Hi, my name is SuHyun Bae. I’m Master student in Mathematical Data Science at Korea University. Also a researcher of AIML@K.

My research interests are mainly in NLP and utilizing Language Models.

I have been interested in creating friendly AI since I was young. So I decided to study mathematics to understand the mathematical principles underlying AI. My goal is to analyze and design models rigorously.

Research Experience

• 2023.07 - 2023.12 | AI Grand Challenge 2023
  Generated and refined a dataset using GPT 3.5 via openai-API to fine-tune a model for answering multi-answer questions. Our lab won 7th place

• 2024.02 - 03 | Practicing RAG
  Conducted research on Retrieval-Augmented Generation (RAG) technology and applied it to implement a model for the DACON competition, “한솔데코 도배하자 Q&A”

• 2024.03 - 06 | KCC 2024 and Hallucination
  Classified hallucination types in QA task into 5 categories and identified that the natural language evaluation metrics BLEU, METEOR, and ROUGE each excel at detecting certain types of hallucinations while struggling with others. This was published at KCC 2024

• 2024.07 - 2025.01 | Internship at SK Magic
  Product planning and development at SK magic, focusing on algorithm development for new product and researching on-device LLM ecosystems and data infrastructure for product integration

• 2025.03 - PRESENT | Grokking and Hallucination
  Interested in the phenomenon of grokking and researching its potential role in reducing hallucinations

Research Focus

My research focus is on understanding the phenomenon of grokking in language models and progressively expanding its application. First, grokking has primarily been observed in domains such as modular arithmetic and group operations. I aim to extend this to logical structures that use natural language, such as syllogisms, to determine whether grokking can occur in these contexts. Second, current studies on grokking have been largely limited to simple transformer-based models. I plan to investigate whether grokking can also be observed in small-scale language models (sLLMs) such as LLaMA 3.2 1B and DeepSeek 1.78B.

Research Goal

My ultimate research goal is to verify that Language Models can generalize certain computations and logical structures through the phenomenon of grokking. Furthermore, I aim to demonstrate that this understanding can significantly reduce hallucination when Language Models generate answers using reasoning chains such as Chain of Thought (CoT) or induction.