<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
    <channel>
        <title>Prm - 标签 - vllbc02&#39;s blogs</title>
        <link>https://blog.vllbc.top/tags/prm/</link>
        <description>Prm - 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>Mon, 08 Sep 2025 00:00:00 &#43;0000</lastBuildDate><atom:link href="https://blog.vllbc.top/tags/prm/" rel="self" type="application/rss+xml" /><item>
    <title>BEYOND CORRECTNESS：HARMONIZING PROCESS AND OUTCOME REWARDS THROUGH RL TRAINING</title>
    <link>https://blog.vllbc.top/beyond-correctnessharmonizing-process-and-outcome-rewards-through-rl-training/</link>
    <pubDate>Mon, 08 Sep 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/beyond-correctnessharmonizing-process-and-outcome-rewards-through-rl-training/</guid>
    <description><![CDATA[<p>利用orm和prm筛选数据，然后用筛选好的数据进行训练，最近好多这种筛选的思路，如<a
href="../LLM-RL/Sample%20More%20to%20Think%20Less：Group%20Filtered%20Policy%20Optimization%20for%20Concise%20Reasoning.md">Sample
More to Think Less：Group Filtered Policy Optimization for Concise
Reasoning</a>和<a
href="../Agent/rStar2-Agent：Agentic%20Reasoning%20Technical%20Report.md">rStar2-Agent：Agentic
Reasoning Technical Report</a></p>]]></description>
</item>
<item>
    <title>Your Reward Function for RL is Your Best PRM for Search：Unifying RL and Search-Based TTS</title>
    <link>https://blog.vllbc.top/your-reward-function-for-rl-is-your-best-prm-for-searchunifying-rl-and-search-based-tts/</link>
    <pubDate>Sat, 23 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/your-reward-function-for-rl-is-your-best-prm-for-searchunifying-rl-and-search-based-tts/</guid>
    <description><![CDATA[<p>利用专家轨迹和逆向强化学习优化一个prm，用于grpo中计算奖励。</p>
<hr />
<p>首先，让我们进入对这篇论文的整体深度解读。</p>
<h3
id="论文深度解读一个免费且动态的推理过程裁判">论文深度解读：一个“免费”且“动态”的推理过程裁判</h3>
<p>当前，为了提升大型语言模型（LLM）在数学、编程、科学推理等复杂任务上的表现，学术界和工业界主要探索两大类<strong>测试时扩展（Test-Time
Scaling,
TTS）</strong>技术。您可以将TTS理解为，在模型推理（回答问题）时，不满足于只生成一个答案，而是通过投入更多的计算资源来探索多种可能性，从而找到更优的解。</p>]]></description>
</item>
<item>
    <title>CAPO：Towards Enhancing LLM Reasoning through Verifiable Generative Credit Assignment</title>
    <link>https://blog.vllbc.top/capotowards-enhancing-llm-reasoning-through-verifiable-generative-credit-assignment/</link>
    <pubDate>Thu, 07 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/capotowards-enhancing-llm-reasoning-through-verifiable-generative-credit-assignment/</guid>
    <description><![CDATA[<p>信用分配在推理任务中。对比了传统prm和RLVR。让大模型对step的正确与否打标签，然后针对step计算reward，然后加权计算最终token的reward计算adv。就是不知道针对adv进行加权计算是否更直接更好一点。很容易想到的方法。</p>]]></description>
</item>
<item>
    <title>Good Learners Think Their Thinking：Generative PRM Makes Large Reasoning Model More Efficient Math Learne</title>
    <link>https://blog.vllbc.top/good-learners-think-their-thinkinggenerative-prm-makes-large-reasoning-model-more-efficient-math-learne/</link>
    <pubDate>Sat, 02 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/good-learners-think-their-thinkinggenerative-prm-makes-large-reasoning-model-more-efficient-math-learne/</guid>
    <description><![CDATA[<p>一篇关于PRM的方法（最近prm有点复兴的趋势）</p>
<hr />
<h3
id="论文深度解读从结果导向到过程学习的范式革命">论文深度解读：从“结果导向”到“过程学习”的范式革命</h3>
<p>在大型语言模型（LLM） стремительно (qián suǒ wèi
yǒu)发展的今天，我们已经见证了它们在语言理解和生成方面的惊人能力。然而，要让模型真正具备强大的<strong>逻辑推理</strong>能力，尤其是在数学、编程等需要严谨步骤的领域，仅仅依赖于预测下一个词的预训练目标是远远不够的。因此，<strong>强化学习（Reinforcement
Learning,
RL）</strong>，特别是基于结果反馈的RL方法，成为了提升模型推理能力的主流技术路径。</p>]]></description>
</item>
<item>
    <title>PROCESS REINFORCEMENT THROUGH IMPLICIT REWARDS</title>
    <link>https://blog.vllbc.top/process-reinforcement-through-implicit-rewards/</link>
    <pubDate>Wed, 16 Jul 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/process-reinforcement-through-implicit-rewards/</guid>
    <description><![CDATA[<h3 id="论文深入解读">论文深入解读</h3>
<p>这篇名为《Process Reinforcement through Implicit
Rewards》(通过隐式奖励进行过程强化)
的论文，由来自清华大学、上海人工智能实验室、UIUC
等顶尖机构的研究者共同完成，为大语言模型（LLM）的强化学习（RL）领域带来了一个极具价值和创新性的解决方案——<strong>PRIME</strong>
框架。其核心贡献在于，它成功地将<strong>过程监督 (Process
Supervision)</strong> 的高效率与<strong>结果监督 (Outcome
Supervision)</strong>
的低成本相结合，解决了在复杂推理任务（如数学和编程）中应用强化学习时面临的关键瓶颈。</p>]]></description>
</item>
</channel>
</rss>
