Investigating how LLM uncertainty guides adaptive test-time compute allocation for search and verification strategies to optimize efficiency and accuracy in reasoning tasks.
This research investigates adaptive test-time compute allocation in Large Language Models, driven by internal uncertainty estimates derived from logits, semantic agreement, or self-consistency. The core aim is to predict the marginal value of applying additional search or verification strategies to enhance efficiency and accuracy, particularly in reasoning tasks. It explores the non-monotonic relationship where optimal compute gain occurs at intermediate uncertainty levels, guiding resource allocation.