Proofbusters at SemEval-2026 Task 11: Neuro-Symbolic Syllogistic Reasoning via LLM-Guided Structure Extraction and Deterministic Validation

AbstractThis paper describes our system for SemEval-2026 Task 11, which evaluates whether language models can perform formal syllogistic reasoning independent of semantic content. Drawing inspiration from Euler’s abstraction in the Königsberg bridges problem where geographical details were stripped away to reveal pure graph structure, three symbolic abstraction strategies are explored to eliminate belief bias. First, terms are replaced with generic placeholders by template abstraction. Second, entities are mapped into explicit constraint-tracking structures by object-oriented abstraction. Third, statements are translated into mathematical set notation by set-theoretic abstraction, with existential import constraints enforced to align with Aristotelian logic. Using Gemini Flash 2.5 and Pro 2.5, the set-theoretic approach achieves 98.95% accuracy with a content bias (TCE) of 2.13 on the English sub-task (overall score: 46.23). Results demonstrate that deeper mathematical abstraction fully stripping semantic content and leveraging formal set notation substantially outperforms template-based approaches in mitigating belief bias, though challenges remain in multi-step constraint composition.