/",
    pages = "3343--3356",
    ISBN = "979-8-89176-414-9",
    abstract = "We present {\textbackslash}textsc{\{}RAGthoven{\}}, our system for SemEval-2026 Task{\textasciitilde}1 (MuWaHaHa), Subtask{\textasciitilde}A (multilingual constrained humor generation in English, Spanish, and Chinese).{\textbackslash}textsc{\{}RAGthoven{\}} decomposes creative text generation into a multi-stage large language model (LLM) pipeline ({\textbackslash}textit{\{}Planner{\}}, {\textbackslash}textit{\{}Writer{\}}, {\textbackslash}textit{\{}Reflector{\}}, {\textbackslash}textit{\{}Judge{\}}) grounded in computational humor theories (Benign Violation Theory, Script-based Semantic Theory of Humor) and iteratively refined through prompt engineering across ten experiments.In our final configuration, we augment the Planner with retrieval-augmented generation (RAG) from a curated joke corpus, seeding generation with diverse joke mechanisms.We additionally explore an agentic variant that exposes the same four pipeline stages as tool-calling agents orchestrated by a model loop with a {\textbackslash}textsc{\{}ConstraintAudit{\}} checker. While it achieves full constraint compliance, human pairwise evaluation did not reveal a significant quality advantage over the simpler non-agentic baseline.{\textbackslash}textsc{\{}RAGthoven{\}} achieves Rank{\textasciitilde}1 in all three languages, with the strongest result in Spanish (Elo 1182, 42 points above the Gemini{\textasciitilde}2.5{\textasciitilde}Flash baseline).However, while the system leads in raw Elo in Spanish, it shares Rank{\textasciitilde}1 with the baseline in all three languages due to overlapping confidence intervals; in English and Chinese the gap narrows further, suggesting that elaborate multi-stage prompt engineering may offer diminishing returns once a strong frontier model is in the loop."
}