reasoningfactbearishSimple repeated sampling with majority voting often outperforms complex reasoning methods when token budgets are equalized across 36 experimental comparisonsMachine Learning02 Aug 2026http://arxiv.org/abs/2607.28576v1