@inproceedings{dehouck-etal-2026-lattice,
    title = "Lattice at {S}em{E}val-2026 Task 1: Why did the prompt engineer break up with their {LLM}? Because zero-shot was zero-fun.",
    author = "Dehouck, Mathieu  and
      Seminck, Olga  and
      Delaborde, Marine  and
      Dupont, Yoann  and
      Durandard, No{\'e}",
    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-1.117/",
    pages = "849--861",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper describes the contribution of theLattice Team to the humor generation MWA-HAHA Sem-Eval shared task on the Englishdata set for subtask A. During the developmentphase, we experimented with two different ap-proaches, but after a quick comparison of theoutputs, it turned out that one was clearly moresuccessful than the other. The winning strategycan be seen as consisting of two phases: first,we used a few-shot framework to let Deepseek-R1 32B generate multiple jokes based on theinput (headlines and word pairs). Second, weset up a voting protocol for Llama-3.1 8B torank the generated jokes and find the funniestone. The other strategy also consisted in twophases: first, we generate many more jokesin a zero-shot way with lighter, faster models,and then we turn back to ranking the generatedjokes, but since we have about ten time morejokes in this second setting, we follow a knock-out tournament procedure in order to find thebest jokes. Our Deepseek-R1 based model isone of the nine systems that shared a first placeon the English data set that received a total of32 valid submissions."
}