{"id":22379,"date":"2025-01-28T16:35:58","date_gmt":"2025-01-28T07:35:58","guid":{"rendered":"http:\/\/www.smallake.kr\/?p=22379"},"modified":"2025-04-22T14:39:00","modified_gmt":"2025-04-22T05:39:00","slug":"deepseek-ai-%eb%aa%a8%eb%8d%b8%ec%97%90-%eb%8c%80%ed%95%9c-%ec%97%ac%eb%9f%ac%ea%b0%80%ec%a7%80-%ec%9d%b4%ed%95%b4","status":"publish","type":"post","link":"https:\/\/smallake.kr\/?p=22379","title":{"rendered":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574"},"content":{"rendered":"<p>1.<br \/>\n DeepSeek\uc640 \uad00\ub828\ud55c \uae00\uc744 \uc77d\uc73c\uba74\uc11c \uc774\ud574\ub3c4\uac00 \ubb34\ucc99 \ub5a8\uc5b4\uc9d1\ub2c8\ub2e4. AI \uae30\uc220\uc5d0 \ub300\ud55c \uc774\ud574\ub3c4 \ub0ae\uace0 Nvidia GPU\uc5d0 \ub300\ud55c \uc774\ud574\ub3c4 \ub0ae\uc2b5\ub2c8\ub2e4. \uadf8\ub798\uc11c A100\uacfc H800\uc774 \ubb34\uc5c7\uc778\uc9c0, \uc5b4\ub5bb\uac8c \ub2e4\ub978\uc9c0\ub3c4 \uc798 \ubaa8\ub985\ub2c8\ub2e4. <a href=\"https:\/\/www.fibermall.com\/ko\/blog\/nvidia-ai-chip.htm\" rel=\"noopener\" target=\"_blank\">\uc5d4\ube44\ub514\uc544 AI \uce69: A100 A800 H100 H800 B200<\/a>\uc5d0 \uc62c\ub77c\uc628 \ube44\uad50\ud45c\ub85c \uc774\ud574\ub97c \ub192\ud614\uc2b5\ub2c8\ub2e4.\uc6b0\uc120&#8230;.<\/p>\n<p><a href=\"https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/chip-comparison.webp\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" src=\"https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/chip-comparison.webp?resize=816%2C287\" alt=\"\" width=\"816\" height=\"287\" class=\"aligncenter size-full wp-image-34848\" srcset=\"https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/chip-comparison.webp?w=816&amp;ssl=1 816w, https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/chip-comparison.webp?resize=300%2C106&amp;ssl=1 300w, https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/chip-comparison.webp?resize=200%2C70&amp;ssl=1 200w\" sizes=\"auto, (max-width: 816px) 100vw, 816px\" \/><\/a><\/p>\n<p>Dropbox \ubd80\uc0ac\uc7a5 Morgan Brown\uc774 \uc4f4 <a href=\"https:\/\/www.linkedin.com\/feed\/update\/urn:li:activity:7289438389001469953\/\" target=\"_blank\" rel=\"noopener\">Why DeepSeek&#8217;s AI Breakthrough May Be the Most Crucial One Yet<\/a>. linkedin\uc5d0 \uc62c\ub9b0 \uae00\uc785\ub2c8\ub2e4. AI Product \ub2f4\ub2f4 \ubd80\uc0ac\uc7a5\uc73c\ub85c\uc11c\uc758 \uacac\ud574\uc785\ub2c8\ub2e4.<\/p>\n<blockquote><p>\ud83d\udd25 Why DeepSeek&#8217;s AI Breakthrough May Be the Most Crucial One Yet.<\/p>\n<p>I finally had a chance to dive into DeepSeek&#8217;s recent r1 model innovations, and it\u2019s hard to overstate the implications. This isn&#8217;t just a technical achievement &#8211; it&#8217;s democratization of AI technology.<\/p>\n<p>Let me explain why this matters for everyone in tech, not just AI teams.<\/p>\n<p>\ud83c\udfaf The Big Picture:<br \/>\nTraditional model development has been like building a skyscraper &#8211; you need massive resources, billions in funding, and years of work. DeepSeek just showed you can build the same thing for 5% of the cost, in a fraction of the time.<\/p>\n<p>Here&#8217;s what they achieved:<br \/>\n\u2022 Matched GPT-4 level performance<br \/>\n\u2022 Cut training costs from $100M+ to $5M<br \/>\n\u2022 Reduced GPU requirements by 98%<br \/>\n\u2022 Made models run on consumer hardware<br \/>\n\u2022 Released everything as open source<\/p>\n<p>\ud83e\udd14 Why This Matters:<\/p>\n<p>1. For Business Leaders:<br \/>\n&#8211; model development &amp; AI implementation costs could drop dramatically<br \/>\n&#8211; Smaller companies can now compete with tech giants<br \/>\n&#8211; ROI calculations for AI projects need complete revision<br \/>\n&#8211; Infrastructure planning can possibly be drastically simplified<\/p>\n<p>2. For Developers &amp; Technical Teams:<br \/>\n&#8211; Advanced AI becomes accessible without massive compute<br \/>\n&#8211; Development cycles can be dramatically shortened<br \/>\n&#8211; Testing and iteration become much more feasible<br \/>\n&#8211; Open source access to state-of-the-art techniques<\/p>\n<p>3. For Product Managers:<br \/>\n&#8211; Features previously considered &#8220;too expensive&#8221; become viable<br \/>\n&#8211; Faster prototyping and development cycles<br \/>\n&#8211; More realistic budgets for AI implementation<br \/>\n&#8211; Better performance metrics for existing solutions<\/p>\n<p>\ud83d\udca1 The Innovation Breakdown:<br \/>\nWhat makes this special isn&#8217;t just one breakthrough &#8211; it&#8217;s five clever innovations working together:<\/p>\n<p>\u2022 Smart number storage (reducing memory needs by 75%)<br \/>\n\u2022 Parallel processing improvements (2x speed increase)<br \/>\n\u2022 Efficient memory management (massive scale improvements)<br \/>\n\u2022 Better resource utilization (near 100% GPU efficiency)<br \/>\n\u2022 Specialist AI system (only using what&#8217;s needed, when needed)<\/p>\n<p>\ud83c\udf1f Real-World Impact:<br \/>\nImagine running ChatGPT-level AI on your gaming computer instead of a data center. That&#8217;s not science fiction anymore &#8211; that&#8217;s what DeepSeek achieved.<\/p>\n<p>\ud83d\udd04 Industry Implications:<br \/>\nThis could reshape the entire AI industry:<br \/>\n&#8211; Hardware manufacturers (looking at you, Nvidia) may need to rethink business models<br \/>\n&#8211; Cloud providers might need to revise their pricing<br \/>\n&#8211; Startups can now compete with tech giants<br \/>\n&#8211; Enterprise AI becomes much more accessible<\/p>\n<p>\ud83d\udcc8 What&#8217;s Next:<br \/>\nI expect we&#8217;ll see:<br \/>\n1. Rapid adoption of these techniques by major players<br \/>\n2. New startups leveraging this more efficient approach<br \/>\n3. Dropping costs for AI implementation<br \/>\n4. More innovative applications as barriers lower<\/p>\n<p>\ud83c\udfaf Key Takeaway:<br \/>\nThe AI playing field is being leveled. What required billions and massive data centers might now be possible with a fraction of the resources. This isn&#8217;t just a technical achievement &#8211; it&#8217;s a democratization of AI technology.<\/p><\/blockquote>\n<p>\ud55c\uae00\ub85c \uc815\ub9ac\ud55c \uae00\uc740 <a href=\"https:\/\/m.ppomppu.co.kr\/new\/bbs_view.php?id=stock&#038;no=286409&#038;page=1\" rel=\"noopener\" target=\"_blank\">\ub4dc\ub86d\ubc15\uc2a4 \ubaa8\uac74 \ube0c\ub77c\uc6b4\uc774 \uc124\uba85\ud55c \ub525\uc2dc\ud06c \uadf8\ub9ac\uace0..<\/a>\uc5d0\uc11c \ubcf4\uc2e4 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud2b8\uc717\uc774\ub098 threads\ub97c \ud1b5\ud574 \ub110\ub9ac \uc62e\uaca8\uc9c0\uace0 \uc788\uae30\ub3c4 \ud569\ub2c8\ub2e4.<\/p>\n<blockquote><p>1.\uba3c\uc800 \ubc30\uacbd\ubd80\ud130 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4. \ud604\uc7ac \ucd5c\ucca8\ub2e8 AI \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\ub294 \ube44\uc6a9\uc740 \uc5c4\uccad\ub098\uac8c \ube44\uc309\ub2c8\ub2e4.OpenAI, Anthropic \uac19\uc740 \ud68c\uc0ac\ub4e4\uc740 \uacc4\uc0b0\uc5d0\ub9cc 1\uc5b5 \ub2ec\ub7ec \uc774\uc0c1\uc744 \uc4f0\uba70, $4\ub9cc\uc9dc\ub9ac GPU \uc218\ucc9c \ub300\uac00 \ud544\uc694\ud55c \ub300\uaddc\ubaa8 \ub370\uc774\ud130 \uc13c\ud130\ub97c \uc6b4\uc601\ud569\ub2c8\ub2e4. \ub9c8\uce58 \uacf5\uc7a5\uc744 \uc6b4\uc601\ud558\uae30 \uc704\ud574 \ubc1c\uc804\uc18c \uc804\uccb4\uac00 \ud544\uc694\ud55c \uc0c1\ud669\uacfc \uac19\uc2b5\ub2c8\ub2e4.<\/p>\n<p>2.\uadf8\ub7f0\ub370 DeepSeek\uc774 \ub098\ud0c0\ub098\uc11c \uc774\ub807\uac8c \ub9d0\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\u201cLOL, \uc6b0\ub9ac\ub77c\uba74 \uc774\uac78 500\ub9cc \ub2ec\ub7ec\ub85c \ud560 \uc218 \uc788\uc744\uac78?\u201d<br \/>\n\uadf8\ub9ac\uace0 \ub9d0\ub85c\ub9cc \ud55c \uac83\uc774 \uc544\ub2c8\ub77c \uc2e4\uc81c\ub85c \ud574\ub0c8\uc2b5\ub2c8\ub2e4.\uadf8\ub4e4\uc758 \ubaa8\ub378\uc740 GPT-4\uc640 Claude\ub97c \ub9ce\uc740 \uc791\uc5c5\uc5d0\uc11c \ub2a5\uac00\ud558\uac70\ub098 \ub300\ub4f1\ud569\ub2c8\ub2e4. AI \uc5c5\uacc4\ub294 (\uc694\uc998 10\ub300\ub4e4\uc774 \ub9d0\ud558\ub4ef) \u2018\ucda9\uaca9\u2019\uc744 \ubc1b\uc558\uc2b5\ub2c8\ub2e4.<\/p>\n<p>3.\uc5b4\ub5bb\uac8c \uac00\ub2a5\ud588\uc744\uae4c\uc694?<br \/>\n\uadf8\ub4e4\uc740 \ubaa8\ub4e0 \uac83\uc744 \ucc98\uc74c\ubd80\ud130 \ub2e4\uc2dc \uc0dd\uac01\ud588\uc2b5\ub2c8\ub2e4.\uc804\ud1b5\uc801\uc778 AI\ub294 \ub9c8\uce58 \ubaa8\ub4e0 \uc22b\uc790\ub97c \uc18c\uc218\uc810 32\uc790\ub9ac\uae4c\uc9c0 \uae30\ub85d\ud558\ub294 \uac83\uacfc \uac19\uc2b5\ub2c8\ub2e4.DeepSeek\uc740 \u201c8\uc790\ub9ac\ub85c\ub9cc \uae30\ub85d\ud558\uba74 \uc5b4\ub5a8\uae4c? \ucda9\ubd84\ud788 \uc815\ud655\ud558\uc796\uc544!\u201d\ub77c\uace0 \uc811\uadfc\ud588\uace0, \uacb0\uacfc\uc801\uc73c\ub85c \uba54\ubaa8\ub9ac \uc0ac\uc6a9\ub7c9\uc774 75% \uac10\uc18c\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<p>4.\uadf8\ub9ac\uace0 \uadf8\ub4e4\uc758 \u201c\uba40\ud2f0 \ud1a0\ud070\u201d \uc2dc\uc2a4\ud15c\ub3c4 \uc8fc\ubaa9\ud560 \ub9cc\ud569\ub2c8\ub2e4.\uc77c\ubc18\uc801\uc778 AI\ub294 \ucd08\ub4f1\ud559\uc0dd\uc774 \uc77d\ub4ef\uc774 \u201cThe\u2026 cat\u2026 sat\u2026\u201d\ucc98\ub7fc \uc77d\uc2b5\ub2c8\ub2e4.\ubc18\uba74, DeepSeek\uc740 \ubb38\uc7a5 \uc804\uccb4\ub97c \ud55c \ubc88\uc5d0 \uc77d\uc2b5\ub2c8\ub2e4. \uacb0\uacfc\uc801\uc73c\ub85c 2\ubc30 \ub354 \ube60\ub974\uace0 90% \uc218\uc900\uc758 \uc815\ud655\ub3c4\ub97c \uc790\ub791\ud569\ub2c8\ub2e4.\uc218\uc2ed\uc5b5 \uac1c\uc758 \ub2e8\uc5b4\ub97c \ucc98\ub9ac\ud560 \ub54c, \uc774\ub7f0 \ud6a8\uc728\uc131\uc740 \ub9e4\uc6b0 \uc911\uc694\ud569\ub2c8\ub2e4.<\/p>\n<p>5.\ud558\uc9c0\ub9cc \uc9c4\uc9dc \uae30\ubc1c\ud55c \uc810\uc740 \u201c\uc804\ubb38\uac00 \uc2dc\uc2a4\ud15c\u201d\uc744 \uad6c\ucd95\ud588\ub2e4\ub294 \uac83\uc785\ub2c8\ub2e4.\ud55c \uac70\ub300\ud55c AI\uac00 \ubaa8\ub4e0 \uac83\uc744 \ub2e4 \uc54c\ub3c4\ub85d \ub9cc\ub4dc\ub294 \ub300\uc2e0(\uc608: \ud55c \uc0ac\ub78c\uc774 \uc758\uc0ac, \ubcc0\ud638\uc0ac, \uc5d4\uc9c0\ub2c8\uc5b4 \uc5ed\ud560\uc744 \ubaa8\ub450 \ud558\ub294 \uac83\ucc98\ub7fc), DeepSeek\uc740 \ud544\uc694\ud55c \uacbd\uc6b0\uc5d0\ub9cc \uc804\ubb38\uac00\ub4e4\uc744 \ud638\ucd9c\ud558\ub3c4\ub85d \uc124\uacc4\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<p>6.\uae30\uc874 \ubaa8\ub378\uc740 1.8\uc870 \uac1c\uc758 \ud30c\ub77c\ubbf8\ud130\uac00 \ud56d\uc0c1 \ud65c\uc131\ud654\ub418\uc5b4\uc57c \ud569\ub2c8\ub2e4.DeepSeek\uc740 6710\uc5b5 \uac1c\uc758 \ud30c\ub77c\ubbf8\ud130 \uc911 \ub2e8\uc9c0 370\uc5b5 \uac1c\ub9cc \ud65c\uc131\ud654\ub429\ub2c8\ub2e4. \ub9c8\uce58 \ud070 \ud300\uc744 \uc6b4\uc601\ud558\ub418 \ud544\uc694\ud55c \uc804\ubb38\uac00\ub9cc \ud638\ucd9c\ud558\ub294 \uac83\uacfc \uac19\uc2b5\ub2c8\ub2e4.<\/p>\n<p>7.\uacb0\uacfc\ub294 \ub180\ub78d\uc2b5\ub2c8\ub2e4:<\/p>\n<blockquote><p>\u2022 \ud6c8\ub828 \ube44\uc6a9: 1\uc5b5 \ub2ec\ub7ec \u2192 500\ub9cc \ub2ec\ub7ec<br \/>\n\u2022 \ud544\uc694\ud55c GPU \uc218: 100,000\ub300 \u2192 2,000\ub300<br \/>\n\u2022 API \ube44\uc6a9: 95% \uc808\uac10<br \/>\n\u2022 \ub370\uc774\ud130 \uc13c\ud130 \ud558\ub4dc\uc6e8\uc5b4 \ub300\uc2e0 \uac8c\uc774\ubc0d GPU\uc5d0\uc11c\ub3c4 \uc2e4\ud589 \uac00\ub2a5<\/p><\/blockquote>\n<p>8.\u201c\uadf8\ub7f0\ub370,\u201d \ub204\uad70\uac00 \ub9d0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \u201c\ubd84\uba85 \ub2e8\uc810\uc774 \uc788\uaca0\uc9c0!\u201d\ub180\ub77c\uc6b4 \uc810\uc740, \ubaa8\ub4e0 \uac83\uc774 \uc624\ud508 \uc18c\uc2a4\ub77c\ub294 \uac83\uc785\ub2c8\ub2e4. \ub204\uad6c\ub098 \uadf8\ub4e4\uc758 \uc791\uc5c5\uc744 \uac80\uc99d\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ucf54\ub4dc\ub294 \uacf5\uac1c\ub418\uc5b4 \uc788\uace0, \uae30\uc220 \ub17c\ubb38\uc740 \ubaa8\ub4e0 \uacfc\uc815\uc744 \uc124\uba85\ud569\ub2c8\ub2e4. \ub9c8\ubc95\uc774 \uc544\ub2c8\ub77c, \ub2e8\uc21c\ud788 \ub9e4\uc6b0 \uc601\ub9ac\ud55c \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1\uc785\ub2c8\ub2e4.<\/p>\n<p>9.\uc65c \uc911\uc694\ud55c\uac00\uc694?<br \/>\n\uc774\ub85c \uc778\ud574 \u201c\ub300\ud615 \uae30\uc220 \uae30\uc5c5\ub9cc\uc774 AI\ub97c \ub2e4\ub8f0 \uc218 \uc788\ub2e4\u201d\ub294 \uae30\uc874\uc758 \ubaa8\ub378\uc774 \uae68\uc84c\uc2b5\ub2c8\ub2e4. \uc774\uc81c \uc218\uc2ed\uc5b5 \ub2ec\ub7ec \uaddc\ubaa8\uc758 \ub370\uc774\ud130 \uc13c\ud130\uac00 \ud544\uc694\ud558\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4. \uc88b\uc740 GPU \uba87 \ub300\ub9cc \uc788\uc73c\uba74 \ub429\ub2c8\ub2e4.<\/p>\n<p>10.Nvidia\uc5d0\uac8c\ub294 \ub450\ub824\uc6b4 \uc774\uc57c\uae30\uc785\ub2c8\ub2e4.\uadf8\ub4e4\uc758 \ube44\uc988\ub2c8\uc2a4 \ubaa8\ub378\uc740 \ucd08\uace0\uac00 GPU\ub97c 90% \ub9c8\uc9c4\uc73c\ub85c \ud310\ub9e4\ud558\ub294 \ub370 \uae30\ubc18\uc744 \ub450\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \ubaa8\ub450\uac00 \uc77c\ubc18 \uac8c\uc774\ubc0d GPU\ub85c AI\ub97c \ub3cc\ub9b4 \uc218 \uc788\uac8c \ub41c\ub2e4\uba74\u2026 \ubb38\uc81c\ub294 \uba85\ud655\ud569\ub2c8\ub2e4.<\/p>\n<p>11. \uadf8\ub9ac\uace0 \uc911\uc694\ud55c \uc810\uc740 DeepSeek\uc774 \uc774\ub97c 200\uba85 \uc774\ud558\uc758 \ud300\uc73c\ub85c \ud574\ub0c8\ub2e4\ub294 \uac83\uc785\ub2c8\ub2e4. \ud55c\ud3b8 Meta\uc758 \ud300\uc740 DeepSeek \uc804\uccb4 \ud6c8\ub828 \uc608\uc0b0\ubcf4\ub2e4 \ub354 \ub9ce\uc740 \uc5f0\ubd09\uc744 \ubc1b\uc73c\uba70 \uc791\uc5c5\uc744 \ud558\uace0 \uc788\uc9c0\ub9cc, \uadf8\ub4e4\uc758 \ubaa8\ub378\uc740 DeepSeek\ub9cc\ud07c \uc88b\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4.<\/p>\n<p>12.\uc774\ub294 \uc804\ud615\uc801\uc778 \ud30c\uad34\uc801 \ud601\uc2e0\uc758 \uc774\uc57c\uae30\uc785\ub2c8\ub2e4.<\/p>\n<p>\uae30\uc874 \uae30\uc5c5\ub4e4\uc740 \uae30\uc874 \ud504\ub85c\uc138\uc2a4\ub97c \ucd5c\uc801\ud654\ud558\ub294 \ub370 \ucd08\uc810\uc744 \ub9de\ucd94\ub294 \ubc18\uba74, \ud30c\uad34\uc801\uc778 \ud601\uc2e0 \uae30\uc5c5\ub4e4\uc740 \uadfc\ubcf8\uc801\uc778 \uc811\uadfc \ubc29\uc2dd\uc744 \ub2e4\uc2dc \uc0dd\uac01\ud569\ub2c8\ub2e4.<\/p>\n<p>DeepSeek\uc740 \u201c\ub354 \ub9ce\uc740 \ud558\ub4dc\uc6e8\uc5b4\ub97c \ud22c\uc785\ud558\uae30\ubcf4\ub2e4 \ub354 \ub611\ub611\ud558\uac8c \uc811\uadfc\ud558\uba74 \uc5b4\ub5a8\uae4c?\u201d\ub77c\uace0 \ubb3c\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n<p>13.\uadf8 \uc601\ud5a5\uc740 \ud07d\ub2c8\ub2e4:<\/p>\n<blockquote><p>\u2022 AI \uac1c\ubc1c\uc774 \ub354 \uc811\uadfc \uac00\ub2a5\ud574\uc9d0<br \/>\n\u2022 \uacbd\uc7c1\uc774 \uae09\uaca9\ud788 \uc99d\uac00<br \/>\n\u2022 \ub300\ud615 \uae30\uc220 \uae30\uc5c5\ub4e4\uc758 \u201c\uc9c4\uc785 \uc7a5\ubcbd\u201d\uc774 \uc791\uc740 \uc6c5\ub369\uc774\ucc98\ub7fc \ubcf4\uc784<br \/>\n\u2022 \ud558\ub4dc\uc6e8\uc5b4 \uc694\uad6c \uc0ac\ud56d(\ubc0f \ube44\uc6a9)\uc774 \uae09\uac10<\/p><\/blockquote>\n<p>14.\ubb3c\ub860, OpenAI\uc640 Anthropic \uac19\uc740 \ub300\uae30\uc5c5\ub4e4\uc774 \uac00\ub9cc\ud788 \uc788\uc9c0\ub294 \uc54a\uc744 \uac83\uc785\ub2c8\ub2e4.\uadf8\ub4e4\uc740 \uc544\ub9c8\ub3c4 \uc774\ubbf8 \uc774\ub7ec\ud55c \ud601\uc2e0\uc744 \uad6c\ud604\ud558\uace0 \uc788\uc744 \uac83\uc785\ub2c8\ub2e4.\uadf8\ub7ec\ub098 \ud6a8\uc728\uc131\uc758 \ub7a8\ud504\ub294 \uc774\uc81c \ubcd1 \ubc16\uc73c\ub85c \ub098\uc654\uc73c\uba70, \u201c\ub354 \ub9ce\uc740 GPU\ub97c \ud22c\uc785\ud558\uc790\u201d\ub77c\ub294 \uc811\uadfc \ubc29\uc2dd\uc73c\ub85c \ub3cc\uc544\uac08 \uc218\ub294 \uc5c6\uc2b5\ub2c8\ub2e4.<\/p>\n<p>15.\ub9c8\uc9c0\ub9c9 \uc0dd\uac01:<br \/>\n\uc774 \uc21c\uac04\uc740 \uc6b0\ub9ac\uac00 \ub098\uc911\uc5d0 \ubcc0\uace1\uc810\uc73c\ub85c \uae30\uc5b5\ud560 \uac00\ub2a5\uc131\uc774 \ub192\uc2b5\ub2c8\ub2e4.\ub9c8\uce58 PC\uac00 \uba54\uc778\ud504\ub808\uc784\uc744 \ub35c \uc911\uc694\ud558\uac8c \ub9cc\ub4e4\uac70\ub098, \ud074\ub77c\uc6b0\ub4dc \ucef4\ud4e8\ud305\uc774 \ubaa8\ub4e0 \uac83\uc744 \ubc14\uafe8\ub358 \uac83\ucc98\ub7fc\uc694.AI\ub294 \ub354 \uc811\uadfc \uac00\ub2a5\ud558\uace0, \ud6e8\uc52c \uc800\ub834\ud574\uc9c8 \uac83\uc785\ub2c8\ub2e4.\uc774 \ubcc0\ud654\uac00 \ud604\uc7ac \ud50c\ub808\uc774\uc5b4\ub4e4\uc5d0\uac8c \uc5b4\ub5a4 \uc601\ud5a5\uc744 \ubbf8\uce60\uc9c0\ub294 \uc18d\ub3c4\uc758 \ubb38\uc81c\uc77c \ubfd0\uc785\ub2c8\ub2e4.<\/p><\/blockquote>\n<p>2.<br \/>\n\ud398\uc774\uc2a4\ubd81\uc744 \ud1b5\ud574 AI\uc640 \uad00\ub828\ud55c \ucd5c\uc2e0 \uc5f0\uad6c\ub17c\ubb38\uc744 \uc18c\uac1c\ud558\uc2dc\ub294 <a href=\"https:\/\/www.facebook.com\/seungjoon.choi\" target=\"_blank\" rel=\"noopener\">\ucd5c\uc2b9\uc900<\/a>\ub2d8\uc774 DeepSeek\ub17c\ubb38\uc744 \ubc88\uc5ed\ud574\uc11c \uc62c\ub824\uc8fc\uc168\uc2b5\ub2c8\ub2e4.<\/p>\n<iframe src=\"\/\/docs.google.com\/viewer?url=http%3A%2F%2Fsmallake.kr%2Fwp-content%2Fuploads%2F2025%2F01%2FDeepSeek_R1.pdf&hl=ko&embedded=true\" class=\"gde-frame\" style=\"width:100%; height:500px; border: none;\" scrolling=\"no\"><\/iframe>\n<p class=\"gde-text\"><a href=\"http:\/\/smallake.kr\/wp-content\/uploads\/2025\/01\/DeepSeek_R1.pdf\" class=\"gde-link\">Download (PDF, 1.27MB)<\/a><\/p>\n<blockquote><p>DeepSeek-R1: \uac15\ud654 \ud559\uc2b5\uc744 \ud1b5\ud55c \ub300\uaddc\ubaa8 \uc5b8\uc5b4 \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825 \uc720\ub3c4<\/p>\n<p>\uc694\uc57d<\/p>\n<p>\uc6b0\ub9ac\ub294 \uccab \ubc88\uc9f8 \uc138\ub300\uc758 \ucd94\ub860 \ubaa8\ub378\uc778 DeepSeek-R1-Zero\uc640 DeepSeek-R1\uc744 \uc18c\uac1c\ud569\ub2c8\ub2e4. \uc0ac\uc804 \ub2e8\uacc4\ub85c \uac10\ub3c5\ub41c \uc138\ubc00 \uc870\uc815(SFT) \uc5c6\uc774 \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5(RL)\uc744 \ud1b5\ud574 \ud6c8\ub828\ub41c DeepSeek-R1-Zero\ub294 \ub6f0\uc5b4\ub09c \ucd94\ub860 \ub2a5\ub825\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. RL\uc744 \ud1b5\ud574 DeepSeek-R1-Zero\ub294 \uc790\uc5f0\uc2a4\ub7fd\uac8c \uac15\ub825\ud558\uace0 \ud765\ubbf8\ub85c\uc6b4 \uc5ec\ub7ec \ucd94\ub860 \ud589\ub3d9\uc744 \ubc1c\ud604\ud569\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \uac00\ub3c5\uc131 \uc800\ud558\uc640 \uc5b8\uc5b4 \ud63c\ud569 \ub4f1\uc758 \ubb38\uc81c\uc5d0 \uc9c1\uba74\ud558\uac8c \ub429\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uace0 \ucd94\ub860 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 \ub2e4\ub2e8\uacc4 \ud6c8\ub828\uacfc RL \uc774\uc804\uc758 \ucf5c\ub4dc \uc2a4\ud0c0\ud2b8 \ub370\uc774\ud130\ub97c \ud1b5\ud569\ud55c DeepSeek-R1\uc744 \ub3c4\uc785\ud569\ub2c8\ub2e4. DeepSeek-R1\uc740 \ucd94\ub860 \uc791\uc5c5\uc5d0\uc11c OpenAI-o1-1217\uacfc \ube44\uad50 \uac00\ub2a5\ud55c \uc131\ub2a5\uc744 \ub2ec\uc131\ud569\ub2c8\ub2e4. \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\ub97c \uc9c0\uc6d0\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 Qwen \ubc0f Llama\ub97c \uae30\ubc18\uc73c\ub85c DeepSeek-R1-Zero, DeepSeek-R1, \uadf8\ub9ac\uace0 6\uac1c\uc758 \ubc00\uc9d1 \ubaa8\ub378(1.5B, 7B, 8B, 14B, 32B, 70B)\uc744 \uc624\ud508 \uc18c\uc2a4\ud654\ud569\ub2c8\ub2e4.<br \/>\n&#8211; AIME 2024 (Pass@1)<br \/>\n&#8211; Codeforces (Percentile)<br \/>\n&#8211; GPQA Diamond (Pass@1)<br \/>\n&#8211; MATH-500 (Pass@1)<br \/>\n&#8211; MMLU (Pass@1)<br \/>\n&#8211; SWE-bench Verified (Resolved)<\/p>\n<p>1. \uc11c\ub860<br \/>\n\ucd5c\uadfc \uba87 \ub144 \ub3d9\uc548, \ub300\uaddc\ubaa8 \uc5b8\uc5b4 \ubaa8\ub378(LLMs)\uc740 \uae09\uc18d\ud55c \ubc18\ubcf5\uacfc \uc9c4\ud654\ub97c \uac70\uce58\uace0 \uc788\uc2b5\ub2c8\ub2e4 (Anthropic, 2024; Google, 2024; OpenAI, 2024a), \uc778\uacf5 \uc77c\ubc18 \uc9c0\ub2a5(AGI)\uc5d0 \ub300\ud55c \uaca9\ucc28\ub97c \uc810\uc810 \uc904\uc5ec\uac00\uace0 \uc788\uc2b5\ub2c8\ub2e4.<br \/>\n\ucd5c\uadfc, \uc0ac\ud6c4 \ud6c8\ub828\uc774 \uc804\uccb4 \ud6c8\ub828 \ud30c\uc774\ud504\ub77c\uc778\uc758 \uc911\uc694\ud55c \uad6c\uc131 \uc694\uc18c\ub85c \ub4f1\uc7a5\ud588\uc2b5\ub2c8\ub2e4. \uc774\ub294 \ucd94\ub860 \uc791\uc5c5\uc758 \uc815\ud655\uc131\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uace0, \uc0ac\ud68c\uc801 \uac00\uce58\uc640 \uc77c\uce58\uc2dc\ud0a4\uba70, \uc0ac\uc6a9\uc790 \uc120\ud638\uc5d0 \uc801\uc751\ud558\ub294 \uac83\uc73c\ub85c \ub098\ud0c0\ub0ac\uc73c\uba70, \uc0ac\uc804 \ud6c8\ub828\uc5d0 \ube44\ud574 \uc0c1\ub300\uc801\uc73c\ub85c \uc801\uc740 \uacc4\uc0b0 \uc790\uc6d0\uc744 \uc694\uad6c\ud569\ub2c8\ub2e4. \ucd94\ub860 \ub2a5\ub825\uc758 \ub9e5\ub77d\uc5d0\uc11c, OpenAI\uc758 o1(OpenAI, 2024b) \uc2dc\ub9ac\uc988 \ubaa8\ub378\uc740 \ucd94\ub860 \uacfc\uc815\uc758 \uccb4\uc778-\uc624\ube0c-\uc18c\ud2b8(Chain-of-Thought) \uae38\uc774\ub97c \ub298\ub824 \ucd94\ub860 \uc2dc\uac04 \ud655\uc7a5\uc744 \ub3c4\uc785\ud55c \ucd5c\ucd08\uc758 \ubaa8\ub378\uc774\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \uc811\uadfc\ubc95\uc740 \uc218\ud559, \ucf54\ub529, \uacfc\ud559\uc801 \ucd94\ub860\uacfc \uac19\uc740 \ub2e4\uc591\ud55c \ucd94\ub860 \uc791\uc5c5\uc5d0\uc11c \uc0c1\ub2f9\ud55c \ud5a5\uc0c1\uc744 \ub2ec\uc131\ud588\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \ud6a8\uacfc\uc801\uc778 \ud14c\uc2a4\ud2b8 \uc2dc\uac04 \ud655\uc7a5\uc758 \ubb38\uc81c\ub294 \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\uc5d0\uac8c \uc5ec\uc804\ud788 \uc5f4\ub824 \uc788\ub294 \uc9c8\ubb38\uc785\ub2c8\ub2e4. \uba87\uba87 \uc774\uc804 \uc5f0\uad6c\ub4e4\uc740 \uacfc\uc815 \uae30\ubc18 \ubcf4\uc0c1 \ubaa8\ub378(Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023), \uac15\ud654 \ud559\uc2b5(Kumar et al., 2024), \uadf8\ub9ac\uace0 \ubaac\ud14c \uce74\ub97c\ub85c \ud2b8\ub9ac \ud0d0\uc0c9\uacfc \ube54 \ud0d0\uc0c9\uacfc \uac19\uc740 \uac80\uc0c9 \uc54c\uace0\ub9ac\uc998(Feng et al., 2024; Trinh et al., 2024; Xin et al., 2024)\uc744 \ud3ec\ud568\ud55c \ub2e4\uc591\ud55c \uc811\uadfc \ubc29\uc2dd\uc744 \ud0d0\uad6c\ud588\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \uc774\ub7ec\ud55c \ubc29\ubc95\ub4e4 \uc911 \uc5b4\ub290 \uac83\ub3c4 OpenAI\uc758 o1 \uc2dc\ub9ac\uc988 \ubaa8\ub378\uacfc \ube44\uad50\ud560 \ub9cc\ud55c \uc77c\ubc18 \ucd94\ub860 \uc131\ub2a5\uc744 \ub2ec\uc131\ud558\uc9c0 \ubabb\ud588\uc2b5\ub2c8\ub2e4.<br \/>\n\uc774 \ub17c\ubb38\uc5d0\uc11c, \uc6b0\ub9ac\ub294 \uc21c\uc218\ud55c \uac15\ud654 \ud559\uc2b5(RL)\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc5b8\uc5b4 \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud55c \uccab \ubc88\uc9f8 \ub2e8\uacc4\ub97c \ubc1f\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\uc758 \ubaa9\ud45c\ub294 \uac10\ub3c5\ub41c \ub370\uc774\ud130 \uc5c6\uc774 LLM\uc774 \uc21c\uc218 RL \uacfc\uc815\uc744 \ud1b5\ud574 \uc790\uae30 \uc9c4\ud654\ub97c \ud1b5\ud574 \ucd94\ub860 \ub2a5\ub825\uc744 \uac1c\ubc1c\ud560 \uc218 \uc788\ub294 \uc7a0\uc7ac\ub825\uc744 \ud0d0\uad6c\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uad6c\uccb4\uc801\uc73c\ub85c, \uc6b0\ub9ac\ub294 DeepSeek-V3-Base\ub97c \ubca0\uc774\uc2a4 \ubaa8\ub378\ub85c \uc0ac\uc6a9\ud558\uace0, GRPO(Shao et al., 2024)\ub97c RL \ud504\ub808\uc784\uc6cc\ud06c\ub85c \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378\uc758 \ucd94\ub860 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0b5\ub2c8\ub2e4. \ud6c8\ub828 \uc911, DeepSeek-R1-Zero\ub294 \uc790\uc5f0\uc2a4\ub7fd\uac8c \uac15\ub825\ud558\uace0 \ud765\ubbf8\ub85c\uc6b4 \uc5ec\ub7ec \ucd94\ub860 \ud589\ub3d9\uc744 \ubc1c\ud604\ud588\uc2b5\ub2c8\ub2e4. \uc218\ucc9c \ubc88\uc758 RL \ub2e8\uacc4\ub97c \uac70\uce5c \ud6c4, DeepSeek-R1-Zero\ub294 \ucd94\ub860 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, AIME 2024\uc5d0\uc11c pass@1 \uc810\uc218\ub294 15.6%\uc5d0\uc11c 71.0%\ub85c \uc99d\uac00\ud588\uc73c\uba70, \ub2e4\uc218 \ud22c\ud45c\ub97c \ud1b5\ud574 \uc810\uc218\uac00 86.7%\ub85c \ub354\uc6b1 \ud5a5\uc0c1\ub418\uc5b4 OpenAI-o1-0912\uc758 \uc131\ub2a5\uacfc \uc77c\uce58\ud569\ub2c8\ub2e4.<br \/>\n\uadf8\ub7ec\ub098 DeepSeek-R1-Zero\ub294 \uac00\ub3c5\uc131 \uc800\ud558\uc640 \uc5b8\uc5b4 \ud63c\ud569\uacfc \uac19\uc740 \ubb38\uc81c\uc5d0 \uc9c1\uba74\ud558\uac8c \ub429\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uace0 \ucd94\ub860 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 \uc18c\ub7c9\uc758 \ucf5c\ub4dc \uc2a4\ud0c0\ud2b8 \ub370\uc774\ud130\uc640 \ub2e4\ub2e8\uacc4 \ud6c8\ub828 \ud30c\uc774\ud504\ub77c\uc778\uc744 \ud1b5\ud569\ud55c DeepSeek-R1\uc744 \ub3c4\uc785\ud569\ub2c8\ub2e4. \uad6c\uccb4\uc801\uc73c\ub85c, \uc6b0\ub9ac\ub294 DeepSeek-V3-Base \ubaa8\ub378\uc744 \uc138\ubc00 \uc870\uc815\ud558\uae30 \uc704\ud574 \uc218\ucc9c \uac1c\uc758 \ucf5c\ub4dc \uc2a4\ud0c0\ud2b8 \ub370\uc774\ud130\ub97c \uc218\uc9d1\ud558\ub294 \uac83\uc73c\ub85c \uc2dc\uc791\ud569\ub2c8\ub2e4. \uadf8 \ud6c4, DeepSeek-R1-Zero\uc640 \uac19\uc740 \ucd94\ub860 \uc9c0\ud5a5 RL\uc744 \uc218\ud589\ud569\ub2c8\ub2e4. RL \uacfc\uc815\uc5d0\uc11c \uc218\ub834\uc5d0 \uac00\uae4c\uc6cc\uc9c8 \ub54c, \uc6b0\ub9ac\ub294 RL \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0\uc11c \uac70\ubd80 \uc0d8\ud50c\ub9c1\uc744 \ud1b5\ud574 \uc0c8\ub85c\uc6b4 SFT \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \uae00\uc4f0\uae30, \uc0ac\uc2e4 QA, \uc790\uae30 \uc778\uc2dd\uacfc \uac19\uc740 \ub3c4\uba54\uc778\uc5d0\uc11c DeepSeek-V3\uc758 \uac10\ub3c5\ub41c \ub370\uc774\ud130\uc640 \uacb0\ud569\ud558\uc5ec DeepSeek-V3-Base \ubaa8\ub378\uc744 \uc7ac\ud6c8\ub828\ud569\ub2c8\ub2e4. \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub85c \uc138\ubc00 \uc870\uc815\ud55c \ud6c4, \uccb4\ud06c\ud3ec\uc778\ud2b8\ub294 \ubaa8\ub4e0 \uc2dc\ub098\ub9ac\uc624\uc758 \ud504\ub86c\ud504\ud2b8\ub97c \uace0\ub824\ud558\uc5ec \ucd94\uac00 RL \uacfc\uc815\uc744 \uac70\uce69\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ub2e8\uacc4\ub97c \uac70\uce5c \ud6c4, \uc6b0\ub9ac\ub294 DeepSeek-R1\ub85c \uc9c0\uce6d\ub418\ub294 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc5bb\uc5c8\uc73c\uba70, \uc774\ub294 OpenAI-o1-1217\uacfc \ub3d9\ub4f1\ud55c \uc131\ub2a5\uc744 \ub2ec\uc131\ud569\ub2c8\ub2e4.<br \/>\n\uc6b0\ub9ac\ub294 \ub610\ud55c DeepSeek-R1\uc5d0\uc11c \uc791\uc740 \ubc00\uc9d1 \ubaa8\ub378\ub85c\uc758 \uc99d\ub958\ub97c \ud0d0\uad6c\ud569\ub2c8\ub2e4. Qwen2.5-32B(Qwen, 2024b)\ub97c \ubca0\uc774\uc2a4 \ubaa8\ub378\ub85c \uc0ac\uc6a9\ud558\uc5ec, DeepSeek-R1\uc5d0\uc11c \uc9c1\uc811 \uc99d\ub958\ud558\ub294 \uac83\uc740 RL\uc744 \uc801\uc6a9\ud558\ub294 \uac83\ubcf4\ub2e4 \ub354 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc785\ub2c8\ub2e4. \uc774\ub294 \ub354 \ud070 \ubca0\uc774\uc2a4 \ubaa8\ub378\uc774 \ubc1c\uacac\ud55c \ucd94\ub860 \ud328\ud134\uc774 \ucd94\ub860 \ub2a5\ub825 \ud5a5\uc0c1\uc5d0 \uc911\uc694\ud568\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uc99d\ub958\ub41c Qwen\uacfc Llama(Dubey et al., 2024) \uc2dc\ub9ac\uc988\ub97c \uc624\ud508 \uc18c\uc2a4\ud654\ud569\ub2c8\ub2e4. \ud2b9\ud788, \uc6b0\ub9ac\uc758 \uc99d\ub958\ub41c 14B \ubaa8\ub378\uc740 \ucd5c\ucca8\ub2e8 \uc624\ud508 \uc18c\uc2a4 QwQ-32B-Preview(Qwen, 2024a)\uc744 \ud070 \ud3ed\uc73c\ub85c \ub2a5\uac00\ud558\uba70, \uc99d\ub958\ub41c 32B \ubc0f 70B \ubaa8\ub378\uc740 \ubc00\uc9d1 \ubaa8\ub378 \uc911 \ucd94\ub860 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c \uc0c8\ub85c\uc6b4 \uae30\ub85d\uc744 \uc138\uc6c1\ub2c8\ub2e4.<\/p>\n<p>1.1. \uae30\uc5ec\ub3c4<br \/>\n\uc0ac\ud6c4 \ud6c8\ub828: \ubca0\uc774\uc2a4 \ubaa8\ub378\uc5d0 \ub300\ud55c \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5<br \/>\n&#8211; \uc6b0\ub9ac\ub294 \uc0ac\uc804 \ub2e8\uacc4\ub85c \uac10\ub3c5\ub41c \uc138\ubc00 \uc870\uc815(SFT)\uc5d0 \uc758\uc874\ud558\uc9c0 \uc54a\uace0 \ubca0\uc774\uc2a4 \ubaa8\ub378\uc5d0 \uc9c1\uc811 \uac15\ud654 \ud559\uc2b5(RL)\uc744 \uc801\uc6a9\ud569\ub2c8\ub2e4. \uc774 \uc811\uadfc\ubc95\uc744 \ud1b5\ud574 \ubaa8\ub378\uc740 \ubcf5\uc7a1\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud55c \uccb4\uc778-\uc624\ube0c-\uc18c\ud2b8(CoT)\ub97c \ud0d0\uc0c9\ud560 \uc218 \uc788\uc73c\uba70, \uadf8 \uacb0\uacfc DeepSeek-R1-Zero\uac00 \uac1c\ubc1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. DeepSeek-R1-Zero\ub294 \uc790\uae30 \uac80\uc99d, \ubc18\uc131, \uae34 CoT \uc0dd\uc131\uacfc \uac19\uc740 \ub2a5\ub825\uc744 \ubcf4\uc5ec\uc8fc\uba70, \uc774\ub294 \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\uc5d0 \uc911\uc694\ud55c \uc774\uc815\ud45c\ub97c \uc138\uc6c1\ub2c8\ub2e4. \ud2b9\ud788, \uc774\ub294 SFT \uc5c6\uc774 \uc21c\uc218\ud558\uac8c RL\uc744 \ud1b5\ud574 LLM\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \uc720\ub3c4\ud560 \uc218 \uc788\uc74c\uc744 \uac80\uc99d\ud55c \ucd5c\ucd08\uc758 \uacf5\uac1c \uc5f0\uad6c\uc785\ub2c8\ub2e4. \uc774 \ub3cc\ud30c\uad6c\ub294 \ud5a5\ud6c4 \uc774 \ubd84\uc57c\uc758 \ubc1c\uc804\uc5d0 \uae38\uc744 \uc5f4\uc5b4\uc90d\ub2c8\ub2e4.<br \/>\n&#8211; \uc6b0\ub9ac\ub294 DeepSeek-R1\uc744 \uac1c\ubc1c\ud558\uae30 \uc704\ud55c \ud30c\uc774\ud504\ub77c\uc778\uc744 \uc18c\uac1c\ud569\ub2c8\ub2e4. \uc774 \ud30c\uc774\ud504\ub77c\uc778\uc740 \ud5a5\uc0c1\ub41c \ucd94\ub860 \ud328\ud134\uc744 \ubc1c\uacac\ud558\uace0 \uc778\uac04\uc758 \uc120\ud638\ub3c4\uc5d0 \ub9de\ucd94\uae30 \uc704\ud55c \ub450 \ub2e8\uacc4\uc758 RL\uacfc \ubaa8\ub378\uc758 \ucd94\ub860 \ubc0f \ube44\ucd94\ub860 \ub2a5\ub825\uc758 \uc2dc\ub4dc \uc5ed\ud560\uc744 \ud558\ub294 \ub450 \ub2e8\uacc4\uc758 SFT\ub97c \ud1b5\ud569\ud569\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uc774 \ud30c\uc774\ud504\ub77c\uc778\uc774 \ub354 \ub098\uc740 \ubaa8\ub378\uc744 \uc0dd\uc131\ud568\uc73c\ub85c\uc368 \uc0b0\uc5c5\uc5d0 \uae30\uc5ec\ud560 \uac83\uc774\ub77c \ubbff\uc2b5\ub2c8\ub2e4.<br \/>\n\uc99d\ub958: \uc791\uc740 \ubaa8\ub378\ub3c4 \uac15\ub825\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4<br \/>\n&#8211; \uc6b0\ub9ac\ub294 \ub354 \ud070 \ubaa8\ub378\uc758 \ucd94\ub860 \ud328\ud134\uc744 \uc791\uc740 \ubaa8\ub378\ub85c \uc99d\ub958\ud560 \uc218 \uc788\uc74c\uc744 \uc785\uc99d\ud569\ub2c8\ub2e4. \uc774\ub294 \uc791\uc740 \ubaa8\ub378\uc5d0 RL\uc744 \uc801\uc6a9\ud558\uc5ec \ubc1c\uacac\ub41c \ucd94\ub860 \ud328\ud134\ubcf4\ub2e4 \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ubc1c\ud718\ud569\ub2c8\ub2e4. \uc624\ud508 \uc18c\uc2a4\uc778 DeepSeek-R1\uacfc \uadf8 API\ub294 \ud5a5\ud6c4 \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\uac00 \ub354 \ub098\uc740 \uc791\uc740 \ubaa8\ub378\uc744 \uc99d\ub958\ud558\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub420 \uac83\uc785\ub2c8\ub2e4.<br \/>\n&#8211; DeepSeek-R1\uc774 \uc0dd\uc131\ud55c \ucd94\ub860 \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\uc5d0\uc11c \ub110\ub9ac \uc0ac\uc6a9\ub418\ub294 \uc5ec\ub7ec \ubc00\uc9d1 \ubaa8\ub378\uc744 \uc138\ubc00 \uc870\uc815\ud588\uc2b5\ub2c8\ub2e4. \ud3c9\uac00 \uacb0\uacfc\ub294 \uc99d\ub958\ub41c \uc791\uc740 \ubc00\uc9d1 \ubaa8\ub378\ub4e4\uc774 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubc1c\ud718\ud568\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, DeepSeek-R1-Distill-Qwen-7B\ub294 AIME 2024\uc5d0\uc11c 55.5%\ub97c \ub2ec\uc131\ud558\uc5ec QwQ-32B-Preview\ub97c \ub2a5\uac00\ud569\ub2c8\ub2e4. \ub610\ud55c, DeepSeek-R1-Distill-Qwen-32B\ub294 AIME 2024\uc5d0\uc11c 72.6%, MATH-500\uc5d0\uc11c 94.3%, LiveCodeBench\uc5d0\uc11c 57.2%\ub97c \uae30\ub85d\ud558\uc600\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uacb0\uacfc\ub294 \uc774\uc804\uc758 \uc624\ud508 \uc18c\uc2a4 \ubaa8\ub378\ub4e4\uc744 \ud06c\uac8c \ub2a5\uac00\ud558\uba70, o1-mini\uc640 \ube44\uad50\ud574\ub3c4 \uc190\uc0c9\uc774 \uc5c6\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 Qwen2.5\uc640 Llama3 \uc2dc\ub9ac\uc988\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \uc99d\ub958\ub41c 1.5B, 7B, 8B, 14B, 32B, 70B \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \ucee4\ubba4\ub2c8\ud2f0\uc5d0 \uc624\ud508 \uc18c\uc2a4\ud654\ud569\ub2c8\ub2e4.<\/p>\n<p>1.2. \ud3c9\uac00 \uacb0\uacfc \uc694\uc57d<br \/>\n&#8211; \ucd94\ub860 \uc791\uc5c5:<br \/>\n1. DeepSeek-R1\uc740 AIME 2024\uc5d0\uc11c Pass@1 \uc810\uc218 79.8%\ub97c \ub2ec\uc131\ud558\uc5ec OpenAI-o1-1217\uc744 \uc57d\uac04 \ub2a5\uac00\ud588\uc2b5\ub2c8\ub2e4. MATH-500\uc5d0\uc11c\ub294 97.3%\ub77c\ub294 \uc778\uc0c1\uc801\uc778 \uc810\uc218\ub97c \uae30\ub85d\ud558\uc5ec OpenAI-o1-1217\uacfc \ub3d9\ub4f1\ud55c \uc131\ub2a5\uc744 \ubcf4\uc600\uc73c\uba70, \ub2e4\ub978 \ubaa8\ub378\ub4e4\ubcf4\ub2e4 \ud06c\uac8c \ub6f0\uc5b4\ub0ac\uc2b5\ub2c8\ub2e4.<br \/>\n2. \ucf54\ub529 \uad00\ub828 \uc791\uc5c5\uc5d0\uc11c\ub294 DeepSeek-R1\uc774 Codeforces\uc5d0\uc11c 2,029 Elo \ub4f1\uae09\uc744 \ub2ec\uc131\ud558\uc5ec \uacbd\uc7c1 \ucc38\uac00\uc790\uc758 96.3%\ub97c \ub2a5\uac00\ud558\ub294 \uc804\ubb38\uac00 \uc218\uc900\uc758 \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc8fc\uc5c8\uc2b5\ub2c8\ub2e4. \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1 \uad00\ub828 \uc791\uc5c5\uc5d0\uc11c\ub294 DeepSeek-R1\uc774 DeepSeek-V3\ubcf4\ub2e4 \uc57d\uac04 \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ubcf4\uc5ec \uc2e4\uc81c \uc138\uacc4\uc758 \uac1c\ubc1c\uc790 \uc791\uc5c5\uc5d0 \ub3c4\uc6c0\uc774 \ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<br \/>\n&#8211; \uc9c0\uc2dd:<br \/>\nMMLU, MMLU-Pro, GPQA Diamond\uc640 \uac19\uc740 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c DeepSeek-R1\uc740 MMLU\uc5d0\uc11c 90.8%, MMLU-Pro\uc5d0\uc11c 84.0%, GPQA Diamond\uc5d0\uc11c 71.5%\uc758 \ub6f0\uc5b4\ub09c \uc810\uc218\ub97c \uae30\ub85d\ud558\uc5ec DeepSeek-V3\ub97c \ud06c\uac8c \ub2a5\uac00\ud588\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c OpenAI-o1-1217\ubcf4\ub2e4\ub294 \uc57d\uac04 \ub0ae\uc740 \uc131\ub2a5\uc744 \ubcf4\uc600\uc9c0\ub9cc, \ub2e4\ub978 \ud3d0\uc1c4\ud615 \ubaa8\ub378\ub4e4\uc744 \ub2a5\uac00\ud558\uc5ec \uad50\uc721 \uc791\uc5c5\uc5d0\uc11c\uc758 \uacbd\uc7c1 \uc6b0\uc704\ub97c \uc785\uc99d\ud588\uc2b5\ub2c8\ub2e4. \uc0ac\uc2e4 \uae30\ubc18 \ubca4\uce58\ub9c8\ud06c\uc778 SimpleQA\uc5d0\uc11c\ub294 DeepSeek-R1\uc774 DeepSeek-V3\ub97c \ub2a5\uac00\ud558\uc5ec \uc0ac\uc2e4 \uae30\ubc18 \ucffc\ub9ac\ub97c \ucc98\ub9ac\ud558\ub294 \ub2a5\ub825\uc744 \ubcf4\uc5ec\uc8fc\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c\ub3c4 OpenAI-o1\uc740 DeepSeek-R1\uc744 \ub2a5\uac00\ud558\ub294 \uc720\uc0ac\ud55c \ucd94\uc138\ub97c \ubcf4\uc600\uc2b5\ub2c8\ub2e4.<br \/>\n&#8211; \uae30\ud0c0:<br \/>\nDeepSeek-R1\uc740 \ucc3d\uc758\uc801 \uae00\uc4f0\uae30, \uc77c\ubc18 \uc9c8\ubb38 \uc751\ub2f5, \ud3b8\uc9d1, \uc694\uc57d \ub4f1 \ub2e4\uc591\ud55c \uc791\uc5c5\uc5d0\uc11c\ub3c4 \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubc1c\ud718\ud569\ub2c8\ub2e4. AlpacaEval 2.0\uc5d0\uc11c \uae38\uc774 \uc870\uc808 \uc2b9\ub960 87.6%, ArenaHard\uc5d0\uc11c \uc2b9\ub960 92.3%\ub97c \uae30\ub85d\ud558\uc5ec \ube44\uc2dc\ud5d8 \uc9c0\ud5a5\uc801\uc778 \ucffc\ub9ac\ub97c \uc9c0\ub2a5\uc801\uc73c\ub85c \ucc98\ub9ac\ud558\ub294 \uac15\ub825\ud55c \ub2a5\ub825\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \ub610\ud55c, DeepSeek-R1\uc740 \uae34 \ucee8\ud14d\uc2a4\ud2b8 \uc774\ud574\uac00 \ud544\uc694\ud55c \uc791\uc5c5\uc5d0\uc11c\ub3c4 \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc774\uba70, \uae34 \ucee8\ud14d\uc2a4\ud2b8 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c DeepSeek-V3\ub97c \uc0c1\ub2f9\ud788 \ub2a5\uac00\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<p>2. \uc811\uadfc \ubc29\ubc95<br \/>\n2.1. \uac1c\uc694<br \/>\n\uc774\uc804 \uc5f0\uad6c\ub4e4\uc740 \ubaa8\ub378 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574 \ub300\ub7c9\uc758 \uac10\ub3c5 \ub370\uc774\ud130\ub97c \ud06c\uac8c \uc758\uc874\ud574\uc654\uc2b5\ub2c8\ub2e4. \ubcf8 \uc5f0\uad6c\uc5d0\uc11c\ub294 \uac10\ub3c5\ub41c \uc138\ubc00 \uc870\uc815(SFT) \uc5c6\uc774\ub3c4 \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5(RL)\uc744 \ud1b5\ud574 \ucd94\ub860 \ub2a5\ub825\uc744 \ud06c\uac8c \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc74c\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \ub354\uc6b1\uc774, \uc18c\ub7c9\uc758 \ucf5c\ub4dc \uc2a4\ud0c0\ud2b8 \ub370\uc774\ud130\ub97c \ud3ec\ud568\ud568\uc73c\ub85c\uc368 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c \uc139\uc158\uc5d0\uc11c\ub294 \ub2e4\uc74c\uc744 \uc81c\uc2dc\ud569\ub2c8\ub2e4: (1) SFT \ub370\uc774\ud130 \uc5c6\uc774 \ubca0\uc774\uc2a4 \ubaa8\ub378\uc5d0 \uc9c1\uc811 RL\uc744 \uc801\uc6a9\ud55c DeepSeek-R1-Zero, (2) \uc218\ucc9c \uac1c\uc758 \uae34 \uccb4\uc778-\uc624\ube0c-\uc18c\ud2b8(CoT) \uc608\uc81c\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc138\ubc00 \uc870\uc815\ub41c \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0\uc11c RL\uc744 \uc2dc\uc791\ud558\ub294 DeepSeek-R1, (3) DeepSeek-R1\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \uc791\uc740 \ubc00\uc9d1 \ubaa8\ub378\ub85c \uc99d\ub958\ud558\ub294 \ubc29\ubc95.<br \/>\n2.2. DeepSeek-R1-Zero: \ubca0\uc774\uc2a4 \ubaa8\ub378\uc5d0 \ub300\ud55c \uac15\ud654 \ud559\uc2b5<br \/>\n\uac15\ud654 \ud559\uc2b5\uc740 \uc774\uc804 \uc5f0\uad6c\ub4e4\uc5d0\uc11c \ucd94\ub860 \uc791\uc5c5\uc5d0\uc11c \uc0c1\ub2f9\ud55c \ud6a8\uacfc\ub97c \uc785\uc99d\ud588\uc2b5\ub2c8\ub2e4 (Shao et al., 2024; Wang et al., 2023). \uadf8\ub7ec\ub098 \uc774\ub7ec\ud55c \uc5f0\uad6c\ub4e4\uc740 \uac10\ub3c5 \ub370\uc774\ud130\uc5d0 \ud06c\uac8c \uc758\uc874\ud588\uc73c\uba70, \uc774\ub294 \uc218\uc9d1\ud558\ub294 \ub370 \uc2dc\uac04\uc774 \ub9ce\uc774 \uc18c\uc694\ub429\ub2c8\ub2e4. \uc774 \uc139\uc158\uc5d0\uc11c\ub294 \uc21c\uc218\ud55c \uac15\ud654 \ud559\uc2b5 \uacfc\uc815\uc744 \ud1b5\ud574 \uac10\ub3c5 \ub370\uc774\ud130 \uc5c6\uc774 LLM\uc774 \ucd94\ub860 \ub2a5\ub825\uc744 \uac1c\ubc1c\ud560 \uc218 \uc788\ub294 \uc7a0\uc7ac\ub825\uc744 \ud0d0\uad6c\ud569\ub2c8\ub2e4. \uba3c\uc800 \uc6b0\ub9ac\uc758 \uac15\ud654 \ud559\uc2b5 \uc54c\uace0\ub9ac\uc998\uc5d0 \ub300\ud55c \uac04\ub7b5\ud55c \uac1c\uc694\ub97c \uc81c\uacf5\ud55c \ud6c4, \uba87 \uac00\uc9c0 \ud765\ubbf8\ub85c\uc6b4 \uacb0\uacfc\ub97c \uc81c\uc2dc\ud558\uba70, \uc774\ub294 \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\uc5d0 \uc720\uc6a9\ud55c \ud1b5\ucc30\uc744 \uc81c\uacf5\ud560 \uac83\uc73c\ub85c \uae30\ub300\ud569\ub2c8\ub2e4.<\/p>\n<p>2.2.1. \uac15\ud654 \ud559\uc2b5 \uc54c\uace0\ub9ac\uc998<br \/>\n\uadf8\ub8f9 \uc0c1\ub300 \uc815\ucc45 \ucd5c\uc801\ud654(Group Relative Policy Optimization, GRPO)<br \/>\nRL\uc758 \ud6c8\ub828 \ube44\uc6a9\uc744 \uc808\uac10\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 \uc815\ucc45 \ubaa8\ub378\uacfc \ub3d9\uc77c\ud55c \ud06c\uae30\uc758 \ube44\ud3c9\uac00 \ubaa8\ub378\uc744 \uc0dd\ub7b5\ud558\uace0 \uadf8\ub8f9 \uc810\uc218\ub85c\ubd80\ud130 \uae30\uc900\uc120\uc744 \ucd94\uc815\ud558\ub294 GRPO(Shao et al., 2024)\ub97c \ucc44\ud0dd\ud569\ub2c8\ub2e4. \uad6c\uccb4\uc801\uc73c\ub85c, \uac01 \uc9c8\ubb38\uc5d0 \ub300\ud574 GRPO\ub294 \uae30\uc874 \uc815\ucc45\uc5d0\uc11c \ucd9c\ub825 \uadf8\ub8f9\uc744 \uc0d8\ud50c\ub9c1\ud55c \ud6c4, \ub2e4\uc74c \ubaa9\ud45c\ub97c \ucd5c\ub300\ud654\ud558\uc5ec \uc815\ucc45 \ubaa8\ub378\ub97c \ucd5c\uc801\ud654\ud569\ub2c8\ub2e4:<\/p>\n<p>2.2.2. \ubcf4\uc0c1 \ubaa8\ub378\ub9c1<br \/>\n\ubcf4\uc0c1\uc740 \ud6c8\ub828 \uc2e0\ud638\uc758 \uc6d0\ucc9c\uc73c\ub85c, RL\uc758 \ucd5c\uc801\ud654 \ubc29\ud5a5\uc744 \uacb0\uc815\ud569\ub2c8\ub2e4. DeepSeek-R1-Zero\ub97c \ud6c8\ub828\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 \uc8fc\ub85c \ub450 \uac00\uc9c0 \uc720\ud615\uc758 \ubcf4\uc0c1\uc73c\ub85c \uad6c\uc131\ub41c \uaddc\uce59 \uae30\ubc18 \ubcf4\uc0c1 \uc2dc\uc2a4\ud15c\uc744 \ucc44\ud0dd\ud588\uc2b5\ub2c8\ub2e4:<br \/>\n&#8211; \uc815\ud655\ub3c4 \ubcf4\uc0c1: \uc815\ud655\ub3c4 \ubcf4\uc0c1 \ubaa8\ub378\uc740 \uc751\ub2f5\uc774 \uc62c\ubc14\ub978\uc9c0 \ud3c9\uac00\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uacb0\uc815\ub860\uc801 \uacb0\uacfc\ub97c \uac16\ub294 \uc218\ud559 \ubb38\uc81c\uc758 \uacbd\uc6b0, \ubaa8\ub378\uc740 \uc9c0\uc815\ub41c \ud615\uc2dd(\uc608: \uc0c1\uc790 \uc548\uc5d0 \ucd5c\uc885 \ub2f5)\uc744 \uc81c\uacf5\ud574\uc57c \ud558\uba70, \uc774\ub97c \ud1b5\ud574 \uc62c\ubc14\ub984\uc744 \uc2e0\ub8b0\ud560 \uc218 \uc788\ub294 \uaddc\uce59 \uae30\ubc18\uc73c\ub85c \uac80\uc99d\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc720\uc0ac\ud558\uac8c, LeetCode \ubb38\uc81c\uc758 \uacbd\uc6b0, \ucef4\ud30c\uc77c\ub7ec\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc0ac\uc804 \uc815\uc758\ub41c \ud14c\uc2a4\ud2b8 \ucf00\uc774\uc2a4\uc5d0 \uae30\ubc18\ud55c \ud53c\ub4dc\ubc31\uc744 \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<br \/>\n&#8211; \ud615\uc2dd \ubcf4\uc0c1: \uc815\ud655\ub3c4 \ubcf4\uc0c1 \ubaa8\ub378 \uc678\uc5d0\ub3c4, \uc6b0\ub9ac\ub294 \ubaa8\ub378\uc774 \uc0ac\uace0 \uacfc\uc815\uc744 &#8221;\uc640 &#8221; \ud0dc\uadf8 \uc0ac\uc774\uc5d0 \ub450\ub3c4\ub85d \uac15\uc81c\ud558\ub294 \ud615\uc2dd \ubcf4\uc0c1 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<br \/>\n\uc6b0\ub9ac\ub294 DeepSeek-R1-Zero\ub97c \uac1c\ubc1c\ud560 \ub54c \uacb0\uacfc\ub098 \uacfc\uc815 \uc2e0\uacbd\ub9dd \ubcf4\uc0c1 \ubaa8\ub378\uc744 \uc801\uc6a9\ud558\uc9c0 \uc54a\uc558\uc2b5\ub2c8\ub2e4. \uadf8 \uc774\uc720\ub294 \uc2e0\uacbd\ub9dd \ubcf4\uc0c1 \ubaa8\ub378\uc774 \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5 \uacfc\uc815\uc5d0\uc11c \ubcf4\uc0c1 \ud574\ud0b9\uc758 \ubb38\uc81c\uac00 \ubc1c\uc0dd\ud560 \uc218 \uc788\uc73c\uba70, \ubcf4\uc0c1 \ubaa8\ub378\uc744 \uc7ac\ud6c8\ub828\ud558\ub294 \ub370 \ucd94\uac00\uc801\uc778 \ud6c8\ub828 \uc790\uc6d0\uc774 \ud544\uc694\ud558\uace0 \uc804\uccb4 \ud6c8\ub828 \ud30c\uc774\ud504\ub77c\uc778\uc744 \ubcf5\uc7a1\ud558\uac8c \ub9cc\ub4e4\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4.<\/p>\n<p>2.2.3. \ud6c8\ub828 \ud15c\ud50c\ub9bf<br \/>\nDeepSeek-R1-Zero\ub97c \ud6c8\ub828\ud558\uae30 \uc704\ud574, \uc800\ud76c\ub294 \uae30\ubcf8 \ubaa8\ub378\uc774 \uc8fc\uc5b4\uc9c4 \uc9c0\uce68\uc744 \ub530\ub97c \uc218 \uc788\ub3c4\ub85d \uc720\ub3c4\ud558\ub294 \uac04\ub2e8\ud55c \ud15c\ud50c\ub9bf\uc744 \uc124\uacc4\ud558\ub294 \uac83\ubd80\ud130 \uc2dc\uc791\ud569\ub2c8\ub2e4. Table 1\uc5d0\uc11c \ubcfc \uc218 \uc788\ub4ef\uc774, \uc774 \ud15c\ud50c\ub9bf\uc740 DeepSeek-R1-Zero\uac00 \uba3c\uc800 \ucd94\ub860 \uacfc\uc815\uc744 \uc0dd\uc131\ud55c \ud6c4 \ucd5c\uc885 \ub2f5\ubcc0\uc744 \ub3c4\ucd9c\ud558\ub3c4\ub85d \uc694\uad6c\ud569\ub2c8\ub2e4. \uc800\ud76c\ub294 \uc758\ub3c4\uc801\uc73c\ub85c \uc774\ub7ec\ud55c \uad6c\uc870\uc801 \ud615\uc2dd\uc5d0 \ub300\ud55c \uc81c\ud55c\ub9cc\uc744 \ub450\uace0, \ubc18\uc131\uc801 \uc0ac\uace0\ub97c \ud544\uc218\uc801\uc73c\ub85c \uc694\uad6c\ud558\uac70\ub098 \ud2b9\uc815 \ubb38\uc81c \ud574\uacb0 \uc804\ub7b5\uc744 \uc7a5\ub824\ud558\ub294 \ub4f1\uc758 \ub0b4\uc6a9\uc801\uc778 \ud3b8\ud5a5\uc744 \ubc30\uc81c\ud558\uc600\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uac15\ud654 \ud559\uc2b5(RL) \uacfc\uc815\uc5d0\uc11c \ubaa8\ub378\uc774 \uc790\uc5f0\uc2a4\ub7fd\uac8c \ubc1c\uc804\ud558\ub294 \ubaa8\uc2b5\uc744 \uc815\ud655\ud558\uac8c \uad00\ucc30\ud560 \uc218 \uc788\ub3c4\ub85d \ud558\uc600\uc2b5\ub2c8\ub2e4.<\/p>\n<p>2.2.4. DeepSeek-R1-Zero\uc758 \uc131\ub2a5, \uc790\uae30 \uc9c4\ud654 \uacfc\uc815 \ubc0f \uc544\ud558 \uc21c\uac04<\/p>\n<p>\uadf8\ub9bc 2\ub294 \uac15\ud654 \ud559\uc2b5(RL) \ud6c8\ub828 \uacfc\uc815 \uc804\ubc18\uc5d0 \uac78\uccd0 AIME 2024 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c DeepSeek-R1-Zero\uc758 \uc131\ub2a5 \uada4\uc801\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uadf8\ub9bc\uc5d0\uc11c \ubcfc \uc218 \uc788\ub4ef\uc774, DeepSeek-R1-Zero\ub294 RL \ud6c8\ub828\uc774 \uc9c4\ud589\ub428\uc5d0 \ub530\ub77c \uc131\ub2a5\uc774 \uafb8\uc900\ud558\uace0 \uc9c0\uc18d\uc801\uc73c\ub85c \ud5a5\uc0c1\ub418\ub294 \ubaa8\uc2b5\uc744 \ubcf4\uc785\ub2c8\ub2e4. \ud2b9\ud788, AIME 2024\uc5d0\uc11c\uc758 \ud3c9\uade0 pass@1 \uc810\uc218\uac00 \ucd08\uae30 15.6%\uc5d0\uc11c \uc778\uc0c1\uc801\uc778 71.0%\ub85c \ud06c\uac8c \uc99d\uac00\ud558\uc5ec OpenAI-o1-0912\uc640 \uc720\uc0ac\ud55c \uc131\ub2a5 \uc218\uc900\uc5d0 \ub3c4\ub2ec\ud55c \uac83\uc774 \uc8fc\ubaa9\ud560 \ub9cc\ud569\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc0c1\ub2f9\ud55c \uac1c\uc120\uc740 \uc2dc\uac04\uc774 \uc9c0\ub0a8\uc5d0 \ub530\ub77c \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ucd5c\uc801\ud654\ud558\ub294 \uc6b0\ub9ac\uc758 RL \uc54c\uace0\ub9ac\uc998\uc758 \ud6a8\uc728\uc131\uc744 \uac15\uc870\ud569\ub2c8\ub2e4.<br \/>\n\ud45c 2\ub294 \ub2e4\uc591\ud55c \ucd94\ub860 \uad00\ub828 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c DeepSeek-R1-Zero\uc640 OpenAI\uc758 o1-0912 \ubaa8\ub378\uc744 \ube44\uad50 \ubd84\uc11d\ud55c \uac83\uc785\ub2c8\ub2e4.<br \/>\n\ud45c 2 | DeepSeek-R1-Zero\uc640 OpenAI o1 \ubaa8\ub378\uc758 \ucd94\ub860 \uad00\ub828 \ubca4\uce58\ub9c8\ud06c \ube44\uad50<br \/>\n\uadf8\ub9bc 2 | \ud6c8\ub828 \uc911 DeepSeek-R1-Zero\uc758 AIME \uc815\ud655\ub3c4. \uac01 \uc9c8\ubb38\uc5d0 \ub300\ud574 16\uac1c\uc758 \uc751\ub2f5\uc744 \uc0d8\ud50c\ub9c1\ud558\uace0 \uc804\uccb4 \ud3c9\uade0 \uc815\ud655\ub3c4\ub97c \uacc4\uc0b0\ud558\uc5ec \uc548\uc815\uc801\uc778 \ud3c9\uac00\ub97c \ubcf4\uc7a5\ud569\ub2c8\ub2e4.<br \/>\nDeepSeek-R1-Zero\ub294 \uc9c0\ub3c4 \ubbf8\uc138 \uc870\uc815 \ub370\uc774\ud130 \uc5c6\uc774\ub3c4 \uac15\ub825\ud55c \ucd94\ub860 \ub2a5\ub825\uc744 \ud68d\ub4dd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 RL\ub9cc\uc73c\ub85c\ub3c4 \ubaa8\ub378\uc774 \ud6a8\uacfc\uc801\uc73c\ub85c \ud559\uc2b5\ud558\uace0 \uc77c\ubc18\ud654\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \uac15\uc870\ud558\ub294 \uc8fc\ubaa9\ud560 \ub9cc\ud55c \uc131\uacfc\uc785\ub2c8\ub2e4. \ub610\ud55c, DeepSeek-R1-Zero\uc758 \uc131\ub2a5\uc740 \ub2e4\uc218\uacb0 \ud22c\ud45c \ubc29\uc2dd\uc744 \uc801\uc6a9\ud568\uc73c\ub85c\uc368 \ub354\uc6b1 \ud5a5\uc0c1\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, AIME \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c \ub2e4\uc218\uacb0 \ud22c\ud45c\ub97c \uc801\uc6a9\ud558\uba74 DeepSeek-R1-Zero\uc758 \uc131\ub2a5\uc774 71.0%\uc5d0\uc11c 86.7%\ub85c \uc0c1\uc2b9\ud558\uc5ec OpenAI-o1-0912\uc758 \uc131\ub2a5\uc744 \ub2a5\uac00\ud558\uac8c \ub429\ub2c8\ub2e4. \ub2e4\uc218\uacb0 \ud22c\ud45c\uc758 \uc720\ubb34\uc5d0 \uad00\uacc4\uc5c6\uc774 DeepSeek-R1-Zero\uac00 \uc774\ub7ec\ud55c \uacbd\uc7c1\ub825 \uc788\ub294 \uc131\ub2a5\uc744 \ub2ec\uc131\ud560 \uc218 \uc788\ub2e4\ub294 \uac83\uc740 \ubaa8\ub378\uc758 \uac15\ub825\ud55c \uae30\ucd08 \ub2a5\ub825\uacfc \ucd94\ub860 \uacfc\uc81c\uc5d0\uc11c\uc758 \ucd94\uac00 \ubc1c\uc804 \uac00\ub2a5\uc131\uc744 \uac15\uc870\ud569\ub2c8\ub2e4.<\/p>\n<p>DeepSeek-R1-Zero\uc758 \uc790\uae30 \uc9c4\ud654 \uacfc\uc815<br \/>\nDeepSeek-R1-Zero\uc758 \uc790\uae30 \uc9c4\ud654 \uacfc\uc815\uc740 RL\uc774 \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \uc790\uc728\uc801\uc73c\ub85c \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc8fc\ub294 \ud765\ubbf8\ub85c\uc6b4 \uc0ac\ub840\uc785\ub2c8\ub2e4. RL\uc744 \uae30\ubcf8 \ubaa8\ub378\uc5d0\uc11c \uc9c1\uc811 \uc2dc\uc791\ud568\uc73c\ub85c\uc368, \uc9c0\ub3c4 \ubbf8\uc138 \uc870\uc815 \ub2e8\uacc4\uc758 \uc601\ud5a5\uc744 \ubc1b\uc9c0 \uc54a\uace0 \ubaa8\ub378\uc758 \uc9c4\ud589 \uc0c1\ud669\uc744 \uba74\ubc00\ud788 \ubaa8\ub2c8\ud130\ub9c1\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uc811\uadfc \ubc29\uc2dd\uc740 \ubaa8\ub378\uc774 \uc2dc\uac04\uc774 \uc9c0\ub0a8\uc5d0 \ub530\ub77c \ud2b9\ud788 \ubcf5\uc7a1\ud55c \ucd94\ub860 \uacfc\uc81c\ub97c \ucc98\ub9ac\ud558\ub294 \ub2a5\ub825 \uce21\uba74\uc5d0\uc11c \uc5b4\ub5bb\uac8c \uc9c4\ud654\ud558\ub294\uc9c0 \uba85\ud655\ud558\uac8c \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<br \/>\n\uadf8\ub9bc 3\uc5d0 \ub098\ud0c0\ub09c \ubc14\uc640 \uac19\uc774, DeepSeek-R1-Zero\uc758 \uc0ac\uace0 \uc2dc\uac04\uc740 \ud6c8\ub828 \uacfc\uc815 \uc804\ubc18\uc5d0 \uac78\uccd0 \uafb8\uc900\ud788 \uc99d\uac00\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \ud5a5\uc0c1\uc740 \uc678\ubd80 \uc870\uc815\uc758 \uacb0\uacfc\uac00 \uc544\ub2c8\ub77c \ubaa8\ub378 \ub0b4\uc5d0\uc11c\uc758 \ubcf8\uc9c8\uc801\uc778 \ubc1c\uc804\uc785\ub2c8\ub2e4. DeepSeek-R1-Zero\ub294 \ud14c\uc2a4\ud2b8 \uc2dc\uc810\uc758 \uacc4\uc0b0\uc744 \ud655\uc7a5\ud568\uc73c\ub85c\uc368 \uc810\uc810 \ub354 \ubcf5\uc7a1\ud55c \ucd94\ub860 \uacfc\uc81c\ub97c \ud574\uacb0\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \uc790\uc5f0\uc2a4\ub7fd\uac8c \uc2b5\ub4dd\ud569\ub2c8\ub2e4. \uc774 \uacc4\uc0b0\uc740 \uc218\ubc31\uc5d0\uc11c \uc218\ucc9c \uac1c\uc758 \ucd94\ub860 \ud1a0\ud070\uc744 \uc0dd\uc131\ud558\ub294 \uac83\uc744 \ud3ec\ud568\ud558\uc5ec, \ubaa8\ub378\uc774 \uc0ac\uace0 \uacfc\uc815\uc744 \ub354 \uae4a\uc774 \ud0d0\uad6c\ud558\uace0 \uc815\uc81c\ud560 \uc218 \uc788\ub3c4\ub85d \ud569\ub2c8\ub2e4.<br \/>\n\uc774 \uc790\uae30 \uc9c4\ud654\uc758 \uac00\uc7a5 \uc8fc\ubaa9\ud560 \ub9cc\ud55c \uce21\uba74 \uc911 \ud558\ub098\ub294 \ud14c\uc2a4\ud2b8 \uc2dc\uc810\uc758 \uacc4\uc0b0\uc774 \uc99d\uac00\ud568\uc5d0 \ub530\ub77c \uc815\uad50\ud55c \ud589\ub3d9\uc774 \ub098\ud0c0\ub09c\ub2e4\ub294 \uc810\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \ubaa8\ub378\uc774 \uc774\uc804 \ub2e8\uacc4\ub97c \uc7ac\uac80\ud1a0\ud558\uace0 \uc7ac\ud3c9\uac00\ud558\ub294 \ubc18\uc131(reflection)\uc774\ub098 \ubb38\uc81c \ud574\uacb0\uc744 \uc704\ud55c \ub300\uccb4 \uc811\uadfc \ubc29\uc2dd\uc744 \ud0d0\uad6c\ud558\ub294 \ud589\ub3d9 \ub4f1\uc774 \uc790\ubc1c\uc801\uc73c\ub85c \ub098\ud0c0\ub0a9\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ud589\ub3d9\uc740 \uba85\uc2dc\uc801\uc73c\ub85c \ud504\ub85c\uadf8\ub798\ubc0d\ub41c \uac83\uc774 \uc544\ub2c8\ub77c, \ubaa8\ub378\uc774 \uac15\ud654 \ud559\uc2b5 \ud658\uacbd\uacfc \uc0c1\ud638 \uc791\uc6a9\ud55c \uacb0\uacfc\ub85c \ub098\ud0c0\ub0a9\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc790\ubc1c\uc801\uc778 \ubc1c\uc804\uc740 DeepSeek-R1-Zero\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \ud06c\uac8c \ud5a5\uc0c1\uc2dc\ucf1c, \ub354 \ub3c4\uc804\uc801\uc778 \uacfc\uc81c\ub97c \ub354 \ud6a8\uc728\uc801\uc774\uace0 \uc815\ud655\ud558\uac8c \ud574\uacb0\ud560 \uc218 \uc788\uac8c \ud569\ub2c8\ub2e4.<\/p>\n<p>DeepSeek-R1-Zero\uc758 \uc544\ud558 \uc21c\uac04<br \/>\nDeepSeek-R1-Zero\uc758 \ud6c8\ub828 \uc911\uc5d0 \uad00\ucc30\ub41c \ud2b9\ud788 \ud765\ubbf8\ub85c\uc6b4 \ud604\uc0c1 \uc911 \ud558\ub098\ub294 &#8220;\uc544\ud558 \uc21c\uac04&#8221;\uc758 \ubc1c\uc0dd\uc785\ub2c8\ub2e4. \ud45c 3\uc5d0 \ub098\ud0c0\ub09c \ubc14\uc640 \uac19\uc774, \uc774 \uc21c\uac04\uc740 \ubaa8\ub378\uc758 \uc911\uac04 \ubc84\uc804\uc5d0\uc11c \ubc1c\uc0dd\ud569\ub2c8\ub2e4. \uc774 \ub2e8\uacc4\uc5d0\uc11c DeepSeek-R1-Zero\ub294 \ucd08\uae30 \uc811\uadfc \ubc29\uc2dd\uc744 \uc7ac\ud3c9\uac00\ud568\uc73c\ub85c\uc368 \ubb38\uc81c\uc5d0 \ub354 \ub9ce\uc740 \uc0ac\uace0 \uc2dc\uac04\uc744 \ud560\ub2f9\ud558\ub294 \ubc29\ubc95\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774 \ud589\ub3d9\uc740 \ubaa8\ub378\uc758 \ud5a5\uc0c1\ub41c \ucd94\ub860 \ub2a5\ub825\uc744 \uc99d\uba85\ud560 \ubfd0\ub9cc \uc544\ub2c8\ub77c, \uac15\ud654 \ud559\uc2b5\uc774 \uc5b4\ub5bb\uac8c \uc608\uc0c1\uce58 \ubabb\ud55c \uc815\uad50\ud55c \uacb0\uacfc\ub97c \uc774\ub04c\uc5b4\ub0bc \uc218 \uc788\ub294\uc9c0 \ubcf4\uc5ec\uc8fc\ub294 \ub9e4\ud639\uc801\uc778 \uc608\uc785\ub2c8\ub2e4.<br \/>\n\uc774 \uc21c\uac04\uc740 \ubaa8\ub378\ubfd0\ub9cc \uc544\ub2c8\ub77c \uadf8 \ud589\ub3d9\uc744 \uad00\ucc30\ud558\ub294 \uc5f0\uad6c\uc790\ub4e4\uc5d0\uac8c\ub3c4 &#8220;\uc544\ud558 \uc21c\uac04&#8221;\uc785\ub2c8\ub2e4. \uc774\ub294 \uac15\ud654 \ud559\uc2b5\uc758 \ud798\uacfc \uc544\ub984\ub2e4\uc6c0\uc744 \uac15\uc870\ud569\ub2c8\ub2e4: \ubb38\uc81c \ud574\uacb0 \ubc29\ubc95\uc744 \uba85\uc2dc\uc801\uc73c\ub85c \uac00\ub974\uce58\ub294 \ub300\uc2e0, \uc801\uc808\ud55c \uc778\uc13c\ud2f0\ube0c\ub97c \uc81c\uacf5\ud568\uc73c\ub85c\uc368 \ubaa8\ub378\uc774 \uc790\uc728\uc801\uc73c\ub85c \uace0\uae09 \ubb38\uc81c \ud574\uacb0 \uc804\ub7b5\uc744 \uac1c\ubc1c\ud558\uac8c \ub429\ub2c8\ub2e4. &#8220;\uc544\ud558 \uc21c\uac04&#8221;\uc740 RL\uc774 \uc778\uacf5\uc9c0\ub2a5 \uc2dc\uc2a4\ud15c\uc758 \uc0c8\ub85c\uc6b4 \uc218\uc900\uc758 \uc9c0\ub2a5\uc744 \uc5f4\uc5b4\uc904 \uc7a0\uc7ac\ub825\uc744 \uac15\ub825\ud558\uac8c \uc0c1\uae30\uc2dc\ucf1c\uc8fc\uba70, \ubbf8\ub798\uc5d0 \ub354 \uc790\uc728\uc801\uc774\uace0 \uc801\uc751\ub825 \uc788\ub294 \ubaa8\ub378\uc758 \uae38\uc744 \ub2e6\uc544\uc90d\ub2c8\ub2e4.<br \/>\n\ud45c 3 | DeepSeek-R1-Zero\uc758 \uc911\uac04 \ubc84\uc804\uc5d0\uc11c \ub098\ud0c0\ub09c \ud765\ubbf8\ub85c\uc6b4 &#8220;\uc544\ud558 \uc21c\uac04&#8221; \ubaa8\ub378\uc774 \uc778\uac04 \uce5c\ud654\uc801\uc778 \uc5b4\uc870\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc7ac\uace0\ud558\ub294 \ubc29\ubc95\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774\ub294 \uc6b0\ub9ac\uc5d0\uac8c\ub3c4 &#8220;\uc544\ud558 \uc21c\uac04&#8221;\uc73c\ub85c, \uac15\ud654 \ud559\uc2b5\uc758 \ud798\uacfc \uc544\ub984\ub2e4\uc6c0\uc744 \ubaa9\uaca9\ud560 \uc218 \uc788\uac8c \ud574\uc90d\ub2c8\ub2e4.<\/p>\n<p>DeepSeek-R1-Zero\uc758 \ub2e8\uc810<br \/>\nDeepSeek-R1-Zero\ub294 \uac15\ub825\ud55c \ucd94\ub860 \ub2a5\ub825\uc744 \ubc1c\ud718\ud558\uace0 \uc790\uc728\uc801\uc73c\ub85c \uc608\uc0c1\uce58 \ubabb\ud55c \uac15\ub825\ud55c \ucd94\ub860 \ud589\ub3d9\uc744 \uac1c\ubc1c\ud558\uc9c0\ub9cc, \uba87 \uac00\uc9c0 \ubb38\uc81c\uc810\uc5d0 \uc9c1\uba74\ud574 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, DeepSeek-R1-Zero\ub294 \uac00\ub3c5\uc131\uc774 \ub5a8\uc5b4\uc9c0\uac70\ub098 \uc5b8\uc5b4\uac00 \ud63c\ud569\ub418\ub294 \ub4f1\uc758 \ubb38\uc81c\ub97c \uacaa\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ucd94\ub860 \uacfc\uc815\uc744 \ub354 \uc77d\uae30 \uc27d\uac8c \ub9cc\ub4e4\uace0 \uc774\ub97c \uacf5\uac1c \ucee4\ubba4\ub2c8\ud2f0\uc640 \uacf5\uc720\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 \uc778\uac04 \uce5c\ud654\uc801\uc778 \ucd08\uae30 \ub370\uc774\ud130\ub97c \ud65c\uc6a9\ud55c RL \ubc29\ubc95\uc778 DeepSeek-R1\uc744 \ud0d0\uad6c\ud569\ub2c8\ub2e4.<\/p>\n<p>2.3. DeepSeek-R1: \ucd08\uae30 \ub370\uc774\ud130\ub97c \ud65c\uc6a9\ud55c \uac15\ud654 \ud559\uc2b5<br \/>\nDeepSeek-R1-Zero\uc758 \uc720\ub9dd\ud55c \uacb0\uacfc\uc5d0 \uc601\uac10\uc744 \ubc1b\uc544 \ub450 \uac00\uc9c0 \uc790\uc5f0\uc2a4\ub7ec\uc6b4 \uc9c8\ubb38\uc774 \uc81c\uae30\ub429\ub2c8\ub2e4: 1) \uc18c\ub7c9\uc758 \uace0\ud488\uc9c8 \ub370\uc774\ud130\ub97c \ucd08\uae30 \ub2e8\uacc4\ub85c \ud3ec\ud568\ud568\uc73c\ub85c\uc368 \ucd94\ub860 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0a4\uac70\ub098 \uc218\ub834 \uc18d\ub3c4\ub97c \uac00\uc18d\ud654\ud560 \uc218 \uc788\ub294\uac00? 2) \uba85\ud655\ud558\uace0 \uc77c\uad00\ub41c \uc0ac\uace0\uc758 \uc0ac\uc2ac(CoT)\uc744 \uc0dd\uc131\ud560 \ubfd0\ub9cc \uc544\ub2c8\ub77c \uac15\ub825\ud55c \uc77c\ubc18 \ub2a5\ub825\uc744 \ubcf4\uc5ec\uc8fc\ub294 \uc0ac\uc6a9\uc790 \uce5c\ud654\uc801\uc778 \ubaa8\ub378\uc744 \uc5b4\ub5bb\uac8c \ud6c8\ub828\uc2dc\ud0ac \uc218 \uc788\ub294\uac00? \uc774\ub7ec\ud55c \uc9c8\ubb38\uc5d0 \ub2f5\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 DeepSeek-R1\uc744 \ud6c8\ub828\uc2dc\ud0a4\uae30 \uc704\ud55c \ud30c\uc774\ud504\ub77c\uc778\uc744 \uc124\uacc4\ud588\uc2b5\ub2c8\ub2e4. \uc774 \ud30c\uc774\ud504\ub77c\uc778\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \ub124 \ub2e8\uacc4\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4.<\/p>\n<p>2.3.1. \ucd08\uae30 \uc2dc\uc791 (Cold Start)<br \/>\nDeepSeek-R1-Zero\uc640 \ub2ec\ub9ac, \uae30\ubcf8 \ubaa8\ub378\uc5d0\uc11c RL \ud6c8\ub828\uc758 \ucd08\uae30 \ubd88\uc548\uc815\ud55c \ucd08\uae30 \uc2dc\uc791 \ub2e8\uacc4\ub97c \ubc29\uc9c0\ud558\uae30 \uc704\ud574, DeepSeek-R1\uc5d0\uc11c\ub294 \uc18c\ub7c9\uc758 \uae34 \uc0ac\uace0\uc758 \uc0ac\uc2ac(CoT) \ub370\uc774\ud130\ub97c \uad6c\uc131\ud558\uace0 \uc218\uc9d1\ud558\uc5ec \ucd08\uae30 RL \uc561\ud130\ub85c\uc11c \ubaa8\ub378\uc744 \ubbf8\uc138 \uc870\uc815\ud569\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ub370\uc774\ud130\ub97c \uc218\uc9d1\ud558\uae30 \uc704\ud574 \uba87 \uac00\uc9c0 \uc811\uadfc \ubc29\uc2dd\uc744 \ud0d0\uad6c\ud588\uc2b5\ub2c8\ub2e4: \uae34 CoT\ub97c \uc608\uc81c\ub85c \uc0ac\uc6a9\ud558\ub294 \uba87 \uc0f7 \ud504\ub86c\ud504\ud2b8, \ubc18\uc131\uacfc \uac80\uc99d\uc744 \ud3ec\ud568\ud55c \uc0c1\uc138\ud55c \ub2f5\ubcc0\uc744 \uc0dd\uc131\ud558\ub3c4\ub85d \ubaa8\ub378\uc744 \uc9c1\uc811 \ud504\ub86c\ud504\ud2b8, \uc77d\uae30 \uc26c\uc6b4 \ud615\uc2dd\uc73c\ub85c DeepSeek-R1-Zero\uc758 \ucd9c\ub825\uc744 \uc218\uc9d1, \uadf8\ub9ac\uace0 \uc778\uac04 \uc8fc\uc11d\uc790\uac00 \ud6c4\ucc98\ub9ac\ub97c \ud1b5\ud574 \uacb0\uacfc\ub97c \uc815\uc81c\ud558\ub294 \ubc29\ubc95 \ub4f1\uc774 \uc788\uc2b5\ub2c8\ub2e4.<br \/>\n\ubcf8 \uc5f0\uad6c\uc5d0\uc11c\ub294 \uc57d \uc218\ucc9c \uac1c\uc758 \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\ub97c \uc218\uc9d1\ud558\uc5ec RL\uc758 \uc2dc\uc791\uc810\uc73c\ub85c DeepSeek-V3-Base\ub97c \ubbf8\uc138 \uc870\uc815\ud588\uc2b5\ub2c8\ub2e4. DeepSeek-R1-Zero\uc640 \ube44\uad50\ud560 \ub54c, \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\uc758 \uc7a5\uc810\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n<p>&#8211; \uac00\ub3c5\uc131: DeepSeek-R1-Zero\uc758 \uc8fc\uc694 \ud55c\uacc4\ub294 \uadf8 \ub0b4\uc6a9\uc774 \uc885\uc885 \uc77d\uae30\uc5d0 \uc801\ud569\ud558\uc9c0 \uc54a\ub2e4\ub294 \uc810\uc785\ub2c8\ub2e4. \uc751\ub2f5\uc774 \uc5ec\ub7ec \uc5b8\uc5b4\ub97c \ud63c\ud569\ud558\uac70\ub098 \uc0ac\uc6a9\uc790\ub97c \uc704\ud55c \ub2f5\ubcc0\uc744 \uac15\uc870\ud558\uae30 \uc704\ud55c \ub9c8\ud06c\ub2e4\uc6b4 \ud615\uc2dd\uc774 \ubd80\uc871\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubc18\uba74, DeepSeek-R1\uc744 \uc704\ud55c \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud560 \ub54c\ub294 \uac01 \uc751\ub2f5 \ub05d\uc5d0 \uc694\uc57d\uc744 \ud3ec\ud568\ud558\ub294 \uc77d\uae30 \uc26c\uc6b4 \ud328\ud134\uc744 \uc124\uacc4\ud558\uace0, \uc77d\uae30 \ubd88\ud3b8\ud55c \uc751\ub2f5\uc740 \ud544\ud130\ub9c1\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c \ucd9c\ub825 \ud615\uc2dd\uc744 &#8216;|special_token||special_token|<\/p>\n<p>&#8216;\ub85c \uc815\uc758\ud558\uba70, reasoning_process\ub294 \uc9c8\uc758\uc5d0 \ub300\ud55c CoT\uc774\uace0, summary\ub294 \ucd94\ub860 \uacb0\uacfc\ub97c \uc694\uc57d\ud558\ub294 \ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<br \/>\n&#8211; \uc7a0\uc7ac\ub825: \uc778\uac04\uc758 \uc0ac\uc804 \uc9c0\uc2dd\uc744 \ubc18\uc601\ud558\uc5ec \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\uc758 \ud328\ud134\uc744 \uc2e0\uc911\ud558\uac8c \uc124\uacc4\ud568\uc73c\ub85c\uc368 DeepSeek-R1-Zero\uc5d0 \ube44\ud574 \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \uad00\ucc30\ud588\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \ubc18\ubcf5\uc801\uc778 \ud6c8\ub828\uc774 \ucd94\ub860 \ubaa8\ub378\uc5d0 \ub354 \ub098\uc740 \ubc29\ubc95\uc774\ub77c\uace0 \ubbff\uc2b5\ub2c8\ub2e4.<\/p>\n<div>\n<p>2.3.2. \ucd94\ub860 \uc9c0\ud5a5 \uac15\ud654 \ud559\uc2b5 (Reasoning-oriented Reinforcement Learning)<br \/>\nDeepSeek-V3-Base\ub97c \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\ub85c \ubbf8\uc138 \uc870\uc815\ud55c \ud6c4, DeepSeek-R1-Zero\uc5d0\uc11c \uc0ac\uc6a9\ub41c \uac83\uacfc \ub3d9\uc77c\ud55c \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5 \ud6c8\ub828 \uacfc\uc815\uc744 \uc801\uc6a9\ud569\ub2c8\ub2e4. \uc774 \ub2e8\uacc4\ub294 \ud2b9\ud788 \ucf54\ub529, \uc218\ud559, \uacfc\ud559, \ub17c\ub9ac\uc801 \ucd94\ub860\uacfc \uac19\uc740 \uba85\ud655\ud55c \ud574\uacb0\ucc45\uc774 \uc788\ub294 \uc798 \uc815\uc758\ub41c \ubb38\uc81c\ub97c \ud3ec\ud568\ud558\ub294 \ucd94\ub860 \uc9d1\uc57d\uc801 \uacfc\uc81c\uc5d0\uc11c \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ub370 \uc911\uc810\uc744 \ub461\ub2c8\ub2e4. \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c CoT\uac00 \uc885\uc885 \uc5ec\ub7ec \uc5b8\uc5b4\ub97c \ud63c\ud569\ud558\ub294 \uacbd\ud5a5\uc744 \ubcf4\uc774\ub294 \uac83\uc744 \uad00\ucc30\ud588\uc2b5\ub2c8\ub2e4. \ud2b9\ud788 RL \ud504\ub86c\ud504\ud2b8\uac00 \uc5ec\ub7ec \uc5b8\uc5b4\ub97c \ud3ec\ud568\ud560 \ub54c \uc774\ub7ec\ud55c \ubb38\uc81c\uac00 \ub450\ub4dc\ub7ec\uc84c\uc2b5\ub2c8\ub2e4. \uc774 \ubb38\uc81c\ub97c \uc644\ud654\ud558\uae30 \uc704\ud574, RL \ud6c8\ub828 \uc911 \uc5b8\uc5b4 \uc77c\uad00\uc131 \ubcf4\uc0c1\uc744 \ub3c4\uc785\ud588\uc2b5\ub2c8\ub2e4. \uc774\ub294 CoT\uc5d0\uc11c \ub300\uc0c1 \uc5b8\uc5b4 \ub2e8\uc5b4\uc758 \ube44\uc728\ub85c \uacc4\uc0b0\ub429\ub2c8\ub2e4. \uc808\uc81c \uc2e4\ud5d8(ablation experiments) \uacb0\uacfc, \uc774\ub7ec\ud55c \uc815\ub82c\uc740 \ubaa8\ub378\uc758 \uc131\ub2a5\uc5d0 \uc57d\uac04\uc758 \uc800\ud558\ub97c \ucd08\ub798\ud558\uc9c0\ub9cc, \uc774\ub294 \uc778\uac04\uc758 \uc120\ud638\uc640 \uc77c\uce58\ud558\uc5ec \uac00\ub3c5\uc131\uc744 \ub192\uc785\ub2c8\ub2e4. \ub9c8\uc9c0\ub9c9\uc73c\ub85c, \uc6b0\ub9ac\ub294 \ucd94\ub860 \uacfc\uc81c\uc758 \uc815\ud655\ub3c4\uc640 \uc5b8\uc5b4 \uc77c\uad00\uc131 \ubcf4\uc0c1\uc744 \uc9c1\uc811 \ud569\uc0b0\ud558\uc5ec \ucd5c\uc885 \ubcf4\uc0c1\uc744 \ud615\uc131\ud569\ub2c8\ub2e4. \uadf8\ub7f0 \ub2e4\uc74c, \ubaa8\ub378\uc774 \ucd94\ub860 \uacfc\uc81c\uc5d0\uc11c \uc218\ub834\ud560 \ub54c\uae4c\uc9c0 \ubbf8\uc138 \uc870\uc815\ub41c \ubaa8\ub378\uc5d0 \ub300\ud574 \uac15\ud654 \ud559\uc2b5(RL) \ud6c8\ub828\uc744 \uc801\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>2.3.3. \uac70\ubd80 \uc0d8\ud50c\ub9c1 \ubc0f \uc9c0\ub3c4 \ubbf8\uc138 \uc870\uc815 (Rejection Sampling and Supervised Fine-Tuning)<br \/>\n\ucd94\ub860 \uc9c0\ud5a5 RL\uc774 \uc218\ub834\ud558\uba74, \uacb0\uacfc\ub85c \ub098\uc628 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc774\ud6c4 \ub77c\uc6b4\ub4dc\ub97c \uc704\ud55c SFT(Supervised Fine-Tuning) \ub370\uc774\ud130\ub97c \uc218\uc9d1\ud569\ub2c8\ub2e4. \ucd08\uae30 \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\uac00 \uc8fc\ub85c \ucd94\ub860\uc5d0 \ucd08\uc810\uc744 \ub9de\ucd98 \ubc18\uba74, \uc774 \ub2e8\uacc4\uc5d0\uc11c\ub294 \uae00\uc4f0\uae30, \uc5ed\ud560 \uc5f0\uae30, \uae30\ud0c0 \uc77c\ubc18 \ubaa9\uc801 \uacfc\uc81c\uc5d0\uc11c \ubaa8\ub378\uc758 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574 \ub2e4\ub978 \ub3c4\uba54\uc778\uc758 \ub370\uc774\ud130\ub97c \ud3ec\ud568\ud569\ub2c8\ub2e4. \uad6c\uccb4\uc801\uc73c\ub85c, \uc6b0\ub9ac\ub294 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0 \uc544\ub798\uc5d0 \uc124\uba85\ub41c \ub300\ub85c \ubaa8\ub378\uc744 \ubbf8\uc138 \uc870\uc815\ud569\ub2c8\ub2e4.<br \/>\n&#8211; \ucd94\ub860 \ub370\uc774\ud130: \ucd94\ub860 \ud504\ub86c\ud504\ud2b8\ub97c \uc120\ubcc4\ud558\uace0, \uc704\uc758 RL \ud6c8\ub828 \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0\uc11c \uac70\ubd80 \uc0d8\ud50c\ub9c1\uc744 \uc218\ud589\ud558\uc5ec \ucd94\ub860 \uada4\uc801\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc774\uc804 \ub2e8\uacc4\uc5d0\uc11c\ub294 \uc8fc\ub85c \uaddc\uce59 \uae30\ubc18 \ubcf4\uc0c1\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud3c9\uac00\ud560 \uc218 \uc788\ub294 \ub370\uc774\ud130\ub9cc \ud3ec\ud568\ud588\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \uc774 \ub2e8\uacc4\uc5d0\uc11c\ub294 \uc77c\ubd80 \ub370\uc774\ud130\ub97c \uc0dd\uc131 \ubcf4\uc0c1 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud655\uc7a5\ud569\ub2c8\ub2e4. \uc774\ub294 DeepSeek-V3\uc5d0 \uc2e4\uc81c \uc815\ub2f5\uacfc \ubaa8\ub378 \uc608\uce21\uc744 \uc785\ub825\ud558\uc5ec \ud310\ub2e8\ud558\uac8c \ud558\ub294 \ubc29\uc2dd\uc785\ub2c8\ub2e4. \ub610\ud55c, \ubaa8\ub378 \ucd9c\ub825\uc774 \ub54c\ub54c\ub85c \ud63c\ub780\uc2a4\ub7fd\uace0 \uc77d\uae30 \uc5b4\ub824\uc6b4 \uacbd\uc6b0\uac00 \uc788\uc5b4, \ud63c\ud569 \uc5b8\uc5b4, \uae34 \ub2e8\ub77d, \ucf54\ub4dc \ube14\ub85d\uc774 \ud3ec\ud568\ub41c \uc0ac\uace0\uc758 \uc0ac\uc2ac\uc744 \ud544\ud130\ub9c1\ud588\uc2b5\ub2c8\ub2e4. \uac01 \ud504\ub86c\ud504\ud2b8\uc5d0 \ub300\ud574 \uc5ec\ub7ec \uc751\ub2f5\uc744 \uc0d8\ud50c\ub9c1\ud558\uace0 \uc62c\ubc14\ub978 \uc751\ub2f5\ub9cc \uc720\uc9c0\ud588\uc2b5\ub2c8\ub2e4. \ucd1d \uc57d 60\ub9cc \uac1c\uc758 \ucd94\ub860 \uad00\ub828 \ud6c8\ub828 \uc0d8\ud50c\uc744 \uc218\uc9d1\ud588\uc2b5\ub2c8\ub2e4.<br \/>\n&#8211; \ube44\ucd94\ub860 \ub370\uc774\ud130: \uae00\uc4f0\uae30, \uc0ac\uc2e4 \uae30\ubc18 QA, \uc790\uae30 \uc778\uc2dd, \ubc88\uc5ed\uacfc \uac19\uc740 \ube44\ucd94\ub860 \ub370\uc774\ud130\uc758 \uacbd\uc6b0, DeepSeek-V3 \ud30c\uc774\ud504\ub77c\uc778\uc744 \ucc44\ud0dd\ud558\uace0 DeepSeek-V3\uc758 SFT \ub370\uc774\ud130\uc14b \uc77c\ubd80\ub97c \uc7ac\uc0ac\uc6a9\ud569\ub2c8\ub2e4. \ud2b9\uc815 \ube44\ucd94\ub860 \uacfc\uc81c\uc5d0 \ub300\ud574, \uc6b0\ub9ac\ub294 \uc9c8\ubb38\uc5d0 \ub2f5\ubcc0\ud558\uae30 \uc804\uc5d0 \uc7a0\uc7ac\uc801\uc778 \uc0ac\uace0\uc758 \uc0ac\uc2ac(CoT)\uc744 \uc0dd\uc131\ud558\ub3c4\ub85d DeepSeek-V3\ub97c \ud638\ucd9c\ud569\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \u201c\uc548\ub155\ud558\uc138\uc694\u201d\uc640 \uac19\uc740 \uac04\ub2e8\ud55c \uc9c8\uc758\uc5d0\ub294 CoT\ub97c \uc81c\uacf5\ud558\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4. \ucd5c\uc885\uc801\uc73c\ub85c, \ucd94\ub860\uacfc \uad00\ub828 \uc5c6\ub294 \uc57d 20\ub9cc \uac1c\uc758 \ud6c8\ub828 \uc0d8\ud50c\uc744 \uc218\uc9d1\ud588\uc2b5\ub2c8\ub2e4.<br \/>\n\uc6b0\ub9ac\ub294 \uc57d 80\ub9cc \uac1c\uc758 \uc0d8\ud50c\ub85c \uad6c\uc131\ub41c \uc704\uc5d0\uc11c \uc120\ubcc4\ub41c \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec DeepSeek-V3-Base\ub97c \ub450 \ubc88\uc758 \uc5d0\ud3ec\ud06c \ub3d9\uc548 \ubbf8\uc138 \uc870\uc815\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>2.3.4. \ubaa8\ub4e0 \uc2dc\ub098\ub9ac\uc624\ub97c \uc704\ud55c \uac15\ud654 \ud559\uc2b5 (Reinforcement Learning for all Scenarios)<br \/>\n\ubaa8\ub378\uc744 \uc778\uac04\uc758 \uc120\ud638\uc5d0 \ub354\uc6b1 \ub9de\ucd94\uae30 \uc704\ud574, \ubaa8\ub378\uc758 \uc720\uc6a9\uc131\uacfc \ubb34\ud574\uc131\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ub3d9\uc2dc\uc5d0 \ucd94\ub860 \ub2a5\ub825\uc744 \ub2e4\ub4ec\ub294 2\ucc28 \uac15\ud654 \ud559\uc2b5 \ub2e8\uacc4\ub97c \uad6c\ud604\ud588\uc2b5\ub2c8\ub2e4. \uad6c\uccb4\uc801\uc73c\ub85c, \uc6b0\ub9ac\ub294 \ubcf4\uc0c1 \uc2e0\ud638\uc640 \ub2e4\uc591\ud55c \ud504\ub86c\ud504\ud2b8 \ubd84\ud3ec\uc758 \uc870\ud569\uc744 \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0b5\ub2c8\ub2e4. \ucd94\ub860 \ub370\uc774\ud130\uc758 \uacbd\uc6b0, \uc218\ud559, \ucf54\ub4dc, \ub17c\ub9ac\uc801 \ucd94\ub860 \ub3c4\uba54\uc778\uc5d0\uc11c \ud559\uc2b5 \uacfc\uc815\uc744 \uc548\ub0b4\ud558\uae30 \uc704\ud574 \uaddc\uce59 \uae30\ubc18 \ubcf4\uc0c1\uc744 \uc0ac\uc6a9\ud558\ub294 DeepSeek-R1-Zero\uc5d0\uc11c \uc124\uba85\ud55c \ubc29\ubc95\ub860\uc744 \uc900\uc218\ud569\ub2c8\ub2e4. \uc77c\ubc18 \ub370\uc774\ud130\uc758 \uacbd\uc6b0, \ubcf5\uc7a1\ud558\uace0 \ubbf8\ubb18\ud55c \uc2dc\ub098\ub9ac\uc624\uc5d0\uc11c \uc778\uac04\uc758 \uc120\ud638\ub97c \ud3ec\ucc29\ud558\uae30 \uc704\ud574 \ubcf4\uc0c1 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 DeepSeek-V3 \ud30c\uc774\ud504\ub77c\uc778\uc744 \uae30\ubc18\uc73c\ub85c \uc120\ud638 \uc30d\uacfc \ud6c8\ub828 \ud504\ub86c\ud504\ud2b8\uc758 \uc720\uc0ac\ud55c \ubd84\ud3ec\ub97c \ucc44\ud0dd\ud569\ub2c8\ub2e4. \uc720\uc6a9\uc131\uc5d0 \ub300\ud574\uc11c\ub294 \ucd5c\uc885 \uc694\uc57d\uc5d0\ub9cc \uc9d1\uc911\ud558\uc5ec, \uc751\ub2f5\uc758 \uc720\uc6a9\uc131\uacfc \uad00\ub828\uc131\uc744 \uac15\uc870\ud558\uba74\uc11c \uae30\ubcf8 \ucd94\ub860 \uacfc\uc815\uc5d0 \ub300\ud55c \uac04\uc12d\uc744 \ucd5c\uc18c\ud654\ud569\ub2c8\ub2e4. \ubb34\ud574\uc131\uc5d0 \ub300\ud574\uc11c\ub294, \ucd94\ub860 \uacfc\uc815\uacfc \uc694\uc57d\uc744 \ubaa8\ub450 \ud3ec\ud568\ud55c \ubaa8\ub378\uc758 \uc804\uccb4 \uc751\ub2f5\uc744 \ud3c9\uac00\ud558\uc5ec, \uc0dd\uc131 \uacfc\uc815\uc5d0\uc11c \ubc1c\uc0dd\ud560 \uc218 \uc788\ub294 \uc7a0\uc7ac\uc801 \uc704\ud5d8, \ud3b8\ud5a5 \ub610\ub294 \uc720\ud574\ud55c \ub0b4\uc6a9\uc744 \uc2dd\ubcc4\ud558\uace0 \uc644\ud654\ud569\ub2c8\ub2e4. \uad81\uadf9\uc801\uc73c\ub85c, \ubcf4\uc0c1 \uc2e0\ud638\uc640 \ub2e4\uc591\ud55c \ub370\uc774\ud130 \ubd84\ud3ec\uc758 \ud1b5\ud569\uc740 \ucd94\ub860\uc5d0 \ud0c1\uc6d4\ud558\uba74\uc11c\ub3c4 \uc720\uc6a9\uc131\uacfc \ubb34\ud574\uc131\uc744 \uc6b0\uc120\uc2dc\ud558\ub294 \ubaa8\ub378\uc744 \ud6c8\ub828\ud560 \uc218 \uc788\uac8c \ud569\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>2.4. \uc99d\ub958: \uc791\uc740 \ubaa8\ub378\uc5d0 \ucd94\ub860 \ub2a5\ub825 \ubd80\uc5ec (Distillation: Empower Small Models with Reasoning Capability)<br \/>\n\ub354 \ud6a8\uc728\uc801\uc778 \uc791\uc740 \ubaa8\ub378\ub4e4\uc5d0\uac8c DeepSeek-R1\uacfc \uac19\uc740 \ucd94\ub860 \ub2a5\ub825\uc744 \ubd80\uc5ec\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 Qwen(Qwen, 2024b)\uacfc Llama(AI@Meta, 2024)\uc640 \uac19\uc740 \uc624\ud508 \uc18c\uc2a4 \ubaa8\ub378\uc744 DeepSeek-R1\uc5d0\uc11c \uc120\ubcc4\ub41c 80\ub9cc \uac1c\uc758 \uc0d8\ud50c\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc9c1\uc811 \ubbf8\uc138 \uc870\uc815\ud588\uc2b5\ub2c8\ub2e4(\u00a72.3.3\uc5d0 \uc790\uc138\ud788 \uc124\uba85\ub428). \uc6b0\ub9ac\uc758 \ubc1c\uacac\uc5d0 \ub530\ub974\uba74, \uc774 \ub2e8\uc21c\ud55c \uc99d\ub958 \ubc29\ubc95\uc740 \uc791\uc740 \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \ud06c\uac8c \ud5a5\uc0c1\uc2dc\ud0b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c \uc0ac\uc6a9\ud55c \uae30\ubcf8 \ubaa8\ub378\uc740 Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, Qwen2.5-14B, Qwen2.5-32B, Llama-3.1-8B, \ubc0f Llama-3.3-70B-Instruct\uc785\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 Llama-3.3\uc758 \ucd94\ub860 \ub2a5\ub825\uc774 Llama-3.1\ubcf4\ub2e4 \uc57d\uac04 \ub354 \uc6b0\uc218\ud558\uae30 \ub54c\ubb38\uc5d0 \uc774\ub97c \uc120\ud0dd\ud588\uc2b5\ub2c8\ub2e4.<br \/>\n\uc99d\ub958\ub41c \ubaa8\ub378\uc758 \uacbd\uc6b0, RL \ub2e8\uacc4\ub97c \ud3ec\ud568\ud558\uc9c0 \uc54a\uace0 SFT\ub9cc \uc801\uc6a9\ud588\uc2b5\ub2c8\ub2e4. \uc774\ub294 RL\uc744 \ud3ec\ud568\ud558\uba74 \ubaa8\ub378 \uc131\ub2a5\uc774 \ud06c\uac8c \ud5a5\uc0c1\ub420 \uc218 \uc788\uc9c0\ub9cc, \uc6b0\ub9ac\uc758 \uc8fc\uc694 \ubaa9\ud45c\ub294 \uc99d\ub958 \uae30\uc220\uc758 \ud6a8\uacfc\ub97c \uc785\uc99d\ud558\ub294 \uac83\uc774\uba70, RL \ub2e8\uacc4\uc758 \ud0d0\uc0c9\uc740 \ub354 \ub113\uc740 \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\uc5d0 \ub9e1\uae30\uae30 \uc704\ud568\uc785\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>3. \uc2e4\ud5d8 (Experiment)<\/p>\n<\/div>\n<div>\n<p>\ubca4\uce58\ub9c8\ud06c<\/p>\n<\/div>\n<div>\n<p>\uc6b0\ub9ac\ub294 MMLU(Hendrycks et al., 2020), MMLU-Redux(Gema et al., 2024), MMLU-Pro(Wang et al., 2024), C-Eval(Huang et al., 2023), CMMLU(Li et al., 2023), IFEval(Zhou et al., 2023), FRAMES(Krishna et al., 2024), GPQA Diamond(Rein et al., 2023), SimpleQA(OpenAI, 2024c), C-SimpleQA(He et al., 2024), SWE-Bench Verified(OpenAI, 2024d), Aider 1, LiveCodeBench(Jain et al., 2024) (2024-08 \u2013 2025-01), Codeforces 2, \uc911\uad6d \uad6d\uac00 \uace0\ub4f1\ud559\uad50 \uc218\ud559 \uc62c\ub9bc\ud53c\uc544\ub4dc(CNMO 2024)3, \ubc0f American Invitational Mathematics Examination 2024 (AIME 2024)(MAA, 2024)\uc5d0\uc11c \ubaa8\ub378\uc744 \ud3c9\uac00\ud588\uc2b5\ub2c8\ub2e4. \ud45c\uc900 \ubca4\uce58\ub9c8\ud06c \uc678\uc5d0\ub3c4, \uc6b0\ub9ac\ub294 LLM\uc744 \uc2ec\uc0ac\uc704\uc6d0\uc73c\ub85c \uc0ac\uc6a9\ud558\uc5ec \uac1c\ubc29\ud615 \uc0dd\uc131 \uacfc\uc81c\ub97c \ud3c9\uac00\ud588\uc2b5\ub2c8\ub2e4. \uad6c\uccb4\uc801\uc73c\ub85c, \uc6b0\ub9ac\ub294 AlpacaEval 2.0(Dubois et al., 2024)\uacfc Arena-Hard(Li et al., 2024)\uc758 \uc6d0\ub798 \uad6c\uc131\uc744 \uc900\uc218\ud558\uba70, \uc774\ub294 \uc30d\ubcc4 \ube44\uad50\ub97c \uc704\ud574 GPT-4-Turbo-1106\uc744 \uc2ec\uc0ac\uc704\uc6d0\uc73c\ub85c \ud65c\uc6a9\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 \uae38\uc774 \ud3b8\ud5a5\uc744 \ud53c\ud558\uae30 \uc704\ud574 \ucd5c\uc885 \uc694\uc57d\ub9cc \ud3c9\uac00\uc5d0 \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc99d\ub958\ub41c \ubaa8\ub378\uc758 \uacbd\uc6b0, AIME 2024, MATH-500, GPQA Diamond, Codeforces, \ubc0f LiveCodeBench\uc5d0\uc11c \ub300\ud45c\uc801\uc778 \uacb0\uacfc\ub97c \ubcf4\uace0\ud569\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>\ud3c9\uac00 \ud504\ub86c\ud504\ud2b8<br \/>\nDeepSeek-V3\uc5d0\uc11c \uc124\uc815\ud55c \ub300\ub85c, MMLU, DROP, GPQA Diamond, \ubc0f SimpleQA\uc640 \uac19\uc740 \ud45c\uc900 \ubca4\uce58\ub9c8\ud06c\ub294 simple-evals \ud504\ub808\uc784\uc6cc\ud06c\uc758 \ud504\ub86c\ud504\ud2b8\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud3c9\uac00\ub429\ub2c8\ub2e4. MMLU-Redux\uc758 \uacbd\uc6b0, zero-shot \uc124\uc815\uc5d0\uc11c Zero-Eval \ud504\ub86c\ud504\ud2b8 \ud615\uc2dd(Lin, 2024)\uc744 \ucc44\ud0dd\ud588\uc2b5\ub2c8\ub2e4. MMLU-Pro, C-Eval \ubc0f CLUE-WSC\uc758 \uacbd\uc6b0, \uc6d0\ub798 \ud504\ub86c\ud504\ud2b8\uac00 \uba87 \uc0f7\uc774\ubbc0\ub85c \uc774\ub97c \uc57d\uac04 \uc218\uc815\ud558\uc5ec zero-shot \uc124\uc815\uc73c\ub85c \ubcc0\uacbd\ud588\uc2b5\ub2c8\ub2e4. few-shot\uc758 CoT\ub294 DeepSeek-R1\uc758 \uc131\ub2a5\uc744 \uc800\ud558\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\ub978 \ub370\uc774\ud130\uc14b\uc740 \uc0dd\uc131\uc790\uc5d0 \uc758\ud574 \uc81c\uacf5\ub41c \uae30\ubcf8 \ud504\ub86c\ud504\ud2b8\ub85c \uc6d0\ub798 \ud3c9\uac00 \ud504\ub85c\ud1a0\ucf5c\uc744 \ub530\ub985\ub2c8\ub2e4. \ucf54\ub4dc \ubc0f \uc218\ud559 \ubca4\uce58\ub9c8\ud06c\uc758 \uacbd\uc6b0, HumanEval-Mul \ub370\uc774\ud130\uc14b\uc740 \uc5ec\ub35f \uac00\uc9c0 \uc8fc\ub958 \ud504\ub85c\uadf8\ub798\ubc0d \uc5b8\uc5b4(Python, Java, C++, C#, JavaScript, TypeScript, PHP, \ubc0f Bash)\ub97c \ub2e4\ub8f9\ub2c8\ub2e4. LiveCodeBench\uc758 \ubaa8\ub378 \uc131\ub2a5\uc740 CoT \ud615\uc2dd\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud3c9\uac00\ub418\uba70, 2024\ub144 8\uc6d4\ubd80\ud130 2025\ub144 1\uc6d4\uae4c\uc9c0 \uc218\uc9d1\ub41c \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. Codeforces \ub370\uc774\ud130\uc14b\uc740 10\uac1c\uc758 Div.2 \ub300\ud68c \ubb38\uc81c\uc640 \uc804\ubb38\uac00\uac00 \uc81c\uc791\ud55c \ud14c\uc2a4\ud2b8 \ucf00\uc774\uc2a4\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud3c9\uac00\ub418\uba70, \uc774\ud6c4 \uc608\uc0c1 \ub4f1\uae09\uacfc \uacbd\uc7c1\uc790\uc758 \ube44\uc728\uc744 \uacc4\uc0b0\ud569\ub2c8\ub2e4. SWE-Bench \uac80\uc99d \uacb0\uacfc\ub294 agentless \ud504\ub808\uc784\uc6cc\ud06c(Xia et al., 2024)\ub97c \ud1b5\ud574 \uc5bb\uc5c8\uc2b5\ub2c8\ub2e4. AIDER \uad00\ub828 \ubca4\uce58\ub9c8\ud06c\ub294 &#8220;diff&#8221; \ud615\uc2dd\uc744 \uc0ac\uc6a9\ud558\uc5ec \uce21\uc815\ud588\uc2b5\ub2c8\ub2e4. DeepSeek-R1 \ucd9c\ub825\uc740 \uac01 \ubca4\uce58\ub9c8\ud06c\ub2f9 \ucd5c\ub300 32,768 \ud1a0\ud070\uc73c\ub85c \uc81c\ud55c\ub429\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>\uae30\uc900 \ubaa8\ub378 (Baselines)<br \/>\n\uc6b0\ub9ac\ub294 DeepSeek-V3, Claude-Sonnet-3.5-1022, GPT-4o-0513, OpenAI-o1-mini, \ubc0f OpenAI-o1-1217\uc744 \ud3ec\ud568\ud55c \uc5ec\ub7ec \uac15\ub825\ud55c \uae30\uc900 \ubaa8\ub378\uacfc \ube44\uad50 \ud3c9\uac00\ub97c \uc218\ud589\ud588\uc2b5\ub2c8\ub2e4. \uc911\uad6d \ubcf8\ud1a0\uc5d0\uc11c OpenAI-o1-1217 API\uc5d0 \uc811\uadfc\ud558\uae30 \uc5b4\ub824\uc6cc, \uacf5\uc2dd \ubcf4\uace0\uc11c\ub97c \uae30\ubc18\uc73c\ub85c \uadf8 \uc131\ub2a5\uc744 \ubcf4\uace0\ud588\uc2b5\ub2c8\ub2e4. \uc99d\ub958\ub41c \ubaa8\ub378\uc758 \uacbd\uc6b0, \uc624\ud508 \uc18c\uc2a4 \ubaa8\ub378\uc778 QwQ-32B-Preview(Qwen, 2024a)\uc640\ub3c4 \ube44\uad50\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>\uc0dd\uc131 \uc124\uc815 (Generation Setup)<br \/>\n\ubaa8\ub4e0 \ubaa8\ub378\uc5d0 \ub300\ud574 \ucd5c\ub300 \uc0dd\uc131 \uae38\uc774\ub294 32,768 \ud1a0\ud070\uc73c\ub85c \uc124\uc815\ud588\uc2b5\ub2c8\ub2e4. \uc0d8\ud50c\ub9c1\uc774 \ud544\uc694\ud55c \ubca4\uce58\ub9c8\ud06c\uc758 \uacbd\uc6b0, \uc628\ub3c4\ub294 0.6, top-p \uac12\uc740 0.95\ub85c \uc124\uc815\ud558\uace0, pass@1\uc744 \ucd94\uc815\ud558\uae30 \uc704\ud574 \uac01 \uc9c8\uc758\ub2f9 64\uac1c\uc758 \uc751\ub2f5\uc744 \uc0dd\uc131\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>3.1. DeepSeek-R1 \ud3c9\uac00<br \/>\nMMLU, MMLU-Pro, \ubc0f GPQA Diamond\uc640 \uac19\uc740 \uad50\uc721 \uc9c0\ud5a5 \uc9c0\uc2dd \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c DeepSeek-R1\uc740 DeepSeek-V3\ubcf4\ub2e4 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc600\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ud5a5\uc0c1\uc740 \uc8fc\ub85c STEM \uad00\ub828 \uc9c8\ubb38\uc758 \uc815\ud655\ub3c4 \ud5a5\uc0c1\uc5d0 \uae30\uc778\ud558\uba70, \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5(RL)\uc744 \ud1b5\ud574 \uc0c1\ub2f9\ud55c \uc774\ub4dd\uc744 \uc5bb\uc5c8\uc2b5\ub2c8\ub2e4. \ub610\ud55c, DeepSeek-R1\uc740 FRAMES\uc640 \uac19\uc740 \uc7a5\uae30 \ubb38\ub9e5 \uc758\uc874 QA \uacfc\uc81c\uc5d0\uc11c\ub3c4 \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc5ec, \ubb38\uc11c \ubd84\uc11d \ub2a5\ub825\uc774 \uac15\ud568\uc744 \uc785\uc99d\ud588\uc2b5\ub2c8\ub2e4. \uc774\ub294 AI \uae30\ubc18 \uac80\uc0c9 \ubc0f \ub370\uc774\ud130 \ubd84\uc11d \uacfc\uc81c\uc5d0\uc11c \ucd94\ub860 \ubaa8\ub378\uc758 \uc7a0\uc7ac\ub825\uc744 \uac15\uc870\ud569\ub2c8\ub2e4. \uc0ac\uc2e4 \uae30\ubc18 \ubca4\uce58\ub9c8\ud06c\uc778 SimpleQA\uc5d0\uc11c\ub294 DeepSeek-R1\uc774 DeepSeek-V3\ubcf4\ub2e4 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc600\uc73c\uba70, \uc774\ub294 \uc0ac\uc2e4 \uae30\ubc18 \uc9c8\uc758\ub97c \ucc98\ub9ac\ud558\ub294 \ub2a5\ub825\uc744 \ub098\ud0c0\ub0c5\ub2c8\ub2e4. \uc720\uc0ac\ud55c \ucd94\uc138\uac00 OpenAI-o1\uc774 GPT-4o\ub97c \uc774 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c \ub2a5\uac00\ud558\ub294 \uac83\uc5d0\uc11c\ub3c4 \uad00\ucc30\ub429\ub2c8\ub2e4. \uadf8\ub7ec\ub098 DeepSeek-R1\uc740 \uc548\uc804 RL \ud6c4 \ud2b9\uc815 \uc9c8\uc758\uc5d0 \ub300\ud55c \uc751\ub2f5\uc744 \uac70\ubd80\ud558\ub294 \uacbd\ud5a5 \ub54c\ubb38\uc5d0 \uc911\uad6d SimpleQA \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c\ub294 DeepSeek-V3\ubcf4\ub2e4 \uc131\ub2a5\uc774 \ub5a8\uc5b4\uc9d1\ub2c8\ub2e4. \uc548\uc804 RL \uc5c6\uc774 DeepSeek-R1\uc740 70% \uc774\uc0c1\uc758 \uc815\ud655\ub3c4\ub97c \ub2ec\uc131\ud560 \uc218 \uc788\uc5c8\uc2b5\ub2c8\ub2e4.<br \/>\nDeepSeek-R1\uc740 \ud615\uc2dd \uc9c0\uc2dc\ub97c \ub530\ub974\ub294 \ubaa8\ub378\uc758 \ub2a5\ub825\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud55c \ubca4\uce58\ub9c8\ud06c\uc778 IF-Eval\uc5d0\uc11c\ub3c4 \uc778\uc0c1\uc801\uc778 \uacb0\uacfc\ub97c \ubcf4\uc600\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ud5a5\uc0c1\uc740 \uc9c0\ub3c4 \ubbf8\uc138 \uc870\uc815(SFT) \ubc0f RL \ud6c8\ub828\uc758 \ub9c8\uc9c0\ub9c9 \ub2e8\uacc4\uc5d0\uc11c \uc9c0\uc2dc \ub530\ub974\uae30 \ub370\uc774\ud130\ub97c \ud3ec\ud568\ud55c \uac83\uacfc \uad00\ub828\uc774 \uc788\uc2b5\ub2c8\ub2e4. \ub610\ud55c, AlpacaEval2.0\uacfc ArenaHard\uc5d0\uc11c\ub3c4 \uc8fc\ubaa9\ud560 \ub9cc\ud55c \uc131\ub2a5\uc744 \ubcf4\uc774\uba70, DeepSeek-R1\uc758 \uae00\uc4f0\uae30 \uacfc\uc81c \ubc0f \uac1c\ubc29\ud615 \ub3c4\uba54\uc778 \uc9c8\ubb38 \uc751\ub2f5\uc5d0\uc11c\uc758 \uac15\uc810\uc744 \ub098\ud0c0\ub0c8\uc2b5\ub2c8\ub2e4. DeepSeek-R1\uc774 DeepSeek-V3\ub97c \ud06c\uac8c \ub2a5\uac00\ud558\ub294 \uac83\uc740 \ub300\uaddc\ubaa8 RL\uc758 \uc77c\ubc18\ud654 \uc774\uc810, \uc989 \ucd94\ub860 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0ac \ubfd0\ub9cc \uc544\ub2c8\ub77c \ub2e4\uc591\ud55c \ub3c4\uba54\uc778\uc5d0\uc11c\uc758 \uc131\ub2a5\uc744 \uac1c\uc120\ud568\uc744 \uac15\uc870\ud569\ub2c8\ub2e4. \ub354\uc6b1\uc774, DeepSeek-R1\uc774 \uc0dd\uc131\ud55c \uc694\uc57d \uae38\uc774\ub294 \uac04\uacb0\ud558\uba70, ArenaHard\uc5d0\uc11c\ub294 \ud3c9\uade0 689 \ud1a0\ud070, AlpacaEval 2.0\uc5d0\uc11c\ub294 2,218\uc790\ub85c GPT \uae30\ubc18 \ud3c9\uac00 \uc911 \uae38\uc774 \ud3b8\ud5a5\uc744 \ub3c4\uc785\ud558\uc9c0 \uc54a\uc74c\uc744 \ub098\ud0c0\ub0b4\uc5b4 \ub2e4\uc591\ud55c \uacfc\uc81c\uc5d0\uc11c\uc758 \uacac\uace0\ud568\uc744 \ub354\uc6b1 \uac15\ud654\ud569\ub2c8\ub2e4.<br \/>\n\uc218\ud559 \uacfc\uc81c\uc5d0\uc11c\ub294 DeepSeek-R1\uc774 OpenAI-o1-1217\uacfc \ub3d9\ub4f1\ud55c \uc131\ub2a5\uc744 \ubcf4\uc774\uba70 \ub2e4\ub978 \ubaa8\ub378\ub4e4\uc744 \ud06c\uac8c \ub2a5\uac00\ud588\uc2b5\ub2c8\ub2e4. LiveCodeBench\uc640 Codeforces\uc640 \uac19\uc740 \ucf54\ub529 \uc54c\uace0\ub9ac\uc998 \uacfc\uc81c\uc5d0\uc11c\ub3c4 \uc720\uc0ac\ud55c \ucd94\uc138\uac00 \ub098\ud0c0\ub098, \ucd94\ub860 \uc911\uc2ec \ubaa8\ub378\uc774 \uc774\ub7ec\ud55c \ubca4\uce58\ub9c8\ud06c\ub97c \uc9c0\ubc30\ud588\uc2b5\ub2c8\ub2e4. \uacf5\ud559 \uc9c0\ud5a5 \ucf54\ub529 \uacfc\uc81c\uc5d0\uc11c\ub294 OpenAI-o1-1217\uc774 Aider\uc5d0\uc11c DeepSeek-R1\uc744 \ub2a5\uac00\ud588\uc9c0\ub9cc, SWE Verified\uc5d0\uc11c\ub294 \ube44\uc2b7\ud55c \uc131\ub2a5\uc744 \ubcf4\uc600\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uad00\ub828 RL \ud6c8\ub828 \ub370\uc774\ud130\uc758 \uc591\uc774 \ud604\uc7ac \ub9e4\uc6b0 \uc81c\ud55c\uc801\uc774\uae30 \ub54c\ubb38\uc5d0 DeepSeek-R1\uc758 \uacf5\ud559 \uc131\ub2a5\uc774 \ub2e4\uc74c \ubc84\uc804\uc5d0\uc11c \ud5a5\uc0c1\ub420 \uac83\uc73c\ub85c \ubbff\uc2b5\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>3.2. \uc99d\ub958\ub41c \ubaa8\ub378 \ud3c9\uac00 (Distilled Model Evaluation)<br \/>\n\ud45c 5\uc5d0\uc11c \ubcfc \uc218 \uc788\ub4ef\uc774, \ub2e8\uc21c\ud788 DeepSeek-R1\uc758 \ucd9c\ub825\uc744 \uc99d\ub958\ud568\uc73c\ub85c\uc368, \ud6a8\uc728\uc801\uc778 DeepSeek-R1-7B(\uc989, DeepSeek-R1-Distill-Qwen-7B)\ub294 \ubaa8\ub4e0 \uce21\uba74\uc5d0\uc11c GPT-4o-0513\uacfc \uac19\uc740 \ube44\ucd94\ub860 \ubaa8\ub378\uc744 \ub2a5\uac00\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. DeepSeek-R1-14B\ub294 \ubaa8\ub4e0 \ud3c9\uac00 \uc9c0\ud45c\uc5d0\uc11c QwQ-32B-Preview\ub97c \ub2a5\uac00\ud588\uc73c\uba70, DeepSeek-R1-32B \ubc0f DeepSeek-R1-70B\ub294 \ub300\ubd80\ubd84\uc758 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c o1-mini\ub97c \ud06c\uac8c \ucd08\uacfc\ud588\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uacb0\uacfc\ub294 \uc99d\ub958\uc758 \uac15\ub825\ud55c \uc7a0\uc7ac\ub825\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \ub610\ud55c, \uc774\ub7ec\ud55c \uc99d\ub958\ub41c \ubaa8\ub378\uc5d0 RL\uc744 \uc801\uc6a9\ud558\uba74 \uc0c1\ub2f9\ud55c \ucd94\uac00 \ud5a5\uc0c1\uc774 \uc774\ub8e8\uc5b4\uc9d0\uc744 \ubc1c\uacac\ud588\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uc774\uac83\uc774 \ub354 \ub9ce\uc740 \ud0d0\uc0c9\uc744 \ud544\uc694\ub85c \ud55c\ub2e4\uace0 \ubbff\uc73c\uba70, \uc5ec\uae30\uc11c\ub294 \ub2e8\uc21c\ud55c SFT-\uc99d\ub958\ub41c \ubaa8\ub378\uc758 \uacb0\uacfc\ub9cc\uc744 \uc81c\uc2dc\ud569\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>4. \ub17c\uc758 (Discussion)<br \/>\n4.1. \uc99d\ub958 vs. \uac15\ud654 \ud559\uc2b5 (Distillation vs. Reinforcement Learning)<br \/>\n\ud45c 6\uc5d0\uc11c \ubcfc \uc218 \uc788\ub4ef\uc774, DeepSeek-R1\uc744 \uc99d\ub958\ud568\uc73c\ub85c\uc368 \uc791\uc740 \ubaa8\ub378\uc774 \uc778\uc0c1\uc801\uc778 \uacb0\uacfc\ub97c \ub2ec\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \uc5ec\uc804\ud788 \ud558\ub098\uc758 \uc9c8\ubb38\uc774 \ub0a8\uc2b5\ub2c8\ub2e4: \uc99d\ub958 \uc5c6\uc774 \ubcf8 \ub17c\ubb38\uc5d0\uc11c \ub17c\uc758\ub41c \ub300\uaddc\ubaa8 RL \ud6c8\ub828\uc744 \ud1b5\ud574 \ubaa8\ub378\uc774 \ub3d9\ub4f1\ud55c \uc131\ub2a5\uc744 \ub2ec\uc131\ud560 \uc218 \uc788\uc744\uae4c\uc694?<br \/>\n\uc774 \uc9c8\ubb38\uc5d0 \ub2f5\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 Qwen-32B-Base\uc5d0 \ub300\ud574 \uc218\ud559, \ucf54\ub4dc, STEM \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud558\uc5ec 10K \uc2a4\ud15d \uc774\uc0c1 \ub300\uaddc\ubaa8 RL \ud6c8\ub828\uc744 \uc218\ud589\ud588\uace0, \uadf8 \uacb0\uacfc DeepSeek-R1-Zero-Qwen-32B\ub97c \uc5bb\uc5c8\uc2b5\ub2c8\ub2e4. \uadf8\ub9bc 6\uc5d0 \ub098\ud0c0\ub09c \uc2e4\ud5d8 \uacb0\uacfc\ub294, 32B \uae30\ubcf8 \ubaa8\ub378\uc774 \ub300\uaddc\ubaa8 RL \ud6c8\ub828 \ud6c4 QwQ-32B-Preview\uc640 \ub3d9\ub4f1\ud55c \uc131\ub2a5\uc744 \ub2ec\uc131\ud568\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uadf8\ub7ec\ub098 DeepSeek-R1\uc5d0\uc11c \uc99d\ub958\ub41c DeepSeek-R1-Distill-Qwen-32B\ub294 \ubaa8\ub4e0 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c DeepSeek-R1-Zero-Qwen-32B\ubcf4\ub2e4 \ud604\uc800\ud788 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc785\ub2c8\ub2e4. \ub530\ub77c\uc11c \ub450 \uac00\uc9c0 \uacb0\ub860\uc744 \ub3c4\ucd9c\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4: \uccab\uc9f8, \ub354 \uac15\ub825\ud55c \ubaa8\ub378\uc744 \uc791\uc740 \ubaa8\ub378\ub85c \uc99d\ub958\ud558\ub294 \uac83\uc740 \uc6b0\uc218\ud55c \uacb0\uacfc\ub97c \uc0b0\ucd9c\ud558\uc9c0\ub9cc, \ubcf8 \ub17c\ubb38\uc5d0\uc11c \uc5b8\uae09\ud55c \ub300\uaddc\ubaa8 RL\uc5d0 \uc758\uc874\ud558\ub294 \uc791\uc740 \ubaa8\ub378\uc740 \ub9c9\ub300\ud55c \uacc4\uc0b0\ub825\uc744 \ud544\uc694\ub85c \ud558\uba70 \uc99d\ub958\uc758 \uc131\ub2a5\uc5d0 \ubbf8\uce58\uc9c0 \ubabb\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub458\uc9f8, \uc99d\ub958 \uc804\ub7b5\uc740 \uacbd\uc81c\uc801\uc774\uace0 \ud6a8\uacfc\uc801\uc774\uc9c0\ub9cc, \uc9c0\ub2a5\uc758 \uacbd\uacc4\ub97c \ub118\uc5b4\uc11c\ub294 \uac83\uc740 \uc5ec\uc804\ud788 \ub354 \uac15\ub825\ud55c \uae30\ubcf8 \ubaa8\ub378\uacfc \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5\uc744 \ud544\uc694\ub85c \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>4.2. \uc2e4\ud328\ud55c \uc2dc\ub3c4 (Unsuccessful Attempts)<br \/>\nDeepSeek-R1\uc744 \uac1c\ubc1c\ud558\ub294 \ucd08\uae30 \ub2e8\uacc4\uc5d0\uc11c, \uc6b0\ub9ac\ub294 \uc2e4\ud328\uc640 \uc88c\uc808\uc744 \uacaa\uae30\ub3c4 \ud588\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 \uc774\ub7ec\ud55c \uc2e4\ud328 \uacbd\ud5d8\uc744 \uacf5\uc720\ud558\uc5ec \ud1b5\ucc30\uc744 \uc81c\uacf5\ud558\uace0\uc790 \ud558\uc9c0\ub9cc, \uc774\ub294 \uc774\ub7ec\ud55c \uc811\uadfc \ubc29\uc2dd\uc774 \ud6a8\uacfc\uc801\uc778 \ucd94\ub860 \ubaa8\ub378\uc744 \uac1c\ubc1c\ud560 \uc218 \uc5c6\uc74c\uc744 \uc758\ubbf8\ud558\uc9c0\ub294 \uc54a\uc2b5\ub2c8\ub2e4.<br \/>\n&#8211; \ud504\ub85c\uc138\uc2a4 \ubcf4\uc0c1 \ubaa8\ub378 (Process Reward Model, PRM): PRM\uc740 \ubaa8\ub378\uc744 \ub354 \ub098\uc740 \ucd94\ub860 \uc811\uadfc \ubc29\uc2dd\uc73c\ub85c \uc774\ub044\ub294 \ud569\ub9ac\uc801\uc778 \ubc29\ubc95\uc785\ub2c8\ub2e4(Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023). \uadf8\ub7ec\ub098 \uc2e4\uc81c\ub85c PRM\uc5d0\ub294 \uc138 \uac00\uc9c0 \uc8fc\uc694 \ud55c\uacc4\uac00 \uc788\uc5b4 \uad81\uadf9\uc801\uc778 \uc131\uacf5\uc744 \uc800\ud574\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uccab\uc9f8, \uc77c\ubc18 \ucd94\ub860\uc5d0\uc11c \uc138\ubd84\ud654\ub41c \ub2e8\uacc4\ub97c \uba85\uc2dc\uc801\uc73c\ub85c \uc815\uc758\ud558\uae30 \uc5b4\ub835\uc2b5\ub2c8\ub2e4. \ub458\uc9f8, \ud604\uc7ac \uc911\uac04 \ub2e8\uacc4\uac00 \uc62c\ubc14\ub978\uc9c0 \uc5ec\ubd80\ub97c \uacb0\uc815\ud558\ub294 \uac83\uc774 \uc5b4\ub824\uc6b4 \uacfc\uc81c\uc785\ub2c8\ub2e4. \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud55c \uc790\ub3d9 \uc8fc\uc11d\uc740 \ub9cc\uc871\uc2a4\ub7ec\uc6b4 \uacb0\uacfc\ub97c \ub0b4\uc9c0 \ubabb\ud560 \uc218 \uc788\uc73c\uba70, \uc218\ub3d9 \uc8fc\uc11d\uc740 \ud655\uc7a5\uc744 \ubc29\ud574\ud569\ub2c8\ub2e4. \uc14b\uc9f8, \ubaa8\ub378 \uae30\ubc18 PRM\uc744 \ub3c4\uc785\ud558\uba74 \ubcf4\uc0c1 \ud574\ud0b9(Gao et al., 2022)\uc774 \ubd88\uac00\ud53c\ud558\uac8c \ubc1c\uc0dd\ud558\uba70, \ubcf4\uc0c1 \ubaa8\ub378\uc744 \uc7ac\ud6c8\ub828\ud558\ub294 \ub370 \ucd94\uac00\uc801\uc778 \ud6c8\ub828 \uc790\uc6d0\uc774 \ud544\uc694\ud558\uace0 \uc804\uccb4 \ud6c8\ub828 \ud30c\uc774\ud504\ub77c\uc778\uc744 \ubcf5\uc7a1\ud558\uac8c \ub9cc\ub4ed\ub2c8\ub2e4. \uacb0\ub860\uc801\uc73c\ub85c, PRM\uc740 \ubaa8\ub378\uc774 \uc0dd\uc131\ud55c \uc0c1\uc704-N \uc751\ub2f5\uc744 \uc7ac\uc815\ub82c\ud558\uac70\ub098 \uc548\ub0b4\ub41c \uac80\uc0c9\uc744 \uc9c0\uc6d0\ud558\ub294 \ub370 \uc88b\uc740 \ub2a5\ub825\uc744 \ubcf4\uc774\uc9c0\ub9cc(Snell et al., 2024), \uc6b0\ub9ac\uc758 \uc2e4\ud5d8\uc5d0\uc11c \ub300\uaddc\ubaa8 \uac15\ud654 \ud559\uc2b5 \uacfc\uc815 \uc911 \ucd94\uac00\uc801\uc778 \uacc4\uc0b0 \uc624\ubc84\ud5e4\ub4dc\uc5d0 \ube44\ud574 \uadf8 \uc774\uc810\uc740 \uc81c\ud55c\uc801\uc785\ub2c8\ub2e4.<br \/>\n&#8211; \ubaac\ud14c \uce74\ub97c\ub85c \ud2b8\ub9ac \ud0d0\uc0c9 (Monte Carlo Tree Search, MCTS): AlphaGo(Silver et al., 2017b) \ubc0f AlphaZero(Silver et al., 2017a)\uc5d0 \uc601\uac10\uc744 \ubc1b\uc544, \uc6b0\ub9ac\ub294 \ud14c\uc2a4\ud2b8 \uc2dc\uc810\uc758 \uacc4\uc0b0 \ud655\uc7a5\uc131\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574 \ubaac\ud14c \uce74\ub97c\ub85c \ud2b8\ub9ac \ud0d0\uc0c9(MCTS)\uc744 \uc0ac\uc6a9\ud558\ub294 \uac83\uc744 \ud0d0\uad6c\ud588\uc2b5\ub2c8\ub2e4. \uc774 \uc811\uadfc \ubc29\uc2dd\uc740 \ub2f5\ubcc0\uc744 \ub354 \uc791\uc740 \ubd80\ubd84\uc73c\ub85c \ub098\ub204\uc5b4 \ubaa8\ub378\uc774 \uccb4\uacc4\uc801\uc73c\ub85c \ud574\uacb0 \uacf5\uac04\uc744 \ud0d0\uc0c9\ud560 \uc218 \uc788\ub3c4\ub85d \ud569\ub2c8\ub2e4. \uc774\ub97c \ucd09\uc9c4\ud558\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 \ubaa8\ub378\uc5d0\uac8c \uac80\uc0c9\uc5d0 \ud544\uc694\ud55c \ud2b9\uc815 \ucd94\ub860 \ub2e8\uacc4\ub97c \ub098\ud0c0\ub0b4\ub294 \uc5ec\ub7ec \ud0dc\uadf8\ub97c \uc0dd\uc131\ud558\ub3c4\ub85d \ud504\ub86c\ud504\ud2b8\ud569\ub2c8\ub2e4. \ud6c8\ub828\uc744 \uc704\ud574, \uba3c\uc800 \uc218\uc9d1\ub41c \ud504\ub86c\ud504\ud2b8\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc0ac\uc804 \ud6c8\ub828\ub41c \uac00\uce58 \ubaa8\ub378\uc5d0 \uc758\ud574 \uc548\ub0b4\ub41c MCTS\ub97c \ud1b5\ud574 \ub2f5\ubcc0\uc744 \ucc3e\uc2b5\ub2c8\ub2e4. \uadf8 \ud6c4, \uacb0\uacfc\ub85c \ub098\uc628 \uc9c8\ubb38-\ub2f5\ubcc0 \uc30d\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc561\ud130 \ubaa8\ub378\uacfc \uac00\uce58 \ubaa8\ub378\uc744 \ubaa8\ub450 \ud6c8\ub828\uc2dc\ud0a4\uace0, \uacfc\uc815\uc744 \ubc18\ubcf5\uc801\uc73c\ub85c \uac1c\uc120\ud569\ub2c8\ub2e4.<br \/>\n\uadf8\ub7ec\ub098 \uc774 \uc811\uadfc \ubc29\uc2dd\uc740 \ud6c8\ub828\uc744 \ud655\uc7a5\ud560 \ub54c \uc5ec\ub7ec \uac00\uc9c0 \ub3c4\uc804\uc5d0 \uc9c1\uba74\ud569\ub2c8\ub2e4. \uccab\uc9f8, \uccb4\uc2a4\uc640 \ub2ec\ub9ac \uac80\uc0c9 \uacf5\uac04\uc774 \uc0c1\ub300\uc801\uc73c\ub85c \uc798 \uc815\uc758\ub418\uc9c0 \uc54a\uc740 \ud1a0\ud070 \uc0dd\uc131\uc740 \uc9c0\uc218\uc801\uc73c\ub85c \ub354 \ud070 \uac80\uc0c9 \uacf5\uac04\uc744 \uc81c\uc2dc\ud569\ub2c8\ub2e4. \uc774\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uac01 \ub178\ub4dc\uc5d0 \ub300\ud55c \ucd5c\ub300 \ud655\uc7a5 \ud55c\uacc4\ub97c \uc124\uc815\ud588\uc9c0\ub9cc, \uc774\ub294 \ubaa8\ub378\uc774 \uc9c0\uc5ed \ucd5c\uc801\uc810\uc5d0 \uac07\ud788\uac8c \ub9cc\ub4e4 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub458\uc9f8, \uac00\uce58 \ubaa8\ub378\uc740 \uac80\uc0c9 \uacfc\uc815\uc758 \uac01 \ub2e8\uacc4\ub97c \uc548\ub0b4\ud558\uae30 \ub54c\ubb38\uc5d0 \uc0dd\uc131 \ud488\uc9c8\uc5d0 \uc9c1\uc811\uc801\uc778 \uc601\ud5a5\uc744 \ubbf8\uce69\ub2c8\ub2e4. \uc138\ubd84\ud654\ub41c \uac00\uce58 \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\ub294 \uac83\uc740 \ubcf8\uc9c8\uc801\uc73c\ub85c \uc5b4\ub824\uc6cc, \ubaa8\ub378\uc774 \ubc18\ubcf5\uc801\uc73c\ub85c \ud5a5\uc0c1\ub418\uae30 \uc5b4\ub835\uac8c \ub9cc\ub4ed\ub2c8\ub2e4. AlphaGo\uc758 \ud575\uc2ec \uc131\uacf5\uc740 \uc810\uc9c4\uc801\uc73c\ub85c \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574 \uac00\uce58 \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\ub294 \ub370 \uc758\uc874\ud588\uc9c0\ub9cc, \ud1a0\ud070 \uc0dd\uc131\uc758 \ubcf5\uc7a1\uc131\uc73c\ub85c \uc778\ud574 \uc6b0\ub9ac\uc758 \uc124\uc815\uc5d0\uc11c\ub294 \uc774 \uc6d0\uce59\uc744 \ubcf5\uc81c\ud558\uae30 \uc5b4\ub835\uc2b5\ub2c8\ub2e4.<br \/>\n\uacb0\ub860\uc801\uc73c\ub85c, MCTS\ub294 \uc0ac\uc804 \ud6c8\ub828\ub41c \uac00\uce58 \ubaa8\ub378\uacfc \uacb0\ud569\ud560 \ub54c \ucd94\ub860 \uc911 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc9c0\ub9cc, \uc790\uac00 \uac80\uc0c9\uc744 \ud1b5\ud574 \ubaa8\ub378 \uc131\ub2a5\uc744 \ubc18\ubcf5\uc801\uc73c\ub85c \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \uac83\uc740 \uc5ec\uc804\ud788 \uc911\uc694\ud55c \ub3c4\uc804 \uacfc\uc81c\ub85c \ub0a8\uc544 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>5. \uacb0\ub860, \ud55c\uacc4 \ubc0f \ud5a5\ud6c4 \uc5f0\uad6c (Conclusion, Limitation, and Future Work)<br \/>\n\ubcf8 \uc5f0\uad6c\uc5d0\uc11c\ub294 \uac15\ud654 \ud559\uc2b5(RL)\uc744 \ud1b5\ud574 \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \uc5ec\uc815\uc744 \uacf5\uc720\ud588\uc2b5\ub2c8\ub2e4. DeepSeek-R1-Zero\ub294 \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\uc5d0 \uc758\uc874\ud558\uc9c0 \uc54a\ub294 \uc21c\uc218 RL \uc811\uadfc \ubc29\uc2dd\uc744 \ub300\ud45c\ud558\uba70, \ub2e4\uc591\ud55c \uacfc\uc81c\uc5d0\uc11c \uac15\ub825\ud55c \uc131\ub2a5\uc744 \ub2ec\uc131\ud588\uc2b5\ub2c8\ub2e4. DeepSeek-R1\uc740 \ucd08\uae30 \uc2dc\uc791 \ub370\uc774\ud130\uc640 \ubc18\ubcf5\uc801\uc778 RL \ubbf8\uc138 \uc870\uc815\uc744 \ud65c\uc6a9\ud558\uc5ec \ub354 \uac15\ub825\ud55c \uc131\ub2a5\uc744 \ubc1c\ud718\ud569\ub2c8\ub2e4. \uad81\uadf9\uc801\uc73c\ub85c, DeepSeek-R1\uc740 \ub2e4\uc591\ud55c \uacfc\uc81c\uc5d0\uc11c OpenAI-o1-1217\uacfc \ub3d9\ub4f1\ud55c \uc131\ub2a5\uc744 \ub2ec\uc131\ud569\ub2c8\ub2e4.<br \/>\n\uc6b0\ub9ac\ub294 \ub610\ud55c \ucd94\ub860 \ub2a5\ub825\uc744 \uc791\uc740 \ubc00\uc9d1 \ubaa8\ub378\ub85c \uc99d\ub958\ud558\ub294 \ubc29\ubc95\uc744 \ud0d0\uad6c\ud588\uc2b5\ub2c8\ub2e4. DeepSeek-R1\uc744 \uad50\uc0ac \ubaa8\ub378\ub85c \uc0ac\uc6a9\ud558\uc5ec 80\ub9cc \uac1c\uc758 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \uc5ec\ub7ec \uc791\uc740 \ubc00\uc9d1 \ubaa8\ub378\uc744 \ubbf8\uc138 \uc870\uc815\ud588\uc2b5\ub2c8\ub2e4. \uadf8 \uacb0\uacfc\ub294 \uc720\ub9dd\ud569\ub2c8\ub2e4: DeepSeek-R1-Distill-Qwen-1.5B\ub294 AIME\uc5d0\uc11c 28.9%, MATH\uc5d0\uc11c 83.9%\ub85c GPT-4o\uc640 Claude-3.5-Sonnet\uc744 \uc218\ud559 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c \ub2a5\uac00\ud588\uc2b5\ub2c8\ub2e4. \ub2e4\ub978 \ubc00\uc9d1 \ubaa8\ub378\ub4e4\ub3c4 \ub3d9\uc77c\ud55c \uae30\ubcf8 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \ub2e4\ub978 \uc9c0\uc2dc \uc870\uc815 \ubaa8\ub378\ub4e4\uc744 \ud06c\uac8c \ub2a5\uac00\ud558\uba70 \uc778\uc0c1\uc801\uc778 \uacb0\uacfc\ub97c \ub2ec\uc131\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n<\/div>\n<div>\n<p>\ud5a5\ud6c4, \uc6b0\ub9ac\ub294 DeepSeek-R1\uc5d0 \ub300\ud574 \ub2e4\uc74c\uacfc \uac19\uc740 \ubc29\ud5a5\uc73c\ub85c \uc5f0\uad6c\ub97c \ud22c\uc790\ud560 \uacc4\ud68d\uc785\ub2c8\ub2e4:<br \/>\n&#8211; \uc77c\ubc18 \ub2a5\ub825 (General Capability): \ud604\uc7ac, DeepSeek-R1\uc758 \ub2a5\ub825\uc740 \ud568\uc218 \ud638\ucd9c, \ub2e4\uc911 \ud134, \ubcf5\uc7a1\ud55c \uc5ed\ud560 \uc5f0\uae30, JSON \ucd9c\ub825\uacfc \uac19\uc740 \uacfc\uc81c\uc5d0\uc11c DeepSeek-V3\ubcf4\ub2e4 \ubd80\uc871\ud569\ub2c8\ub2e4. \uc55e\uc73c\ub85c, \uc6b0\ub9ac\ub294 \uae34 CoT\ub97c \ud65c\uc6a9\ud558\uc5ec \uc774\ub7ec\ud55c \ubd84\uc57c\uc758 \uacfc\uc81c\ub97c \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ubc29\ubc95\uc744 \ud0d0\uad6c\ud560 \uacc4\ud68d\uc785\ub2c8\ub2e4.<br \/>\n&#8211; \uc5b8\uc5b4 \ud63c\ud569 (Language Mixing): DeepSeek-R1\uc740 \ud604\uc7ac \uc911\uad6d\uc5b4\uc640 \uc601\uc5b4\uc5d0 \ucd5c\uc801\ud654\ub418\uc5b4 \uc788\uc5b4, \ub2e4\ub978 \uc5b8\uc5b4\ub85c \ub41c \uc9c8\uc758\ub97c \ucc98\ub9ac\ud560 \ub54c \uc5b8\uc5b4 \ud63c\ud569 \ubb38\uc81c\uac00 \ubc1c\uc0dd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, DeepSeek-R1\uc740 \uc9c8\uc758\uac00 \uc601\uc5b4 \ub610\ub294 \uc911\uad6d\uc5b4\uac00 \uc544\ub2cc \ub2e4\ub978 \uc5b8\uc5b4\uc77c\uc9c0\ub77c\ub3c4 \ucd94\ub860\uacfc \uc751\ub2f5\uc5d0 \uc601\uc5b4\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \ud5a5\ud6c4 \uc5c5\ub370\uc774\ud2b8\uc5d0\uc11c \uc774 \ud55c\uacc4\ub97c \ud574\uacb0\ud560 \uacc4\ud68d\uc785\ub2c8\ub2e4.<br \/>\n&#8211; \ud504\ub86c\ud504\ud2b8 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1 (Prompting Engineering): DeepSeek-R1\uc744 \ud3c9\uac00\ud560 \ub54c, \uc6b0\ub9ac\ub294 \ud504\ub86c\ud504\ud2b8\uc5d0 \ubbfc\uac10\ud568\uc744 \uad00\ucc30\ud588\uc2b5\ub2c8\ub2e4. few-shot \ud504\ub86c\ud504\ud2b8\ub294 \uc77c\uad00\ub418\uac8c \uc131\ub2a5\uc744 \uc800\ud558\uc2dc\ud0b5\ub2c8\ub2e4. \ub530\ub77c\uc11c, \uc6b0\ub9ac\ub294 \uc0ac\uc6a9\uc790\uc5d0\uac8c \ubb38\uc81c\ub97c \uc9c1\uc811 \uc124\uba85\ud558\uace0 zero-shot \uc124\uc815\uc744 \uc0ac\uc6a9\ud558\uc5ec \ucd9c\ub825 \ud615\uc2dd\uc744 \uc9c0\uc815\ud558\ub3c4\ub85d \uad8c\uc7a5\ud569\ub2c8\ub2e4.<br \/>\n&#8211; \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1 \uacfc\uc81c (Software Engineering Tasks): RL \uacfc\uc815\uc758 \ud6a8\uc728\uc131\uc5d0 \uc601\ud5a5\uc744 \ubbf8\uce58\ub294 \uae34 \ud3c9\uac00 \uc2dc\uac04 \ub54c\ubb38\uc5d0, \ub300\uaddc\ubaa8 RL\uc774 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1 \uacfc\uc81c\uc5d0 \uad11\ubc94\uc704\ud558\uac8c \uc801\uc6a9\ub418\uc9c0 \uc54a\uc558\uc2b5\ub2c8\ub2e4. \uacb0\uacfc\uc801\uc73c\ub85c, DeepSeek-R1\uc740 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c DeepSeek-V3\uc5d0 \ube44\ud574 \ud070 \ud5a5\uc0c1\uc744 \ubcf4\uc774\uc9c0 \uc54a\uc558\uc2b5\ub2c8\ub2e4. \ud5a5\ud6c4 \ubc84\uc804\uc5d0\uc11c\ub294 \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uac70\ubd80 \uc0d8\ud50c\ub9c1\uc744 \uad6c\ud604\ud558\uac70\ub098 RL \uacfc\uc815 \uc911 \ube44\ub3d9\uae30 \ud3c9\uac00\ub97c \ud1b5\ud569\ud558\uc5ec \ud6a8\uc728\uc131\uc744 \ud5a5\uc0c1\uc2dc\ud0ac \uacc4\ud68d\uc785\ub2c8\ub2e4.\n<\/p><\/div>\n<\/blockquote>\n<p>\uae40\uc131\uc644\ub2d8\uc740 \uc544\ub798\uc640 \uac19\uc774 \uc704 \ub17c\ubb38\uc744 \uc694\uc57d\ud569\ub2c8\ub2e4.<\/p>\n<blockquote><p>DeepSeek-R1-Zero: \uc0ac\uc804 \uc9c0\ub3c4 \ubbf8\uc138 \uc870\uc815(SFT) \uc5c6\uc774 \ub300\uaddc\ubaa8 \uac15\ud654\ud559\uc2b5(RL)\uc73c\ub85c \ud6c8\ub828\ub41c \ubaa8\ub378\ub85c, \uad04\ubaa9\ud560 \ub9cc\ud55c \ucd94\ub860 \ub2a5\ub825\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. RL\uc744 \ud1b5\ud574 \uc790\uc5f0\uc2a4\ub7fd\uac8c \ub2e4\uc591\ud558\uace0 \uac15\ub825\ud55c \ucd94\ub860 \ud589\ub3d9\uc744 \ud559\uc2b5\ud588\uc9c0\ub9cc, \uac00\ub3c5\uc131 \uc800\ud558 \ubc0f \uc5b8\uc5b4 \ud63c\ud569\uacfc \uac19\uc740 \ubb38\uc81c\uc810\ub3c4 \ub098\ud0c0\ub0ac\uc2b5\ub2c8\ub2e4.<br \/>\nDeepSeek-R1: R1-Zero\uc758 \ubb38\uc81c\uc810\uc744 \ud574\uacb0\ud558\uace0 \ucd94\ub860 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574 \ub2e4\ub2e8\uacc4 \ud6c8\ub828 \ubc0f RL \uc774\uc804\uc758 \ucf5c\ub4dc \uc2a4\ud0c0\ud2b8 \ub370\uc774\ud130\ub97c \ub3c4\uc785\ud588\uc2b5\ub2c8\ub2e4. OpenAI-o1-1217\uc5d0 \ud544\uc801\ud558\ub294 \ucd94\ub860 \uc131\ub2a5\uc744 \ub2ec\uc131\ud588\uc2b5\ub2c8\ub2e4.<br \/>\n\uacf5\uac1c: \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0 \uc9c0\uc6d0\uc744 \uc704\ud574 DeepSeek-R1-Zero, DeepSeek-R1, \uadf8\ub9ac\uace0 Qwen\uacfc Llama \uae30\ubc18\uc73c\ub85c DeepSeek-R1\uc5d0\uc11c \ucd94\ucd9c\ud55c 6\uac1c\uc758 \ubc00\uc9d1 \ubaa8\ub378(1.5B, 7B, 8B, 14B, 32B, 70B)\uc744 \uc624\ud508\uc18c\uc2a4\ub85c \uacf5\uac1c\ud569\ub2c8\ub2e4.<br \/>\n\ud575\uc2ec \uae30\uc5ec:<br \/>\n\uc0ac\uc804 \ud6c8\ub828: SFT \uc5c6\uc774 \uae30\ubcf8 \ubaa8\ub378\uc5d0 \uc9c1\uc811 RL\uc744 \uc801\uc6a9\ud558\uc5ec CoT(Chain-of-Thought)\ub97c \ud0d0\uc0c9\ud558\uace0 \ubcf5\uc7a1\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ucf30\uc2b5\ub2c8\ub2e4. \uc21c\uc218 RL\ub9cc\uc73c\ub85c LLM\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc74c\uc744 \uc99d\uba85\ud55c \uccab \uc5f0\uad6c\uc785\ub2c8\ub2e4.<br \/>\n\ucd94\ucd9c(Distillation): \ub300\ud615 \ubaa8\ub378\uc758 \ucd94\ub860 \ud328\ud134\uc744 \uc18c\ud615 \ubaa8\ub378\ub85c \ucd94\ucd9c\ud558\uc5ec \uc18c\ud615 \ubaa8\ub378 \uc790\uccb4 RL \ud559\uc2b5\ubcf4\ub2e4 \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ub2ec\uc131\ud588\uc2b5\ub2c8\ub2e4. \uc624\ud508\uc18c\uc2a4 DeepSeek-R1\uacfc API\ub294 \ud5a5\ud6c4 \ub354 \ub098\uc740 \uc18c\ud615 \ubaa8\ub378 \ucd94\ucd9c\uc5d0 \ub3c4\uc6c0\uc774 \ub420 \uac83\uc785\ub2c8\ub2e4.<br \/>\n\uc18c\ud615 \ubaa8\ub378\uc758 \uac15\ud654: DeepSeek-R1\uc5d0\uc11c \uc0dd\uc131\ub41c \ucd94\ub860 \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc5f0\uad6c \ucee4\ubba4\ub2c8\ud2f0\uc5d0\uc11c \ub110\ub9ac \uc0ac\uc6a9\ub418\ub294 \uc5ec\ub7ec \ubc00\uc9d1 \ubaa8\ub378\uc744 \ubbf8\uc138 \uc870\uc815\ud588\uc2b5\ub2c8\ub2e4. \ucd94\ucd9c\ub41c \uc18c\ud615 \ubc00\uc9d1 \ubaa8\ub378\uc740 \ubca4\uce58\ub9c8\ud06c\uc5d0\uc11c \ub9e4\uc6b0 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc600\uc2b5\ub2c8\ub2e4. \ud2b9\ud788 DeepSeek-R1-Distill-Qwen-7B\ub294 AIME 2024\uc5d0\uc11c 55.5%\ub97c \ub2ec\uc131\ud558\uc5ec QwQ-32B-Preview\ub97c \ub2a5\uac00\ud588\uc2b5\ub2c8\ub2e4. DeepSeek-R1-Distill-Qwen-32B\ub294 AIME 2024\uc5d0\uc11c 72.6%, MATH-500\uc5d0\uc11c 94.3%, LiveCodeBench\uc5d0\uc11c 57.2%\ub97c \uae30\ub85d\ud558\uba70 \uc774\uc804 \uc624\ud508\uc18c\uc2a4 \ubaa8\ub378\ub4e4\uc744 \ud06c\uac8c \uc55e\uc11c\uace0 o1-mini\uc5d0 \ud544\uc801\ud558\ub294 \uc131\ub2a5\uc744 \ubcf4\uc600\uc2b5\ub2c8\ub2e4.<br \/>\n\uc811\uadfc \ubc29\uc2dd:<br \/>\nDeepSeek-R1-Zero: SFT \uc5c6\uc774 \uae30\ubcf8 \ubaa8\ub378\uc5d0 \uc9c1\uc811 RL\uc744 \uc801\uc6a9. \uaddc\uce59 \uae30\ubc18 \ubcf4\uc0c1 \uc2dc\uc2a4\ud15c(\uc815\ud655\ub3c4, \ud615\uc2dd)\uc744 \uc0ac\uc6a9\ud558\uc5ec GRPO(Group Relative Policy Optimization) \uc54c\uace0\ub9ac\uc998\uc73c\ub85c \ud6c8\ub828. \uc790\uccb4 \uac80\uc99d, \ubc18\uc131, \uae34 CoT \uc0dd\uc131 \ub4f1 \ud765\ubbf8\ub85c\uc6b4 \ucd94\ub860 \ud589\ub3d9\uc774 \uc790\uc5f0\uc2a4\ub7fd\uac8c \ub4f1\uc7a5\ud588\uc9c0\ub9cc \uac00\ub3c5\uc131 \ubc0f \uc5b8\uc5b4 \ud63c\ud569 \ubb38\uc81c \ubc1c\uc0dd.<br \/>\nDeepSeek-R1: \ucf5c\ub4dc \uc2a4\ud0c0\ud2b8 \ub370\uc774\ud130\ub97c \ud65c\uc6a9\ud55c RL \uc801\uc6a9.<br \/>\n\ucf5c\ub4dc \uc2a4\ud0c0\ud2b8: \uc218\ucc9c \uac1c\uc758 \uae34 CoT \uc608\uc81c\ub85c \ubaa8\ub378\uc744 \ubbf8\uc138 \uc870\uc815\ud558\uc5ec \ucd08\uae30 RL \ud589\uc704\uc790(actor)\ub85c \uc0ac\uc6a9. \uac00\ub3c5\uc131 \ud5a5\uc0c1 \ubc0f \uc7a0\uc7ac\ub825 \ud5a5\uc0c1.<br \/>\n\ucd94\ub860 \uc9c0\ud5a5 RL: \ucf5c\ub4dc \uc2a4\ud0c0\ud2b8 \ub370\uc774\ud130\ub85c \ubbf8\uc138 \uc870\uc815\ub41c \ubaa8\ub378\uc5d0 R1-Zero\uc640 \ub3d9\uc77c\ud55c \ub300\uaddc\ubaa8 RL \ud6c8\ub828 \uc801\uc6a9. \uc5b8\uc5b4 \ud63c\ud569 \ubb38\uc81c \uc644\ud654\ub97c \uc704\ud574 \uc5b8\uc5b4 \uc77c\uad00\uc131 \ubcf4\uc0c1 \ub3c4\uc785.<br \/>\n\uac70\ubd80 \uc0d8\ud50c\ub9c1 \ubc0f \uc9c0\ub3c4 \ubbf8\uc138 \uc870\uc815: RL \ud6c8\ub828\uc5d0\uc11c \uc5bb\uc740 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc0ac\uc6a9\ud558\uc5ec \uac70\ubd80 \uc0d8\ud50c\ub9c1\uc744 \ud1b5\ud574 SFT \ub370\uc774\ud130 \uc218\uc9d1. \ucd94\ub860 \ub370\uc774\ud130 \uc678\uc5d0\ub3c4 \uc4f0\uae30, \uc0ac\uc2e4 \uc9c8\uc758\uc751\ub2f5, \uc790\uae30 \uc778\uc2dd \ub4f1 \ub2e4\uc591\ud55c \ub3c4\uba54\uc778\uc758 \ub370\uc774\ud130 \ud3ec\ud568.<br \/>\n\ubaa8\ub4e0 \uc2dc\ub098\ub9ac\uc624\uc5d0 \ub300\ud55c RL: \ubaa8\ub378\uc758 \uc720\uc6a9\uc131\uacfc \ubb34\ud574\uc131\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uba74\uc11c \ucd94\ub860 \ub2a5\ub825\uc744 \uac1c\uc120\ud558\uae30 \uc704\ud574 2\ucc28 RL \ub2e8\uacc4 \uc801\uc6a9. \ucd94\ub860 \ub370\uc774\ud130\uc5d0\ub294 \uaddc\uce59 \uae30\ubc18 \ubcf4\uc0c1, \uc77c\ubc18 \ub370\uc774\ud130\uc5d0\ub294 \ubcf4\uc0c1 \ubaa8\ub378 \uc0ac\uc6a9.<br \/>\n\ucd94\ucd9c: DeepSeek-R1\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 Qwen, Llama\uc640 \uac19\uc740 \uc18c\ud615 \ubc00\uc9d1 \ubaa8\ub378\ub85c \ucd94\ucd9c. 800k \uc0d8\ud50c \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud558\uc5ec SFT \uc801\uc6a9. \ub2e8\uc21c \ucd94\ucd9c\ub9cc\uc73c\ub85c\ub3c4 \uc18c\ud615 \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825\uc774 \ud06c\uac8c \ud5a5\uc0c1\ub428\uc744 \ud655\uc778.<br \/>\n\uacb0\ub860 \ubc0f \ud5a5\ud6c4 \uc5f0\uad6c \ubc29\ud5a5:<br \/>\nDeepSeek-R1\uc740 RL\uc744 \ud1b5\ud574 \ubaa8\ub378 \ucd94\ub860 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ub370 \uc131\uacf5\ud588\uc9c0\ub9cc, \ud568\uc218 \ud638\ucd9c, \ub2e4\uc911 \ud134 \ub300\ud654, \ubcf5\uc7a1\ud55c \uc5ed\ud560\uadf9, JSON \ucd9c\ub825 \ub4f1 \uc77c\ubc18\uc801\uc778 \ub2a5\ub825, \uc5b8\uc5b4 \ud63c\ud569 \ubb38\uc81c, \ud504\ub86c\ud504\ud2b8 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1, \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1 \uc791\uc5c5 \ub4f1\uc5d0\uc11c \uac1c\uc120\uc758 \uc5ec\uc9c0\uac00 \uc788\uc2b5\ub2c8\ub2e4. \ud5a5\ud6c4 \uc5f0\uad6c\uc5d0\uc11c\ub294 \uc774\ub7ec\ud55c \ubd80\ubd84\uc5d0 \uc9d1\uc911\ud558\uc5ec DeepSeek-R1\uc758 \uc131\ub2a5\uc744 \ub354\uc6b1 \ubc1c\uc804\uc2dc\ud0ac \uacc4\ud68d\uc785\ub2c8\ub2e4.<\/p><\/blockquote>\n<p>3.<br \/>\n<a href=\"https:\/\/www.facebook.com\/OccucySesamelStreet\" target=\"_blank\" rel=\"noopener\">\uc591\ubcd1\ucc2c<\/a>\ub2d8\uc774 <a href=\"https:\/\/www.nature.com\/articles\/d41586-025-00229-6?\" target=\"_blank\" rel=\"noopener\">China\u2019s cheap, open AI model DeepSeek thrills scientists<\/a>\uc758 \uc77c\ubd80\ub97c \ubc88\uc5ed\ud558\uc5ec \ud398\uc774\uc2a4\ubd81\uc5d0 \uc62c\ub838\uc2b5\ub2c8\ub2e4. \uadf8\uc911 \uc77c\ubd80\uc785\ub2c8\ub2e4.<\/p>\n<blockquote><p>1. \ub3c4\uc804 \ubaa8\ub378<br \/>\nR1\uc740 \uc911\uad6d\uc5d0\uc11c \uc77c\uace0 \uc788\ub294 \ub300\ud615 \uc5b8\uc5b4 \ubaa8\ub378(LLM) \ubd90\uc758 \uc77c\ubd80\ub2e4. \ud5e4\uc9c0\ud380\ub4dc\uc5d0\uc11c \ubd84\uc0ac\ud55c \ub525\uc2dd\uc740 \uc9c0\ub09c\ub2ec \uc800\uc608\uc0b0\uc73c\ub85c \uad6c\ucd95\ub418\uc5c8\uc74c\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 \uc8fc\uc694 \uacbd\uc7c1\uc5c5\uccb4\ub97c \ub2a5\uac00\ud558\ub294 V3\ub77c\ub294 \ucc57\ubd07\uc744 \ucd9c\uc2dc\ud558\uba74\uc11c \ube44\uad50\uc801 \ub4e3\ubcf4\uc7a1 \uc0c1\ud0dc\uc5d0\uc11c \ubc97\uc5b4\ub0ac\ub2e4. \uc804\ubb38\uac00\ub4e4\uc740 \ub525\uc2dd\uc774 \ubaa8\ub378 \ud6c8\ub828\uc5d0 \ud544\uc694\ud55c \ud558\ub4dc\uc6e8\uc5b4\ub97c \uc784\ub300\ud558\ub294 \ub370 \uc57d 600\ub9cc \ub2ec\ub7ec\uac00 \uc18c\uc694\ub410\uc744 \uac83\uc73c\ub85c \ucd94\uc815\ud558\ub294 \ubc18\uba74, 11\ubc30\uc758 \ucef4\ud4e8\ud305 \ub9ac\uc18c\uc2a4\ub97c \uc0ac\uc6a9\ud55c \uba54\ud0c0(Meta)\uc758 Llama 3.1 405B\uc758 \uacbd\uc6b0 6\ucc9c\ub9cc \ub2ec\ub7ec \uc774\uc0c1\uc774 \uc18c\uc694\ub41c \uac83\uc73c\ub85c \ucd94\uc815\ud558\uace0 \uc788\ub2e4.<br \/>\n\ub525\uc2dd\uc5d0 \ub300\ud55c \uc18c\ubb38 \uc911 \ud558\ub098\ub294, AI \ucc98\ub9ac\ub97c \uc704\ud574 \uc124\uacc4\ub41c \ucd5c\uace0\uc758 \ucef4\ud4e8\ud130 \uce69\uc5d0 \ub300\ud55c \uc911\uad6d \uae30\uc5c5\uc758 \uc811\uadfc\uc744 \uc81c\ud55c\ud558\ub294 \ubbf8\uad6d\uc758 \uc218\ucd9c \uaddc\uc81c\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 R1\uc744 \ub9cc\ub4dc\ub294 \ub370 \uc131\uacf5\ud588\ub2e4\ub294 \uac83\uc774\ub2e4. \u201c\uadf8\uac8c \uc911\uad6d\uc5d0\uc11c \ub098\uc654\ub2e4\ub294 \uc0ac\uc2e4\uc740 &#8216;\ucef4\ud4e8\ud305 \uaddc\ubaa8&#8217;\ubcf4\ub2e4 &#8216;\uc790\uc6d0\uc758 \ud6a8\uc728\uc131&#8217;\uc774 \ub354 \uc911\uc694\ud558\ub2e4\ub294 \uac83\uc744 \ubcf4\uc5ec\uc900\ub2e4\u201d\ub77c\uace0 \uc6cc\uc2f1\ud134\uc8fc \uc2dc\uc560\ud2c0\uc758 AI \uc5f0\uad6c\uc6d0 \ud504\ub791\uc218\uc544 \ucf5c\ub808(Fran\u00e7ois Chollet)\ub294 \ub9d0\ud55c\ub2e4.<br \/>\n\uc6cc\uc2f1\ud134\uc8fc \ubca8\ubdf0\uc758 \uae30\uc220 \uc804\ubb38\uac00\uc774\uc790 \ub300\ub9cc\uc5d0 \ubcf8\uc0ac\ub97c \ub454 \ubab0\uc785\ud615 \uae30\uc220 \ud68c\uc0ac HTC\uc5d0\uc11c \uc77c\ud558\ub294 \uc568\ube48 \uc655 \uadf8\ub808\uc774\ub9b0(Alvin Wang Graylin)\uc740 X\uc5d0 \ub2e4\uc74c\uacfc \uac19\uc774 \uc37c\ub2e4. \u201c\ub525\uc2dd\uc758 \ubc1c\uc804\uc740 \ud55c\ub54c \ubbf8\uad6d\uc774 \ucc28\uc9c0\ud588\ub358 \uac83\uc73c\ub85c \uc5ec\uaca8\uc9c0\ub294 \uc120\ub450\uad8c\uacfc\uc758 \uaca9\ucc28\uac00 \uc0c1\ub2f9\ud788 \uc881\uc544\uc84c\uc74c\uc744 \uc2dc\uc0ac\ud55c\ub2e4. \uc591\uad6d(\uf978\u570b)\uc740 \ud604\uc7ac\uc758 \uc2b9\uc0b0 \uc5c6\ub294 \uad70\ube44 \uacbd\uc7c1 \ubc29\uc2dd\uc744 \uacc4\uc18d\ud558\ub294 \ub300\uc2e0 \uace0\uae09 AI \uad6c\ucd95\uc744 \uc704\ud55c \ud611\ub825\uc801 \uc811\uadfc \ubc29\uc2dd\uc744 \ucd94\uad6c\ud574\uc57c \ud55c\ub2e4.\u201d<\/p>\n<p>2. \uc0ac\uace0(\u601d\u8003)\uc758 \uc0ac\uc2ac<br \/>\nLLM\uc740 \uc218\uc2ed\uc5b5 \uac1c\uc758 \ud14d\uc2a4\ud2b8 \uc0d8\ud50c\uc744 \ud559\uc2b5\ud558\uc5ec &#8216;\ud1a0\ud070&#8217;\uc774\ub77c\ub294 \ub2e8\uc5b4 \ubd80\ubd84\uc73c\ub85c \uc798\ub77c\ub0b4\uace0 \ub370\uc774\ud130\uc5d0\uc11c \ud328\ud134\uc744 \ud559\uc2b5\ud55c\ub2e4. \uc774\ub7ec\ud55c \uc5f0\uad00\uc131\uc744 \ud1b5\ud574 \ubaa8\ub378\uc740 \ubb38\uc7a5\uc758 \ud6c4\uc18d \ud1a0\ud070\uc744 \uc608\uce21\ud560 \uc218 \uc788\ub2e4. \uadf8\ub7ec\ub098 LLM\uc740 &#8216;\ud658\uac01(hallucination)&#8217;\uc774\ub77c\ub294 \ud604\uc0c1\uc744 \ud1b5\ud574 \uc0ac\uc2e4\uc744 \ub9cc\ub4e4\uc5b4\ub0b4\ub294 \uacbd\ud5a5\uc774 \uc788\uc73c\uba70(https:\/\/www.nature.com\/articles\/d41586-025-00068-5), \ubb38\uc81c\ub97c \ucd94\ub860\ud558\ub294 \ub370 \uc5b4\ub824\uc6c0\uc744 \uacaa\ub294 \uacbd\uc6b0\uac00 \ub9ce\ub2e4.<br \/>\no1\uacfc \ub9c8\ucc2c\uac00\uc9c0\ub85c, R1\uc740 &#8216;\uc0ac\uace0\uc758 \uc0ac\uc2ac(chain of thought)&#8217; \ubc29\ubc95\uc744 \uc0ac\uc6a9\ud558\uc5ec \ub354 \ubcf5\uc7a1\ud55c \uc791\uc5c5(\uc608: \ub54c\ub54c\ub85c \uc811\uadfc \ubc29\uc2dd\uc744 \uc5ed\ucd94\uc801\ud558\uace0 \ud3c9\uac00\ud568)\uc744 \ud574\uacb0\ud558\ub294 LLM\uc758 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0a8\ub2e4. \ub525\uc2dd\uc740 \uac15\ud654 \ud559\uc2b5(reinforcement learning)\uc744 \ud1b5\ud574 V3\ub97c &#8216;\ubbf8\uc138 \uc870\uc815&#8217;\ud568\uc73c\ub85c\uc368 R1\uc744 \ub9cc\ub4e4\uc5c8\ub294\ub370, \uc774\ub294 \ubaa8\ub378\uc774 &#8216;\uc815\ub2f5\uc5d0 \ub3c4\ub2ec\ud558\ub294 \uac83&#8217;\uacfc &#8216;\uc0ac\uace0\ub97c \uc124\uba85\ud558\ub294 \ubc29\uc2dd\uc73c\ub85c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \uac83&#8217;\uc5d0 \ub300\ud574 \ubcf4\uc0c1\uc744 \uc81c\uacf5\ud588\ub2e4.<br \/>\n&#8220;\ub525\uc2dd\uc740 \ucef4\ud4e8\ud305 \uc131\ub2a5\uc774 \uc81c\ud55c\ub418\uc5b4 \uc788\uc5c8\uae30 \ub54c\ubb38\uc5d0 \uc54c\uace0\ub9ac\uc998\uc801\uc73c\ub85c \ud601\uc2e0\ud560 \uc218 \uc788\uc5c8\ub2e4&#8221;\ub77c\uace0 \uc601\uad6d \uc5d0\ub4e0\ubc84\ub7ec \ub300\ud559\uad50\uc758 AI \uc5f0\uad6c\uc6d0 \uc6ec\ub2e4 \ub9ac(Wenda Li)\ub294 \ub9d0\ud55c\ub2e4. \uac15\ud654 \ud559\uc2b5\uc744 \ud558\ub294 \ub3d9\uc548, \uc5f0\uad6c\ud300\uc740 \uac01 \ub2e8\uacc4\uc5d0\uc11c \ubaa8\ub378\uc758 \uc9c4\ud589 \uc0c1\ud669\uc744 \ucd94\uc815\ud558\ub294 \ub300\uc2e0 \ubcc4\ub3c4\uc758 \ub124\ud2b8\uc6cc\ud06c\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud3c9\uac00\ud588\ub2e4. \uc774\ub294 \ud6c8\ub828 \ubc0f \uc6b4\uc601 \ube44\uc6a9\uc744 \uc904\uc774\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub418\uc5c8\ub2e4\uace0 \uc601\uad6d \ucf00\uc784\ube0c\ub9ac\uc9c0 \ub300\ud559\uad50\uc758 \ucef4\ud4e8\ud130 \uacfc\ud559\uc790 \ub9c8\ud14c\uc57c \uc7a0\ub2c9(Mateja Jamnik)\uc740 \ub9d0\ud55c\ub2e4. \uc5f0\uad6c\uc9c4\uc740 \ub610\ud55c \ubaa8\ub378\uc774 \uac01 \uc791\uc5c5\uacfc \uad00\ub828\ub41c \ubd80\ubd84\ub9cc \ud65c\uc131\ud654\ud560 \uc218 \uc788\ub294 &#8216;\uc804\ubb38\uac00 \ud63c\ud569(mixture-of-experts)&#8217; \uc544\ud0a4\ud14d\ucc98\ub97c \uc0ac\uc6a9\ud588\ub2e4.<\/p>\n<p>\ubaa8\ub378\uacfc \ud568\uaed8 \uc81c\uacf5\ub418\ub294 \uae30\uc220 \ub17c\ubb38(https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-R1)\uc5d0 \ubcf4\uace0\ub41c \ubca4\uce58\ub9c8\ud06c \ud14c\uc2a4\ud2b8\uc5d0\uc11c, \ub525\uc2dd-R1\uc740 UC \ubc84\ud074\ub9ac\uc758 \uc5f0\uad6c\uc9c4\uc774 \ub9cc\ub4e0 \uc218\ud559 \ubb38\uc81c \uc138\ud2b8 MATH-500\uc5d0\uc11c 97.3%\ub97c \uae30\ub85d\ud588\uc73c\uba70, Codeforces \ub300\ud68c\uc5d0\uc11c \uc778\uac04 \ucc38\uac00\uc790\uc758 96.3%\ubcf4\ub2e4 \uc6b0\uc218\ud55c \uc131\uc801\uc744 \uac70\ub450\uc5c8\ub2e4. \uc774\ub294 o1\uc758 \ub2a5\ub825\uacfc \ub300\ub4f1\ud55c \uc218\uc900\uc774\uba70, o3\ub294 \ube44\uad50 \ub300\uc0c1\uc5d0 \ud3ec\ud568\ub418\uc9c0 \uc54a\uc558\ub2e4(\uc544\ub798 \ub313\uae00\uc758 &#8216;AI \ub77c\uc774\ubc8c&#8217; \ucc38\uc870).<br \/>\n\ubca4\uce58\ub9c8\ud06c\uac00 \ubaa8\ub378\uc758 \uc9c4\uc815\ud55c \ucd94\ub860 \ub610\ub294 \uc77c\ubc18\ud654 \ub2a5\ub825\uc744 \ud3ec\ucc29\ud558\ub294\uc9c0, \uc544\ub2c8\uba74 \ub2e8\uc21c\ud788 \uadf8\ub7ec\ud55c \ud14c\uc2a4\ud2b8\ub97c \ud1b5\uacfc\ud558\uae30 \uc704\ud55c \uac83\uc778\uc9c0\ub294 \uc54c\uae30 \uc5b4\ub835\ub2e4. \ud558\uc9c0\ub9cc R1\uc740 \uac1c\ubc29\ub418\uc5b4 \uc788\uae30 \ub54c\ubb38\uc5d0 \uc5f0\uad6c\uc790\ub4e4\uc774 \uadf8 \uc0ac\uace0\uc758 \uc0ac\uc2ac\uc5d0 \uc811\uadfc\ud560 \uc218 \uc788\ub2e4\uace0 \ucf00\uc784\ube0c\ub9ac\uc9c0 \ub300\ud559\uc758 \ucef4\ud4e8\ud130 \uacfc\ud559\uc790 \ub9c8\ub974\ucf54 \ub3c4\uc2a4 \uc0b0\ud1a0\uc2a4(Marco Dos Santos)\ub294 \ub9d0\ud55c\ub2e4. \u201c\uc774\ub97c \ud1b5\ud574 \ubaa8\ub378\uc758 \ucd94\ub860 \uacfc\uc815\uc744 \ub354 \uc798 \ud574\uc11d\ud560 \uc218 \uc788\ub2e4&#8221;\ub77c\uace0 \uadf8\ub294 \ub9d0\ud55c\ub2e4.<br \/>\n\uc774\ubbf8 \uacfc\ud559\uc790\ub4e4\uc740 R1\uc758 \ub2a5\ub825\uc744 \ud14c\uc2a4\ud2b8\ud558\uace0 \uc788\ub2e4. \ud06c\ub80c\uc740 \ub450 \uacbd\uc7c1 \ubaa8\ub378\uc5d0 3,000\uac1c\uc758 \uc5f0\uad6c \uc544\uc774\ub514\uc5b4\ub97c \uc5bc\ub9c8\ub098 \ud765\ubbf8\ub85c\uc6b4\uc9c0\uc5d0 \ub530\ub77c \ubd84\ub958\ud558\ub3c4\ub85d \ud558\uace0 \uadf8 \uacb0\uacfc\ub97c \uc0ac\ub78c\uc774 \ub9cc\ub4e0 \uc21c\uc704\uc640 \ube44\uad50\ud588\ub2e4. \uc774 \uce21\uc815\uc5d0\uc11c R1\uc740 o1\uc5d0 \ube44\ud574 \uc57d\uac04 \uc800\uc870\ud55c \uc131\uc801\uc744 \ubcf4\uc600\ub2e4. \ud558\uc9c0\ub9cc R1\uc740 \uc591\uc790 \uad11\ud559\uc758 \ud2b9\uc815 \uacc4\uc0b0\uc5d0\uc11c o1\uc744 \uc55e\uc130\ub2e4\uace0 \ud06c\ub80c\uc740 \ub9d0\ud55c\ub2e4. \u201c\uc774\ub294 \ub9e4\uc6b0 \uc778\uc0c1\uc801\uc778 \uacb0\uacfc\ub2e4\u201d\ub77c\uace0 \uadf8\ub294 \ub9d0\ud55c\ub2e4.<\/p><\/blockquote>\n<p>4.<br \/>\n<a href=\"https:\/\/www.facebook.com\/friends\/?profile_id=100001721966098\" rel=\"noopener\" target=\"_blank\">\uc774\uc9c4\uc6d0<\/a>\ub2d8\uc774 \ud398\uc774\uc2a4\ubd81\uc5d0 \uc62c\ub9b0 \uae00\uc785\ub2c8\ub2e4. <strong>DeepSeek-R1\uc740 \uc9c4\uc9dc\ub85c \uac00\uc131\ube44 \ubaa8\ub378\uc778\uac00?<\/strong>\ub77c\ub294 \uc8fc\uc81c\uc785\ub2c8\ub2e4.<\/p>\n<blockquote><p>\uc774\uc81c \uc9c4\uc2e4\uc774 \ubb34\uc5c7\uc778\uc9c0\ub294 \ubcc4\ub85c \uc911\uc694\ud558\uc9c0 \uc54a\uc744 \ub9cc\ud07c DeepSeek\uc740 \uc774\ubbf8 \uc131\uacf5\ud55c \ubaa8\ub378\uc774 \ub418\uc5b4\ubc84\ub9b0 \uac83 \uac19\uc9c0\ub9cc\u2026 \uc81c\uac00 \uc0dd\uac01\ud558\ub294 \ubc14\ub97c \uc815\ub9ac\ud574\uc11c \uc801\uc5b4\ubd05\ub2c8\ub2e4. DeepSeek-R1\uc774 \ubc1c\ud45c\ub41c \uc774\ud6c4 Nvidia \uc8fc\uac00\uac00 20%\uac00\uae4c\uc774 \ud3ed\ub77d\ud558\uace0, \ub2e4\ub978 \ube45\ud14c\ud06c \uae30\uc5c5\ub4e4\uc758 \uc8fc\uac00\ub3c4 \uc904\uc904\uc774 \ud3ed\ub77d\ud55c \uac83\uc744 \ubcf4\uba74 \uc774\uc81c \uc5c4\uccad\ub098\uac8c \ub9ce\uc740 GPU\uc640 computing power\ub97c \uc55e\uc138\uc6b0\uc9c0 \uc54a\uc544\ub3c4 \uc88b\uc740 \ubaa8\ub378\uc744 \ub9cc\ub4e4 \uc218 \uc788\ub2e4. \uadf8\ub798\uc11c \ube45\ud14c\ud06c\ub4e4\ub9cc\uc758 \uc601\uc5ed\uc774\uc5c8\ub358 foundation model\uc744 \ub204\uad6c\ub098 \ub9cc\ub4e4 \uc218 \uc788\uc744 \uac70\ub2e4. \ub77c\ub294 \uc758\ubbf8\ub85c \ubc1b\uc544\ub4e4\uc774\ub294 \uac83 \uac19\uc740\ub370\uc694\u2026 \ud558\ub098\uc529 \uc9da\uc5b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4 (\uc77c\ubd80 \uc798\ubabb\ub41c \ub0b4\uc6a9\uc774 \uc788\uc744 \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4)<br \/>\n 1. \uc6b0\uc120 \uac00\uc131\ube44\uac00 \uc88b\ub2e4\ub294 \ub9d0\uc774 \ud6c8\ub828\uc5d0\uc11c\uc758 \uac00\uc131\ube44\uc778\uc9c0 \ucd94\ub860 \ud639\uc740 \uc11c\ube44\uc2a4\uc5d0\uc11c\uc758 \uac00\uc131\ube44\uc778\uc9c0\ub97c \uba85\ud655\ud788 \ud574\uc57c\ud560 \ud544\uc694\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uc9c0\uae08\uae4c\uc9c0 \ub098\uc624\ub294 \ub0b4\uc6a9\uc744 \uc815\ub9ac\ud574\ubcf4\uba74 DeepSeek\uc774 H800 GPU\ub85c 557.6\ub9cc\ub2ec\ub7ec(\uc57d 80\uc5b5\uc6d0)\ub85c \ud559\uc2b5\uc744 \ud588\ub2e4\ub294 \uac8c \uc81c\uc77c \ud654\uc81c\uac00 \ub418\uace0 \uc788\uace0(\uadf8\ub9bc1), API \ube44\uc6a9\ub3c4 o1 \ub300\ube44\ud574\uc11c 1\/10 \uc774\ud558\ub85c \uc11c\ube44\uc2a4 \ud558\uace0 \uc788\ub2e4\ub294 \uc0ac\uc2e4(\uadf8\ub9bc2)\ub3c4 \uc8fc\ubaa9\ubc1b\uace0 \uc788\uc2b5\ub2c8\ub2e4.<br \/>\n 2.\ud559\uc2b5\ubd80\ud130 \uc0dd\uac01\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774\uac83\ub3c4 \ub450\uac00\uc9c0\ub85c \ub098\ub220\uc11c \uc0dd\uac01\ud574\ubcfc \uc218 \uc788\ub294\ub370, \uccab\ubc88\uc9f8\ub294 \ud559\uc2b5\ube44\uc6a9\uc774 80\uc5b5\ubc16\uc5d0(?) \uc548\ub4e4\uc5c8\ub2e4\ub294 \ubaa8\ub378\uc740 \uc774\ubc88\uc5d0 \ud654\uc81c\uac00 \ub41c DeepSeek-R1\uc774 \uc544\ub2c8\uace0 DeepSeek-V3\uc785\ub2c8\ub2e4. \uadf8\ub7fc R1\uc740 \ud559\uc2b5\ud558\ub294\ub370 \uc5bc\ub9c8\uac00 \ub4e4\uc5c8\uc744\uae4c\uc694? \uadf8\uac74 \uc54c \uc218 \uc5c6\uc9c0\ub9cc R1\uc740 DeepSeek-V3\uac00 \uc788\ub2e4\ub294 \uc804\uc81c \ud558\uc5d0\uc11c \uc2dc\uc791\ub429\ub2c8\ub2e4 \uadf8\ub7ec\ub2c8\uae4c \uc77c\ub2e8 V3\ub97c \ud559\uc2b5\ud558\ub294 \ube44\uc6a9\uc774 \ub4e4\uace0 \uac70\uae30\uc5d0 \ucd94\uac00\ub85c \ube44\uc6a9\uc774 \ub4ed\ub2c8\ub2e4. R1\uc744 \uc5b4\ub5bb\uac8c \ud559\uc2b5\ud588\ub294\uc9c0 technical report\ub97c \uc0b4\ud3b4\ubcf4\uba74, 1) \uc6b0\uc120 V3\ub97c \uac00\uc9c0\uace0 GPRO\ub77c\ub294 \ubc29\ubc95\uc73c\ub85c \uac15\ud654\ud559\uc2b5\uc744 \ud1b5\ud574\uc11c reasoning\uc744 \ud560 \uc218 \uc788\ub3c4\ub85d \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc774 DeepSeek-R1-Zero \uc785\ub2c8\ub2e4. 2) \uc774\uc81c DeepSeek-R1-Zero\ub97c \uac00\uc9c0\uace0 reasoning\uc744 \ud559\uc2b5\ud560 CoT(Chain of Thought) data\ub97c \uc0dd\uc131\ud574\ub0b8 \ub2e4\uc74c \uc0ac\ub78c\ub4e4\uc774 \ubd99\uc5b4\uc11c \ub370\uc774\ud130\ub97c \uc5f4\uc2ec\ud788 \uac00\uacf5\ud569\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \uc774 data\ub97c \uc774\uc6a9\ud574\uc11c SFT(Supervised Fine-Tuning)\uc73c\ub85c DeepSeek-V3-base \ubaa8\ub378\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4(cold start) 3) \uc774\ub807\uac8c \ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \ub2e4\uc2dc 1)\uc5d0\uc11c \ud55c \uac83\uacfc \uac19\uc740 \ubc29\uc2dd\uc73c\ub85c \uac15\ud654\ud559\uc2b5\uc744 \ud1b5\ud574\uc11c reasoning\uc744 \uc798 \ud560 \uc218 \uc788\ub3c4\ub85d \ud559\uc2b5\uc2dc\ud0b5\ub2c8\ub2e4. 4) 3)\uc5d0\uc11c \ud559\uc2b5\ud560 \ub54c \uc0ac\uc6a9\ud55c checkpoint\ub97c \uc774\uc6a9\ud558\uc5ec reasoning data(60k)\ub97c \uc0dd\uc131\ud558\uace0, DeepSeek-V3\ub97c \ud559\uc2b5\ud560 \ub54c \uc37c\ub358 SFT data\ub4f1\uc744 \ud65c\uc6a9\ud558\uc5ec non-reasoning data(20k)\ub97c \ub9cc\ub4e0 \ub2e4\uc74c \ub610 SFT\ub85c \ud559\uc2b5\ud569\ub2c8\ub2e4. 5) \ub9c8\uc9c0\ub9c9\uc73c\ub85c \uac15\ud654\ud559\uc2b5\uc744 \uc0ac\uc6a9\ud558\uc5ec helpfulness\uc640 harmlessness\ub97c \ud5a5\uc0c1\uc2dc\ud0a4\uace0 \ucd94\ub860\ub2a5\ub825\uc744 \uc815\uad50\ud654\ud558\uac8c\ub429\ub2c8\ub2e4. \uad49\uc7a5\ud788 \ubcf5\uc7a1\ud558\uc9c0\uc694? H800\uc73c\ub85c 80\uc5b5 \ub4e4\uc5ec\uc11c \ud559\uc2b5\ud55c \uc774\ud6c4\uc5d0 \uc774\ub807\uac8c \ub9ce\uc740 \uacfc\uc815\uc744 \ucd94\uac00\ub85c \ud559\uc2b5\ud574\uc57c DeepSeek-R1\uc774 \ub9cc\ub4e4\uc5b4\uc9d1\ub2c8\ub2e4. \uc774\uac8c \uc9c4\uc9dc \ube44\uc6a9 \ud6a8\uc728\uc801\uc778 \ud559\uc2b5\uc778\uac00\uc694? \uc5b4\ucc28\ud53c \uc0c1\ub300\uc801\uc778 \ube44\uad50\ub77c\uc11c \uc81c\uac00 \ub9de\ub2e4 \ud2c0\ub9ac\ub2e4 \uc598\uae30\ud560 \uc218\ub294 \uc5c6\uc9c0\ub9cc, \uc0ac\ub78c\ub4e4\uc774 \uc598\uae30\ud558\ub294 \uac83\ubcf4\ub2e4\ub294 \ud6e8\uc52c \ub354 \ub9ce\uc740 \ube44\uc6a9\uc774 \ub4e4\uc5b4\uac14\ub2e4\ub294 \uac78 \uc54c \uc218 \uc788\uc2b5\ub2c8\ub2e4.<br \/>\n 3.\uadf8\ub7fc\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 \uc791\ub144\uc5d0 \ub098\uc628 DeepSeek-V3\ub3c4 \ucda9\ubd84\ud788 \uc88b\uc740 \ubaa8\ub378\uc774\uae34 \ud569\ub2c8\ub2e4. \uadf8\ub7fc \uc774\uac70\ub77c\ub3c4 \uc801\uc740 \ube44\uc6a9 \ub4e4\uc5ec\uc11c \ud559\uc2b5\ud588\uc73c\uba74 \uc88b\uc740\uac70 \uc544\ub2c8\ub0d0? \ub77c\uace0 \uc0dd\uac01\ud560 \uc218 \uc788\uaca0\uc8e0. \uadf8\ub7f0\ub370 \uc774 \ubd80\ubd84\ub3c4 \ub2e4\uc2dc \uc0dd\uac01\ud574\ubd10\uc57c\ud560 \uac83\uc740 1\ud68c \ud559\uc2b5\ud558\ub294\ub370 \ub4dc\ub294 \ube44\uc6a9\uc774 \uc801\ub2e4\uace0 \ud574\uc11c \uc2e4\uc81c\ub85c \uc774 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uc5b4\ub0b4\ub294\ub370 \ub4e4\uc5b4\uac04 \ube44\uc6a9\uc774 \uc801\ub2e4\uace0\ub294 \ud560 \uc218 \uc5c6\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4. \ud55c\ubc29\uc5d0 \ud559\uc2b5\uc774 \uc798 \ub41c\ub2e4\uba74 80\uc5b5\uc73c\ub85c \uac00\ub2a5\ud560 \uc218 \uc788\uc9c0\ub9cc, \uc88b\uc740 \uacb0\uacfc\ub97c \uc704\ud574\uc11c\ub294 \ub2f9\uc5f0\ud788 \uc5ec\ub7ec\ubc88\uc758 \uc2dc\ud589\ucc29\uc624\uac00 \uc788\uc5b4\uc57c \ud558\uace0 \ucd5c\uace0\uc758 \ud559\uc2b5 \ub808\uc2dc\ud53c\ub97c \ucc3e\uc544\ub0b4\ub294 \uacfc\uc815\uc774 \ud544\uc694\ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c \uc5bc\ub9c8\ub098 \ub9ce\uc740 GPU\uc640 computing power\uac00 \uc0ac\uc6a9\ub418\uc5c8\ub294\uc9c0\ub294 \uc544\ubb34\ub3c4 \ubaa8\ub985\ub2c8\ub2e4. \uc2e4\uc81c\ub85c DeepSeek\uc5d0 H100\uc774 5\ub9cc\ub300 \uc788\ub2e4\ub294 \uc598\uae30\ub3c4 \uc788\uad6c\uc694&#8230; \uc544\ub9c8\ub3c4 \uc5c4\uccad\ub09c \ube44\uc6a9\uc744 \ub4e4\uc5ec\uc11c \ud559\uc2b5\ud558\ub294 \uacfc\uc815\uc774 \uc788\uc5c8\uc744 \uac70\ub77c\uace0 \uc0dd\uac01\ud569\ub2c8\ub2e4. \ud558\uc9c0\ub9cc 1\ud68c \ud559\uc2b5\ud558\ub294\ub370 \ub4dc\ub294 \ube44\uc6a9\uc744 \uc904\uc778 \uac83 \uc790\uccb4\ub294 \ud6cc\ub96d\ud55c \uc5c5\uc801\uc774\ub77c\uace0 \uc0dd\uac01\ud569\ub2c8\ub2e4. \uc81c\uac00 \uac71\uc815\ud558\ub294 \uac74 \uc774\uc81c \uc6b0\ub9ac\ub3c4 80\uc5b5 \uc788\uc73c\uba74 \uc774\ub7f0\uac70 \ub9cc\ub4e4 \uc218 \uc788\uaca0\ub124\ub77c\uace0 \uc0dd\uac01\ud558\ub294 \uc0ac\ub78c\ub4e4\uc774 \ub9ce\uc544\uc9c0\ub294 \uac83\uc785\ub2c8\ub2e4.<br \/>\n 4. \uc774\uc81c \ucd94\ub860(\uc11c\ube44\uc2a4)\ub85c \ub118\uc5b4\uac00\ubcfc\uae4c\uc694? \uc0ac\uc2e4 \ucd94\ub860 \uc11c\ube44\uc2a4\uc5d0\uc11c \ube44\uc6a9\uc774 \uc2e4\uc81c\ub85c \uc5bc\ub9c8\ub098 \ub4dc\ub294\uc9c0\ub294 API \uac00\uaca9\uc73c\ub85c \ud310\ub2e8\ud558\uae30 \uc27d\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4. \ud68c\uc0ac\ub9c8\ub2e4 \uc774\uc724\uc774 \ucc28\uc9c0\ud558\ub294 \ube44\uc728\uc774 \ub2e4\ub974\uace0, \ub610 \uc5b4\ub5a4 \ud68c\uc0ac\ub294 \uc190\ud574\ub97c \uac10\uc218\ud558\uba74\uc11c\uae4c\uc9c0 \uc2dc\uc7a5\uc810\uc720\uc728 \ub4f1\uc744 \uc704\ud574\uc11c \uc11c\ube44\uc2a4\ub97c \ud560 \uc218 \uc788\uae30 \ub54c\ubb38\uc774\uc8e0. \uc2e4\uc81c\ub85c together.ai\uc5d0 \uac00\ubcf4\uc2dc\uba74 \ud604\uc7ac \ub2e4\uc591\ud55c \uc624\ud508\ubaa8\ub378\ub4e4\uc744 \uc11c\ube44\uc2a4 \ud558\uace0 \uc788\ub294\ub370, (\uadf8\ub9bc3)\uc5d0\uc11c \ubcf4\uc2dc\ub294 \uac83\ucc98\ub7fc DeepSeek-R1\uc740 \uac00\uaca9\uc774 \ub9e4\uc6b0\ub9e4\uc6b0 \ube44\uc309\ub2c8\ub2e4. LLaMA 405B \ubaa8\ub378\ubcf4\ub2e4 2\ubc30 \ube44\uc2fc \uac00\uaca9\uc73c\ub85c \uc11c\ube44\uc2a4 \ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud55c \ud68c\uc0ac\uc5d0\uc11c \uc5ec\ub7ec\uac1c\uc758 \ubaa8\ub378\uc744 \uc11c\ube59\ud558\uace0 \uc788\ub2e4\uba74 \ubaa8\ub378\ubcc4\ub85c \uc774\uc724\uc744 \ube44\uc2b7\ud558\uac8c \ub0a8\uae38 \uac00\ub2a5\uc131\uc774 \ub192\uace0 \uadf8\ub807\ub2e4\uba74 \uc774 \ube44\uc6a9\uc774 \ub354 \uc2e4\uc81c \ube44\uc6a9\uacfc \uac00\uae5d\ub2e4\uace0 \uc0dd\uac01\ud558\ub294\uac8c \ud0c0\ub2f9\ud558\uc9c0 \uc54a\uc744\uae4c\uc694?<br \/>\n 5. DeepSeek-R1\uc774 \uc2f8\uac8c \uc11c\ube59\ud560 \uc218 \uc788\ub294 \uc774\uc720\ub85c (\uc798\ubabb)\uc54c\ub824\uc9c4 \uac83 \uc911 \ub300\ud45c\uc801\uc778 \uac83\uc774 MoE(Mixture of Expert) \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\ub2e4\ub294 \uac83\uc785\ub2c8\ub2e4. DeepSeek model\uc740 671B\uc758 \ud30c\ub77c\ubbf8\ud130\ub97c \uac00\uc9c0\uace0 \uc788\ub294\ub370 \uadf8 \uc911\uc5d0 1\ubc88\uc758 inference\uc5d0\uc11c\ub294 37B\ub9cc \ud65c\uc131\ud654\ub41c\ub2e4\uace0 \ud569\ub2c8\ub2e4. \uadf8\ub7f0\ub370 \ud074\ub77c\uc6b0\ub4dc\uc5d0\uc11c \ub2e4\uc591\ud55c request\ub97c \ubc1b\uc544\uc11c \ub3d9\uc2dc\uc5d0 \ucc98\ub9ac\ud558\ub294 \uacbd\uc6b0\uc5d0\ub294 \uacb0\uad6d 671B \uc804\uccb4\ub97c \uc5ec\ub7ec\ub300\uc758 GPU\uc5d0 \ub098\ub220\uc11c \ub2e4 \uac00\uc9c0\uace0 \uc788\uc5b4\uc57c \ud569\ub2c8\ub2e4. \uac01 request\ub9c8\ub2e4 \uc0ac\uc6a9\ud558\ub294 expert\uac00 \ub2e4 \ub2e4\ub974\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4. \ub354\uad70\ub2e4\ub098 expert\ub4e4\uc774 GPU \uc5ec\ub7ec\uac1c\uc5d0 \ud769\uc5b4\uc838 \uc788\uac8c \ub418\uae30 \ub54c\ubb38\uc5d0, GPU\uac04\uc5d0 data\ub97c \uc8fc\uace0\ubc1b\uc544\uc57c \ud558\ub294 \uc77c\ub4e4\uc774 \ub354 \ub9ce\uc544\uc9c0\uac8c \ub429\ub2c8\ub2e4. \ub530\ub77c\uc11c MoE \uad6c\uc870 \uc790\uccb4\ub294 \ucd94\ub860(inference)\uc5d0 \uc0c1\ub300\uc801\uc73c\ub85c \ube44\ud6a8\uc728\uc801\uc778 \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \uc774\ub7f0 \uc774\uc720\ub85c \uc0c1\ub300\uc801\uc778 \ucd94\ub860 \uc18d\ub3c4(tokens\/sec)\ub3c4 \ub9e4\uc6b0 \ub290\ub824\uc9c8 \uc218\ubc16\uc5d0 \uc5c6\uc2b5\ub2c8\ub2e4. (\uadf8\ub9bc4)\ub97c \ubcf4\uc2dc\uba74 R1\uc774 o1 \ub300\ube44\ud574\uc11c \ucd94\ub860 \uc18d\ub3c4\uac00 40%\uc774\uc0c1 \ub290\ub9b0 \uac83\uc744 \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.<br \/>\n 6. \uacb0\ub860\uc801\uc73c\ub85c \ud559\uc2b5\uc774\ub4e0 \ucd94\ub860\uc774\ub4e0 \uac00\uc131\ube44\uac00 \uc88b\uc740 \ubaa8\ub378\uc778\uac00\uc5d0 \ub300\ud55c \uc9c8\ubb38\uc5d0 \uc800\ub294 \uadf8\ub807\uc9c0 \uc54a\ub2e4\uace0 \uc0dd\uac01\ud569\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \uadf8\ub807\ub2e4\uace0 DeepSeek\uc744 \ud3c9\uac00\uc808\ud558\ud560 \ud544\uc694\ub294 \uc5c6\ub2e4\uace0 \uc0dd\uac01\ud569\ub2c8\ub2e4. \uc131\ub2a5\uc774 \uc88b\uc740 \ubaa8\ub378\uc778 \uac74 \ud655\uc2e4\ud558\uace0, \uc774\ub7f0 \ubaa8\ub378\ub4e4\uc774 \uc624\ud508\ub418\uc5c8\ub2e4\ub294 \uac83\uc740 \ud070 \uc758\ubbf8\uac00 \uc788\ub2e4\uace0 \uc0dd\uac01\ud569\ub2c8\ub2e4.<br \/>\n 7. \ub9c8\uc9c0\ub9c9\uc73c\ub85c \uc5c4\ubc00\ud558\uac8c \ub9d0\ud574\uc11c DeepSeek\uc740 \uc624\ud508\uc18c\uc2a4\ub294 \uc544\ub2c8\uace0 \uc624\ud508\ubaa8\ub378\uc785\ub2c8\ub2e4(weight\ub9cc \uacf5\uac1c, \ud559\uc2b5\ub370\uc774\ud130\ub098 \ub808\uc2dc\ud53c\ub294 \uacf5\uac1c\ud558\uc9c0 \uc54a\uc74c). \uadf8\ub798\uc11c \uc774\uc81c \ub204\uad6c\ub098 \uc774\ub7f0 \ubaa8\ub378\uc744 \uc27d\uac8c \ub9cc\ub4e4 \uc218 \uc788\ub290\ub0d0\ub294 \uc9c8\ubb38\uc5d0\ub294 \uc544\ub2c8\ub77c\uace0 \ud560 \uc218 \uc788\uc744 \uac83 \uac19\uc2b5\ub2c8\ub2e4. \uc704\uc5d0 \ud559\uc2b5 \ubc29\ubc95\uc744 \uac04\ub2e8\ud558\uac8c \uc815\ub9ac\ud55c \ubd80\ubd84\uc5d0\uc11c\ub3c4 \uc54c \uc218 \uc788\ub2e4\uc2dc\ud53c technical report\uc5d0 \uc790\uc138\ud788 \uc801\ud788\uc9c0 \uc54a\uc740 \uc218\ub9ce\uc740 \ub0b4\uc6a9\uc774 \uc788\uc744 \uac83\uc774\uace0, \uacf5\uac1c\ub41c \ub0b4\uc6a9\ub9cc\uc73c\ub85c\ub3c4 \uc218\ub9ce\uc740 \ub178\ub825\uacfc \ub178\ud558\uc6b0\uac00 \ud544\uc694\ud558\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4. DeepSeek-R1\uc740 openAI\uc758 o1 \ubaa8\ub378\uc774 \uc5b4\ub5bb\uac8c reasoning\uc744 \uc798\ud558\ub294\uc9c0\uc5d0 \ub300\ud574\uc11c \ubaa8\ub450\uac00 \uad81\uae08\ud574\ud588\ub358 \ubd80\ubd84\uc5d0 \ub300\ud55c \ub2f5\uc744 \uc5b4\ub290\uc815\ub3c4 \uc54c\ub824\uc92c\uace0, \ub204\uad70\uac00\ub294 \uc774\ub7ec\ud55c \ubc29\ubc95\uc744 \uc774\uc6a9\ud574\uc11c \ube44\uc2b7\ud558\uac70\ub098 \ub354 \ub098\uc740 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uc5b4\ub0bc \uc218 \uc788\ub294 \ubc1c\ud310\uc744 \ub9c8\ub828\ud574\uc92c\ub2e4\uace0 \uc0dd\uac01\ud569\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 DeepSeek\uacfc \uac19\uc740 \ud68c\uc0ac\ub4e4\uc774 \ub9ce\uc774 \ub098\uc624\uae30\ub97c \uae30\ub300\ud574\ubd05\ub2c8\ub2e4.<\/p><\/blockquote>\n<p>\ucc38\uace0\ub85c Deepseek\uac00 \uad6c\ud604\ud55c 3FS\ub77c\ub294 \ubd84\uc0b0\ud30c\uc77c\uc2dc\uc2a4\ud15c\uc744 \uc18c\uac1c\ud558\ub294 \uae00\uc785\ub2c8\ub2e4.<\/p>\n<p><a href=\"https:\/\/maknee.github.io\/blog\/2025\/3FS-Performance-Journal-1\/\" rel=\"noopener\" target=\"_blank\">An Intro to DeepSeek&#8217;s Distributed File System<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. DeepSeek\uc640 \uad00\ub828\ud55c \uae00\uc744 \uc77d\uc73c\uba74\uc11c \uc774\ud574\ub3c4\uac00 \ubb34\ucc99 \ub5a8\uc5b4\uc9d1\ub2c8\ub2e4. AI \uae30\uc220\uc5d0 \ub300\ud55c \uc774\ud574\ub3c4 \ub0ae\uace0 Nvidia GPU\uc5d0 \ub300\ud55c \uc774\ud574\ub3c4 \ub0ae\uc2b5\ub2c8\ub2e4. \uadf8\ub798\uc11c A100\uacfc H800\uc774 \ubb34\uc5c7\uc778\uc9c0, \uc5b4\ub5bb\uac8c \ub2e4\ub978\uc9c0\ub3c4 \uc798 \ubaa8\ub985\ub2c8\ub2e4. \uc5d4\ube44\ub514\uc544 AI \uce69: A100 A800 H100 H800 B200\uc5d0 \uc62c\ub77c\uc628 \ube44\uad50\ud45c\ub85c \uc774\ud574\ub97c \ub192\ud614\uc2b5\ub2c8\ub2e4.\uc6b0\uc120&#8230;. Dropbox \ubd80\uc0ac\uc7a5 Morgan Brown\uc774 \uc4f4 Why DeepSeek&#8217;s AI Breakthrough May Be the Most Crucial One Yet. linkedin\uc5d0 \uc62c\ub9b0 \uae00\uc785\ub2c8\ub2e4. AI Product \ub2f4\ub2f4&#8230;<br \/><a class=\"read-more-button\" href=\"https:\/\/smallake.kr\/?p=22379\">Read more<\/a><\/p>\n","protected":false},"author":1,"featured_media":34845,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"nf_dc_page":"","om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"DeepSeek \ub17c\ubb38\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2},"jetpack_post_was_ever_published":false},"categories":[2684],"tags":[4165],"class_list":["post-22379","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-2684","tag-deepseek"],"aioseo_notices":[],"aioseo_head":"\n\t\t<!-- All in One SEO 4.9.8 - aioseo.com -->\n\t<meta name=\"robots\" content=\"max-image-preview:large\" \/>\n\t<meta name=\"author\" content=\"smallake\"\/>\n\t<meta name=\"keywords\" content=\"deepseek\" \/>\n\t<link rel=\"canonical\" href=\"https:\/\/smallake.kr\/?p=22379\" \/>\n\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 4.9.8\" \/>\n\t\t<meta property=\"og:locale\" content=\"ko_KR\" \/>\n\t\t<meta property=\"og:site_name\" content=\"\uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218 | \uc0b4\uc544\uc628 \ub0a0\uc758 \ud754\uc801, \uc0b4\uc544\uac08 \ub0a0\uc758 \uae30\ub85d\" \/>\n\t\t<meta property=\"og:type\" content=\"article\" \/>\n\t\t<meta property=\"og:title\" content=\"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574 | \uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218\" \/>\n\t\t<meta property=\"og:url\" content=\"https:\/\/smallake.kr\/?p=22379\" \/>\n\t\t<meta property=\"article:published_time\" content=\"2025-01-28T07:35:58+00:00\" \/>\n\t\t<meta property=\"article:modified_time\" content=\"2025-04-22T05:39:00+00:00\" \/>\n\t\t<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/iumgroup\" \/>\n\t\t<meta name=\"twitter:card\" content=\"summary\" \/>\n\t\t<meta name=\"twitter:site\" content=\"@smallake\" \/>\n\t\t<meta name=\"twitter:title\" content=\"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574 | \uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218\" \/>\n\t\t<meta name=\"twitter:creator\" content=\"@smallake\" \/>\n\t\t<script type=\"application\/ld+json\" class=\"aioseo-schema\">\n\t\t\t{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#article\",\"name\":\"DeepSeek AI \\ubaa8\\ub378\\uc5d0 \\ub300\\ud55c \\uc5ec\\ub7ec\\uac00\\uc9c0 \\uc774\\ud574 | \\uadf8\\ub300\\uc548\\uc758\\uc791\\uc740\\ud638\\uc218\",\"headline\":\"DeepSeek AI \\ubaa8\\ub378\\uc5d0 \\ub300\\ud55c \\uc5ec\\ub7ec\\uac00\\uc9c0 \\uc774\\ud574\",\"author\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/?author=1#author\"},\"publisher\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/#person\"},\"image\":{\"@type\":\"ImageObject\",\"url\":\"https:\\\/\\\/i0.wp.com\\\/smallake.kr\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/FE73buGX0AQOLgx-e1738048201645.jpeg?fit=1080%2C810&ssl=1\",\"width\":1080,\"height\":810},\"datePublished\":\"2025-01-28T16:35:58+09:00\",\"dateModified\":\"2025-04-22T14:39:00+09:00\",\"inLanguage\":\"ko-KR\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#webpage\"},\"isPartOf\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#webpage\"},\"articleSection\":\"\\ud540\\ud14c\\ud06c,\\ube14\\ub85d\\uccb4\\uc778 \\ubc0f AI, DeepSeek\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#breadcrumblist\",\"itemListElement\":[{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/smallake.kr#listItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/smallake.kr\",\"nextItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?cat=2684#listItem\",\"name\":\"\\ud540\\ud14c\\ud06c,\\ube14\\ub85d\\uccb4\\uc778 \\ubc0f AI\"}},{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?cat=2684#listItem\",\"position\":2,\"name\":\"\\ud540\\ud14c\\ud06c,\\ube14\\ub85d\\uccb4\\uc778 \\ubc0f AI\",\"item\":\"https:\\\/\\\/smallake.kr\\\/?cat=2684\",\"nextItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#listItem\",\"name\":\"DeepSeek AI \\ubaa8\\ub378\\uc5d0 \\ub300\\ud55c \\uc5ec\\ub7ec\\uac00\\uc9c0 \\uc774\\ud574\"},\"previousItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/smallake.kr#listItem\",\"name\":\"Home\"}},{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#listItem\",\"position\":3,\"name\":\"DeepSeek AI \\ubaa8\\ub378\\uc5d0 \\ub300\\ud55c \\uc5ec\\ub7ec\\uac00\\uc9c0 \\uc774\\ud574\",\"previousItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?cat=2684#listItem\",\"name\":\"\\ud540\\ud14c\\ud06c,\\ube14\\ub85d\\uccb4\\uc778 \\ubc0f AI\"}}]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/#person\",\"name\":\"smallake\",\"image\":{\"@type\":\"ImageObject\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#personImage\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/85bc072c859f1b0c21b1f8ac82394cf817695e7d0f250d2dcfabf8ce7217b415?s=96&r=g\",\"width\":96,\"height\":96,\"caption\":\"smallake\"}},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?author=1#author\",\"url\":\"https:\\\/\\\/smallake.kr\\\/?author=1\",\"name\":\"smallake\",\"image\":{\"@type\":\"ImageObject\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#authorImage\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/85bc072c859f1b0c21b1f8ac82394cf817695e7d0f250d2dcfabf8ce7217b415?s=96&r=g\",\"width\":96,\"height\":96,\"caption\":\"smallake\"}},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#webpage\",\"url\":\"https:\\\/\\\/smallake.kr\\\/?p=22379\",\"name\":\"DeepSeek AI \\ubaa8\\ub378\\uc5d0 \\ub300\\ud55c \\uc5ec\\ub7ec\\uac00\\uc9c0 \\uc774\\ud574 | \\uadf8\\ub300\\uc548\\uc758\\uc791\\uc740\\ud638\\uc218\",\"inLanguage\":\"ko-KR\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/#website\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#breadcrumblist\"},\"author\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/?author=1#author\"},\"creator\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/?author=1#author\"},\"image\":{\"@type\":\"ImageObject\",\"url\":\"https:\\\/\\\/i0.wp.com\\\/smallake.kr\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/FE73buGX0AQOLgx-e1738048201645.jpeg?fit=1080%2C810&ssl=1\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379\\\/#mainImage\",\"width\":1080,\"height\":810},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/?p=22379#mainImage\"},\"datePublished\":\"2025-01-28T16:35:58+09:00\",\"dateModified\":\"2025-04-22T14:39:00+09:00\"},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/smallake.kr\\\/#website\",\"url\":\"https:\\\/\\\/smallake.kr\\\/\",\"name\":\"\\uadf8\\ub300\\uc548\\uc758 \\uc791\\uc740 \\ud638\\uc218\",\"description\":\"\\uc0b4\\uc544\\uc628 \\ub0a0\\uc758 \\ud754\\uc801, \\uc0b4\\uc544\\uac08 \\ub0a0\\uc758 \\uae30\\ub85d\",\"inLanguage\":\"ko-KR\",\"publisher\":{\"@id\":\"https:\\\/\\\/smallake.kr\\\/#person\"}}]}\n\t\t<\/script>\n\t\t<!-- All in One SEO -->\n\n","aioseo_head_json":{"title":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574 | \uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218","description":"","canonical_url":"https:\/\/smallake.kr\/?p=22379","robots":"max-image-preview:large","keywords":"deepseek","webmasterTools":{"miscellaneous":""},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/smallake.kr\/?p=22379#article","name":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574 | \uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218","headline":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574","author":{"@id":"https:\/\/smallake.kr\/?author=1#author"},"publisher":{"@id":"https:\/\/smallake.kr\/#person"},"image":{"@type":"ImageObject","url":"https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/FE73buGX0AQOLgx-e1738048201645.jpeg?fit=1080%2C810&ssl=1","width":1080,"height":810},"datePublished":"2025-01-28T16:35:58+09:00","dateModified":"2025-04-22T14:39:00+09:00","inLanguage":"ko-KR","mainEntityOfPage":{"@id":"https:\/\/smallake.kr\/?p=22379#webpage"},"isPartOf":{"@id":"https:\/\/smallake.kr\/?p=22379#webpage"},"articleSection":"\ud540\ud14c\ud06c,\ube14\ub85d\uccb4\uc778 \ubc0f AI, DeepSeek"},{"@type":"BreadcrumbList","@id":"https:\/\/smallake.kr\/?p=22379#breadcrumblist","itemListElement":[{"@type":"ListItem","@id":"https:\/\/smallake.kr#listItem","position":1,"name":"Home","item":"https:\/\/smallake.kr","nextItem":{"@type":"ListItem","@id":"https:\/\/smallake.kr\/?cat=2684#listItem","name":"\ud540\ud14c\ud06c,\ube14\ub85d\uccb4\uc778 \ubc0f AI"}},{"@type":"ListItem","@id":"https:\/\/smallake.kr\/?cat=2684#listItem","position":2,"name":"\ud540\ud14c\ud06c,\ube14\ub85d\uccb4\uc778 \ubc0f AI","item":"https:\/\/smallake.kr\/?cat=2684","nextItem":{"@type":"ListItem","@id":"https:\/\/smallake.kr\/?p=22379#listItem","name":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574"},"previousItem":{"@type":"ListItem","@id":"https:\/\/smallake.kr#listItem","name":"Home"}},{"@type":"ListItem","@id":"https:\/\/smallake.kr\/?p=22379#listItem","position":3,"name":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574","previousItem":{"@type":"ListItem","@id":"https:\/\/smallake.kr\/?cat=2684#listItem","name":"\ud540\ud14c\ud06c,\ube14\ub85d\uccb4\uc778 \ubc0f AI"}}]},{"@type":"Person","@id":"https:\/\/smallake.kr\/#person","name":"smallake","image":{"@type":"ImageObject","@id":"https:\/\/smallake.kr\/?p=22379#personImage","url":"https:\/\/secure.gravatar.com\/avatar\/85bc072c859f1b0c21b1f8ac82394cf817695e7d0f250d2dcfabf8ce7217b415?s=96&r=g","width":96,"height":96,"caption":"smallake"}},{"@type":"Person","@id":"https:\/\/smallake.kr\/?author=1#author","url":"https:\/\/smallake.kr\/?author=1","name":"smallake","image":{"@type":"ImageObject","@id":"https:\/\/smallake.kr\/?p=22379#authorImage","url":"https:\/\/secure.gravatar.com\/avatar\/85bc072c859f1b0c21b1f8ac82394cf817695e7d0f250d2dcfabf8ce7217b415?s=96&r=g","width":96,"height":96,"caption":"smallake"}},{"@type":"WebPage","@id":"https:\/\/smallake.kr\/?p=22379#webpage","url":"https:\/\/smallake.kr\/?p=22379","name":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574 | \uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218","inLanguage":"ko-KR","isPartOf":{"@id":"https:\/\/smallake.kr\/#website"},"breadcrumb":{"@id":"https:\/\/smallake.kr\/?p=22379#breadcrumblist"},"author":{"@id":"https:\/\/smallake.kr\/?author=1#author"},"creator":{"@id":"https:\/\/smallake.kr\/?author=1#author"},"image":{"@type":"ImageObject","url":"https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/FE73buGX0AQOLgx-e1738048201645.jpeg?fit=1080%2C810&ssl=1","@id":"https:\/\/smallake.kr\/?p=22379\/#mainImage","width":1080,"height":810},"primaryImageOfPage":{"@id":"https:\/\/smallake.kr\/?p=22379#mainImage"},"datePublished":"2025-01-28T16:35:58+09:00","dateModified":"2025-04-22T14:39:00+09:00"},{"@type":"WebSite","@id":"https:\/\/smallake.kr\/#website","url":"https:\/\/smallake.kr\/","name":"\uadf8\ub300\uc548\uc758 \uc791\uc740 \ud638\uc218","description":"\uc0b4\uc544\uc628 \ub0a0\uc758 \ud754\uc801, \uc0b4\uc544\uac08 \ub0a0\uc758 \uae30\ub85d","inLanguage":"ko-KR","publisher":{"@id":"https:\/\/smallake.kr\/#person"}}]},"og:locale":"ko_KR","og:site_name":"\uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218 | \uc0b4\uc544\uc628 \ub0a0\uc758 \ud754\uc801, \uc0b4\uc544\uac08 \ub0a0\uc758 \uae30\ub85d","og:type":"article","og:title":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574 | \uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218","og:url":"https:\/\/smallake.kr\/?p=22379","article:published_time":"2025-01-28T07:35:58+00:00","article:modified_time":"2025-04-22T05:39:00+00:00","article:publisher":"https:\/\/www.facebook.com\/iumgroup","twitter:card":"summary","twitter:site":"@smallake","twitter:title":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574 | \uadf8\ub300\uc548\uc758\uc791\uc740\ud638\uc218","twitter:creator":"@smallake"},"aioseo_meta_data":{"post_id":"22379","title":null,"description":null,"keywords":null,"keyphrases":{"focus":{"keyphrase":"","score":0,"analysis":{"keyphraseInTitle":{"score":0,"maxScore":9,"error":1}}},"additional":[]},"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":"","og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"Article","isEnabled":true},"graphs":[],"defaultGraph":"","defaultPostTypeGraph":""},"schema_type":null,"schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":"-1","robots_max_videopreview":"-1","robots_max_imagepreview":"large","priority":null,"frequency":"default","location":null,"local_seo":null,"breadcrumb_settings":null,"limit_modified_date":false,"ai":null,"created":"2020-12-21 05:27:06","updated":"2025-06-05 14:19:11","seo_analyzer_scan_date":null},"aioseo_breadcrumb":"<div class=\"aioseo-breadcrumbs\"><span class=\"aioseo-breadcrumb\">\n\t\t\t<a href=\"https:\/\/smallake.kr\" title=\"Home\">Home<\/a>\n\t\t<\/span><span class=\"aioseo-breadcrumb-separator\">&raquo;<\/span><span class=\"aioseo-breadcrumb\">\n\t\t\t<a href=\"https:\/\/smallake.kr\/?cat=2684\" title=\"\ud540\ud14c\ud06c,\ube14\ub85d\uccb4\uc778 \ubc0f AI\">\ud540\ud14c\ud06c,\ube14\ub85d\uccb4\uc778 \ubc0f AI<\/a>\n\t\t<\/span><span class=\"aioseo-breadcrumb-separator\">&raquo;<\/span><span class=\"aioseo-breadcrumb\">\n\t\t\tDeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574\n\t\t<\/span><\/div>","aioseo_breadcrumb_json":[{"label":"Home","link":"https:\/\/smallake.kr"},{"label":"\ud540\ud14c\ud06c,\ube14\ub85d\uccb4\uc778 \ubc0f AI","link":"https:\/\/smallake.kr\/?cat=2684"},{"label":"DeepSeek AI \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5ec\ub7ec\uac00\uc9c0 \uc774\ud574","link":"https:\/\/smallake.kr\/?p=22379"}],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"https:\/\/i0.wp.com\/smallake.kr\/wp-content\/uploads\/2025\/01\/FE73buGX0AQOLgx-e1738048201645.jpeg?fit=1080%2C810&ssl=1","jetpack_shortlink":"https:\/\/wp.me\/p2yULC-5OX","jetpack_sharing_enabled":true,"jetpack_likes_enabled":false,"_links":{"self":[{"href":"https:\/\/smallake.kr\/index.php?rest_route=\/wp\/v2\/posts\/22379","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/smallake.kr\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/smallake.kr\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/smallake.kr\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/smallake.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=22379"}],"version-history":[{"count":0,"href":"https:\/\/smallake.kr\/index.php?rest_route=\/wp\/v2\/posts\/22379\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/smallake.kr\/index.php?rest_route=\/wp\/v2\/media\/34845"}],"wp:attachment":[{"href":"https:\/\/smallake.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=22379"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/smallake.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=22379"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/smallake.kr\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=22379"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}