---
id: 20261114-T0-10
title: "排行榜也能被刷：基于偏好的模型评测易受操纵"
title_en: "Leaderboard Hacking: Preference-Based Model Evaluations Are Vulnerable to Manipulation"
url: https://ai.daily.yangsir.net/daily/20261114-T0-10
issue_date: 2026-11-14
publish_date: 2026-11-13T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.14769
---

# 排行榜也能被刷：基于偏好的模型评测易受操纵

研究者指出，Arena式语言模型排名——广泛用于在线排行榜、论文和生产环境模型评估——容易受到操纵。这类排名通过匿名模型的成对比较来评估性能，但基于偏好的评测机制存在被针对性策略影响的空间。研究揭示了排行榜结果可能被系统性扭曲的风险，提醒使用这类排名做模型选型时需要谨慎。

## English Version

**Leaderboard Hacking: Preference-Based Model Evaluations Are Vulnerable to Manipulation**

Researchers show that Arena-style language model rankings, widely used in online leaderboards, papers, and production model evaluation, are vulnerable to manipulation. These rankings assess performance through pairwise comparisons of anonymous models, but preference-based evaluation leaves room for targeted strategies to influence outcomes. The work exposes the risk that leaderboard results can be systematically distorted, a caution for anyone using such rankings to select models.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.14769)

**详情页**：https://ai.daily.yangsir.net/daily/20261114-T0-10

---

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