---
id: 20260407-T0-14
title: "单智能体LLM在同等思考token预算下多跳推理更优"
title_en: "Single-Agent LLMs Outperform Multi-Agent in Multi-Hop Reasoning"
url: https://ai.daily.yangsir.net/daily/20260407-T0-14
issue_date: 2026-04-07
publish_date: 2026-04-06T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2604.02460
---

# 单智能体LLM在同等思考token预算下多跳推理更优

arXiv研究发现，在计算资源（思考token）相同的情况下，单智能体LLM系统在多跳推理任务上表现优于多智能体系统。研究澄清了先前多智能体系统性能提升的误解——这些提升主要来自增加的计算量，而非架构优势。当公平分配计算资源时，单智能体系统能达到或超过多智能体性能，这对设计高效LLM推理系统具有重要启示。

## English Version

**Single-Agent LLMs Outperform Multi-Agent in Multi-Hop Reasoning**

arXiv research finds that single-agent LLM systems outperform multi-agent systems on multi-hop reasoning tasks when given equal computational budgets (thinking tokens). The study clarifies previous misconceptions about multi-agent performance gains, which were primarily due to increased computation rather than architectural advantages. With fair resource allocation, single agents can match or exceed multi-agent performance.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2604.02460)

**详情页**：https://ai.daily.yangsir.net/daily/20260407-T0-14

---

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