---
id: 20260423-T0-12
title: "2D早出策略让LLM推理速度提升30%"
title_en: "2D Early Exit Strategy Boosts LLM Inference Speed by 30%"
url: https://ai.daily.yangsir.net/daily/20260423-T0-12
issue_date: 2026-04-23
publish_date: 2026-04-22T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2604.18592
---

# 2D早出策略让LLM推理速度提升30%

研究人员提出了一种二维早出策略，通过分层和分句协同退出机制，让大语言模型在分类任务中实现逐步激活和提前退出。该策略在增量处理句子时动态决定深度，实验显示推理速度提升30%，同时保持85%的准确率。开发者可在低算力设备上部署此方案，适用于实时问答和内容分类场景。

## English Version

**2D Early Exit Strategy Boosts LLM Inference Speed by 30%**

Researchers introduce a 2D early exit strategy coordinating layer-wise and sentence-wise exiting for LLM classification tasks. By processing inputs incrementally and dynamically activating deeper layers, the method achieves 30% faster inference while maintaining 85% accuracy. This enables deployment on low-power devices for real-time Q&A and content classification.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2604.18592)

**详情页**：https://ai.daily.yangsir.net/daily/20260423-T0-12

---

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