Uncertainty-Aware Budget Allocation for Adaptive Test-Time Reasoning

Researchers propose Uncertainty-Aware Budget Allocation for test-time reasoning in language models.

The method gives each question one initial sample, measures uncertainty from its output, then greedily allocates extra samples where they are most needed.

Across six models and five benchmarks in math, logic, and preference tasks, it outperforms uniform sampling baselines.

The result suggests that steering limited compute toward uncertain questions can improve reasoning accuracy without extra systems or model calls.