---
id: 20260423-T0-11
title: "研究提出高效自我进化LLM方法，数据利用率提升90%"
title_en: "Self-Evolving LLMs Achieve 90% Data Efficiency"
url: https://ai.daily.yangsir.net/daily/20260423-T0-11
issue_date: 2026-04-23
publish_date: 2026-04-22T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2604.18639
---

# 研究提出高效自我进化LLM方法，数据利用率提升90%

最新研究提出'Easy Samples'方法，通过数据高效强化学习实现LLM自我进化。该方法避免了传统RL的高标注成本和无监督学习的性能局限，在保持模型性能的同时将数据需求降低90%。实验显示，该方法在多个基准测试中优于现有技术，为降低大模型训练成本提供了新方向。

## English Version

**Self-Evolving LLMs Achieve 90% Data Efficiency**

Researchers have developed 'Easy Samples,' a data-efficient reinforcement learning method that enables self-evolving LLMs. This approach avoids the high annotation costs of traditional RL and performance limitations of unsupervised methods, reducing data needs by 90% while maintaining model performance. The method outperforms existing techniques across multiple benchmarks, offering a new direction for reducing LLM training costs.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2604.18639)

**详情页**：https://ai.daily.yangsir.net/daily/20260423-T0-11

---

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