{
    "componentChunkName": "component---src-templates-post-tsx",
    "path": "/llm-abs-202311/",
    "result": {"data":{"logo":null,"markdownRemark":{"html":"<h1>LLM Paper Abstract - 2023.11</h1>\n<p>LLM 관련해서 워낙 많은 논문들이 나와서, 최근에 읽은 논문들에 대해 간단하게 요약한 리스트입니다.<br>\n아래 리스트중에는 가볍게 읽어본 논문들이 포함되어 있어서 요약에 틀린 내용이 있을 수 있습니다.</p>\n<h2>Abstract</h2>\n<h3>11월 4주차</h3>\n<ul>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.11045\"><strong><code class=\"language-text\">Orca 2: Teaching Small Language Models How to Reason</code></strong></a> : 마이크로소프트에서 공개한 Orca-2 모델 관련 논문. 7B, 13B 모델의 reasoning 능력을 개선해보려는 시도와 결과에 대해 리포트 함.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.14904\"><strong><code class=\"language-text\">LLM-ASSISTED CODE CLEANING FOR TRAINING ACCURATE CODE GENERATORS</code></strong></a> : 코드 생성 AI를 학습시킬때, 리팩토링을 통해 코드 가독성을 높인 코드로 학습시키면 모델 성능이 크게 좋아진다는 논문.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.16989\"><strong><code class=\"language-text\">ChatGPT’s One-year Anniversary: Are Open-Source Large Language Models Catching up?</code></strong></a> : ChatGPT 1주년 기념으로, 오픈소스 LLM들이 ChatGPT, GPT4들을 얼마나 따라잡았는지를 평가한 논문</p>\n</li>\n<li>\n<p><a href=\"https://www.nature.com/articles/s41586-023-06735-9\"><strong><code class=\"language-text\">Scaling deep learning for materials discovery</code></strong></a> : Google DeepMind 논문에서 신소재 검색을 위한 AI 모델 ‘GNoME’을 통해 38만개의 신소재를 발견했다는 논문. 그동안 인류가 발견했던 2만개와 비교하면 20배에 달하는 능력이라고..</p>\n</li>\n</ul>\n<h3>11월 3주차</h3>\n<ul>\n<li><a href=\"https://arxiv.org/abs/2311.09277\"><strong><code class=\"language-text\">Contrastive Chain-of-Thought Prompting</code></strong></a> : CoT의 한계를 지적한 논문. CoT가 비록 논리적인 추론에 대해서 중요한 부분이긴 하지만, 프롬프트 작성 시에 잘못 작성 했을 시에 대해 피할 수 없고, 이에 따라 많은 오류를 발생시킴. 따라서 instruction에 대한 negative sample과 positive sample을 동시에 입력으로 넣어 contrastive한 방식으로 추론할 수 있는 기법을 만들어낸 논문.</li>\n</ul>\n<h3>11월 2주차</h3>\n<ul>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.09210\"><strong><code class=\"language-text\">Chain-of-Note: Enhancing Robustness in Retrieval-Augmented Language Models</code></strong></a> : Retrieval-augmented language models에서 정확한 정보를 retrieval해도 아직 부족한 점이 많다는 점에서 나온 논문. Chain-of-Noting(CoN)이라는 방식을 활용하여서 retrieval된 도큐먼트에 대한 순차적인 읽기를 생성해서 질문과의 연관성을 평가하여 최종 답변을 도출한다는 논문. RAG 분야 성능을 높이려고 노력한 논문.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.00871\"><strong><code class=\"language-text\">Pretraining Data Mixtures Enable Narrow Model Selection Capabilities in Transformer Models</code></strong></a> : 추론 시 사전학습된 데이터의 distribution과 너무 차이가 나는 prediction은 잘 하지 못한다는 내용의 논문. 즉 in-context learning은 학습 데이터와 많이 차이 나지 않는 distribution 범위에서만 가능하다는 논문.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2306.15895\"><strong><code class=\"language-text\">Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias</code></strong></a> : ChatGPT와 같은 LLM을 데이터 생성에 사용하는 법에 관련한 논문. 다양성과 편향에 대해서도 같이 다룸.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.02462\"><strong><code class=\"language-text\">Levels of AGI: Operationalizing Progress on the Path to AGI</code></strong></a> : 구글, 딥마인드에서 AI 능력을 5단계로 구분하고 레벨을 매겨 AGI로 가는 길에 대해 제시한 논문. 내용 자체는 뻔한 내용. ChatGPT를 1단계로 설정.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.01906\"><strong><code class=\"language-text\">SIMPLIFYING TRANSFORMER BLOCKS</code></strong></a> : 트랜스포머 구조를 단순화하여 15%의 학습 속도 개선과 15%의 파라미터를 줄이는 방향으로 개선안을 제안한 논문.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2306.04563\"><strong><code class=\"language-text\">ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models</code></strong></a> : ChatGPT의 Joke에 관련한 논문. ChatGPT가 Joke를 잘하는듯 보이지만, 사실 ChatGPT가 하는 대부분(90%)의 joke가 25개의 joke에 한정되어있다고 함.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2308.04948\"><strong><code class=\"language-text\">EXTRAPOLATING LARGE LANGUAGE MODELS TO NON-ENGLISH BY ALIGNING LANGUAGES</code></strong></a> : LLaMA에서 non-english 능력을 이끌어내기 위한 방법에 대해 제시한 논문. 번역 태스크로 instruction 튜닝을 하는 등의 방법을 사용함.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.03285v2\"><strong><code class=\"language-text\">S-LORA: SERVING THOUSANDS OF CONCURRENT LORA ADAPTERS</code></strong></a> : LoRA Adapter가 여러개일 때 서빙하는 방법에 대한 논문.</p>\n</li>\n<li>\n<p><a href=\"https://arxiv.org/abs/2311.04205\"><strong><code class=\"language-text\">Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves</code></strong></a> : LLM에게 질문을 rephrase 시키고 답변시켰을 때 더 좋은 결과가 있었다는 논문.</p>\n</li>\n</ul>\n<h3>11월 1주차</h3>\n<ul>\n<li><a href=\"https://arxiv.org/abs/2309.00071\"><strong><code class=\"language-text\">YaRN: Efficient Context Window Extension of Large Language Models</code></strong></a> : YaRN(Yet another RoPE extensioN method)이라는 Rotary Embedding을 개선한 버전을 제안한 논문. 기존 pre-training시 사용한 max length보다 더 길게 확장할 수 있음을 보임 (LLaMA를 128k 토큰까지 확장)</li>\n</ul>","htmlAst":{"type":"root","children":[{"type":"element","tagName":"h1","properties":{},"children":[{"type":"text","value":"LLM Paper Abstract - 2023.11"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"text","value":"LLM 관련해서 워낙 많은 논문들이 나와서, 최근에 읽은 논문들에 대해 간단하게 요약한 리스트입니다."},{"type":"element","tagName":"br","properties":{},"children":[]},{"type":"text","value":"\n아래 리스트중에는 가볍게 읽어본 논문들이 포함되어 있어서 요약에 틀린 내용이 있을 수 있습니다."}]},{"type":"text","value":"\n"},{"type":"element","tagName":"h2","properties":{},"children":[{"type":"text","value":"Abstract"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"h3","properties":{},"children":[{"type":"text","value":"11월 4주차"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"ul","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.11045"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Orca 2: Teaching Small Language Models How to Reason"}]}]}]},{"type":"text","value":" : 마이크로소프트에서 공개한 Orca-2 모델 관련 논문. 7B, 13B 모델의 reasoning 능력을 개선해보려는 시도와 결과에 대해 리포트 함."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.14904"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"LLM-ASSISTED CODE CLEANING FOR TRAINING ACCURATE CODE GENERATORS"}]}]}]},{"type":"text","value":" : 코드 생성 AI를 학습시킬때, 리팩토링을 통해 코드 가독성을 높인 코드로 학습시키면 모델 성능이 크게 좋아진다는 논문."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.16989"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"ChatGPT’s One-year Anniversary: Are Open-Source Large Language Models Catching up?"}]}]}]},{"type":"text","value":" : ChatGPT 1주년 기념으로, 오픈소스 LLM들이 ChatGPT, GPT4들을 얼마나 따라잡았는지를 평가한 논문"}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://www.nature.com/articles/s41586-023-06735-9"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Scaling deep learning for materials discovery"}]}]}]},{"type":"text","value":" : Google DeepMind 논문에서 신소재 검색을 위한 AI 모델 ‘GNoME’을 통해 38만개의 신소재를 발견했다는 논문. 그동안 인류가 발견했던 2만개와 비교하면 20배에 달하는 능력이라고.."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"h3","properties":{},"children":[{"type":"text","value":"11월 3주차"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"ul","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.09277"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Contrastive Chain-of-Thought Prompting"}]}]}]},{"type":"text","value":" : CoT의 한계를 지적한 논문. CoT가 비록 논리적인 추론에 대해서 중요한 부분이긴 하지만, 프롬프트 작성 시에 잘못 작성 했을 시에 대해 피할 수 없고, 이에 따라 많은 오류를 발생시킴. 따라서 instruction에 대한 negative sample과 positive sample을 동시에 입력으로 넣어 contrastive한 방식으로 추론할 수 있는 기법을 만들어낸 논문."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"h3","properties":{},"children":[{"type":"text","value":"11월 2주차"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"ul","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.09210"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Chain-of-Note: Enhancing Robustness in Retrieval-Augmented Language Models"}]}]}]},{"type":"text","value":" : Retrieval-augmented language models에서 정확한 정보를 retrieval해도 아직 부족한 점이 많다는 점에서 나온 논문. Chain-of-Noting(CoN)이라는 방식을 활용하여서 retrieval된 도큐먼트에 대한 순차적인 읽기를 생성해서 질문과의 연관성을 평가하여 최종 답변을 도출한다는 논문. RAG 분야 성능을 높이려고 노력한 논문."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.00871"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Pretraining Data Mixtures Enable Narrow Model Selection Capabilities in Transformer Models"}]}]}]},{"type":"text","value":" : 추론 시 사전학습된 데이터의 distribution과 너무 차이가 나는 prediction은 잘 하지 못한다는 내용의 논문. 즉 in-context learning은 학습 데이터와 많이 차이 나지 않는 distribution 범위에서만 가능하다는 논문."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2306.15895"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias"}]}]}]},{"type":"text","value":" : ChatGPT와 같은 LLM을 데이터 생성에 사용하는 법에 관련한 논문. 다양성과 편향에 대해서도 같이 다룸."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.02462"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Levels of AGI: Operationalizing Progress on the Path to AGI"}]}]}]},{"type":"text","value":" : 구글, 딥마인드에서 AI 능력을 5단계로 구분하고 레벨을 매겨 AGI로 가는 길에 대해 제시한 논문. 내용 자체는 뻔한 내용. ChatGPT를 1단계로 설정."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.01906"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"SIMPLIFYING TRANSFORMER BLOCKS"}]}]}]},{"type":"text","value":" : 트랜스포머 구조를 단순화하여 15%의 학습 속도 개선과 15%의 파라미터를 줄이는 방향으로 개선안을 제안한 논문."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2306.04563"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models"}]}]}]},{"type":"text","value":" : ChatGPT의 Joke에 관련한 논문. ChatGPT가 Joke를 잘하는듯 보이지만, 사실 ChatGPT가 하는 대부분(90%)의 joke가 25개의 joke에 한정되어있다고 함."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2308.04948"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"EXTRAPOLATING LARGE LANGUAGE MODELS TO NON-ENGLISH BY ALIGNING LANGUAGES"}]}]}]},{"type":"text","value":" : LLaMA에서 non-english 능력을 이끌어내기 위한 방법에 대해 제시한 논문. 번역 태스크로 instruction 튜닝을 하는 등의 방법을 사용함."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.03285v2"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"S-LORA: SERVING THOUSANDS OF CONCURRENT LORA ADAPTERS"}]}]}]},{"type":"text","value":" : LoRA Adapter가 여러개일 때 서빙하는 방법에 대한 논문."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"p","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2311.04205"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves"}]}]}]},{"type":"text","value":" : LLM에게 질문을 rephrase 시키고 답변시켰을 때 더 좋은 결과가 있었다는 논문."}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"h3","properties":{},"children":[{"type":"text","value":"11월 1주차"}]},{"type":"text","value":"\n"},{"type":"element","tagName":"ul","properties":{},"children":[{"type":"text","value":"\n"},{"type":"element","tagName":"li","properties":{},"children":[{"type":"element","tagName":"a","properties":{"href":"https://arxiv.org/abs/2309.00071"},"children":[{"type":"element","tagName":"strong","properties":{},"children":[{"type":"element","tagName":"code","properties":{"className":["language-text"]},"children":[{"type":"text","value":"YaRN: Efficient Context Window Extension of Large Language Models"}]}]}]},{"type":"text","value":" : YaRN(Yet another RoPE extensioN method)이라는 Rotary Embedding을 개선한 버전을 제안한 논문. 기존 pre-training시 사용한 max length보다 더 길게 확장할 수 있음을 보임 (LLaMA를 128k 토큰까지 확장)"}]},{"type":"text","value":"\n"}]}],"data":{"quirksMode":false}},"excerpt":"LLM Paper Abstract - 2023.11 LLM…","fields":{"readingTime":{"text":"6 min read"}},"frontmatter":{"title":"LLM Paper Abstract - 2023.11","userDate":"5 December 2023","date":"2023-12-05T10:00:00.000Z","tags":["nlp"],"excerpt":null,"image":{"childImageSharp":{"gatsbyImageData":{"layout":"fullWidth","backgroundColor":"#080808","images":{"fallback":{"src":"/static/388c93716b829f310da6d75d5b6d10cd/3832f/llm-abs-202310.jpg","srcSet":"/static/388c93716b829f310da6d75d5b6d10cd/5224e/llm-abs-202310.jpg 750w,\n/static/388c93716b829f310da6d75d5b6d10cd/49438/llm-abs-202310.jpg 1080w,\n/static/388c93716b829f310da6d75d5b6d10cd/eede5/llm-abs-202310.jpg 1366w,\n/static/388c93716b829f310da6d75d5b6d10cd/3832f/llm-abs-202310.jpg 1920w","sizes":"100vw"},"sources":[{"srcSet":"/static/388c93716b829f310da6d75d5b6d10cd/f7f77/llm-abs-202310.webp 750w,\n/static/388c93716b829f310da6d75d5b6d10cd/27610/llm-abs-202310.webp 1080w,\n/static/388c93716b829f310da6d75d5b6d10cd/f698a/llm-abs-202310.webp 1366w,\n/static/388c93716b829f310da6d75d5b6d10cd/e4c47/llm-abs-202310.webp 1920w","type":"image/webp","sizes":"100vw"}]},"width":1,"height":0.75}}},"author":[{"id":"Soohwan Kim","bio":"Co-founder/A.I. engineer at TUNiB.","avatar":{"children":[{"gatsbyImageData":{"layout":"fullWidth","backgroundColor":"#282838","images":{"fallback":{"src":"/static/a9e6b445142b247ee4cfa66155398bb2/0d6f4/soohwan.png","srcSet":"/static/a9e6b445142b247ee4cfa66155398bb2/248f9/soohwan.png 40w,\n/static/a9e6b445142b247ee4cfa66155398bb2/fd435/soohwan.png 80w,\n/static/a9e6b445142b247ee4cfa66155398bb2/0d6f4/soohwan.png 120w","sizes":"100vw"},"sources":[{"srcSet":"/static/a9e6b445142b247ee4cfa66155398bb2/e7f45/soohwan.webp 40w,\n/static/a9e6b445142b247ee4cfa66155398bb2/589ec/soohwan.webp 80w,\n/static/a9e6b445142b247ee4cfa66155398bb2/71a38/soohwan.webp 120w","type":"image/webp","sizes":"100vw"}]},"width":1,"height":0.6833333333333333}}]}}]}},"relatedPosts":{"totalCount":50,"edges":[{"node":{"id":"6529d72e-80e7-5d61-a672-d66276a3f641","excerpt":"MoE(Mixture of Experts) 기초부터 DeepSeek 혁신까지 작년 이맘때쯤 DeepSeek-V3가 저비용으로 엄청난 성능을 보이면서 화제가 되었습니다. 그 핵심 기술인 MoE(Mixture of Experts…","frontmatter":{"title":"MoE(Mixture of Experts) 기초부터 DeepSeek 혁신까지","date":"2026-01-15T01:11:55.000Z"},"fields":{"readingTime":{"text":"8 min read"},"slug":"/developed-moe/"}}},{"node":{"id":"8b49d3ef-4ce3-568c-9d71-240ff17fc3e0","excerpt":"BERT는 사실 Diffusion 모델이였다?! 최근 굉장히 흥미로운 글을 읽게되어 공유합니다. 원문 : link BERT와 Diffusion이 같은 방식이다?! NLP 연구자들에게 BERT는 너무 익숙한 모델입니다. 201…","frontmatter":{"title":"BERT는 사실 Diffusion 모델이였다?!","date":"2025-10-21T12:00:00.000Z"},"fields":{"readingTime":{"text":"8 min read"},"slug":"/bert_diffusion/"}}},{"node":{"id":"7c7e1676-ea84-58de-970c-ddec9aa10660","excerpt":"RLHF는 수다쟁이를 만든다?! (Does RLHF Breed Verbose Chatterboxes?!) RLHF(Reinforcement Learning from Human Feedback)는 OpenAI의 ChatGPT…","frontmatter":{"title":"RLHF는 수다쟁이를 만든다?! (Does RLHF Breed Verbose Chatterboxes?!)","date":"2024-03-13T01:11:55.000Z"},"fields":{"readingTime":{"text":"7 min read"},"slug":"/rlhf-vervosity/"}}},{"node":{"id":"2b9d3e22-1796-5fab-878d-5941d2e76e9d","excerpt":"LLM Paper Abstract - 2023.12 LLM…","frontmatter":{"title":"LLM Paper Abstract - 2023.12","date":"2024-01-05T10:00:00.000Z"},"fields":{"readingTime":{"text":"5 min read"},"slug":"/llm-abs-202312/"}}},{"node":{"id":"f43fa33b-917b-5c00-8462-937d99592ad7","excerpt":"What it MoE? (Mixture of Experts) 현존 최강 LLM인 GPT-4에서 “MoE (Mixture of Experts)” 방식을 채택하여 사용하고 있다고 알려졌는데요, 최근 AI계의 뜨거운 감자 Mistral AI…","frontmatter":{"title":"What is MoE? (Mixture of Experts)","date":"2023-12-22T01:11:55.000Z"},"fields":{"readingTime":{"text":"9 min read"},"slug":"/moe/"}}}]}},"pageContext":{"slug":"/llm-abs-202311/","prev":{"excerpt":"Self Branding…","frontmatter":{"title":"Self Branding","tags":["thinking","book","review"],"date":"2023-11-25T10:00:00.000Z","draft":false,"excerpt":null,"image":{"childImageSharp":{"gatsbyImageData":{"layout":"fullWidth","placeholder":{"fallback":"data:image/jpeg;base64,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"},"images":{"fallback":{"src":"/static/a92d428eeb886278b11ba3d261014fd9/1b035/self-brand.jpg","srcSet":"/static/a92d428eeb886278b11ba3d261014fd9/7284f/self-brand.jpg 750w,\n/static/a92d428eeb886278b11ba3d261014fd9/1b035/self-brand.jpg 1000w","sizes":"100vw"},"sources":[{"srcSet":"/static/a92d428eeb886278b11ba3d261014fd9/57584/self-brand.webp 750w,\n/static/a92d428eeb886278b11ba3d261014fd9/b7ef5/self-brand.webp 1000w","type":"image/webp","sizes":"100vw"}]},"width":1,"height":0.667}}},"author":[{"id":"Soohwan Kim","bio":"Co-founder/A.I. engineer at TUNiB.","avatar":{"childImageSharp":{"gatsbyImageData":{"layout":"fullWidth","placeholder":{"fallback":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAOCAYAAAAvxDzwAAAACXBIWXMAABYlAAAWJQFJUiTwAAADjElEQVQ4y23Oy48TdQDA8TExEaTttvSBfdHddru7bWc6bWd+8+i0u23n0SdlBRQXZGHDgiiLDwgqXjCIJEaC8cLBmHgwUU9EYzRqQuLJi/4NJB6NF+8evgbOHr7XT76SM5ix7s4QnRHnnm9yZzvBTIswN8JcP5bC67ZoWi7CdtFsn5bloZo+dSNAMYbIIqCmD6mJIVXdR3J6I7r9IQ3T59pZlW+uxzndCbPdC/HZXoq3TikowsfqBOjOmFZ7SrM9p2kfRbXn1O0jKNbsSbJ5BMnu9Gl3+qiiz4OPSvz0QRLHFGwOZH6+neLHT5YRbRerM0J0JrTas/8FZXNKzZgg2es+ZtvDdbs8+iHJl+9lWVP7OG2br26k+PvXJCfmFqoYIpwhTXtMwxpTN0coxhjFmCKbE2rGmKoYPQY9qi2fyzsq//4e5sNXi5TlLqbd5t6lPH/9kuDTG2usqgGi7dO0hzSsIXUzQDaGKGKMLEZUxZCKCJB02+PUpMH9vQwP7ybYnSxyfNrjwumA3dEiD94v8cfnK7xzWsbpujQtH9X0qBs+iuEjC5+a7lPVPSqai9SyelyYFbjkh5hqC3TkKOutLON2nkYpgtuMcT6IszdPslo8SLlcpmUH1A0XRQyQhUtNuFQ1l4o2QFI1h3OTVXbcg+grIfSVBRqlEGopgrYSRi0e4KiIse1mScSeopA/iLADmoZL3Rgg631q+oCq1qeq95FOzhQefpzn7eMpBmqcoBVj04pzwkky0eNMRYIdL8F3twusLe2nslrE6g5pGX1U8/FhD1nvUW1tUJK7SH9+n+DR1zF2vQSvH0lz88wiV45V2J1WuHqsxLXNAnuzNJdfWmK5/BylUgFhb6DZAxR9g7rpIos+Wttja2uE9M9vYb69k+DqfJErWzLOuoVqdKlpDo4jODms4zlFckuHObx4iHwuRi6fYqmYY201x/LqMrnFDNl8gRvvukhNdT9394rcOlNFMxrUhYFmW9imTEsp0mhWKC2nWSokyWSiFA7HSKcXiEb34fayyLVniUQk8rmn2drKIqlqiNcuVrh2ymQS1Nh8sc/GoMEbl7u8eb5D1yqTzkRJJsOkkiGymQVShyLE4yHWKnEq1RjpTIhmI8r2y3mkyloI38uha1n6bpVXrkw5+4LD/S/2uHnrIopaJhZ9hmQqwqHkAXJP8BDhyD46ZoWdMxOGcx/PNfECk/8AbxXdRjRliPoAAAAASUVORK5CYII="},"images":{"fallback":{"src":"/static/a9e6b445142b247ee4cfa66155398bb2/7cf1f/soohwan.png","srcSet":"/static/a9e6b445142b247ee4cfa66155398bb2/34f77/soohwan.png 750w,\n/static/a9e6b445142b247ee4cfa66155398bb2/a94f6/soohwan.png 1080w,\n/static/a9e6b445142b247ee4cfa66155398bb2/7cf1f/soohwan.png 1148w","sizes":"100vw"},"sources":[{"srcSet":"/static/a9e6b445142b247ee4cfa66155398bb2/38420/soohwan.webp 750w,\n/static/a9e6b445142b247ee4cfa66155398bb2/7470d/soohwan.webp 1080w,\n/static/a9e6b445142b247ee4cfa66155398bb2/b5ef6/soohwan.webp 1148w","type":"image/webp","sizes":"100vw"}]},"width":1,"height":0.6829268292682927}}}}]},"fields":{"readingTime":{"text":"12 min read"},"layout":"","slug":"/self-branding/"}},"next":{"excerpt":"Welcome to the Gemini Era Google이 드디어 OpenAI에 대한 반격을 시작한걸까요? 구글이 ‘Gemini’라는 이름의 초거대 모델을 갑작스럽게 공개했습니다. MMLU 벤치마크에서 GPT-4를 넘으며, 32개 중 3…","frontmatter":{"title":"Welcome to the Gemini Era","tags":["news"],"date":"2023-12-07T10:00:00.000Z","draft":false,"excerpt":null,"image":{"childImageSharp":{"gatsbyImageData":{"layout":"fullWidth","placeholder":{"fallback":"data:image/png;base64,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"},"images":{"fallback":{"src":"/static/a3a835addaf3b2535fb6d7c81f1dc68c/6e89c/gemini.png","srcSet":"/static/a3a835addaf3b2535fb6d7c81f1dc68c/6e89c/gemini.png 744w","sizes":"100vw"},"sources":[{"srcSet":"/static/a3a835addaf3b2535fb6d7c81f1dc68c/02bc4/gemini.webp 744w","type":"image/webp","sizes":"100vw"}]},"width":1,"height":0.7741935483870968}}},"author":[{"id":"Soohwan Kim","bio":"Co-founder/A.I. engineer at TUNiB.","avatar":{"childImageSharp":{"gatsbyImageData":{"layout":"fullWidth","placeholder":{"fallback":"data:image/png;base64,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"},"images":{"fallback":{"src":"/static/a9e6b445142b247ee4cfa66155398bb2/7cf1f/soohwan.png","srcSet":"/static/a9e6b445142b247ee4cfa66155398bb2/34f77/soohwan.png 750w,\n/static/a9e6b445142b247ee4cfa66155398bb2/a94f6/soohwan.png 1080w,\n/static/a9e6b445142b247ee4cfa66155398bb2/7cf1f/soohwan.png 1148w","sizes":"100vw"},"sources":[{"srcSet":"/static/a9e6b445142b247ee4cfa66155398bb2/38420/soohwan.webp 750w,\n/static/a9e6b445142b247ee4cfa66155398bb2/7470d/soohwan.webp 1080w,\n/static/a9e6b445142b247ee4cfa66155398bb2/b5ef6/soohwan.webp 1148w","type":"image/webp","sizes":"100vw"}]},"width":1,"height":0.6829268292682927}}}}]},"fields":{"readingTime":{"text":"3 min read"},"layout":"","slug":"/gemini/"}},"primaryTag":"nlp"}},
    "staticQueryHashes": ["3170763342","3229353822"]}