@inproceedings{advani-2026-lakshadvani,
    title = "lakshadvani at {S}em{E}val-2026 Task 11: A Neuro-Symbolic Approach to Content-Independent Syllogistic Reasoning",
    author = "Advani, Laksh",
    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-1.80/",
    pages = "561--566",
    ISBN = "979-8-89176-414-9",
    abstract = "We describe our system for SemEval-2026 Task 11 on disentangling content from formal reasoning. The content effect in syllogistic reasoning, where models judge validity based on conclusion plausibility rather than logical structure, persists even with explicit instructions to ignore real-world knowledge. We find that this bias is better addressed structurally than through prompting: by restricting the LLM to a translation role (mapping natural language to abstract variables) and delegating all deductive reasoning to a deterministic checker over the 24 valid Aristotelian forms, we eliminate content bias entirely on Subtask 1 (100.0 combined, TCE=0.0, 4th place).Our Subtask 2 system, which lacks this separation, scores 41.08 (7th place) despite 95.26{\%} accuracy and 99.47{\%} premise retrieval F1, because a TCE of 2.94 incurs a 58{\%} penalty. A three-way ablation on training data using GPT-5 confirms the pattern:Vanilla LLM: 78{\%} accuracy / TCE=19LLM + Aristotelian Rules in Prompt: 90{\%} accuracy / TCE=5LLM + Symbolic Checker: 97{\%} accuracy / TCE=3"
}