<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
    <channel>
        <title>Cot_compress - 标签 - vllbc02&#39;s blogs</title>
        <link>https://blog.vllbc.top/tags/cot_compress/</link>
        <description>Cot_compress - 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>Tue, 09 Sep 2025 00:00:00 &#43;0000</lastBuildDate><atom:link href="https://blog.vllbc.top/tags/cot_compress/" rel="self" type="application/rss+xml" /><item>
    <title>Less is More Tokens：Efficient Math Reasoning via Difficulty-Aware Chain-of-Thought Distillation</title>
    <link>https://blog.vllbc.top/less-is-more-tokensefficient-math-reasoning-via-difficulty-aware-chain-of-thought-distillation/</link>
    <pubDate>Tue, 09 Sep 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/less-is-more-tokensefficient-math-reasoning-via-difficulty-aware-chain-of-thought-distillation/</guid>
    <description><![CDATA[<p>利用dpo来压缩思维链，和<a
href="Pruning%20Long%20Chain-of-Thought%20of%20Large%20Reasoning%20Models%20via%20Small-Scale%20Preference%20Optimization.md">Pruning
Long Chain-of-Thought of Large Reasoning Models via Small-Scale
Preference
Optimization</a>思路很像。本文的方法是先生成，后压缩，然后利用压缩后和压缩前的作为偏好数据训练。</p>]]></description>
</item>
<item>
    <title>ThinkDial：An Open Recipe for Controlling Reasoning Effort in Large Language Models</title>
    <link>https://blog.vllbc.top/thinkdialan-open-recipe-for-controlling-reasoning-effort-in-large-language-models/</link>
    <pubDate>Fri, 29 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/thinkdialan-open-recipe-for-controlling-reasoning-effort-in-large-language-models/</guid>
    <description><![CDATA[<p>自适应思考，让模型自己选择思考模式（sft阶段赋予）。主要在于其设计的奖励函数，其发现了推理长度hacking，设计了奖励来引导。</p>
<hr />
<h3
id="论文深度解读揭开可控推理的开放秘方"><strong>论文深度解读：揭开可控推理的“开放秘方”</strong></h3>
<p>大型语言模型，尤其是通过<strong>思维链（Chain-of-thought,
CoT）</strong>
进行复杂推理的模型，其强大的能力背后是高昂的计算成本。模型在思考时生成的每一个“token”（词或字符）都意味着实实在在的算力消耗和时间延迟。然而，并非所有任务都需要模型“全力以赴”。一个简单的客服问答和一个需要深度逻辑推演的数学难题，对推理深度的要求截然不同。当前，像OpenAI的gpt-oss系列模型已经展示了通过“高、中、低”等离散模式来控制推理投入的先进能力，用户可以根据需求选择模式，从而在效果和成本之间取得平衡。但这套技术是闭源的，开源社区一直缺少一套行之有效的方法论来复现这种能力。</p>]]></description>
</item>
<item>
    <title>Think in Blocks：Adaptive Reasoning from Direct Response to Deep Reasoning</title>
    <link>https://blog.vllbc.top/think-in-blocksadaptive-reasoning-from-direct-response-to-deep-reasoning/</link>
    <pubDate>Sat, 23 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/think-in-blocksadaptive-reasoning-from-direct-response-to-deep-reasoning/</guid>
    <description><![CDATA[<p>缩减思维链长度。</p>
<hr />
<p>这篇论文的核心目标是解决当前大语言模型（LLMs）普遍存在的“<strong>过度思考</strong>”（Overthinking）问题。具体来说，许多强大的模型，无论面对“1+1=?”这样的简单问题，还是复杂的微积分难题，都会生成长度相近、步骤繁琐的思维链（Chain-of-Thought,
CoT），这不仅造成了巨大的计算资源浪费，也显著增加了响应延迟，限制了其在实际应用中的效率。</p>]]></description>
</item>
<item>
    <title>Promoting Efficient Reasoning with Verifiable Stepwise Reward</title>
    <link>https://blog.vllbc.top/promoting-efficient-reasoning-with-verifiable-stepwise-reward/</link>
    <pubDate>Mon, 18 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/promoting-efficient-reasoning-with-verifiable-stepwise-reward/</guid>
    <description><![CDATA[<p>思想和<a
href="Good%20Learners%20Think%20Their%20Thinking：Generative%20PRM%20Makes%20Large%20Reasoning%20Model%20More%20Efficient%20Math%20Learne.md">Good
Learners Think Their Thinking：Generative PRM Makes Large Reasoning
Model More Efficient Math
Learne</a>非常相似，都是对step重新划分然后针对step计算奖励</p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<h3 id="论文深度解读">论文深度解读</h3>
<p>这篇论文的核心目标是解决大型推理模型（<strong>Large Reasoning Models,
LRMs</strong>）在处理复杂任务时普遍存在的“<strong>过度思考</strong>”（Overthinking）问题。所谓“过度思考”，指的是模型在面对一些相对简单的问题时，仍然会生成过长、冗余的推理链（<strong>Chain-of-Thought,
CoT</strong>），这不仅大大增加了计算成本和推理时间，有时甚至会因为引入不必要的干扰而降低准确率。这个问题已经成为制约LRMs在真实世界中大规模、经济地部署的关键障碍。</p>]]></description>
</item>
<item>
    <title>Pruning Long Chain-of-Thought of Large Reasoning Models via Small-Scale Preference Optimization</title>
    <link>https://blog.vllbc.top/pruning-long-chain-of-thought-of-large-reasoning-models-via-small-scale-preference-optimization/</link>
    <pubDate>Mon, 18 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/pruning-long-chain-of-thought-of-large-reasoning-models-via-small-scale-preference-optimization/</guid>
    <description><![CDATA[<p>一开始还以为是GRPO的变式，没想到是DPO的变式。对于回答正确的回复，短的为winner，长的为loser。数据筛选部分使用的方法是对一个问题多次回答，根据正确率来判断难度。</p>]]></description>
</item>
<item>
    <title>Sample More to Think Less：Group Filtered Policy Optimization for Concise Reasoning</title>
    <link>https://blog.vllbc.top/sample-more-to-think-lessgroup-filtered-policy-optimization-for-concise-reasoning/</link>
    <pubDate>Fri, 15 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/sample-more-to-think-lessgroup-filtered-policy-optimization-for-concise-reasoning/</guid>
    <description><![CDATA[<p>压缩cot的工作，类似<a
href="../Reasoning/LAPO：INTERNALIZING%20REASONING%20EFFICIENCY%20VIA%20LENGTH-ADAPTIVE%20POLICY%20OPTIMIZATION.md">LAPO：INTERNALIZING
REASONING EFFICIENCY VIA LENGTH-ADAPTIVE POLICY
OPTIMIZATION</a>，这篇叫做GFPO。</p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<p>这篇论文直面了当前大模型领域一个非常棘手且普遍的问题：在通过强化学习（Reinforcement
Learning）进行对齐训练时，模型为了追求更高的奖励（reward），往往会产生越来越长、越来越啰嗦的回答。这种现象被称为<strong>“长度膨胀”（Length
Inflation）</strong>。尽管更长的回答在处理复杂问题时有时是必要的，但很多时候模型只是在生成重复、冗长的“填充性”文本，这不仅极大地增加了推理的计算成本和延迟，有时甚至会因为引入过多无关信息而降低回答的准确性。</p>]]></description>
</item>
<item>
    <title>TRAIN LONG, THINK SHORT：CURRICULUM LEARNING FOR EFFICIENT REASONING</title>
    <link>https://blog.vllbc.top/train-long-think-shortcurriculum-learning-for-efficient-reasoning/</link>
    <pubDate>Thu, 14 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/train-long-think-shortcurriculum-learning-for-efficient-reasoning/</guid>
    <description><![CDATA[<p>课程学习，逐步压缩思维链，思想类似于<a
href="LAPO：INTERNALIZING%20REASONING%20EFFICIENCY%20VIA%20LENGTH-ADAPTIVE%20POLICY%20OPTIMIZATION.md">LAPO：INTERNALIZING
REASONING EFFICIENCY VIA LENGTH-ADAPTIVE POLICY OPTIMIZATION</a></p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<h3
id="论文深度解读从长篇大论到言简意赅的智能进化">论文深度解读：从“长篇大论”到“言简意赅”的智能进化</h3>
<p>这篇论文（可以在<a
href="https://arxiv.org/abs/2508.08940">arXiv:2508.08940v1</a>上访问）的核心贡献在于，它巧妙地将<strong>课程学习（Curriculum
Learning）</strong>
的思想引入到大型语言模型的推理能力训练中，旨在解决一个日益突出的矛盾：我们既希望模型具备强大的、多步骤的复杂推理能力（这通常需要长的“思考链”，即Chain-of-Thought），又希望它在推理时尽可能地节省计算资源（即使用更少的token）。</p>]]></description>
</item>
<item>
    <title>LAPO：INTERNALIZING REASONING EFFICIENCY VIA LENGTH-ADAPTIVE POLICY OPTIMIZATION</title>
    <link>https://blog.vllbc.top/lapointernalizing-reasoning-efficiency-via-length-adaptive-policy-optimization/</link>
    <pubDate>Mon, 04 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/lapointernalizing-reasoning-efficiency-via-length-adaptive-policy-optimization/</guid>
    <description><![CDATA[<p>一言以蔽之就是对于一个问题找出n个正确回答的长度中位数，然后将这个数放到prompt提示大模型，进行GRPO的训练。两个阶段有不同的奖励。</p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<p>首先，我将为您呈现对这篇论文的整体性深度解读。</p>]]></description>
</item>
</channel>
</rss>
