공대생 도전 일지

https://arxiv.org/abs/2303.17651 Self-Refine: Iterative Refinement with Self-FeedbackLike humans, large language models (LLMs) do not always generate the best output on their first try. Motivated by how humans refine their written text, we introduce Self-Refine, an approach for improving initial outputs from LLMs through iterative feedbackarxiv.org 아이디어는 굉장히 단순하다. LLM을 통해 생산된 글을 다시 동일한 LLM에 넣어 피..

https://arxiv.org/abs/2405.06682 Self-Reflection in LLM Agents: Effects on Problem-Solving PerformanceIn this study, we investigated the effects of self-reflection in large language models (LLMs) on problem-solving performance. We instructed nine popular LLMs to answer a series of multiple-choice questions to provide a performance baseline. For each incorrarxiv.org CoT는 LLM의 성능을 크게 올리지만 논리, 수학적,..

https://aclanthology.org/2025.coling-main.504/ DORA: Dynamic Optimization Prompt for Continuous Reflection of LLM-based AgentKun Li, Tingzhang Zhao, Wei Zhou, Songlin Hu. Proceedings of the 31st International Conference on Computational Linguistics. 2025.aclanthology.org 기존 Reflcetion은 성능을 올리긴 했지만 iteration이 증가할 수록 성능 향상이 더뎌졌다.위 그래프에서 보듯 Early Stop Reflection문제가 발생하였고, DORA(Dynamic and Optimized..

https://arxiv.org/abs/2310.06271 Towards Mitigating Hallucination in Large Language Models via Self-ReflectionLarge language models (LLMs) have shown promise for generative and knowledge-intensive tasks including question-answering (QA) tasks. However, the practical deployment still faces challenges, notably the issue of "hallucination", where models generate plauarxiv.org 그럴듯하게 들리지만 사실이 아니거나 터무..

https://arxiv.org/abs/2405.19648 Detecting Hallucinations in Large Language Model Generation: A Token Probability ApproachConcerns regarding the propensity of Large Language Models (LLMs) to produce inaccurate outputs, also known as hallucinations, have escalated. Detecting them is vital for ensuring the reliability of applications relying on LLM-generated content. Current mearxiv.org Hallucinat..

https://arxiv.org/abs/2412.11919 RetroLLM: Empowering Large Language Models to Retrieve Fine-grained Evidence within GenerationLarge language models (LLMs) exhibit remarkable generative capabilities but often suffer from hallucinations. Retrieval-augmented generation (RAG) offers an effective solution by incorporating external knowledge, but existing methods still face several limarxiv.org  LLM은..

https://arxiv.org/abs/2501.02772 GeAR: Generation Augmented RetrievalDocument retrieval techniques form the foundation for the development of large-scale information systems. The prevailing methodology is to construct a bi-encoder and compute the semantic similarity. However, such scalar similarity is difficult to reflect earxiv.org 현재 존재하는 Bi-Encoder를 통한 유사도 계산은 정보를 충분히 반영하기 어렵고, 이해하기도 어렵다. 또한 ..

https://arxiv.org/abs/2307.06945 In-context Autoencoder for Context Compression in a Large Language ModelWe propose the In-context Autoencoder (ICAE), leveraging the power of a large language model (LLM) to compress a long context into short compact memory slots that can be directly conditioned on by the LLM for various purposes. ICAE is first pretrained usinarxiv.org긴 컨텍스트를 이겨내기 위해 다양한 접근 방법이 있..

2025.02.25 - [인공지능/논문 리뷰 or 진행] - GRIT 생성과 Embedding을 동시에 Generative Representational Instruction Tuning - 논문 리뷰일단 시작은 이 논문이겠습니다.생성과 Embedding을 동시에 하는 모델이 있으면 좋겠다 싶었는데 2025 ICLR에 올라간 것 같네요 ㅎㅎ...그래서 이 논문을 인용한 다른 논문들을 확인해보면서 제가 무엇을 할 수 있을지, 좀 더 다른 점을 어떻게 만들 수 있을지 확인해보겠습니다. https://arxiv.org/abs/2403.20327 Gecko: Versatile Text Embeddings Distilled from Large Language ModelsWe present Gecko, a comp..

https://arxiv.org/abs/2402.09906 기존 모델들은 생성만 잘하거나, Embedding만 잘 진행하였습니다.그리하여 둘 다 잘 하도록 두개 다 학습을 진행한 GRIT이 등장합니다.생성과 Representation 모두 진행하여 학습하는 것을 볼 수 있다. 임베딩에서는 양방향 Attention을 사용하고, 생성에서는 단방향 Attention을 진행하는 것을 볼 수 있습니다.진짜 단순한 아이디어였고, Loss도 어려운 수식이 아닙니다.Closed Model인 OpenAI를 이기는 모습을 볼 수 있습니다.Embedding 성능에서 높은 성과를 보이는 것을 볼 수 있습니다. 여기서 8X7B의 성능이 낮은 이유는 배치가 작아졌다는 이유라고 말합니다. 여기선 생성형 능력을 볼 수 있습니다.생성..