---
id: 20260527-T0-08
title: "双二进制低秩适配实现设备端高效微调"
title_en: "Double-Binary Low-Rank Adaptation Enables Efficient On-Device Tuning"
url: https://ai.daily.yangsir.net/daily/20260527-T0-08
issue_date: 2026-05-27
publish_date: 2026-05-26T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2605.24058
---

# 双二进制低秩适配实现设备端高效微调

arXiv论文提出Signs Floats方法，针对设备端LLM微调中的低效问题。该技术通过双二进制量化将适配器大小压缩40%，同时保持性能不下降。研究显示，在手机等资源受限设备上，该方法比传统LoRA节省30%推理时间。开发者可用此技术为移动端应用定制轻量级AI模型，降低硬件依赖。

## English Version

**Double-Binary Low-Rank Adaptation Enables Efficient On-Device Tuning**

arXiv paper introduces Signs Floats method to address inefficiency in on-device LLM fine-tuning. The technique reduces adapter size by 40% using double-binary quantization while maintaining performance. On resource-constrained devices, it cuts inference time by 30% compared to traditional LoRA. Developers can use this to create lightweight AI models for mobile apps with reduced hardware dependency.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2605.24058)

**详情页**：https://ai.daily.yangsir.net/daily/20260527-T0-08

---

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