<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
    <channel>
        <title>Entropy - 标签 - vllbc02&#39;s blogs</title>
        <link>https://blog.vllbc.top/tags/entropy/</link>
        <description>Entropy - 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>Sat, 06 Sep 2025 00:00:00 &#43;0000</lastBuildDate><atom:link href="https://blog.vllbc.top/tags/entropy/" rel="self" type="application/rss+xml" /><item>
    <title>DCPO：DYNAMIC CLIPPING POLICY OPTIMIZATION</title>
    <link>https://blog.vllbc.top/dcpodynamic-clipping-policy-optimization/</link>
    <pubDate>Sat, 06 Sep 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/dcpodynamic-clipping-policy-optimization/</guid>
    <description><![CDATA[<hr />
<h3
id="论文深度解读在精细化探索中寻求效率与稳定的平衡">论文深度解读：在精细化探索中寻求效率与稳定的平衡</h3>
<p>这篇论文的核心，是针对一个日益重要的领域——<strong>基于可验证奖励的强化学习（Reinforcement
Learning from Verifiable Rewards,
RLVR）</strong>——提出了一种更高效、更稳定的训练框架。RLVR旨在通过基于规则的、明确的奖励信号（例如，数学题答案是否正确）来提升LLMs在数学、编程等需要严谨推理领域的表现。然而，现有的主流方法，如<strong>GRPO</strong>和<strong>DAPO</strong>，在实践中暴露了若干关键瓶颈，这正是DCPO试图攻克的难题。</p>]]></description>
</item>
<item>
    <title>From Trial-and-Error to Improvement：A Systematic Analysis of LLM Exploration Mechanisms in RLVR</title>
    <link>https://blog.vllbc.top/from-trial-and-error-to-improvementa-systematic-analysis-of-llm-exploration-mechanisms-in-rlvr/</link>
    <pubDate>Thu, 28 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/from-trial-and-error-to-improvementa-systematic-analysis-of-llm-exploration-mechanisms-in-rlvr/</guid>
    <description><![CDATA[<p>探讨了rlvr中的熵、ppl和尾端推理权重，并在此基础上设计了奖励。</p>
<hr />
<h3 id="论文综合深度解读">论文综合深度解读</h3>
<p>这篇论文的核心，是围绕着如何让大语言模型（LLM）在解决复杂问题（如数学、编程）时，更有效地“思考”和“探索”。传统的<strong>监督微调（SFT）</strong>
教会模型模仿，而<strong>基于人类反馈的强化学习（RLHF）</strong>
教会模型“讨喜”，但这两种方法都难以让模型在多步骤、有唯一正确答案的推理任务上实现根本性的突破。</p>]]></description>
</item>
<item>
    <title>BEYOND PASS@1：SELF-PLAY WITH VARIATIONAL PROBLEM SYNTHESIS SUSTAINS RLVR</title>
    <link>https://blog.vllbc.top/beyond-pass@1self-play-with-variational-problem-synthesis-sustains-rlvr/</link>
    <pubDate>Thu, 21 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/beyond-pass@1self-play-with-variational-problem-synthesis-sustains-rlvr/</guid>
    <description><![CDATA[<p>解决熵坍塌，自我博弈策略。</p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<p>这篇论文的核心思想极其精妙，它试图解决当前大模型在进行强化学习训练时遇到的一个根本性瓶颈：<strong>“竭泽而渔”式的学习</strong>。传统的强化学习方法，尤其是<strong>基于可验证奖励的强化学习（Reinforcement
Learning with Verifiable Rewards,
RLVR）</strong>，在提升模型解决特定问题的“准确率”（即
<strong>Pass@1</strong>）方面效果显著，但这往往以牺牲模型的“思维多样性”为代价。您可以将其想象成一个学生，为了应付考试，只反复刷同一套题库。他可能会对这些题型了如指掌，考试能得高分（高Pass@1），但一旦遇到稍微新颖或需要发散思维的题目，他就束手无策了，因为他的解题思路已经被固化，探索更优解法的潜力（即
<strong>Pass@k</strong>）受到了限制。</p>]]></description>
</item>
<item>
    <title>Beyond the 80 20 Rule：High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning</title>
    <link>https://blog.vllbc.top/beyond-the-80-20-rulehigh-entropy-minority-tokens-drive-effective-reinforcement-learning-for-llm-reasoning/</link>
    <pubDate>Thu, 21 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/beyond-the-80-20-rulehigh-entropy-minority-tokens-drive-effective-reinforcement-learning-for-llm-reasoning/</guid>
    <description><![CDATA[<p>好的，非常荣幸能以领域专家的身份，为您深入解读这篇富有洞见的论文《Beyond
the 80/20 Rule: High-Entropy Minority Tokens Drive Effective
Reinforcement Learning for LLM Reasoning》。</p>
<p>这篇论文的核心思想，用一句话概括，就是<strong>揭示了在大型语言模型（LLM）的推理过程中，并非所有生成的词元（token）都同等重要；一小部分“高熵”词元扮演了决定推理路径的关键“岔路口”角色，而强化学习的有效性，正源于对这些少数关键词元的精准优化。</strong>
作者通过引入“词元熵”这一新颖视角，不仅深刻洞察了带有可验证奖励的强化学习（<strong>RLVR</strong>,
Reinforcement Learning with Verifiable
Rewards）的工作机制，还提出了一种远超传统“二八定律”的、更高效的训练策略。</p>]]></description>
</item>
<item>
    <title>CURE：Critical-Token-Guided Re-concatenation for Entropy-collapse Prevention</title>
    <link>https://blog.vllbc.top/curecritical-token-guided-re-concatenation-for-entropy-collapse-prevention/</link>
    <pubDate>Tue, 19 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/curecritical-token-guided-re-concatenation-for-entropy-collapse-prevention/</guid>
    <description><![CDATA[<p>感觉思想和<a
href="First%20Return,%20Entropy-Eliciting%20Explore.md">First Return,
Entropy-Eliciting Explore</a>和<a
href="../Agent/AGENTIC%20REINFORCED%20POLICY%20OPTIMIZATION.md">AGENTIC
REINFORCED POLICY
OPTIMIZATION</a>很类似，利用熵做额外探索。联合优化使用的GRPO</p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<h3
id="论文深度解读在不确定处寻找确定的进步">论文深度解读：在“不确定”处寻找“确定”的进步</h3>
<p>这篇论文的核心，是解决一个在利用<strong>强化学习与验证者奖励
(Reinforcement Learning with Verifier Reward, RLVR)</strong>
训练大语言模型（LLM）进行复杂推理（如数学解题）时普遍存在的棘手问题。这个问题就是，模型在训练过程中会很快变得“自信”，倾向于生成少数几种高度相似的、缺乏多样性的答案。这种现象被称为<strong>策略熵衰减</strong>。一旦熵衰减发生，模型的探索能力就会急剧下降，无法发现新的、更优的解题路径，导致其性能提升过早地陷入停滞。这就像一个学生只反复练习自己已经掌握的几种解法，却不去探索更巧妙或更通用的方法，最终难以成为顶尖高手。</p>]]></description>
</item>
<item>
    <title>EDGE-GRPO：Entropy-Driven GRPO with Guided Error Correction for Advantage Diversity</title>
    <link>https://blog.vllbc.top/edge-grpoentropy-driven-grpo-with-guided-error-correction-for-advantage-diversity/</link>
    <pubDate>Tue, 19 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/edge-grpoentropy-driven-grpo-with-guided-error-correction-for-advantage-diversity/</guid>
    <description><![CDATA[<p>利用熵来给予不同的reward，变相的改变了优势值，类似于<a
href="ETTRL：BALANCING%20EXPLORATION%20AND%20EXPLOITATION%20IN%20LLM%20TEST-TIME%20REINFORCEMENT%20LEARNING%20VIA%20ENTROPY%20MECHANISM.md">ETTRL：BALANCING
EXPLORATION AND EXPLOITATION IN LLM TEST-TIME REINFORCEMENT LEARNING VIA
ENTROPY MECHANISM</a></p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<p>首先，让我们对这篇论文进行一个全面的、高度浓缩的解读。</p>
<p>这篇论文的核心目标是解决在利用<strong>强化学习</strong>（Reinforcement
Learning, RL）提升大模型（LLM）推理能力时，一个常见于 <strong>GRPO
(Group Relative Policy Optimization)</strong>
算法中的技术瓶颈——<strong>优势坍塌 (Advantage
Collapse)</strong>。简单来说，GRPO通过比较模型对同一个问题生成的多个答案（一个“组”）的好坏来进行优化。但当一个问题太难，模型生成的所有答案都是错的（或者太简单，所有答案都对），它们会得到完全相同的奖励分数。这时，每个答案相对于“组平均水平”的“优势（Advantage）”就变成了零。策略优化的梯度依赖于这个“优势”值，一旦为零，模型就无法从这个样本中学习到任何东西，导致训练效率低下，尤其是在面对有价值的难题时。</p>]]></description>
</item>
<item>
    <title>ETTRL：BALANCING EXPLORATION AND EXPLOITATION IN LLM TEST-TIME REINFORCEMENT LEARNING VIA ENTROPY MECHANISM</title>
    <link>https://blog.vllbc.top/ettrlbalancing-exploration-and-exploitation-in-llm-test-time-reinforcement-learning-via-entropy-mechanism/</link>
    <pubDate>Tue, 19 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/ettrlbalancing-exploration-and-exploitation-in-llm-test-time-reinforcement-learning-via-entropy-mechanism/</guid>
    <description><![CDATA[<p>依旧是利用熵，也是在熵高的位置分叉，类似<a
href="../Agent/AGENTIC%20REINFORCED%20POLICY%20OPTIMIZATION.md">AGENTIC
REINFORCED POLICY OPTIMIZATION</a>和<a
href="CURE：Critical-Token-Guided%20Re-concatenation%20for%20Entropy-collapse%20Prevention.md">CURE：Critical-Token-Guided
Re-concatenation for Entropy-collapse
Prevention</a>。从而共享路径，减少计算。此外基于熵进行了优势函数重塑。根据熵来调整优势值。<a
href="TTRL：Test-Time%20Reinforcement%20Learning.md">TTRL：Test-Time
Reinforcement Learning</a></p>]]></description>
</item>
<item>
    <title>SSRL：SELF-SEARCH REINFORCEMENT LEARNING</title>
    <link>https://blog.vllbc.top/ssrlself-search-reinforcement-learning/</link>
    <pubDate>Mon, 18 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/ssrlself-search-reinforcement-learning/</guid>
    <description><![CDATA[<p>相较于<a
href="Search-R1：Training%20LLMs%20to%20Reason%20and%20Leverage%20Search%20Engines%20with%20Reinforcement%20Learning.md">Search-R1：Training
LLMs to Reason and Leverage Search Engines with Reinforcement
Learning</a>的使用真实的搜索引擎和<a
href="ZEROSEARCH：Incentivize%20the%20Search%20Capability%20of%20LLMs%20without%20Searching.md">ZEROSEARCH：Incentivize
the Search Capability of LLMs without
Searching</a>使用虚拟的训练的大模型作为搜索引擎，这篇论文直接让自己当搜索引擎，更节省成本。</p>]]></description>
</item>
<item>
    <title>GTPO and GRPO-S：Token and Sequence-Level Reward Shaping with Policy Entropy</title>
    <link>https://blog.vllbc.top/gtpo-and-grpo-stoken-and-sequence-level-reward-shaping-with-policy-entropy/</link>
    <pubDate>Fri, 08 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/gtpo-and-grpo-stoken-and-sequence-level-reward-shaping-with-policy-entropy/</guid>
    <description><![CDATA[<p>又是信用分配的论文，不过实验做的比较少，直接把wandb的表放上来了。</p>
<hr />
<h3
id="论文深度解读从大锅饭式奖励到精准制导"><strong>论文深度解读：从“大锅饭”式奖励到“精准制导”</strong></h3>
<p>这篇论文的核心，是解决大型语言模型在进行长链推理（long-chain
reasoning）任务时的一个根本性难题：<strong>信贷分配（Credit
Assignment）</strong>。想象一下，模型在解答一个需要50个步骤的复杂数学题。如果它在第49步都完美无瑕，但在最后一步算错了，导致最终答案错误。那么，传统的强化学习方法，如<strong>GRPO
(Group Relative Policy
Optimization)</strong>，会给这整个50步的解答过程打上一个“0分”的标签。这意味着，前面49个正确的推理步骤，和最后一个错误的计算步骤，受到了同等的“惩罚”。这种“成王败寇”、“大锅饭”式的奖励机制，效率低下且不合理，是提升模型深度推理能力的一大瓶颈。</p>]]></description>
</item>
<item>
    <title>GTPO：Trajectory-Based Policy Optimization in Large Language Models</title>
    <link>https://blog.vllbc.top/gtpotrajectory-based-policy-optimization-in-large-language-models/</link>
    <pubDate>Fri, 08 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/gtpotrajectory-based-policy-optimization-in-large-language-models/</guid>
    <description><![CDATA[<p>针对信用分配问题和熵坍塌问题进行解决。名字虽然和<a
href="GTPO%20and%20GRPO-S：Token%20and%20Sequence-Level%20Reward%20Shaping%20with%20Policy%20Entropy.md">GTPO
and GRPO-S：Token and Sequence-Level Reward Shaping with Policy
Entropy</a>一样，但T的含义不同。也都是针对信用分配提出的方法。</p>]]></description>
</item>
</channel>
</rss>
