---
id: 20261110-T0-12
title: "测试时扩展算力怎么选答案？新方法用成对锦标赛做方面级比较"
title_en: "Aspect-wise Pairwise Tournaments Pick Better Answers from Best-of-n Sampling"
url: https://ai.daily.yangsir.net/daily/20261110-T0-12
issue_date: 2026-11-10
publish_date: 2026-11-09T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.08640
---

# 测试时扩展算力怎么选答案？新方法用成对锦标赛做方面级比较

Best-of-n采样是提升LLM推理的常用手段：模型生成多个候选解，再挑一个。难点在于怎么挑。论文提出方面级成对锦标赛方法，让候选答案两两比较、按不同维度分别评判，最终选出最优解。相比整体打分的单一判断，这种比较方式更能区分质量接近的候选。对依赖测试时算力扩展的推理系统，选择器的精度直接决定多花的算力有没有回报。

## English Version

**Aspect-wise Pairwise Tournaments Pick Better Answers from Best-of-n Sampling**

Best-of-n sampling improves LLM reasoning by generating many candidate solutions and selecting one—but selection is the hard part. This paper proposes aspect-wise pairwise tournaments, pitting candidates against each other in pairs and judging along separate dimensions before picking a winner. Compared with single holistic scores, pairwise comparison better separates candidates of similar quality. For systems relying on test-time compute scaling, selector accuracy determines whether the extra compute pays off.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.08640)

**详情页**：https://ai.daily.yangsir.net/daily/20261110-T0-12

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*