Majority Voting

Majority voting studies how to combine decisions from classifiers, human panels, or multiple model outputs, from social-choice theory to LLM self-consistency. Approaches include Condorcet and PAC-Bayesian analyses, repeated sampling, confidence-weighted voting, minority overrides, and pruning unproductive reasoning paths. Evaluations span math and science questions such as GPQA, knowledge benchmarks such as MMLU-Pro, and code-generation tasks. A central tension is that voting can reduce independent errors but amplify shared errors, silence useful minority views, or even hurt accuracy, so researchers are testing when confidence, diversity, and verification signals can make aggregation safer.