I am a Ph.D. candidate in Industrial Engineering at Seoul National University, advised by Prof. Pilsung Kang in the Data Science and Business Analytics Lab. My research centers on how language models reason, adapt, and act in real-world environments — studying process-level supervision, reinforcement learning for LLMs, and agentic systems that operate under practical constraints.

What I’m Up To

• My Ph.D. research asks how language models can go beyond imitation — reasoning through multi-step problems, learning from process-level signals, and acting reliably in real-world settings.
• I gravitate toward projects that sit close to practice: benchmarks teams can actually maintain, retrieval systems for specialized corpora, and LLM pipelines that have to work under real-world constraints.
• I am currently building a stream-based benchmark for evaluating language model agents on sequential, real-world tasks — measuring how well they reason, use tools, and adapt across a stream of observations.
• I am looking to collaborate with researchers from diverse domains. If your field involves reasoning, retrieval, or decision-making problems that language models might help solve, I would be glad to explore potential joint work.
• I co-organize an NLP paper review study group called Unknown NLP. We publish our session write-ups at unknown-nlp.github.io.

Recent Updates

• May 2026: Updated my CV with new under-review manuscripts and archived publications.
• 2026: AlienLM was accepted to ICML 2026.
• 2026: New manuscripts on belief-shift rewards and semiconductor equipment log retrieval are under review.
• 2025: CheckEval appeared at EMNLP 2025 and Verbosity-Aware Rationale Reduction appeared at ACL 2025.
• 2024: ContAccum, a memory-efficient dense retriever training method, appeared at NeurIPS 2024.

Research Themes

Reasoning Signals & Process Supervision

I study process-level supervision and reinforcement learning for reasoning-oriented models — moving beyond outcome-based rewards toward belief-aware training signals that shape how models reason step by step. I am also interested in rationale reduction and decoding strategies that preserve reasoning quality while keeping inference efficient.

Semantic Embedding & Retrieval

I work on retrieval systems for domain-specific and document-heavy settings, including scientific metadata, industrial logs, and hierarchical business documents. I am especially interested in training embedders that remain useful when data is specialized, sparse, or operationally messy.

Trustworthy & Privacy-Preserving LLM Systems

I build evaluation and deployment methods for LLM systems used in sensitive domains. This includes checklist-based LLM-as-a-judge evaluation, API-boundary privacy via transformed representation spaces, and agent benchmarks where reliability and auditability matter as much as task performance.

Selected Publications

• AlienLM: Alienization of Language for API-Boundary Privacy in Black-Box LLMs. Jaehee Kim Α, Pilsung Kang Ω. ICML 2026.
• Belief-Shift Reward: Sculpting the Flat Reward Landscape via Telescoping Information Gain. Keonwoo Kim Α, Jaehee Kim, Dongyoon Han, Haanju Yoo Ω. Under review, 2026.
• CheckEval: A reliable LLM-as-a-Judge framework for evaluating text generation using checklists. Yukyung Lee Α, JoongHoon Kim, Jaehee Kim, Hyowon Cho, Jaewook Kang, Pilsung Kang Ω, Najoung Kim Ω. EMNLP 2025.
• Verbosity-Aware Rationale Reduction: Sentence-Level Rationale Reduction for Efficient and Effective Reasoning. Joonwon Jang Α, Jaehee Kim, Wonbin Kweon, Seonghyeon Lee, Hwanjo Yu Ω. ACL 2025.
• A Gradient Accumulation Method for Dense Retriever under Memory Constraint. Jaehee Kim Α, Yukyung Lee, Pilsung Kang Ω. NeurIPS 2024.

For the complete list, including workshops and domestic journal papers, see Publications.

Applied AI Projects

• Development of Worker-Friendly Innovative AI Agents for Autonomous Manufacturing (IITP, Apr 2025–Dec 2025): Built multi-agent orchestration, multi-party engine, and RAG pipeline for autonomous manufacturing AI agents assisting factory workers.
• Data Construction and Optimization for Training/Inference of AI Chatbot (Samsung Fire & Marine Insurance, Mar 2025–Mar 2026): Conducting PDF parsing, hierarchical document preprocessing, and chatbot training data construction for insurance-domain documents.
• Developing an Information Retrieval System for Scholarly Achievement and Research Projects in the College of Engineering (College of Engineering, SNU, Jan 2025–Jul 2025): Developed scientific-domain dense retrievers and evaluation pipelines based on research lab and achievement metadata.
• Developing a Large Language Model Evaluation Framework for Financial Domain (KakaoBank, Nov 2023–Aug 2024): Built safety, truthfulness, and mathematical reasoning benchmarks and evaluation frameworks for finance-domain LLMs.
• Developing Customer Content via Data Analysis Techniques 2nd Stage (LG Electronics, Dec 2022–Nov 2023): Performed unsupervised dense retrieval, review clustering, and text/meta information visualization using customer reviews.
• Developing Customer Content via Data Analysis Techniques 1st Stage (LG Electronics, Apr 2022–Nov 2022): Developed customer insight mining pipeline using sentence anomaly detection, pain point discovery, and clustering via self-training.

Collaboration

I enjoy working with researchers across disciplines — NLP, RL, systems, and applied domains — who want to push language models toward more reliable and autonomous behavior. If you see a problem where reasoning, retrieval, or agentic AI might help, feel free to reach out.

Recent Posts