Interpretable Adaptive Sampling for LLM Test-Time Scaling

This paper proposes a lightweight fuzzy controller for adaptive test-time scaling in large language models. Instead of assigning a fixed sampling budget to every prompt, the controller uses interpretable signals such as estimated prompt complexity and model confidence to determine how many candidates to generate. The authors evaluate it under matched decoding and controlled answer-selection settings on question-answering and mathematical reasoning tasks, comparing it with best-of-N, compute-aware scaling, and self-certainty baselines. The reported results show improvements over several baselines while reducing average sample counts, with performance remaining close to a selector-matched full-budget control.