---
id: 20260527-T0-04
title: "LLM推理中存在大量冗余：研究量化思考效率"
title_en: "LLM reasoning shows high redundancy: study quantifies efficiency"
url: https://ai.daily.yangsir.net/daily/20260527-T0-04
issue_date: 2026-05-27
publish_date: 2026-05-26T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2605.23926
---

# LLM推理中存在大量冗余：研究量化思考效率

arXiv论文研究发现，大型语言模型在解决复杂问题时会产生大量冗余推理，包括重复表述、自我验证和循环反思。这种冗余推理虽然提高了问题解决的准确性，但也显著增加了延迟、GPU时间和能耗。研究提出了量化冗余程度的方法。

## English Version

**LLM reasoning shows high redundancy: study quantifies efficiency**

arXiv paper finds LLMs generate extensive redundant reasoning when solving complex problems, including reformulation, self-verification, and circular reflection. This improves accuracy but significantly increases latency, GPU time, and energy consumption.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2605.23926)

**详情页**：https://ai.daily.yangsir.net/daily/20260527-T0-04

---

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