Sequence‑level power sampling has recently been introduced as a training‑free reasoning technique that samples from a sharpened output distribution of a base large language model (LLM). Existing approaches typically apply a uniform sharpening exponent across all queries, ignoring variations in query difficulty and the model's existing competence on each query.\ \ The goal of this work is to endow power sampling with query adaptivity. Theoretical analysis shows that the benefit of additional sharpening is determined by the self‑reward gap between correct and incorrect responses. Guided by this insight, we propose Adaptive Power Sampling (APS), which adjusts the sharpening exponent per query at test time based on the relationship between answer agreement and the model's self‑reward.\ \ Experiments span diverse reasoning tasks such as MATH500, HumanEval, and GPQA. The results demonstrate that APS consistently outperforms fixed‑exponent power sampling without any extra training.\ \ Review