TINY: Rethinking Selection Bias in LLMs: Quantification and Mitigation using Efficient Majority Voting

Selection bias in Large Language Models (LLMs) for multiple-choice question (MCQ) answering occurs when models show a preference for specific answer choices based on factors like their position or symbolic representation, rather than their content. This bias can undermine the fairness and reliability of LLM-based systems. In this paper, we first introduce a granular label-free selection bias metric that enables efficient and robust evaluation of selection bias without requiring the answer distributions. Although majority voting, which aggregates predictions across all possible permutations of answer choices, has proven effective in mitigating this bias, its computational cost increases factorially with the number of choices. We then propose Batch Question-Context KV caching (BAQCKV), an efficient majority voting technique, which reduces computational overhead while maintaining the effectiveness of bias mitigation. Our methods provide an efficient solution for addressing selection bias, enhancing fairness, and improving the reliability of LLM-based MCQ answering systems.