---
id: 20260318-T0-28
title: "新理论解释Grokking延迟：表征相变的关键机制"
title_en: "New Theory Explains Grokking Delay: Representational Phase Transitions"
url: https://ai.daily.yangsir.net/daily/20260318-T0-28
issue_date: 2026-03-18
publish_date: 2026-03-17T04:00:00.000Z
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2603.13331
---

# 新理论解释Grokking延迟：表征相变的关键机制

研究者发表新论文，从基本原理解释Grokking现象——模型在完全记住训练数据后仍需很长时间才突然泛化。通过表征相变理论量化延迟时间，发现模型需经历从记忆到泛化的关键转变。该理论可预测不同任务下的Grokking延迟时长。

## English Version

**New Theory Explains Grokking Delay: Representational Phase Transitions**

Researchers published a first-principles theory explaining Grokking's delay—why models suddenly generalize long after memorizing training data. Through representational phase transitions, they quantify the delay and identify the critical shift from memorization to generalization. The theory can predict Grokking delays for different tasks.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2603.13331)

**详情页**：https://ai.daily.yangsir.net/daily/20260318-T0-28

---

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