arXiv：cs.AI· Bingnan Xiao, Chenhao Yang, Bingcong Li, Wei Ni, Xin Wang·· 6 小时前AI 评分37
面向 LLM 推理的自适应 Power Sampling（APS）
Adaptive Power Sampling for LLM Reasoning
AI 导读
研究者提出自适应 Power Sampling（APS），在测试时按每个查询调整锐化指数，无需额外训练即可提升 LLM 推理能力。其理论依据是进一步锐化的收益取决于正确答案与错误答案之间的自奖励差距。在 MATH500、HumanEval 和 GPQA 上，APS 的表现始终优于固定锐化指数的 power sampling。
正文
Abstract:Sequence-level power sampling has recently emerged as a training-free approach to reasoning by sampling from a sharpened output distribution of a base large language model (LLM). Nevertheless, existing methods typically sharpen the base model distribution uniformly across queries, overlooking variations in query difficulty and in how well the base model already handles each query. The goal of this work is to equip power sampling with query adaptivity. Theoretically, we show that the benefits of further sharpening are determined by the self-reward gap between correct and incorrect responses. Based on this insight, we propose \emph{Adaptive Power Sampling} (APS), which adjusts the sharpening exponent on a per-query basis at test time using the relationship between answer agreement and the model's self-reward. Experiments across diverse reasoning tasks, including MATH500, HumanEval, and GPQA, show that APS consistently outperforms power sampling with a fixed sharpening exponent, without additional training.
| Comments: | 22 pages, 6 figures |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.08563 [cs.AI] |
| (or arXiv:2610.08563v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.08563 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Bingnan Xiao [view email]
[v1]
Tue, 6 Oct 2026 15:43:35 UTC (279 KB)
来源：arXiv：cs.AI · arxiv.org