---
id: 20260320-T0-16
title: "Transformer能学习未见过的规则"
title_en: "Transformers Learn Rules They've Never Seen"
url: https://ai.daily.yangsir.net/daily/20260320-T0-16
issue_date: 2026-03-20
publish_date: 2026-03-19T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2603.17019
---

# Transformer能学习未见过的规则

arXiv研究证明Transformer能超出插值范围学习规则。研究测试了强插值假设，发现Transformer能推断训练中未出现的规则。通过在算术任务中验证，模型能正确执行未见过的运算规则，表明Transformer具备真正的泛化能力，而非简单的相似性插值。这一发现对理解大模型的推理机制具有重要意义。

## English Version

**Transformers Learn Rules They've Never Seen**

An arXiv study proves Transformers can learn rules beyond interpolation range. Researchers tested a strong interpolation hypothesis and found Transformers can infer rules absent from training. Verified through arithmetic tasks, models correctly executed unseen operations, proving Transformers possess genuine generalization ability rather than just similarity-based interpolation. This finding is significant for understanding LLM reasoning mechanisms.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2603.17019)

**详情页**：https://ai.daily.yangsir.net/daily/20260320-T0-16

---

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