<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
    <channel>
        <title>Paper/Agentic-Rl - 标签 - vllbc02&#39;s blogs</title>
        <link>https://blog.vllbc.top/tags/paper/agentic-rl/</link>
        <description>Paper/Agentic-Rl - 标签 - vllbc02&#39;s blogs</description>
        <generator>Hugo -- gohugo.io</generator><language>zh-CN</language><copyright>This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.</copyright><lastBuildDate>Fri, 21 Aug 2026 00:00:00 &#43;0000</lastBuildDate><atom:link href="https://blog.vllbc.top/tags/paper/agentic-rl/" rel="self" type="application/rss+xml" /><item>
    <title>SAO — Single-Rollout Asynchronous Optimization for Agentic RL</title>
    <link>https://blog.vllbc.top/sao/</link>
    <pubDate>Fri, 21 Aug 2026 00:00:00 &#43;0000</pubDate>
    <author>Zhenyu Hou, Yujiang Li, Jie Tang, Yuxiao Dong</author>
    <guid>https://blog.vllbc.top/sao/</guid>
    <description><![CDATA[<h1 id="sao-single-rollout-asynchronous-optimization">SAO: Single-Rollout Asynchronous Optimization</h1>
<h2 id="笔记地图">笔记地图</h2>
<ul>
<li><i class="far fa-check-square fa-fw" aria-hidden="true"></i> 问题定义：同步 vs 异步 RL 的根本矛盾</li>
<li><i class="far fa-check-square fa-fw" aria-hidden="true"></i> 方法：DIS + Single-rollout + 三个 value 改进</li>
<li><i class="far fa-check-square fa-fw" aria-hidden="true"></i> 实验：6 个基准 + ablation + 训练动态 + 在线学习仿真</li>
<li><i class="far fa-square fa-fw" aria-hidden="true"></i> 相关工作：与 GRPO / VAPO / AReaL 的关系</li>
<li><i class="far fa-square fa-fw" aria-hidden="true"></i> 我的疑问：frozen-attention 的代价？DIS 在长 horizon 是否还有效？</li>
</ul>
<blockquote>
<p><strong>阅读进度</strong>：当前 <code>read-status: in-progress</code>，用 [[Meta Bind]] 插件可一键切换</p>]]></description>
</item>
</channel>
</rss>
