PageSourceSearch

https://facerain.me/nonmun-ribyu-search-r1-training-llms-to-reason…search-engines-with-reinforcement-learning/

html facerain.me collected 2026-10-04 08:09:55 UTC 46,597 bytes, 555 lines download raw bytes

1<!DOCTYPE html>
2<html lang="ko">
3<head>
4
5    <title>[논문 리뷰] Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning</title>
6    <meta charset="utf-8" />
7    <meta http-equiv="X-UA-Compatible" content="IE=edge" />
8    <meta name="HandheldFriendly" content="True" />
9    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
10
11    <link rel="preload" as="style" href="/assets/built/screen.css?v=4f24c6eeff" />
12    <link rel="preload" as="script" href="/assets/built/casper.js?v=4f24c6eeff" />
13
14    <link rel="stylesheet" type="text/css" href="/assets/built/screen.css?v=4f24c6eeff" />
15
16    <link rel="canonical" href="https://facerain.me/nonmun-ribyu-search-r1-training-llms-to-reason-and-leverage-search-engines-with-reinforcement-learning/">
17    <meta name="referrer" content="no-referrer-when-downgrade">
18    
19    <meta property="og:site_name" content="FacerAin&#x27;s blog">
20    <meta property="og:type" content="article">
21    <meta property="og:title" content="[논문 리뷰] Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning">
22    <meta property="og:description" content="들어가며
23
24이번 시간에는 LLM이 검색 엔진과 상호작용하며 추론(Reasoning)을 수행할 수 있는 강화 학습 프레임워크 Search-R1을 소개합니다. 최근 OpenAI의 Deep Research나 여러 최신 연구에서 알 수 있듯, LLM의 추론 능력뿐 아니라 실시간 검색과 결합된 Reasoning이 큰 주목을 받고 있습니다.
25하지만 기존의 RAG(Retrieval-Augmented Generation)이나 Tool-Use">
26    <meta property="og:url" content="https://facerain.me/nonmun-ribyu-search-r1-training-llms-to-reason-and-leverage-search-engines-with-reinforcement-learning/">
27    <meta property="og:image" content="https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&amp;ixlib&#x3D;rb-4.0.3&amp;q&#x3D;80&amp;w&#x3D;2000">
28    <meta property="article:published_time" content="2025-03-25T11:54:44.000Z">
29    <meta property="article:modified_time" content="2026-07-07T12:50:11.000Z">
30    <meta property="article:tag" content="Research">
31    
32    <meta property="article:publisher" content="https://www.facebook.com/ghost">
33    <meta name="twitter:card" content="summary_large_image">
34    <meta name="twitter:title" content="[논문 리뷰] Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning">
35    <meta name="twitter:description" content="들어가며
36
37이번 시간에는 LLM이 검색 엔진과 상호작용하며 추론(Reasoning)을 수행할 수 있는 강화 학습 프레임워크 Search-R1을 소개합니다. 최근 OpenAI의 Deep Research나 여러 최신 연구에서 알 수 있듯, LLM의 추론 능력뿐 아니라 실시간 검색과 결합된 Reasoning이 큰 주목을 받고 있습니다.
38하지만 기존의 RAG(Retrieval-Augmented Generation)이나 Tool-Use">
39    <meta name="twitter:url" content="https://facerain.me/nonmun-ribyu-search-r1-training-llms-to-reason-and-leverage-search-engines-with-reinforcement-learning/">
40    <meta name="twitter:image" content="https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&amp;ixlib&#x3D;rb-4.0.3&amp;q&#x3D;80&amp;w&#x3D;2000">
41    <meta name="twitter:label1" content="Written by">
42    <meta name="twitter:data1" content="Yongwoo Song">
43    <meta name="twitter:label2" content="Filed under">
44    <meta name="twitter:data2" content="Research">
45    <meta name="twitter:site" content="@ghost">
46    <meta property="og:image:width" content="1200">
47    <meta property="og:image:height" content="1799">
48    
49    
49<script type="application/ld+json">
50{
51    "@context": "https://schema.org",
52    "@type": "Article",
53    "publisher": {
54        "@type": "Organization",
55        "name": "FacerAin&#x27;s blog",
56        "url": "https://facerain.me/",
57        "logo": {
58            "@type": "ImageObject",
59            "url": "https://facerain.me/favicon.ico",
60            "width": 48,
61            "height": 48
62        }
63    },
64    "author": {
65        "@type": "Person",
66        "name": "Yongwoo Song",
67        "url": "https://facerain.me/author/yongwoo/",
68        "sameAs": []
69    },
70    "headline": "[논문 리뷰] Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning",
71    "url": "https://facerain.me/nonmun-ribyu-search-r1-training-llms-to-reason-and-leverage-search-engines-with-reinforcement-learning/",
72    "datePublished": "2025-03-25T11:54:44.000Z",
73    "dateModified": "2026-07-07T12:50:11.000Z",
74    "image": {
75        "@type": "ImageObject",
76        "url": "https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop=entropy&cs=tinysrgb&fit=max&fm=jpg&ixid=M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&ixlib=rb-4.0.3&q=80&w=2000",
77        "width": 1200,
78        "height": 1799
79    },
80    "keywords": "Research",
81    "description": "들어가며\n\n이번 시간에는 LLM이 검색 엔진과 상호작용하며 추론(Reasoning)을 수행할 수 있는 강화 학습 프레임워크 Search-R1을 소개합니다. 최근 OpenAI의 Deep Research나 여러 최신 연구에서 알 수 있듯, LLM의 추론 능력뿐 아니라 실시간 검색과 결합된 Reasoning이 큰 주목을 받고 있습니다.\n하지만 기존의 RAG(Retrieval-Augmented Generation)이나 Tool-Use 방식은\n\n * 복잡한 다단계 검색에 유연하게 대응하기 어렵거나\n * 대규모 고품질 지도 학습 데이터가 필요하다는 한계점이 있습니다.\n\n이런 문제를 해결하기 위해, 이번 연구에서는 LLM이 스스로 검색 쿼리를 생성하며 Reasoning을 배우는 RL 기반 프레임워크, 바로 Search-R1을 제안했습니다.\n\nSearch-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learn",
82    "mainEntityOfPage": "https://facerain.me/nonmun-ribyu-search-r1-training-llms-to-reason-and-leverage-search-engines-with-reinforcement-learning/"
83}
84    </script>
84
85
86    <meta name="generator" content="Ghost 6.45">
87    <link rel="alternate" type="application/rss+xml" title="FacerAin&#x27;s blog" href="https://facerain.me/rss/">
88    
88<script defer src="https://cdn.jsdelivr.net/ghost/portal@~2.69/umd/portal.min.js" data-i18n="true" data-ghost="https://facerain.me/" data-key="3fb8b569937df4745424d79246" data-api="https://facerain.me/ghost/api/content/" data-locale="ko" crossorigin="anonymous"></script>
88<style id="gh-members-styles">.gh-post-upgrade-cta-content,
89.gh-post-upgrade-cta {
90    display: flex;
91    flex-direction: column;
92    align-items: center;
93    font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, Cantarell, 'Open Sans', 'Helvetica Neue', sans-serif;
94    text-align: center;
95    width: 100%;
96    color: #ffffff;
97    font-size: 16px;
98}
99
100.gh-post-upgrade-cta-content {
101    border-radius: 8px;
102    padding: 40px 4vw;
103}
104
105.gh-post-upgrade-cta h2 {
106    color: #ffffff;
107    font-size: 28px;
108    letter-spacing: -0.2px;
109    margin: 0;
110    padding: 0;
111}
112
113.gh-post-upgrade-cta p {
114    margin: 20px 0 0;
115    padding: 0;
116}
117
118.gh-post-upgrade-cta small {
119    font-size: 16px;
120    letter-spacing: -0.2px;
121}
122
123.gh-post-upgrade-cta a {
124    color: #ffffff;
125    cursor: pointer;
126    font-weight: 500;
127    box-shadow: none;
128    text-decoration: underline;
129}
130
131.gh-post-upgrade-cta a:hover {
132    color: #ffffff;
133    opacity: 0.8;
134    box-shadow: none;
135    text-decoration: underline;
136}
137
138.gh-post-upgrade-cta a.gh-btn {
139    display: block;
140    background: #ffffff;
141    text-decoration: none;
142    margin: 28px 0 0;
143    padding: 8px 18px;
144    border-radius: 4px;
145    font-size: 16px;
146    font-weight: 600;
147}
148
149.gh-post-upgrade-cta a.gh-btn:hover {
150    opacity: 0.92;
151}</style>
152    
152<script defer src="https://cdn.jsdelivr.net/ghost/sodo-search@~1.8/umd/sodo-search.min.js" data-key="3fb8b569937df4745424d79246" data-styles="https://cdn.jsdelivr.net/ghost/sodo-search@~1.8/umd/main.css" data-sodo-search="https://facerain.me/" data-locale="ko" crossorigin="anonymous"></script>
152
153    
154    <link href="https://facerain.me/webmentions/receive/" rel="webmention">
155    
155<script defer src="/public/cards.min.js?v=4f24c6eeff"></script>
155
156    <link rel="stylesheet" type="text/css" href="/public/cards.min.css?v=4f24c6eeff">
157    
157<script defer src="/public/comment-counts.min.js?v=4f24c6eeff" data-ghost-comments-counts-api="https://facerain.me/members/api/comments/counts/"></script>
157
158    
158<script defer src="/public/member-attribution.min.js?v=4f24c6eeff"></script>
158<style>:root {--ghost-accent-color: #000000;}</style>
159    <meta name="google-site-verification" content="_RKINYc3HLIqW1Nraw2D_rmdlnH8bz4itY8YZ5pysIA" />
160<!-- Google tag (gtag.js) -->
vendor: 64 bytes, lines 160-161
160
161<script async src="https://www.googletagmanager.com/gtag/js?id=
161G-1W6H6D1TCR
vendor: 12 bytes, line 161
161"></script>
162<script>
163  
vendor: 134 bytes, lines 163-167
163window.dataLayer = window.dataLayer || [];
164  function gtag(){dataLayer.push(arguments);}
165  gtag('js', new Date());
166
167  gtag('config', '
167G-1W6H6D1TCR
vendor: 4 bytes, line 167
167');
168</script>
168
169
170<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.css" crossorigin="anonymous">
171<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.js" crossorigin="anonymous"></script>
vendor: 1 bytes, line 171
171
172<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/contrib/auto-render.min.js" crossorigin="anonymous" onload="renderMathInElement(document.body);"></script>
172
173
174<link href='https://rsms.me/inter/inter.css' rel='stylesheet' type='text/css'>
175<link href='https://spoqa.github.io/spoqa-han-sans/css/SpoqaHanSans-kr.css' rel='stylesheet' type='text/css'>
176<style>
177body,  
178h1, h2, h3, h4, h5, h6,
179.main-nav a,
180.subscribe-button,
181.page-title,
182.post-meta,
183.read-next-story .post:before,
184.pagination,
185.site-footer,
186.post-full-content,
187.post-card-excerpt,
188.post-full-custom-excerpt,
189[class^="icon-"]:before, [class*=" icon-"]:before
190{
191font-family:
192  "-apple-system", "BlinkMacSystemFont","Apple SD Gothic Neo",
193  "Inter", "Spoqa Han Sans", "Segoe UI", Sans-Serif,
194  "Apple Color Emoji", "Segoe UI Emoji", "Segoe UI Symbol";
195}
196</style>
197
198<script type="text/x-mathjax-config">
199MathJax.Hub.Config({
200  jax: ["input/TeX", "output/SVG"],
201  tex2jax: {inlineMath: [['$$', '$$']],displayMath : [['$$$$','$$$$']]},
202});
203</script>
vendor: 1 bytes, line 203
203
204<script src="//cdn.bootcss.com/mathjax/2.6.1/MathJax.js?config=TeX-AMS_HTML"\></script>
204
205    <link rel="preconnect" href="https://fonts.bunny.net"><link rel="stylesheet" href="https://fonts.bunny.net/css?family=noto-sans:400,700"><style>:root {--gh-font-heading: Noto Sans;--gh-font-body: Noto Sans;}</style>
206
207</head>
208<body class="post-template tag-research gh-font-heading-noto-sans gh-font-body-noto-sans is-head-middle-logo has-cover">
209<div class="viewport">
210
211    <header id="gh-head" class="gh-head outer">
212        <div class="gh-head-inner inner">
213            <div class="gh-head-brand">
214                <a class="gh-head-logo no-image" href="https://facerain.me">
215                        FacerAin&#x27;s blog
216                </a>
217                <button class="gh-search gh-icon-btn" aria-label="Search this site" data-ghost-search><svg xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 24 24" stroke="currentColor" stroke-width="2" width="20" height="20"><path stroke-linecap="round" stroke-linejoin="round" d="M21 21l-6-6m2-5a7 7 0 11-14 0 7 7 0 0114 0z"></path></svg></button>
218                <button class="gh-burger" aria-label="Main Menu"></button>
219            </div>
220
221            <nav class="gh-head-menu">
222                <ul class="nav">
223    <li class="nav-home"><a href="https://facerain.me/">Home</a></li>
224    <li class="nav-engineering"><a href="https://facerain.me/tag/engineering/">Engineering</a></li>
225    <li class="nav-research"><a href="https://facerain.me/tag/research/">Research</a></li>
226    <li class="nav-life"><a href="https://facerain.me/tag/life/">Life</a></li>
227    <li class="nav-github"><a href="https://github.com/FacerAin">Github</a></li>
228</ul>
229
230            </nav>
231
232            <div class="gh-head-actions">
233                    <button class="gh-search gh-icon-btn" aria-label="Search this site" data-ghost-search><svg xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 24 24" stroke="currentColor" stroke-width="2" width="20" height="20"><path stroke-linecap="round" stroke-linejoin="round" d="M21 21l-6-6m2-5a7 7 0 11-14 0 7 7 0 0114 0z"></path></svg></button>
234                    <div class="gh-head-members">
235                                <a class="gh-head-link" href="#/portal/signin" data-portal="signin">Sign in</a>
236                                <a class="gh-head-button" href="#/portal/signup" data-portal="signup">Subscribe</a>
237                    </div>
238            </div>
239        </div>
240    </header>
241
242    <div class="site-content">
243        
244
245
246
247<main id="site-main" class="site-main">
248<article class="article post tag-research ">
249
250    <header class="article-header gh-canvas">
251
252        <div class="article-tag post-card-tags">
253                <span class="post-card-primary-tag">
254                    <a href="/tag/research/">Research</a>
255                </span>
256        </div>
257
258        <h1 class="article-title">[논문 리뷰] Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning</h1>
259
260
261        <div class="article-byline">
262        <section class="article-byline-content">
263
264            <ul class="author-list instapaper_ignore">
265                <li class="author-list-item">
266                    <a href="/author/yongwoo/" class="author-avatar author-profile-image" aria-label="Read more of Yongwoo Song"><svg viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><g fill="none" fill-rule="evenodd"><path d="M3.513 18.998C4.749 15.504 8.082 13 12 13s7.251 2.504 8.487 5.998C18.47 21.442 15.417 23 12 23s-6.47-1.558-8.487-4.002zM12 12c2.21 0 4-2.79 4-5s-1.79-4-4-4-4 1.79-4 4 1.79 5 4 5z" fill="#FFF"/></g></svg>
267</a>
268                </li>
269            </ul>
270
271            <div class="article-byline-meta">
272                <h4 class="author-name"><a href="/author/yongwoo/">Yongwoo Song</a></h4>
273                <div class="byline-meta-content">
274                    <time class="byline-meta-date" datetime="2025-03-25">25 3월 2025</time>
275                        <span class="byline-reading-time"><span class="bull">&bull;</span> 11 min read</span>
276                </div>
277            </div>
278
279        </section>
280        <a href="#/share" class="gh-button gh-button-share">Share</a>
281        </div>
282
283            <figure class="article-image">
284                <img
285                    srcset="https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&amp;ixlib&#x3D;rb-4.0.3&amp;q&#x3D;80&amp;w&#x3D;300 300w,
286                            https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&amp;ixlib&#x3D;rb-4.0.3&amp;q&#x3D;80&amp;w&#x3D;600 600w,
287                            https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&amp;ixlib&#x3D;rb-4.0.3&amp;q&#x3D;80&amp;w&#x3D;1000 1000w,
288                            https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&amp;ixlib&#x3D;rb-4.0.3&amp;q&#x3D;80&amp;w&#x3D;2000 2000w"
289                    sizes="(min-width: 1400px) 1400px, 92vw"
290                    src="https://images.unsplash.com/photo-1742325989789-b42912a531dd?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDN8fHx8fHx8fDE3NDI4ODc1Nzl8&amp;ixlib&#x3D;rb-4.0.3&amp;q&#x3D;80&amp;w&#x3D;2000"
291                    alt="[논문 리뷰] Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning"
292                    fetchpriority="high"
293                    loading="eager"
294                />
295                    <figcaption><span style="white-space: pre-wrap;">Photo by </span><a href="https://unsplash.com/@steelhouseproductions?utm_source=ghost&amp;utm_medium=referral&amp;utm_campaign=api-credit"><span style="white-space: pre-wrap;">Tom Nagel</span></a><span style="white-space: pre-wrap;"> / </span><a href="https://unsplash.com/?utm_source=ghost&amp;utm_medium=referral&amp;utm_campaign=api-credit"><span style="white-space: pre-wrap;">Unsplash</span></a></figcaption>
296            </figure>
297
298    </header>
299
300    <section class="gh-content gh-canvas">
301        <h3 id="%EB%93%A4%EC%96%B4%EA%B0%80%EB%A9%B0">들어가며</h3><p>이번 시간에는 LLM이 검색 엔진과 상호작용하며 추론(Reasoning)을 수행할 수 있는 강화 학습 프레임워크 <strong>Search-R1</strong>을 소개합니다. 최근 OpenAI의 Deep Research나 여러 최신 연구에서 알 수 있듯, LLM의 추론 능력뿐 아니라 <strong>실시간 검색과 결합된 Reasoning</strong>이 큰 주목을 받고 있습니다.<br>하지만 기존의 <strong>RAG(Retrieval-Augmented Generation)</strong>이나 <strong>Tool-Use</strong> 방식은</p><ul><li>복잡한 다단계 검색에 유연하게 대응하기 어렵거나</li><li>대규모 고품질 지도 학습 데이터가 필요하다는 한계점이 있습니다.</li></ul><p>이런 문제를 해결하기 위해, 이번 연구에서는 <strong>LLM이 스스로 검색 쿼리를 생성하며 Reasoning을 배우는 RL 기반 프레임워크</strong>, 바로 <strong>Search-R1</strong>을 제안했습니다.</p><figure class="kg-card kg-bookmark-card"><a class="kg-bookmark-container" href="https://arxiv.org/abs/2503.09516?ref=facerain.me"><div class="kg-bookmark-content"><div class="kg-bookmark-title">Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning</div><div class="kg-bookmark-description">Efficiently acquiring external knowledge and up-to-date information is essential for effective reasoning and text generation in large language models (LLMs). Prompting advanced LLMs with reasoning capabilities during inference to use search engines is not optimal, since the LLM does not learn how to optimally interact with the search engine. This paper introduces Search-R1, an extension of the DeepSeek-R1 model where the LLM learns -- solely through reinforcement learning (RL) -- to autonomously generate (multiple) search queries during step-by-step reasoning with real-time retrieval. Search-R1 optimizes LLM rollouts with multi-turn search interactions, leveraging retrieved token masking for stable RL training and a simple outcome-based reward function. Experiments on seven question-answering datasets show that Search-R1 improves performance by 26% (Qwen2.5-7B), 21% (Qwen2.5-3B), and 10% (LLaMA3.2-3B) over strong baselines. This paper further provides empirical insights into RL optimization methods, LLM choices, and response length dynamics in retrieval-augmented reasoning. The code and model checkpoints are available at https://github.com/PeterGriffinJin/Search-R1.</div><div class="kg-bookmark-metadata"><img class="kg-bookmark-icon" src="https://arxiv.org/static/browse/0.3.4/images/icons/apple-touch-icon.png" alt=""><span class="kg-bookmark-author">arXiv.org</span><span class="kg-bookmark-publisher">Bowen Jin</span></div></div><div class="kg-bookmark-thumbnail"><img src="https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png" alt=""></div></a></figure><figure class="kg-card kg-bookmark-card"><a class="kg-bookmark-container" href="https://github.com/PeterGriffinJin/Search-R1?ref=facerain.me"><div class="kg-bookmark-content"><div class="kg-bookmark-title">GitHub - PeterGriffinJin/Search-R1: Search-R1: An Efficient, Scalable RL Training Framework for Reasoning &amp; Search Engine Calling interleaved LLM based on veRL</div><div class="kg-bookmark-description">Search-R1: An Efficient, Scalable RL Training Framework for Reasoning &amp; Search Engine Calling interleaved LLM based on veRL - PeterGriffinJin/Search-R1</div><div class="kg-bookmark-metadata"><img class="kg-bookmark-icon" src="https://github.githubassets.com/assets/pinned-octocat-093da3e6fa40.svg" alt=""><span class="kg-bookmark-author">GitHub</span><span class="kg-bookmark-publisher">PeterGriffinJin</span></div></div>
301<div class="kg-bookmark-thumbnail"><img src="https://opengraph.githubassets.com/e06fce9804724dd3173758bc21e7213f33c70fd1876c67c072e0c9af1b30604f/PeterGriffinJin/Search-R1" alt=""></div></a></figure><h3 id="abstract-conclusion">Abstract &amp; Conclusion</h3><ul><li>기존 방법론(RAG, Tool-Use)은 <strong>LLM이 검색 엔진과 상호작용하는 법</strong>을 배우지 못함.</li><li>Search-R1은 <strong>강화 학습 기반으로 LLM이 다단계로 검색 쿼리를 생성하면서 Reasoning</strong>을 수행하도록 훈련.</li><li>7개의 QA 데이터셋 평가에서 큰 성능 향상:Qwen2.5-7B: <strong>26%</strong>Qwen2.5-3B: <strong>21%</strong>LLaMA3.2-3B: <strong>10%</strong></li></ul><h3 id="introduction">Introduction</h3><p>LLM이 외부 지식을 활용하기 위해서는 RAG와 Tool-Use 크게 두 가지 방법이 있습니다. 하지만 각 방법은 아래와 같은 한계가 있습니다.</p><p><strong>RAG (Retrieval-Augmented Generation)</strong> - 입력 쿼리를 바탕으로 관련 Passage를 검색하여 Context에 함께 LLM의 Input으로 제공</p><ul><li>검색된 정보가 관련성이 없을 수도 있음</li><li>Multi-hop retrieval 지원하지 않음</li><li>LLM이 Search Engine과 상호작용 하는 방법을 배우지 못함</li></ul><p><strong>Search Engine Tool-Use - </strong>Search Engine을 Tool로 활용하도록 학습</p><ul><li>일반화의 어려움</li><li>확장성 문제: 훈련 기반 방식은 높은 품질의 대규모 annotated trajectories가 필요함</li></ul><p></p><p>따라서 저자들은 이러한 문제를 해결하기 위해 강화학습 기반 방법론인 S<strong>earch-R1을 제안</strong>합니다.</p><ul><li>학습 시 Search Engine과의 상호작용을 학습의 일부로 포함</li><li>Multi-Turn Interleaved Reasoning and Search</li><li>Reward Design</li></ul><h3 id="methodology">Methodology</h3><p>그럼 연구팀은 RL을 바탕으로 어떻게 Search Engine과의 상호작용을 학습 시킬 수 있었을까요?</p><p>Search-R1의 Objective는 다음과 같습니다.</p><p>$$\max_{\pi} E_{x \sim D, y \sim \pi_{\theta}(\cdot | x; R)} \left[ r \phi(x, y)\right] - \beta D_{KL} \left[ \pi_{\theta}(y | x; R) | \pi_{ref}(y | x; R) \right]$$</p><p>일반적인 RL Objective와 크게 다르지 않습니다.</p><p>$$\pi_\theta(\cdot | x; R)$$</p><p>R은 Search Engine을 의미합니다. Policy LLM은 R (Search Engine)과 함께 Reward를 최대화할 수 있도록 학습됩니다. 이를 통해 LLM이 Search Engine과 효과적으로 Interact 하는 방법을 배울 수 있는 것이죠.</p><p>굉장히 나이브하면서도 간단한 접근 방식입니다. 🤔</p><h4 id="ppo-search-engine">PPO + Search Engine</h4><p>앞서 정의한 RL objective는 PPO 혹은 GRPO로 학습 (Policy Optimization) 할 수 있다고 합니다. 식은 다음과 같습니다.</p><p>\[J_{PPO}(\theta) = E_{x \sim D, y \sim \pi_{old}( \cdot | x; R)} \left[ \frac{1}{\sum_{t=1}^{|y|} I(y_t)} \sum_{t=1}^{|y|} I(y_t) \cdot \min \left( \frac{\pi_{\theta}(y_t | x, y_{&lt;t}; R)}{\pi_{old}(y_t | x, y_{&lt;t}; R)} A_t,\ \text{clip} \left( \frac{\pi_{\theta}(y_t | x, y_{&lt;t}; R)}{\pi_{old}(y_t | x, y_{&lt;t}; R)} , 1 - \epsilon,\ 1 + \epsilon \right) A_t \right) \right]\]</p><p>Policy Model에서 Search Engine R을 함께 사용한다는 것을 제외하면 일반적인 PPO의 식과 크게 다르지 않습니다.</p><p>한 가지 주목할 점은 <strong>검색된 토큰에 대해서는 Loss Masking을 진행</strong>했다는 점입니다.</p><p>$I(y_t)$ 에서<strong> LLM에 의해 생성된 토큰인 경우, loss를 계산하고 학습에 반영</strong>하는 한편, <strong>Search Engine에서 검색된 토큰인 경우 loss 계산에서 제외</strong>합니다.</p><p>이를 통해 LLM이 직접 생성한 것이 아닌 토큰은 학습 과정에서 제외하여 <strong>불필요한 학습을 방지하고 학습 안정성을 높일 수 있었다</strong>고 합니다.</p><p></p><h4 id="grpo-search-engine">GRPO + Search Engine</h4><p>최근 Deepseek로 핫한 GRPO도 활용이 가능합니다.</p><figure class="kg-card kg-image-card"><img src="https://facerain.me/content/images/2025/03/image-6.png" class="kg-image" alt="" loading="lazy" width="1388" height="287" srcset="https://facerain.me/content/images/size/w600/2025/03/image-6.png 600w, https://facerain.me/content/images/size/w1000/2025/03/image-6.png 1000w, https://facerain.me/content/images/2025/03/image-6.png 1388w" sizes="(min-width: 720px) 720px"></figure><p></p><p>PPO와 마찬가지로 기존 GRPO 식과 크게 다르지 않으며, 검색된 토큰에 대해서는 Loss Masking을 진행하였습니다.</p><h4 id="text-generation-with-interleaved-multi-turn-search-engine-call">Text Generation with Interleaved Multi-turn Search Engine Call</h4><p>그렇다면 LLM은 어떻게 생성 (추론)을 진행하는 동안 Search Engine과 상호작용을 할 수 있었을까요?</p><p>모델은 3가지 Special Token을 활용합니다.</p><p><strong>&lt;search&gt;&lt;/search&gt;</strong>: 해당 토큰 내부의 텍스트는 검색 쿼리로 파싱하여 Search Engine을 호출</p><p><strong>&lt;information&gt;&lt;/information&gt;</strong>
301: 앞서 Search Engine에 의한 검색 결과</p><p><strong>&lt;answer&gt;&lt;/answer&gt;</strong>: 최종 응답</p><p>그리고 학습을 위해 사용한 Trainiing Template은 아래와 같다고 합니다.</p><p>앞서 살펴본 think, search, answer 토큰이 활용되고 있습니다.</p><pre><code>Answer the given question. You must conduct reasoning inside &lt;think&gt; and &lt;/think&gt;
302first every time you get new information. After reasoning, if you find you lack some
303knowledge, you can call a search engine by &lt;search&gt; query &lt;/search&gt;, and it will
304return the top searched results between &lt;information&gt; and &lt;/information&gt;. You
305can search as many times as you want. If you find no further external knowledge
306needed, you can directly provide the answer inside &lt;answer&gt; and &lt;/answer&gt; without
307detailed illustrations. For example, &lt;answer&gt; Beijing &lt;/answer&gt;. Question: question.</code></pre><h4 id=""></h4><h3 id="reward-modeling">Reward Modeling</h3><p>본 프레임워크에서 Reward Model은 어떻게 설계하였을까요?</p><p>$$r_\phi(x, y) = EM(a_{pred}, a_{gold})$$</p><p>연구팀은 모델이 예측한 답변과 정답 사이의<strong> EM (Exact Match) Score</strong>를 사용하였습니다. 대표적인 다른 Reward Model 방법 중 하나인 Neural Reward Model은 Reward Hacking (모델이 오히려 편법으로 Reward만 높이려 해서 점수가 향상되지 않는 현상)ê³¼ 계산 비용 및 복잡성 문제가 있었다고 합니다.</p><p>또한 Deepseek-R1에서 활용한 format reward는 모델이 이미 구조적 규칙을 잘 따르기 때문에 별도로 적용하지 않았다고 합니다.</p><p></p><h3 id="main-results">Main Results</h3><p>연구팀은 평가를 위해 7가지 QA 데이터셋을 활용하여 평가 벤치마크를 구성하였습니다. (General QA: NQ, TriviaQA / Multi-Hop Question Answering: HotpotQA, 2WikiMultiHopQA, Musique, Banboogle)</p><p>비교 평가를 위해 3가지 비교군을 채택하였습니다.</p><ul><li>Inference without Retrieval<ul><li>External Knoledge 없이 Direct Inference 혹은 CoT 추론</li></ul></li><li>Inference with Retrieval<ul><li>RAG, IRCoT, Search-o1</li></ul></li><li>Fine-Tuning-Based Methods<ul><li>SFT</li><li>R1 (without a Search Engine)</li></ul></li></ul><p>그리고 Qwen-2.5-3B(Base/Instruct), Qwen-2.5-7B(Base/Instruct), Llama-3.2-3B(Base/Instruct)를 활용하여 학습을 진행하였고,</p><p>Retriever로는 E5를 활용하여 Wikipedia dump를 제공했다고 합니다.</p><figure class="kg-card kg-image-card"><img src="https://facerain.me/content/images/2025/03/image-1.png" class="kg-image" alt="" loading="lazy" width="554" height="584"></figure><p>결과는 위와 같습니다.</p><p>대부분의 데이터셋에서 높은 성능을 보이고 있습니다.</p><p></p><h3 id="analysis">Analysis</h3><p>개인적으로 흥미롭게 읽었던 부분입니다!</p><p>연구팀은 Search-R1 Method에서 여러가지 비교 실험을 진행하였는데요,</p><p><strong>GRPO vs PPO</strong></p><ul><li>GRPO는 PPO보다 빠르게 수렴, PPO는  critic model을 사용하기 위한 초기 학습에서 준비 단계가 필요함</li><li>PPO는 GRPO보다 훈련 안정성이 더 높음</li><li>PPO와 GRPO의 최종 훈련 보상은 비슷함</li><li>모델 별로도 다른 양상을 보일 때도 있음</li></ul><figure class="kg-card kg-image-card"><img src="https://facerain.me/content/images/2025/03/image-2.png" class="kg-image" alt="" loading="lazy" width="656" height="124" srcset="https://facerain.me/content/images/size/w600/2025/03/image-2.png 600w, https://facerain.me/content/images/2025/03/image-2.png 656w"></figure><p><strong>Base vs Instruct</strong></p><ul><li>Instruct 모델이 초기 성능도 높고 수렴 속도도 더 빠름</li><li>하지만 훈련이 진행됨에 따라 Base 모델은 Instruct 모델과 비슷한 수준의 성능에 도달할 수 있음</li><li>RL이 일반적인 SFT와 Pre-training 간의 격차를 해소할 수 있다고 주장</li></ul><figure class="kg-card kg-image-card"><img src="https://facerain.me/content/images/2025/03/image-3.png" class="kg-image" alt="" loading="lazy" width="657" height="156" srcset="https://facerain.me/content/images/size/w600/2025/03/image-3.png 600w, https://facerain.me/content/images/2025/03/image-3.png 657w"></figure><p><strong>Response Length Study</strong></p><ul><li>LLaMA3.2-3b-base 모델 기반, NQ 데이터셋으로 훈련</li><li>초기 단계 (~100 스텝)<ul><li>응답 길이가 급격히 감소</li><li>훈련 reward은 약간 증가</li><li>모델이 불필요한 단어를 제거하고 작업 요구 사항에 적응</li></ul></li><li>중간 단계 (100-130 스텝)<ul><li>응답 길이와 훈련 보상이 모두 크게 증가</li><li>LLM이 Search Engine을 호출하는 것을 학습하여 검색 ê²°ê
307³¼ë¡œ 인한 응답이 길어짐</li><li>모델이 검색 결과를 효과적으로 활용하여 훈련 reward가 크게 향상</li></ul></li><li>후기 단계 (130 스텝 이후)<ul><li>응답 길이가 안정화</li><li>훈련 Reward는 계속 약간 증가</li><li>모델이 Search Engine을 효과적으로 사용하는 방법을 학습하고 Search Query를 개선하는데 집중</li></ul></li></ul><figure class="kg-card kg-image-card"><img src="https://facerain.me/content/images/2025/03/image-4.png" class="kg-image" alt="" loading="lazy" width="984" height="305" srcset="https://facerain.me/content/images/size/w600/2025/03/image-4.png 600w, https://facerain.me/content/images/2025/03/image-4.png 984w" sizes="(min-width: 720px) 720px"></figure><p><strong>Study of Retrieved Tokens Loss Masking</strong></p><ul><li>Loss Masking을 적용하면 학습이 더욱 안정적으로 진행되고, 전반적인 성능이 향상됨.</li></ul><figure class="kg-card kg-image-card"><img src="https://facerain.me/content/images/2025/03/image-5.png" class="kg-image" alt="" loading="lazy" width="548" height="68"></figure><h3 id="%EB%A7%88%EC%B9%98%EB%A9%B0">마치며</h3><p>Search-R1은 단순한 RAG나 툴 사용을 넘어서, <strong>LLM이 스스로 Reasoning 과정 속에서 검색을 전략적으로 활용하도록 RL 전략으로 학습시킨</strong> 점이 인상 깊었는데요,<br>최근 LLM의 Reasoning 및 Agent 연구들이 활발해지고 있는데, <strong>Search-R1</strong>은 이 흐름 속에서 Search Engine을 LLM에 어떻게 녹여낼 수 있을지 흥미로운 관점을 제시한다고 생각합니다.</p><p>감사합니다 😊</p>
308    </section>
309
310        <section class="article-comments gh-canvas">
311            
312        
312<script defer src="https://cdn.jsdelivr.net/ghost/comments-ui@~1.5/umd/comments-ui.min.js" data-locale="ko" data-ghost-comments="https://facerain.me/" data-api="https://facerain.me/ghost/api/content/" data-admin="https://facerain.me/ghost/" data-key="3fb8b569937df4745424d79246" data-title="null" data-count="true" data-post-id="67e25a922a4be6054c3df8ef" data-color-scheme="auto" data-avatar-saturation="60" data-accent-color="#000000" data-comments-enabled="all" data-publication="FacerAin's blog" crossorigin="anonymous"></script>
312
313    
314        </section>
315
316</article>
317</main>
318
319
320
321
322            <aside class="read-more-wrap outer">
323                <div class="read-more inner">
324                        
325<article class="post-card post">
326
327    <a class="post-card-image-link" href="/dogseo-sagooeju/">
328
329        <img class="post-card-image"
330            srcset="https://images.unsplash.com/photo-1789921464199-d87f82e3f8d2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDExfHx8fHx8fHwxNzkxMDM3MzIxfA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;300 300w,
331                    https://images.unsplash.com/photo-1789921464199-d87f82e3f8d2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDExfHx8fHx8fHwxNzkxMDM3MzIxfA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;600 600w,
332                    https://images.unsplash.com/photo-1789921464199-d87f82e3f8d2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDExfHx8fHx8fHwxNzkxMDM3MzIxfA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;1000 1000w,
333                    https://images.unsplash.com/photo-1789921464199-d87f82e3f8d2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDExfHx8fHx8fHwxNzkxMDM3MzIxfA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;2000 2000w"
334            sizes="(max-width: 1000px) 400px, 800px"
335            src="https://images.unsplash.com/photo-1789921464199-d87f82e3f8d2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDExfHx8fHx8fHwxNzkxMDM3MzIxfA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;600"
336            alt="[독서] 사고외주"
337            loading="lazy"
338        />
339
340
341    </a>
342
343    <div class="post-card-content">
344
345        <a class="post-card-content-link" href="/dogseo-sagooeju/">
346            <header class="post-card-header">
347                <div class="post-card-tags">
348                </div>
349                <h2 class="post-card-title">
350                    [독서] 사고외주
351                </h2>
352            </header>
353                <div class="post-card-excerpt">사고외주 | 홍진기 | 어크로스 - 예스24“우리는 지금 ‘똑똑한 무능함’에 빠져 있다”과정 없이 결론만 얻는 시대, 인간은 무엇을 잃게 되는가?연세대 홍진기 교수, AI 시대 가장 조용하고 위험한 변화를 말하다매끄러운 문장 사이로, 어느 날부터 학생의 얼굴이 보이지 않기 시작했다. 연세대 화공생명공학…어크로스홍진기
354 * 한줄평: AI에게 사유를 빼앗겨버린 우리에게 건네는 안내서 (5/</div>
355        </a>
356
357        <footer class="post-card-meta">
358            <time class="post-card-meta-date" datetime="2026-10-04">04 10월 2026</time>
359                <span class="post-card-meta-length">8 min read</span>
360                
360<script
361    data-ghost-comment-count="6ac11d6abcedbc0001dc661d"
362    data-ghost-comment-count-empty=""
363    data-ghost-comment-count-singular="comment"
364    data-ghost-comment-count-plural="comments"
365    data-ghost-comment-count-tag="span"
366    data-ghost-comment-count-class-name=""
367    data-ghost-comment-count-autowrap="true"
368>
369</script>
369
370        </footer>
371
372    </div>
373
374</article>
375                        
376<article class="post-card post">
377
378    <a class="post-card-image-link" href="/beonyeog-the-second-half/">
379
380        <img class="post-card-image"
381            srcset="https://images.unsplash.com/photo-1760592150404-adacb88548e2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDJ8fHx8fHx8fDE3NjA2NTUzOTJ8&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;300 300w,
382                    https://images.unsplash.com/photo-1760592150404-adacb88548e2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDJ8fHx8fHx8fDE3NjA2NTUzOTJ8&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;600 600w,
383                    https://images.unsplash.com/photo-1760592150404-adacb88548e2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDJ8fHx8fHx8fDE3NjA2NTUzOTJ8&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;1000 1000w,
384                    https://images.unsplash.com/photo-1760592150404-adacb88548e2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDJ8fHx8fHx8fDE3NjA2NTUzOTJ8&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;2000 2000w"
385            sizes="(max-width: 1000px) 400px, 800px"
386            src="https://images.unsplash.com/photo-1760592150404-adacb88548e2?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDJ8fHx8fHx8fDE3NjA2NTUzOTJ8&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;600"
387            alt="[번역] The Second Half"
388            loading="lazy"
389        />
390
391
392    </a>
393
394    <div class="post-card-content">
395
396        <a class="post-card-content-link" href="/beonyeog-the-second-half/">
397            <header class="post-card-header">
398                <div class="post-card-tags">
399                </div>
400                <h2 class="post-card-title">
401                    [번역] The Second Half
402                </h2>
403            </header>
404                <div class="post-card-excerpt">The Second Halftldr: We’re at AI’s halftime.Shunyu Yao
405
406이번 시간에는 Shunya Yao의 The Second Half의 포스트를 소개합니다. 최근 Deepseek R1, OpenAI O-Series 등 강화학습을 적용한 Reasoning 모델들이 높은 성능으로 두각을 보이고 있습니다. 무엇이 언어 모델의 강화학습을 성공하게 했을까요? 또한 2025년 4월 기준 AI는 지금까지 어떤 길을</div>
407        </a>
408
409        <footer class="post-card-meta">
410            <time class="post-card-meta-date" datetime="2025-10-20">20 10월 2025</time>
411                <span class="post-card-meta-length">23 min read</span>
412                
412<script
413    data-ghost-comment-count="68f17584237318cbc3a0ec08"
414    data-ghost-comment-count-empty=""
415    data-ghost-comment-count-singular="comment"
416    data-ghost-comment-count-plural="comments"
417    data-ghost-comment-count-tag="span"
418    data-ghost-comment-count-class-name=""
419    data-ghost-comment-count-autowrap="true"
420>
421</script>
421
422        </footer>
423
424    </div>
425
426</article>
427                        
428<article class="post-card post">
429
430    <a class="post-card-image-link" href="/paper-review-colpali-efficient-document-retrieval-with-vision-language-models/">
431
432        <img class="post-card-image"
433            srcset="https://images.unsplash.com/photo-1499428665502-503f6c608263?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDcyfHx8fHx8fHwxNzUyNDgwODc3fA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;300 300w,
434                    https://images.unsplash.com/photo-1499428665502-503f6c608263?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDcyfHx8fHx8fHwxNzUyNDgwODc3fA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;600 600w,
435                    https://images.unsplash.com/photo-1499428665502-503f6c608263?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDcyfHx8fHx8fHwxNzUyNDgwODc3fA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;1000 1000w,
436                    https://images.unsplash.com/photo-1499428665502-503f6c608263?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDcyfHx8fHx8fHwxNzUyNDgwODc3fA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;2000 2000w"
437            sizes="(max-width: 1000px) 400px, 800px"
438            src="https://images.unsplash.com/photo-1499428665502-503f6c608263?crop&#x3D;entropy&amp;cs&#x3D;tinysrgb&amp;fit&#x3D;max&amp;fm&#x3D;jpg&amp;ixid&#x3D;M3wxMTc3M3wwfDF8YWxsfDcyfHx8fHx8fHwxNzUyNDgwODc3fA&amp;ixlib&#x3D;rb-4.1.0&amp;q&#x3D;80&amp;w&#x3D;600"
439            alt="[논문 리뷰] ColPali: EFFICIENT DOCUMENT RETRIEVAL WITH VISION LANGUAGE MODELS"
440            loading="lazy"
441        />
442
443
444    </a>
445
446    <div class="post-card-content">
447
448        <a class="post-card-content-link" href="/paper-review-colpali-efficient-document-retrieval-with-vision-language-models/">
449            <header class="post-card-header">
450                <div class="post-card-tags">
451                </div>
452                <h2 class="post-card-title">
453                    [논문 리뷰] ColPali: EFFICIENT DOCUMENT RETRIEVAL WITH VISION LANGUAGE MODELS
454                </h2>
455            </header>
456                <div class="post-card-excerpt">ColPali: Efficient Document Retrieval with Vision Language ModelsDocuments are visually rich structures that convey information through text, but also figures, page layouts, tables, or even fonts. Since modern retrieval systems mainly rely on the textual information they extract from document pages to index documents -often through lengthy and brittle processes-</div>
457        </a>
458
459        <footer class="post-card-meta">
460            <time class="post-card-meta-date" datetime="2025-07-14">14 7월 2025</time>
461                <span class="post-card-meta-length">16 min read</span>
462                
462<script
463    data-ghost-comment-count="6874bc526734f9d66b63dea0"
464    data-ghost-comment-count-empty=""
465    data-ghost-comment-count-singular="comment"
466    data-ghost-comment-count-plural="comments"
467    data-ghost-comment-count-tag="span"
468    data-ghost-comment-count-class-name=""
469    data-ghost-comment-count-autowrap="true"
470>
471</script>
471
472        </footer>
473
474    </div>
475
476</article>
477                </div>
478            </aside>
479
480
481
482    </div>
483
484    <footer class="site-footer outer">
485        <div class="inner">
486            <section class="copyright"><a href="https://facerain.me">FacerAin&#x27;s blog</a> &copy; 2026</section>
487            <div class="site-footer-center">
488                <div class="site-footer-social-links">
489                        <a href="https://x.com/ghost" target="_blank" rel="noopener" aria-label="X">
490<svg class="icon" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg">
491    <path d="M18.2439 2.25H21.5519L14.3249 10.51L22.8269 21.75H16.1699L10.9559 14.933L4.98991 21.75H1.67991L9.40991 12.915L1.25391 2.25H8.07991L12.7929 8.481L18.2439 2.25ZM17.0829 19.77H18.9159L7.08391 4.126H5.11691L17.0829 19.77Z" fill="currentColor"/>
492</svg>                        </a>
493                        <a href="https://www.facebook.com/ghost" target="_blank" rel="noopener" aria-label="Facebook">
494<svg class="icon" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg" fill="currentColor"><path d="M23.9981 11.9991C23.9981 5.37216 18.626 0 11.9991 0C5.37216 0 0 5.37216 0 11.9991C0 17.9882 4.38789 22.9522 10.1242 23.8524V15.4676H7.07758V11.9991H10.1242V9.35553C10.1242 6.34826 11.9156 4.68714 14.6564 4.68714C15.9692 4.68714 17.3424 4.92149 17.3424 4.92149V7.87439H15.8294C14.3388 7.87439 13.8739 8.79933 13.8739 9.74824V11.9991H17.2018L16.6698 15.4676H13.8739V23.8524C19.6103 22.9522 23.9981 17.9882 23.9981 11.9991Z"/></svg>                        </a>
495                </div>
496                <nav class="site-footer-nav">
497                    <ul class="nav">
498    <li class="nav-sign-up"><a href="#/portal/">Sign up</a></li>
499</ul>
500
501                </nav>
502            </div>
503            <div class="gh-powered-by"><a href="https://ghost.org/" target="_blank" rel="noopener">Powered by Ghost</a></div>
504        </div>
505    </footer>
506
507</div>
508
509    <div class="pswp" tabindex="-1" role="dialog" aria-hidden="true">
510    <div class="pswp__bg"></div>
511
512    <div class="pswp__scroll-wrap">
513        <div class="pswp__container">
514            <div class="pswp__item"></div>
515            <div class="pswp__item"></div>
516            <div class="pswp__item"></div>
517        </div>
518
519        <div class="pswp__ui pswp__ui--hidden">
520            <div class="pswp__top-bar">
521                <div class="pswp__counter"></div>
522
523                <button class="pswp__button pswp__button--close" title="Close (Esc)"></button>
524                <button class="pswp__button pswp__button--share" title="Share"></button>
525                <button class="pswp__button pswp__button--fs" title="Toggle fullscreen"></button>
526                <button class="pswp__button pswp__button--zoom" title="Zoom in/out"></button>
527
528                <div class="pswp__preloader">
529                    <div class="pswp__preloader__icn">
530                        <div class="pswp__preloader__cut">
531                            <div class="pswp__preloader__donut"></div>
532                        </div>
533                    </div>
534                </div>
535            </div>
536
537            <div class="pswp__share-modal pswp__share-modal--hidden pswp__single-tap">
538                <div class="pswp__share-tooltip"></div>
539            </div>
540
541            <button class="pswp__button pswp__button--arrow--left" title="Previous (arrow left)"></button>
542            <button class="pswp__button pswp__button--arrow--right" title="Next (arrow right)"></button>
543
544            <div class="pswp__caption">
545                <div class="pswp__caption__center"></div>
546            </div>
547        </div>
548    </div>
549</div>
550<script src="/assets/built/casper.js?v=4f24c6eeff" defer></script>
550
551
552ссс
553
554</body>
555</html>

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.