<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Training on Xu'Blog</title><link>https://xuquant.com/tags/training/</link><description>Recent content in Training on Xu'Blog</description><image><title>Xu'Blog</title><url>https://xuquant.com/og-default.png</url><link>https://xuquant.com/og-default.png</link></image><generator>Hugo -- 0.152.2</generator><language>zh</language><lastBuildDate>Sat, 13 Jun 2026 10:00:00 +0800</lastBuildDate><atom:link href="https://xuquant.com/tags/training/index.xml" rel="self" type="application/rss+xml"/><item><title>从 million 到 billion：VLA 训练 recipe 在量级跃迁上的工程层重构</title><link>https://xuquant.com/posts/autonomous-driving/vla-train-recipe-billion-scale/</link><pubDate>Sat, 13 Jun 2026 10:00:00 +0800</pubDate><guid>https://xuquant.com/posts/autonomous-driving/vla-train-recipe-billion-scale/</guid><description>Toyota Research 刚发出一篇 4000 小时机器人数据 + 50M VL 样本的 co-training 实证研究 (arXiv:2602.01067)，正好覆盖 VLA 训练数据从 million 量级推进到 billion 量级的工程区间。这篇文章把它的五条结论映射到自驾 VLA 上，再结合 Qwen-VLA paper 的训练 recipe 经验，给出一份面向 billion 级 sample 的训练 recipe 思路——数据混料、Phase 结构、loss 设计、compute 与工程四个维度的具体取舍。</description></item></channel></rss>