https://mars160.github.io/ Recent content on 某位老王的小窝 Hugo -- gohugo.io zh-cn © 2026 Abanana Sat, 06 Dec 2025 17:02:12 +0800 https://mars160.github.io/posts/%E9%9D%A2%E5%90%91%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E9%97%A8%E6%8E%A7%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6%E9%9D%9E%E7%BA%BF%E6%80%A7%E7%A8%80%E7%96%8F%E6%80%A7%E5%92%8Cattention-sink-free/ Sat, 06 Dec 2025 17:02:12 +0800 https://mars160.github.io/posts/%E9%9D%A2%E5%90%91%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E9%97%A8%E6%8E%A7%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6%E9%9D%9E%E7%BA%BF%E6%80%A7%E7%A8%80%E7%96%8F%E6%80%A7%E5%92%8Cattention-sink-free/ <h2 class="relative group">面向大语言模型的门控注意力机制：非线性、稀疏性和 Attention-Sink-Free <div id="面向大语言模型的门控注意力机制非线性稀疏性和-attention-sink-free" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e9%9d%a2%e5%90%91%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%9a%84%e9%97%a8%e6%8e%a7%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b6%e9%9d%9e%e7%ba%bf%e6%80%a7%e7%a8%80%e7%96%8f%e6%80%a7%e5%92%8c-attention-sink-free" aria-label="锚点">#</a> </span> </h2> <p><figure><img class="my-0 rounded-md" loading="lazy" decoding="async" fetchpriority="low" alt="" src="https://mars160.github.io/post_imgs/%E9%9D%A2%E5%90%91%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E9%97%A8%E6%8E%A7%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6%EF%BC%9A%E9%9D%9E%E7%BA%BF%E6%80%A7%E3%80%81%E7%A8%80%E7%96%8F%E6%80%A7%E5%92%8CAttention-Sink-Free/VeBEbQu4tokbLoxifqrc69bJnjh.png" ></figure> </p> https://mars160.github.io/posts/saber%E4%B8%80%E7%A7%8D%E9%92%88%E5%AF%B9%E6%89%A9%E6%95%A3%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E8%87%AA%E9%80%82%E5%BA%94%E5%8A%A0%E9%80%9F%E4%B8%8E%E5%9B%9E%E6%BA%AF%E5%A2%9E%E5%BC%BA%E7%9A%84%E9%AB%98%E6%95%88%E9%87%87%E6%A0%B7%E6%96%B9%E6%B3%95/ Fri, 07 Nov 2025 15:56:05 +0800 https://mars160.github.io/posts/saber%E4%B8%80%E7%A7%8D%E9%92%88%E5%AF%B9%E6%89%A9%E6%95%A3%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E8%87%AA%E9%80%82%E5%BA%94%E5%8A%A0%E9%80%9F%E4%B8%8E%E5%9B%9E%E6%BA%AF%E5%A2%9E%E5%BC%BA%E7%9A%84%E9%AB%98%E6%95%88%E9%87%87%E6%A0%B7%E6%96%B9%E6%B3%95/ <h2 class="relative group">Saber：一种针对扩散语言模型的自适应加速与回溯增强的高效采样方法 <div id="saber一种针对扩散语言模型的自适应加速与回溯增强的高效采样方法" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#saber%e4%b8%80%e7%a7%8d%e9%92%88%e5%af%b9%e6%89%a9%e6%95%a3%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%9a%84%e8%87%aa%e9%80%82%e5%ba%94%e5%8a%a0%e9%80%9f%e4%b8%8e%e5%9b%9e%e6%ba%af%e5%a2%9e%e5%bc%ba%e7%9a%84%e9%ab%98%e6%95%88%e9%87%87%e6%a0%b7%e6%96%b9%e6%b3%95" aria-label="锚点">#</a> </span> </h2> <p>Saber: An Efficient Sampling with Adaptive Acceleration and Backtracking Enhanced Remasking for Diffusion Language Model</p> https://mars160.github.io/posts/r-stitch%E7%94%A8%E4%BA%8E%E9%AB%98%E6%95%88%E6%8E%A8%E7%90%86%E7%9A%84%E5%8A%A8%E6%80%81%E8%BD%A8%E8%BF%B9%E6%8B%BC%E6%8E%A5/ Fri, 07 Nov 2025 15:55:51 +0800 https://mars160.github.io/posts/r-stitch%E7%94%A8%E4%BA%8E%E9%AB%98%E6%95%88%E6%8E%A8%E7%90%86%E7%9A%84%E5%8A%A8%E6%80%81%E8%BD%A8%E8%BF%B9%E6%8B%BC%E6%8E%A5/ <h2 class="relative group">R-STITCH：用于高效推理的动态轨迹拼接 <div id="r-stitch用于高效推理的动态轨迹拼接" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#r-stitch%e7%94%a8%e4%ba%8e%e9%ab%98%e6%95%88%e6%8e%a8%e7%90%86%e7%9a%84%e5%8a%a8%e6%80%81%e8%bd%a8%e8%bf%b9%e6%8b%bc%e6%8e%a5" aria-label="锚点">#</a> </span> </h2> <p>R-STITCH: DYNAMIC TRAJECTORY STITCHING FOR EFFICIENT REASONING</p> https://mars160.github.io/posts/%E7%9C%9F-self-spec-dlm/ Fri, 07 Nov 2025 15:55:28 +0800 https://mars160.github.io/posts/%E7%9C%9F-self-spec-dlm/ <h2 class="relative group">真-Self-Spec-DLM <div id="真-self-spec-dlm" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e7%9c%9f-self-spec-dlm" aria-label="锚点">#</a> </span> </h2> <p>SELF SPECULATIVE DECODING FOR DIFFUSION LARGE LANGUAGE MODELS</p> https://mars160.github.io/posts/slowfast%E9%87%87%E6%A0%B7%E5%8A%A0%E9%80%9Fdlm/ Fri, 07 Nov 2025 15:55:08 +0800 https://mars160.github.io/posts/slowfast%E9%87%87%E6%A0%B7%E5%8A%A0%E9%80%9Fdlm/ <h2 class="relative group">SlowFast 采样加速 DLM <div id="slowfast-采样加速-dlm" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#slowfast-%e9%87%87%e6%a0%b7%e5%8a%a0%e9%80%9f-dlm" aria-label="锚点">#</a> </span> </h2> <p>ACCELERATING DIFFUSION LARGE LANGUAGE MODELS WITH SLOWFAST SAMPLING: THE THREE GOLDEN PRINCIPLES</p> https://mars160.github.io/posts/%E7%AE%80%E6%B4%81%E6%8F%90%E7%A4%BA%E9%80%9A%E8%BF%87%E7%94%9F%E6%88%90%E8%BF%87%E7%A8%8B%E4%B8%AD%E7%9A%84%E8%BF%9E%E7%BB%AD%E7%AE%80%E6%B4%81%E6%8F%90%E7%A4%BA%E6%8F%90%E5%8D%87%E6%8E%A8%E7%90%86%E6%95%88%E7%8E%87/ Fri, 10 Oct 2025 15:59:21 +0800 https://mars160.github.io/posts/%E7%AE%80%E6%B4%81%E6%8F%90%E7%A4%BA%E9%80%9A%E8%BF%87%E7%94%9F%E6%88%90%E8%BF%87%E7%A8%8B%E4%B8%AD%E7%9A%84%E8%BF%9E%E7%BB%AD%E7%AE%80%E6%B4%81%E6%8F%90%E7%A4%BA%E6%8F%90%E5%8D%87%E6%8E%A8%E7%90%86%E6%95%88%E7%8E%87/ <h2 class="relative group">简洁提示：通过生成过程中的连续简洁提示提升推理效率 <div id="简洁提示通过生成过程中的连续简洁提示提升推理效率" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e7%ae%80%e6%b4%81%e6%8f%90%e7%a4%ba%e9%80%9a%e8%bf%87%e7%94%9f%e6%88%90%e8%bf%87%e7%a8%8b%e4%b8%ad%e7%9a%84%e8%bf%9e%e7%bb%ad%e7%ae%80%e6%b4%81%e6%8f%90%e7%a4%ba%e6%8f%90%e5%8d%87%e6%8e%a8%e7%90%86%e6%95%88%e7%8e%87" aria-label="锚点">#</a> </span> </h2> <p>ConciseHint: Boosting Efficient Reasoning via Continuous Concise Hints during Generation</p> https://mars160.github.io/posts/prophetdiffusion%E6%A8%A1%E5%9E%8B%E5%9F%BA%E4%BA%8E%E7%BD%AE%E4%BF%A1%E5%BA%A6%E7%9A%84decoding%E6%97%A9%E5%81%9C/ Fri, 10 Oct 2025 15:58:42 +0800 https://mars160.github.io/posts/prophetdiffusion%E6%A8%A1%E5%9E%8B%E5%9F%BA%E4%BA%8E%E7%BD%AE%E4%BF%A1%E5%BA%A6%E7%9A%84decoding%E6%97%A9%E5%81%9C/ <h2 class="relative group">Prophet：Diffusion 模型基于置信度的 Decoding 早停 <div id="prophetdiffusion-模型基于置信度的-decoding-早停" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#prophetdiffusion-%e6%a8%a1%e5%9e%8b%e5%9f%ba%e4%ba%8e%e7%bd%ae%e4%bf%a1%e5%ba%a6%e7%9a%84-decoding-%e6%97%a9%e5%81%9c" aria-label="锚点">#</a> </span> </h2> <p>Prophet: Fast Decoding for Diffusion Language Models</p> https://mars160.github.io/posts/specdiff%E4%BD%BF%E7%94%A8%E6%89%A9%E6%95%A3%E6%A8%A1%E5%9E%8B%E4%BD%9C%E4%B8%BAdraft%E6%A8%A1%E5%9E%8B/ Fri, 10 Oct 2025 15:57:40 +0800 https://mars160.github.io/posts/specdiff%E4%BD%BF%E7%94%A8%E6%89%A9%E6%95%A3%E6%A8%A1%E5%9E%8B%E4%BD%9C%E4%B8%BAdraft%E6%A8%A1%E5%9E%8B/ <h2 class="relative group">SpecDiff：使用扩散模型作为 Draft 模型 <div id="specdiff使用扩散模型作为-draft-模型" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#specdiff%e4%bd%bf%e7%94%a8%e6%89%a9%e6%95%a3%e6%a8%a1%e5%9e%8b%e4%bd%9c%e4%b8%ba-draft-%e6%a8%a1%e5%9e%8b" aria-label="锚点">#</a> </span> </h2> <p>Speculative Diffusion Decoding: Accelerating Language Generation through Diffusion</p> https://mars160.github.io/posts/diffuspec%E8%A7%A3%E9%94%81dlm%E5%81%9A%E6%8A%95%E6%9C%BA%E9%87%87%E6%A0%B7/ Fri, 10 Oct 2025 13:11:36 +0800 https://mars160.github.io/posts/diffuspec%E8%A7%A3%E9%94%81dlm%E5%81%9A%E6%8A%95%E6%9C%BA%E9%87%87%E6%A0%B7/ <h2 class="relative group">DiffuSpec：解锁 DLM 做投机采样 <div id="diffuspec解锁-dlm-做投机采样" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#diffuspec%e8%a7%a3%e9%94%81-dlm-%e5%81%9a%e6%8a%95%e6%9c%ba%e9%87%87%e6%a0%b7" aria-label="锚点">#</a> </span> </h2> <p>DIFFUSPEC: UNLOCKING DIFFUSION LANGUAGE MODELS FOR SPECULATIVE DECODING</p> https://mars160.github.io/posts/%E5%90%8E%E7%AB%AF%E9%83%A8%E7%BD%B2%E7%AC%AC%E4%BA%8C%E6%AD%A5%E6%9C%AC%E5%9C%B0%E5%88%B0%E5%85%AC%E7%BD%91%E5%A6%82%E4%BD%95%E4%BD%BF%E7%94%A8-nginx-%E5%8F%91%E5%B8%83-fastapi-%E6%9C%8D%E5%8A%A1/ Fri, 05 Sep 2025 18:40:15 +0800 https://mars160.github.io/posts/%E5%90%8E%E7%AB%AF%E9%83%A8%E7%BD%B2%E7%AC%AC%E4%BA%8C%E6%AD%A5%E6%9C%AC%E5%9C%B0%E5%88%B0%E5%85%AC%E7%BD%91%E5%A6%82%E4%BD%95%E4%BD%BF%E7%94%A8-nginx-%E5%8F%91%E5%B8%83-fastapi-%E6%9C%8D%E5%8A%A1/ <h2 class="relative group">后端部署第二步：本地到公网——如何使用 Nginx 发布 FastAPI 服务 <div id="后端部署第二步本地到公网如何使用-nginx-发布-fastapi-服务" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e5%90%8e%e7%ab%af%e9%83%a8%e7%bd%b2%e7%ac%ac%e4%ba%8c%e6%ad%a5%e6%9c%ac%e5%9c%b0%e5%88%b0%e5%85%ac%e7%bd%91%e5%a6%82%e4%bd%95%e4%bd%bf%e7%94%a8-nginx-%e5%8f%91%e5%b8%83-fastapi-%e6%9c%8d%e5%8a%a1" aria-label="锚点">#</a> </span> </h2> <p>在当今的开发环境中，快速构建和部署后端服务变得至关重要。FastAPI 作为一个高性能、现代化的 Python 异步 Web 框架，广受开发者喜爱。而 Nginx 则是部署 Web 应用最常见也是最稳定的解决方案之一。</p> https://mars160.github.io/posts/fast-dllm%E9%80%9A%E8%BF%87kv-cache%E5%92%8C%E5%B9%B6%E8%A1%8Cdecoding%E5%8A%A0%E9%80%9Fdllm/ Fri, 05 Sep 2025 02:24:01 +0800 https://mars160.github.io/posts/fast-dllm%E9%80%9A%E8%BF%87kv-cache%E5%92%8C%E5%B9%B6%E8%A1%8Cdecoding%E5%8A%A0%E9%80%9Fdllm/ <h2 class="relative group">Fast-dLLM：通过 KV Cache 和并行 Decoding 加速 dLLM <div id="fast-dllm通过-kv-cache-和并行-decoding-加速-dllm" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#fast-dllm%e9%80%9a%e8%bf%87-kv-cache-%e5%92%8c%e5%b9%b6%e8%a1%8c-decoding-%e5%8a%a0%e9%80%9f-dllm" aria-label="锚点">#</a> </span> </h2> <p>Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding</p> https://mars160.github.io/posts/%E5%90%8E%E7%AB%AF%E9%83%A8%E7%BD%B2%E7%AC%AC%E4%B8%80%E6%AD%A5nginx-%E5%8F%91%E5%B8%83%E6%9C%8D%E5%8A%A1%E5%89%8D%E7%9A%84%E4%B8%87%E5%85%A8%E5%87%86%E5%A4%87/ Wed, 03 Sep 2025 02:36:15 +0800 https://mars160.github.io/posts/%E5%90%8E%E7%AB%AF%E9%83%A8%E7%BD%B2%E7%AC%AC%E4%B8%80%E6%AD%A5nginx-%E5%8F%91%E5%B8%83%E6%9C%8D%E5%8A%A1%E5%89%8D%E7%9A%84%E4%B8%87%E5%85%A8%E5%87%86%E5%A4%87/ <h2 class="relative group">后端部署第一步：Nginx 发布服务前的准备工作 <div id="后端部署第一步nginx-发布服务前的准备工作" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e5%90%8e%e7%ab%af%e9%83%a8%e7%bd%b2%e7%ac%ac%e4%b8%80%e6%ad%a5nginx-%e5%8f%91%e5%b8%83%e6%9c%8d%e5%8a%a1%e5%89%8d%e7%9a%84%e5%87%86%e5%a4%87%e5%b7%a5%e4%bd%9c" aria-label="锚点">#</a> </span> </h2> <p>你是不是刚刚用 Java、Python 或 Go 写出了第一个后端程序？它在本地跑得飞快，通过 <code>localhost:8080</code> 就能访问，功能也都挺顺畅。接下来，自然而然会冒出一个念头：<strong>“我要怎样才能让别人也能访问到它？”</strong></p> https://mars160.github.io/posts/%E8%87%AA%E4%BF%A1%E5%9C%B0%E6%B7%B1%E5%BA%A6%E6%80%9D%E8%80%83/ Tue, 02 Sep 2025 07:22:48 +0800 https://mars160.github.io/posts/%E8%87%AA%E4%BF%A1%E5%9C%B0%E6%B7%B1%E5%BA%A6%E6%80%9D%E8%80%83/ <h2 class="relative group">自信地深度思考 <div id="自信地深度思考" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e8%87%aa%e4%bf%a1%e5%9c%b0%e6%b7%b1%e5%ba%a6%e6%80%9d%e8%80%83" aria-label="锚点">#</a> </span> </h2> <p>DEEP THINK WITH CONFIDENCE</p> https://mars160.github.io/posts/trianglemix%E6%97%A0%E6%8D%9F%E4%B8%94%E9%AB%98%E6%95%88%E7%9A%84%E7%94%A8%E4%BA%8E%E9%A2%84%E5%A1%AB%E5%85%85%E9%98%B6%E6%AE%B5%E7%9A%84%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%A8%A1%E5%BC%8F/ Tue, 02 Sep 2025 07:22:45 +0800 https://mars160.github.io/posts/trianglemix%E6%97%A0%E6%8D%9F%E4%B8%94%E9%AB%98%E6%95%88%E7%9A%84%E7%94%A8%E4%BA%8E%E9%A2%84%E5%A1%AB%E5%85%85%E9%98%B6%E6%AE%B5%E7%9A%84%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%A8%A1%E5%BC%8F/ <h2 class="relative group">TriangleMix：无损且高效的用于预填充阶段的注意力模式 <div id="trianglemix无损且高效的用于预填充阶段的注意力模式" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#trianglemix%e6%97%a0%e6%8d%9f%e4%b8%94%e9%ab%98%e6%95%88%e7%9a%84%e7%94%a8%e4%ba%8e%e9%a2%84%e5%a1%ab%e5%85%85%e9%98%b6%e6%ae%b5%e7%9a%84%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%a8%a1%e5%bc%8f" aria-label="锚点">#</a> </span> </h2> <p>TriangleMix: A Lossless and Efficient Attention Pattern for Long Context Prefilling</p> https://mars160.github.io/posts/asccot%E5%8E%8B%E7%BC%A9%E7%9A%84%E6%BF%80%E6%B4%BB%E5%BC%95%E5%AF%BC-training-free/ Tue, 02 Sep 2025 07:22:41 +0800 https://mars160.github.io/posts/asccot%E5%8E%8B%E7%BC%A9%E7%9A%84%E6%BF%80%E6%B4%BB%E5%BC%95%E5%AF%BC-training-free/ <h2 class="relative group">ASC：CoT 压缩的激活引导 Training free <div id="asccot-压缩的激活引导-training-free" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#asccot-%e5%8e%8b%e7%bc%a9%e7%9a%84%e6%bf%80%e6%b4%bb%e5%bc%95%e5%af%bc-training-free" aria-label="锚点">#</a> </span> </h2> <p>Activation Steering for Chain-of-Thought Compression</p> https://mars160.github.io/posts/seal%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E5%8F%AF%E6%93%8D%E6%8E%A7%E6%8E%A8%E7%90%86-traning-free/ Tue, 02 Sep 2025 07:22:36 +0800 https://mars160.github.io/posts/seal%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E5%8F%AF%E6%93%8D%E6%8E%A7%E6%8E%A8%E7%90%86-traning-free/ <h2 class="relative group">SEAL：大语言模型的可操控推理 Traning Free <div id="seal大语言模型的可操控推理-traning-free" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#seal%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%9a%84%e5%8f%af%e6%93%8d%e6%8e%a7%e6%8e%a8%e7%90%86-traning-free" aria-label="锚点">#</a> </span> </h2> <p>SEAL: Steerable Reasoning Calibration of Large Language Models for Free</p> https://mars160.github.io/posts/hmm%E7%AD%89token%E5%BD%B1%E5%93%8D%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86%E8%83%BD%E5%8A%9B/ Tue, 02 Sep 2025 07:22:33 +0800 https://mars160.github.io/posts/hmm%E7%AD%89token%E5%BD%B1%E5%93%8D%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86%E8%83%BD%E5%8A%9B/ <h2 class="relative group">Hmm 等 Token 影响模型推理能力 <div id="hmm-等-token-影响模型推理能力" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#hmm-%e7%ad%89-token-%e5%bd%b1%e5%93%8d%e6%a8%a1%e5%9e%8b%e6%8e%a8%e7%90%86%e8%83%bd%e5%8a%9b" aria-label="锚点">#</a> </span> </h2> <p>Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning</p> https://mars160.github.io/posts/adacot%E9%80%9A%E8%BF%87%E5%BC%BA%E5%8C%96%E5%AD%A6%E4%B9%A0%E5%AE%9E%E7%8E%B0%E7%9A%84%E5%B8%95%E7%B4%AF%E6%89%98%E6%9C%80%E4%BC%98%E8%87%AA%E9%80%82%E5%BA%94%E9%93%BE%E5%BC%8F%E6%80%9D%E7%BB%B4%E8%A7%A6%E5%8F%91%E5%99%A8/ Tue, 02 Sep 2025 07:22:28 +0800 https://mars160.github.io/posts/adacot%E9%80%9A%E8%BF%87%E5%BC%BA%E5%8C%96%E5%AD%A6%E4%B9%A0%E5%AE%9E%E7%8E%B0%E7%9A%84%E5%B8%95%E7%B4%AF%E6%89%98%E6%9C%80%E4%BC%98%E8%87%AA%E9%80%82%E5%BA%94%E9%93%BE%E5%BC%8F%E6%80%9D%E7%BB%B4%E8%A7%A6%E5%8F%91%E5%99%A8/ <h2 class="relative group">AdaCoT：通过强化学习实现的帕累托最优自适应链式思维触发器 <div id="adacot通过强化学习实现的帕累托最优自适应链式思维触发器" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#adacot%e9%80%9a%e8%bf%87%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e5%ae%9e%e7%8e%b0%e7%9a%84%e5%b8%95%e7%b4%af%e6%89%98%e6%9c%80%e4%bc%98%e8%87%aa%e9%80%82%e5%ba%94%e9%93%be%e5%bc%8f%e6%80%9d%e7%bb%b4%e8%a7%a6%e5%8f%91%e5%99%a8" aria-label="锚点">#</a> </span> </h2> <p>AdaCoT: Pareto-Optimal Adaptive Chain-of-Thought Triggering via Reinforcement Learning</p> https://mars160.github.io/posts/adaptthink_-%E8%AE%A9%E6%A8%A1%E5%9E%8B%E5%86%B3%E5%AE%9A%E6%98%AF%E5%90%A6%E6%80%9D%E8%80%83/ Tue, 02 Sep 2025 07:22:25 +0800 https://mars160.github.io/posts/adaptthink_-%E8%AE%A9%E6%A8%A1%E5%9E%8B%E5%86%B3%E5%AE%9A%E6%98%AF%E5%90%A6%E6%80%9D%E8%80%83/ <h2 class="relative group">AdaptThink: 让模型决定是否思考 <div id="adaptthink-让模型决定是否思考" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#adaptthink-%e8%ae%a9%e6%a8%a1%e5%9e%8b%e5%86%b3%e5%ae%9a%e6%98%af%e5%90%a6%e6%80%9d%e8%80%83" aria-label="锚点">#</a> </span> </h2> <p>AdaptThink: Reasoning Models Can Learn When to Think</p> https://mars160.github.io/posts/specreason%E4%BD%BF%E7%94%A8%E6%8E%A8%E6%B5%8B%E6%80%A7%E6%8E%A8%E7%90%86%E5%AE%9E%E7%8E%B0%E5%8A%A0%E9%80%9F%E6%8E%A8%E7%90%86/ Tue, 02 Sep 2025 07:22:21 +0800 https://mars160.github.io/posts/specreason%E4%BD%BF%E7%94%A8%E6%8E%A8%E6%B5%8B%E6%80%A7%E6%8E%A8%E7%90%86%E5%AE%9E%E7%8E%B0%E5%8A%A0%E9%80%9F%E6%8E%A8%E7%90%86/ <h2 class="relative group">SpecReason：使用推测性推理实现加速推理 <div id="specreason使用推测性推理实现加速推理" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#specreason%e4%bd%bf%e7%94%a8%e6%8e%a8%e6%b5%8b%e6%80%a7%e6%8e%a8%e7%90%86%e5%ae%9e%e7%8e%b0%e5%8a%a0%e9%80%9f%e6%8e%a8%e7%90%86" aria-label="锚点">#</a> </span> </h2> <p>SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning</p> https://mars160.github.io/posts/thinkless%E4%B8%80%E7%A7%8D%E6%97%A0%E9%9C%80%E8%AE%AD%E7%BB%83%E7%9A%84%E6%8E%A8%E7%90%86%E9%AB%98%E6%95%88%E6%96%B9%E6%B3%95%E7%94%A8%E4%BA%8E%E5%87%8F%E5%B0%91%E6%8E%A8%E7%90%86%E5%86%97%E4%BD%99/ Tue, 02 Sep 2025 07:22:20 +0800 https://mars160.github.io/posts/thinkless%E4%B8%80%E7%A7%8D%E6%97%A0%E9%9C%80%E8%AE%AD%E7%BB%83%E7%9A%84%E6%8E%A8%E7%90%86%E9%AB%98%E6%95%88%E6%96%B9%E6%B3%95%E7%94%A8%E4%BA%8E%E5%87%8F%E5%B0%91%E6%8E%A8%E7%90%86%E5%86%97%E4%BD%99/ <h2 class="relative group">ThinkLess：一种无需训练的推理高效方法，用于减少推理冗余 <div id="thinkless一种无需训练的推理高效方法用于减少推理冗余" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#thinkless%e4%b8%80%e7%a7%8d%e6%97%a0%e9%9c%80%e8%ae%ad%e7%bb%83%e7%9a%84%e6%8e%a8%e7%90%86%e9%ab%98%e6%95%88%e6%96%b9%e6%b3%95%e7%94%a8%e4%ba%8e%e5%87%8f%e5%b0%91%e6%8e%a8%e7%90%86%e5%86%97%e4%bd%99" aria-label="锚点">#</a> </span> </h2> <p>ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy</p> https://mars160.github.io/posts/thinkless_-llm-learns-when-to-think/ Tue, 02 Sep 2025 07:22:19 +0800 https://mars160.github.io/posts/thinkless_-llm-learns-when-to-think/ <h2 class="relative group">Thinkless: LLM Learns When to Think <div id="thinkless-llm-learns-when-to-think" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#thinkless-llm-learns-when-to-think" aria-label="锚点">#</a> </span> </h2> <p>Thinkless: LLM Learns When to Think</p> https://mars160.github.io/posts/%E6%80%9D%E7%BB%B4%E6%93%8D%E6%8E%A7%E5%A4%96%E9%83%A8cot%E8%BE%85%E5%8A%A9%E5%A4%A7%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86/ Tue, 02 Sep 2025 07:22:16 +0800 https://mars160.github.io/posts/%E6%80%9D%E7%BB%B4%E6%93%8D%E6%8E%A7%E5%A4%96%E9%83%A8cot%E8%BE%85%E5%8A%A9%E5%A4%A7%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86/ <h2 class="relative group">🧠 思维操控：外部 CoT 辅助大模型推理 <div id="-思维操控外部-cot-辅助大模型推理" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#-%e6%80%9d%e7%bb%b4%e6%93%8d%e6%8e%a7%e5%a4%96%e9%83%a8-cot-%e8%be%85%e5%8a%a9%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%8e%a8%e7%90%86" aria-label="锚点">#</a> </span> </h2> <p>Thought Manipulation: External Thought Can Be Efficient for Large Reasoning Models</p> https://mars160.github.io/posts/%E5%BC%80%E5%A7%8B%E4%BD%9C%E7%AD%94%E5%BC%B9%E6%80%A7%E6%8E%A8%E7%90%86%E5%AE%9E%E7%8E%B0%E7%9A%84%E5%8F%AF%E6%89%A9%E5%B1%95%E7%9A%84cot/ Tue, 02 Sep 2025 07:22:15 +0800 https://mars160.github.io/posts/%E5%BC%80%E5%A7%8B%E4%BD%9C%E7%AD%94%E5%BC%B9%E6%80%A7%E6%8E%A8%E7%90%86%E5%AE%9E%E7%8E%B0%E7%9A%84%E5%8F%AF%E6%89%A9%E5%B1%95%E7%9A%84cot/ <h2 class="relative group">⏰ 开始作答：弹性推理实现的可扩展的 CoT <div id="-开始作答弹性推理实现的可扩展的-cot" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#-%e5%bc%80%e5%a7%8b%e4%bd%9c%e7%ad%94%e5%bc%b9%e6%80%a7%e6%8e%a8%e7%90%86%e5%ae%9e%e7%8e%b0%e7%9a%84%e5%8f%af%e6%89%a9%e5%b1%95%e7%9a%84-cot" aria-label="锚点">#</a> </span> </h2> <p>Scalable Chain of Thoughts via Elastic Reasoning</p> https://mars160.github.io/posts/deer%E5%9F%BA%E4%BA%8Etrial%E7%BD%AE%E4%BF%A1%E5%BA%A6%E7%9A%84%E6%8E%A8%E7%90%86%E6%97%A9%E5%81%9C/ Tue, 02 Sep 2025 07:22:12 +0800 https://mars160.github.io/posts/deer%E5%9F%BA%E4%BA%8Etrial%E7%BD%AE%E4%BF%A1%E5%BA%A6%E7%9A%84%E6%8E%A8%E7%90%86%E6%97%A9%E5%81%9C/ <h2 class="relative group">DEER：基于 Trial 置信度的推理早停 <div id="deer基于-trial-置信度的推理早停" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#deer%e5%9f%ba%e4%ba%8e-trial-%e7%bd%ae%e4%bf%a1%e5%ba%a6%e7%9a%84%e6%8e%a8%e7%90%86%e6%97%a9%e5%81%9c" aria-label="锚点">#</a> </span> </h2> <p>DYNAMIC EARLY EXIT IN REASONING MODELS</p> https://mars160.github.io/posts/softcotprompt%EF%B8%8Fslm%EF%B8%8Fllm/ Tue, 02 Sep 2025 07:22:09 +0800 https://mars160.github.io/posts/softcotprompt%EF%B8%8Fslm%EF%B8%8Fllm/ <h2 class="relative group">SoftCoT：Prompt➡️SLM➡️LLM <div id="softcotpromptslmllm" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#softcotpromptslmllm" aria-label="锚点">#</a> </span> </h2> <p>SoftCoT: Soft Chain-of-Thought for Efficient Reasoning with LLMs</p> https://mars160.github.io/posts/codi_-%E9%80%9A%E8%BF%87%E8%87%AA%E8%92%B8%E9%A6%8F%E5%B0%86cot%E5%8E%8B%E7%BC%A9%E5%88%B0%E8%BF%9E%E7%BB%AD%E7%A9%BA%E9%97%B4%E4%B8%AD/ Tue, 02 Sep 2025 07:22:06 +0800 https://mars160.github.io/posts/codi_-%E9%80%9A%E8%BF%87%E8%87%AA%E8%92%B8%E9%A6%8F%E5%B0%86cot%E5%8E%8B%E7%BC%A9%E5%88%B0%E8%BF%9E%E7%BB%AD%E7%A9%BA%E9%97%B4%E4%B8%AD/ <h2 class="relative group">CODI: 通过自蒸馏将 CoT 压缩到连续空间中 <div id="codi-通过自蒸馏将-cot-压缩到连续空间中" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#codi-%e9%80%9a%e8%bf%87%e8%87%aa%e8%92%b8%e9%a6%8f%e5%b0%86-cot-%e5%8e%8b%e7%bc%a9%e5%88%b0%e8%bf%9e%e7%bb%ad%e7%a9%ba%e9%97%b4%e4%b8%ad" aria-label="锚点">#</a> </span> </h2> <p>CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation</p> https://mars160.github.io/posts/hawkeye%E5%A4%A7%E5%B0%8F%E6%A8%A1%E5%9E%8B%E5%8D%8F%E4%BD%9C%E5%AE%9E%E7%8E%B0%E7%B2%BE%E7%AE%80cot/ Tue, 02 Sep 2025 07:22:04 +0800 https://mars160.github.io/posts/hawkeye%E5%A4%A7%E5%B0%8F%E6%A8%A1%E5%9E%8B%E5%8D%8F%E4%BD%9C%E5%AE%9E%E7%8E%B0%E7%B2%BE%E7%AE%80cot/ <h2 class="relative group">HAWKEYE：大小模型协作实现精简 CoT <div id="hawkeye大小模型协作实现精简-cot" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#hawkeye%e5%a4%a7%e5%b0%8f%e6%a8%a1%e5%9e%8b%e5%8d%8f%e4%bd%9c%e5%ae%9e%e7%8e%b0%e7%b2%be%e7%ae%80-cot" aria-label="锚点">#</a> </span> </h2> <p>Hawkeye:Efficient Reasoning with Model Collaboration</p> https://mars160.github.io/posts/lcirc_-%E9%95%BF%E6%96%87%E6%A1%A3%E5%BE%AA%E7%8E%AF%E5%8E%8B%E7%BC%A9%E6%96%B9%E6%B3%95-llm-training-free/ Tue, 02 Sep 2025 07:22:00 +0800 https://mars160.github.io/posts/lcirc_-%E9%95%BF%E6%96%87%E6%A1%A3%E5%BE%AA%E7%8E%AF%E5%8E%8B%E7%BC%A9%E6%96%B9%E6%B3%95-llm-training-free/ <h2 class="relative group">LCIRC: 长文档循环压缩方法 （LLM Training Free） <div id="lcirc-长文档循环压缩方法-llm-training-free" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#lcirc-%e9%95%bf%e6%96%87%e6%a1%a3%e5%be%aa%e7%8e%af%e5%8e%8b%e7%bc%a9%e6%96%b9%e6%b3%95-llm-training-free" aria-label="锚点">#</a> </span> </h2> <p>LCIRC: A Recurrent Compression Approach for Efficient Long-form Context and Query Dependent Modeling in LLMs</p> https://mars160.github.io/posts/llm%E5%9C%A8%E8%BF%9E%E7%BB%ADlatent%E7%A9%BA%E9%97%B4%E4%B8%AD%E6%8E%A8%E7%90%86/ Tue, 02 Sep 2025 07:21:57 +0800 https://mars160.github.io/posts/llm%E5%9C%A8%E8%BF%9E%E7%BB%ADlatent%E7%A9%BA%E9%97%B4%E4%B8%AD%E6%8E%A8%E7%90%86/ <h2 class="relative group">LLM 在连续 Latent 空间中推理 <div id="llm-在连续-latent-空间中推理" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#llm-%e5%9c%a8%e8%bf%9e%e7%bb%ad-latent-%e7%a9%ba%e9%97%b4%e4%b8%ad%e6%8e%a8%e7%90%86" aria-label="锚点">#</a> </span> </h2> <p>Training Large Language Models to Reason in a Continuous Latent Space</p> https://mars160.github.io/posts/pause-token%E9%9A%90%E5%BC%8Fcot/ Tue, 02 Sep 2025 07:21:55 +0800 https://mars160.github.io/posts/pause-token%E9%9A%90%E5%BC%8Fcot/ <h2 class="relative group">Pause Token：隐式 CoT <div id="pause-token隐式-cot" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#pause-token%e9%9a%90%e5%bc%8f-cot" aria-label="锚点">#</a> </span> </h2> <p>Think before you speak: Training Language Models With Pause Tokens</p> https://mars160.github.io/posts/%E9%9D%A2%E5%90%91%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86%E6%80%9D%E8%80%83%E4%BC%98%E5%8C%96%E7%9A%84test-time-scaling/ Tue, 02 Sep 2025 07:21:53 +0800 https://mars160.github.io/posts/%E9%9D%A2%E5%90%91%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86%E6%80%9D%E8%80%83%E4%BC%98%E5%8C%96%E7%9A%84test-time-scaling/ <h2 class="relative group">面向模型推理思考优化的 Test time scaling <div id="面向模型推理思考优化的-test-time-scaling" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e9%9d%a2%e5%90%91%e6%a8%a1%e5%9e%8b%e6%8e%a8%e7%90%86%e6%80%9d%e8%80%83%e4%bc%98%e5%8c%96%e7%9a%84-test-time-scaling" aria-label="锚点">#</a> </span> </h2> <p>Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning</p> https://mars160.github.io/posts/ccot%E9%80%9A%E8%BF%87%E5%AF%86%E9%9B%86%E8%A1%A8%E7%A4%BA%E5%AE%9E%E7%8E%B0%E9%AB%98%E6%95%88%E6%8E%A8%E7%90%86/ Tue, 02 Sep 2025 07:21:49 +0800 https://mars160.github.io/posts/ccot%E9%80%9A%E8%BF%87%E5%AF%86%E9%9B%86%E8%A1%A8%E7%A4%BA%E5%AE%9E%E7%8E%B0%E9%AB%98%E6%95%88%E6%8E%A8%E7%90%86/ <h2 class="relative group">CCoT：通过密集表示实现高效推理 <div id="ccot通过密集表示实现高效推理" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#ccot%e9%80%9a%e8%bf%87%e5%af%86%e9%9b%86%e8%a1%a8%e7%a4%ba%e5%ae%9e%e7%8e%b0%e9%ab%98%e6%95%88%e6%8e%a8%e7%90%86" aria-label="锚点">#</a> </span> </h2> <p>Compressed Chain of Thought: Efficient Reasoning Through Dense Representations</p> https://mars160.github.io/posts/token%E9%A2%84%E7%AE%97%E6%84%8F%E8%AF%86%E7%9A%84llm%E6%8E%A8%E7%90%86/ Tue, 02 Sep 2025 07:21:47 +0800 https://mars160.github.io/posts/token%E9%A2%84%E7%AE%97%E6%84%8F%E8%AF%86%E7%9A%84llm%E6%8E%A8%E7%90%86/ <h2 class="relative group">Token 预算意识的 llm 推理 <div id="token-预算意识的-llm-推理" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#token-%e9%a2%84%e7%ae%97%e6%84%8f%e8%af%86%e7%9a%84-llm-%e6%8e%a8%e7%90%86" aria-label="锚点">#</a> </span> </h2> <p>Token-Budget-Aware LLM Reasoning</p> https://mars160.github.io/posts/inftythink%E6%89%93%E7%A0%B4%E5%A4%A7%E5%9E%8B%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E9%95%BF%E4%B8%8A%E4%B8%8B%E6%96%87%E6%8E%A8%E7%90%86%E9%95%BF%E5%BA%A6%E9%99%90%E5%88%B6/ Tue, 02 Sep 2025 07:21:43 +0800 https://mars160.github.io/posts/inftythink%E6%89%93%E7%A0%B4%E5%A4%A7%E5%9E%8B%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E9%95%BF%E4%B8%8A%E4%B8%8B%E6%96%87%E6%8E%A8%E7%90%86%E9%95%BF%E5%BA%A6%E9%99%90%E5%88%B6/ <h2 class="relative group">INFTYTHINK：打破大型语言模型长上下文推理长度限制 <div id="inftythink打破大型语言模型长上下文推理长度限制" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#inftythink%e6%89%93%e7%a0%b4%e5%a4%a7%e5%9e%8b%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e9%95%bf%e4%b8%8a%e4%b8%8b%e6%96%87%e6%8e%a8%e7%90%86%e9%95%bf%e5%ba%a6%e9%99%90%e5%88%b6" aria-label="锚点">#</a> </span> </h2> <p>InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models</p> https://mars160.github.io/posts/tokenskip_%E5%8F%AF%E6%8E%A7%E7%9A%84cot%E5%8E%8B%E7%BC%A9-in-llms/ Tue, 02 Sep 2025 07:21:38 +0800 https://mars160.github.io/posts/tokenskip_%E5%8F%AF%E6%8E%A7%E7%9A%84cot%E5%8E%8B%E7%BC%A9-in-llms/ <h2 class="relative group">TokenSkip:可控的 CoT 压缩 in LLMs <div id="tokenskip可控的-cot-压缩-in-llms" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#tokenskip%e5%8f%af%e6%8e%a7%e7%9a%84-cot-%e5%8e%8b%e7%bc%a9-in-llms" aria-label="锚点">#</a> </span> </h2> <p>TokenSkip: Controllable Chain-of-Thought Compression in LLMs</p> https://mars160.github.io/posts/lightthinker_-%E6%AF%8F%E4%B8%AA%E6%83%B3%E6%B3%95%E5%8E%8B%E7%BC%A9%E6%88%90%E4%B8%A4%E4%B8%AAtoken/ Tue, 02 Sep 2025 07:21:35 +0800 https://mars160.github.io/posts/lightthinker_-%E6%AF%8F%E4%B8%AA%E6%83%B3%E6%B3%95%E5%8E%8B%E7%BC%A9%E6%88%90%E4%B8%A4%E4%B8%AAtoken/ <h2 class="relative group">LightThinker: 每个想法压缩成两个 token <div id="lightthinker-每个想法压缩成两个-token" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#lightthinker-%e6%af%8f%e4%b8%aa%e6%83%b3%e6%b3%95%e5%8e%8b%e7%bc%a9%e6%88%90%e4%b8%a4%e4%b8%aa-token" aria-label="锚点">#</a> </span> </h2> <p>如果 lt 确实靠谱</p>