Uncertainty-Aware Budget Allocation for Adaptive Test-Time Reasoning
Uncertainty-Aware Budget Allocation (UAB) reallocates a fixed sampling budget across multiple questions to maximize accuracy. It uses a two-phase approach: Phase 1 collects zero-cost difficulty signals from one generation per question via average negative log-likelihood (ANLL). Phase 2 adaptively distributes the remaining budget greedily based on marginal gain, ensuring samples target the most uncertain questions.