---
id: 20260418-T0-16
title: "MixAtlas：多模态LLM训练数据混合优化方案"
title_en: "MixAtlas: Uncertainty-aware Data Mixture Optimization for Multimodal LLMs"
url: https://ai.daily.yangsir.net/daily/20260418-T0-16
issue_date: 2026-04-18
publish_date: 2026-04-17T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2604.14198
---

# MixAtlas：多模态LLM训练数据混合优化方案

arXiv论文《MixAtlas》提出多模态LLM训练数据不确定性优化方法。该方案通过计算样本不确定性，动态调整不同模态数据的混合比例，提升训练效率。在Image-Text混合任务中，该方法比传统方法提高15%下游任务准确率。

## English Version

**MixAtlas: Uncertainty-aware Data Mixture Optimization for Multimodal LLMs**

The arXiv paper 'MixAtlas' proposes uncertainty-aware data mixture optimization for multimodal LLM training. It dynamically adjusts multimodal data ratios based on sample uncertainty, improving training efficiency. It achieves 15% higher downstream accuracy on Image-Text tasks compared to traditional methods.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2604.14198)

**详情页**：https://ai.daily.yangsir.net/daily/20260418-T0-16

---

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