---
id: 20261117-T0-04
title: "多智能体LLM评测：协调、集成与推理算力常被混为一谈"
title_en: "Multi-Agent LLM Evals Confound Coordination, Ensembling, and Test-Time Compute"
url: https://ai.daily.yangsir.net/daily/20261117-T0-04
issue_date: 2026-11-17
publish_date: 2026-11-16T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.17610
---

# 多智能体LLM评测：协调、集成与推理算力常被混为一谈

一项研究指出，多智能体语言模型的常见评测方式存在混淆：把通信架构与单次贪心推理对比，实际把协调机制、样本数量和推理算力三个变量混在一起。研究者整合了九项主要实证评估，试图拆解各自贡献。结论指向一个方法问题——此前的多智能体优势可能部分来自更多采样或更多算力，而非协调本身。对做多智能体系统的团队，这意味着性能归因需要更严格控制变量。

## English Version

**Multi-Agent LLM Evals Confound Coordination, Ensembling, and Test-Time Compute**

A study finds that common evaluations of multi-agent language models are confounded: comparing communicating architectures against a single greedy pass mixes coordination with sample multiplicity and inference compute. The researchers synthesized nine primary empirical evaluations to separate each factor's contribution. The implication is methodological: some reported multi-agent gains may stem from more samples or more compute rather than coordination itself. Teams building multi-agent systems should control variables more carefully when attributing performance.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.17610)

**详情页**：https://ai.daily.yangsir.net/daily/20261117-T0-04

---

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