Structured elimination with self-consistency and verification for robust multiple-choice reasoning: a large-scale sports training benchmark and cross-domain evaluation

Large language models (LLMs) excel at many QA tasks but still struggle with multiple-choice question answering (MCQA), especially under strong distractors. Humans often solve such questions by eliminating implausible options and verifying the remaining candidates. We propose a model-agnostic structured elimination framework that unifies stepwise elimination, answer verification, and self-consistency. Instantiated with LLaMA-3 (8B) as the primary backbone, the model performs multi-round option elimination, optionally verifies eliminations via internal checks or lightweight evidence retrieval (e.g., Wikipedia), and aggregates multiple sampled elimination chains for robust decisions. We introduce SportsMCQ-5k, a 5,000-question sports training MCQA benchmark, and evaluate on it alongside CommonsenseQA, Social IQa, and MedMCQA. Across datasets, our method consistently improves accuracy over strong 7B–9B open-source baselines by 4–7 points, while ablations confirm the contributions of verification and self-consistency. The proposed framework enhances robustness and interpretability for educational assessment, including sports training and other discipline-specific testing.