---
id: 20260908-T0-06
title: "LLM算术推理：找到共享电路，就能跨格式泛化能力有望提升"
title_en: "Shared Circuits Found to Be Key for LLM Arithmetic Reasoning Generalization"
url: https://ai.daily.yangsir.net/daily/20260908-T0-06
issue_date: 2026-09-08
publish_date: 2026-09-07T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2609.04463
---

# LLM算术推理：找到共享电路，就能跨格式泛化能力有望提升

一篇论文发现，LLM在算术推理中的泛化能力，依赖于是否存在跨输入格式（如从'2+5'到'two plus five'）共享的神经电路。在观察到这种共享电路存在的情况下，模型会表现出远好于预期的跨格式泛化能力；而模型无法跨格式推理时，这些共享电路往往是缺失的。这项研究为如何提升语言模型的根本性推理能力提供了新的微观层面的证据，也为未来模型架构调整和训练算法的设计提供了方向性指导，即寻找并强化这种跨任务的共享结构。

## English Version

**Shared Circuits Found to Be Key for LLM Arithmetic Reasoning Generalization**

A new paper identifies shared neural circuits as the key predictor of whether LLMs can generalize reasoning across input formats, like from '2+5' to 'two plus five'. When these shared circuits emerge, models exhibit robust cross-format generalization; their absence correlates with brittleness. This offers micro-level evidence that can guide architecture and training adjustments to build more fundamentally robust reasoning capabilities in future language models.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2609.04463)

**详情页**：https://ai.daily.yangsir.net/daily/20260908-T0-06

---

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