---
id: 20260607-T0-04
title: "LLM判决易被操控？研究发现决策后交互会显著影响评价结果"
title_en: "LLM Judges Easily Manipulated: Post-Decision Interaction Skews Results"
url: https://ai.daily.yangsir.net/daily/20260607-T0-04
issue_date: 2026-06-07
publish_date: 2026-06-06T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2606.05384
---

# LLM判决易被操控？研究发现决策后交互会显著影响评价结果

最新研究揭示，广泛使用的LLM自动评估系统存在严重缺陷。传统评估假设判断结果是固定输入的稳定属性，但实验证明，在决策后与模型进行交互会显著改变评价结果。这种“决策后交互攻击”可能导致评估结果失真，影响基准测试的可靠性。研究团队通过实验验证了这一现象，并提出了相应的防御建议。对于依赖LLM评估的开发者和研究人员，需要注意这一潜在风险。

## English Version

**LLM Judges Easily Manipulated: Post-Decision Interaction Skews Results**

New research reveals a critical flaw in LLM-as-judge evaluation systems. Contrary to the assumption that judgments are stable, experiments show that post-decision interaction with the model can significantly alter evaluation results. This 'post-decision interaction attack' can lead to skewed benchmarks and unreliable assessments. The team demonstrates this vulnerability and suggests defensive measures. Developers relying on LLM evaluations should be aware of this risk.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2606.05384)

**详情页**：https://ai.daily.yangsir.net/daily/20260607-T0-04

---

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