<?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>训练系统 on MessageDaily</title><link>https://inkeast.github.io/MessageDaily/tags/%E8%AE%AD%E7%BB%83%E7%B3%BB%E7%BB%9F/</link><description>Recent content in 训练系统 on MessageDaily</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Wed, 09 Sep 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://inkeast.github.io/MessageDaily/tags/%E8%AE%AD%E7%BB%83%E7%B3%BB%E7%BB%9F/index.xml" rel="self" type="application/rss+xml"/><item><title>TGOPD：在策略蒸馏前先验证教师——提示级可靠性门控 精读</title><link>https://inkeast.github.io/MessageDaily/posts/2026-09-09-tgopd-teacher-gating-opd-paper-reading/</link><pubDate>Wed, 09 Sep 2026 00:00:00 +0000</pubDate><guid>https://inkeast.github.io/MessageDaily/posts/2026-09-09-tgopd-teacher-gating-opd-paper-reading/</guid><description>AllSpark 团队推出 TGOPD：对每个提示用少量验证器打分的教师探针估计可靠性，通过门控路由密集 OPD 信号或验证器锚定的 GRPO。4B/35B 两尺度三域 6/6 设置全部超过 Vanilla OPD，35B LiveCodeBench 64.0（其他蒸馏方法全部负迁移），探针填充教师空闲算力使 GPU 利用率 9.8%→78.9%。</description></item></channel></rss>