Integrated M.S./Ph.D. Student · Korea University
I work on language models at Korea University's NLP & AI Lab. My research covers private inference, multilingual retrieval, and agent evaluation. I want to make model reasoning more efficient and extend this work to multimodal models.
Open to research internships and collaborations.
AgentHop accepted at NeurIPS 2026, Evaluations and Datasets Track (Poster).
* Equal contribution.
A benchmark of 1,011 scientific questions that tests how agents find and combine evidence across papers. It diagnoses retrieval, synthesis, tool use, and resource management.
My role: Contributed ideas, assisted with experiments, and helped write the paper.
Accepted to NeurIPS 2026, Evaluations and Datasets Track (Poster).
PPFT lets a server-side LLM answer from noisy prompt embeddings without receiving the raw text. We study the trade-off between answer quality and resistance to text reconstruction.
My role: Led the project from the initial idea through experiments, analysis, and writing.
ACL 2026, Volume 1: Long Papers, pp. 25961-25981. Poster presentation. DOI: 10.18653/v1/2026.acl-long.1191.
Uses a sparse autoencoder to remove language-specific features from text embeddings, improving multilingual retrieval without retraining the encoder.
My role: Contributed to experiments, analysis, and writing.
ACL 2026, Volume 1: Long Papers, pp. 36374-36389. Oral presentation. DOI: 10.18653/v1/2026.acl-long.1685.
Adds uncertainty estimates to pretrained models through a lightweight transformation of their output logits, evaluated on image classification and LLM question answering.
My role: Contributed to initial experiments and writing.
CVPR 2026, pp. 6157-6166. Highlight paper; poster presentation.
Tests whether 18 LLMs detect ethically problematic language when it appears in code comments, identifiers, and other parts of Python and C programs.
My role: Contributed to experiments, analysis, and writing.
Mathematics 14(1), Article 189. DOI: 10.3390/math14010189.
Studies how many upper layers can be removed from an LLM used as a text embedder while retaining retrieval quality.
My role: Designed and carried out the study from start to finish.
37th Annual Conference on Human & Cognitive Language Technology (HCLT 2025), pp. 14-18.
A Korean-English RAG model trained to decide when to retrieve, filter irrelevant evidence, and check its answers.
My role: Contributed to experiments and analysis.
37th Annual Conference on Human & Cognitive Language Technology (HCLT 2025), pp. 135-140.
How can a model reach a good answer with less computation? I am interested in reducing unnecessary reasoning steps and context processing while preserving answer quality. I also want to bring these ideas to models that work with both text and images.
My work so far includes learning useful text representations, studying what they reveal about private inputs, and evaluating how models retrieve evidence and use tools.
Graduate researcher at the NLP & AI Lab, advised by Prof. Heuiseok Lim.
Double major in Software Engineering, with project work spanning simulation, autonomous driving, reinforcement learning, and machine learning.
Implemented and debugged finite-element simulation components for beam and shell structures using C++ and MATLAB, including 6-DOF representations and coordinate transformations.
Built perception, planning, and control components using LiDAR, cameras, ROS, MATLAB/Simulink, and Gazebo.
Text embeddings, sparse autoencoders, dense retrieval, RAG, agent evaluation, uncertainty estimation, and privacy evaluation.
Python, PyTorch, Hugging Face Transformers, TensorFlow, LoRA fine-tuning, model evaluation pipelines, data processing, and experiment automation.
C/C++, MATLAB/Simulink, ROS, Gazebo, finite-element simulation, image-processing pipelines, autonomous-driving control, and reinforcement learning.
I am looking for research internships and collaborators working on language models, efficient reasoning, or multimodal learning.
aa007878@korea.ac.kr