Research & Projects
AI Safety, Formal Methods, and Engineering
SolEvolve: An LLM-driven Evolutionary Discovery of Algorithms
An autonomous discovery system where the LLM acts as an active researcher. Rediscovered the "Shortened Golay Code" ($[22,11,6]$) autonomously using SAT-seeded Genetic Algorithms and Black-Box Optimization.
- ✓Autonomous Discovery: Closed-loop system with Generator, Evolver, and Verifier agents.
- ✓Formal Verification: Rigorous mathematical checks with UNSAT certificates.
- ✓Performance: Outperformed algebraic software in discovering Optimal Binary Linear Codes.
Alignment Faking in LLMs: A Case Study
Mathematical formalization of the "Santa Claus" problem: where an LLM's external behavior (compliance) diverges from its internal state due to awareness of monitoring.
Recently Accepted Papers
Application of TimeGPT for Enhancing Water Level Prediction in Gamcheon River, Korea
Investigating the efficacy of Foundation Models (TimeGPT) in hydrological time-series forecasting compared to traditional statistical methods.
Hybrid Multimodal GenAI for Solving Math Problems Containing Various Figures
A novel approach combining visual encoders and language models to solve complex geometry and algebra problems involving diagrams.
Journal Papers
Performance Improvement of LLMs for Regulatory Document Understanding based on Modified RAG Approach
Jae-Hyun Baek, Jon-Lark Kim | JKIIS 2025 (Published)
MekaNet: WSI-based Tiny Object Detection
Jae-Hyun Baek (co-author) | Medical Image Analysis (Under Review)
Symmetric Sudoku-Type Games from Perfect Codes
Jae-Hyun Baek (co-author) | IEEE Transactions on Games (Submitted)
Conference Presentations
Sudoku-type Puzzles from Coding Theory (Invited Talk)
11th Sino-Korea International Conference on Coding Theory | July 2025
Alignment Faking in LLMs: A Case Study (Oral)
Korean Institute of Intelligent Systems | May 2025
Modified RAG Framework for Regulatory Documents
KSIAM Conference | Apr 2025
EntropyMath
Agentic Tool Use Evaluation Leaderboard.
Built pipelines to measure "honest" reasoning capabilities of LLM Agents to solve benchmark contamination.
SOGAMBOT.com
AI Chatbot for Sogang University (Team Leader).
Led the transformation of university-wide data into AI-ready formats. Managed full-stack development and RAG implementation.
Impact: Digital transformation of university administrative data.
GPT-OSS-20B Persona Injection
HuggingFace Community Project.
Created and optimized fine-tuning datasets for persona injection, achieving 100+ downloads/week.
River-GNN Flood Forecasting
Industry-Academia Collaboration (KICT).
Developing AI models for water level prediction and flood safety assessment using GNNs and TimeGPT.
"Creating production-level media content beyond Slop using Generative AI."
AI Commercial Festival
Planned and executed commercial AI content strategies. Built workflows for high-quality media generation.
Media Production Pipeline
Implemented technical pipelines for integrating various GenAI tools into creative workflows.
Graduate Seminar Lecturer
Sogang University | "LLM Trends & Albatross Seminar" for undergraduates & graduates.
High School AI Outreach
Sogang Math Dept x Bokja Girls' High School
- "Viewing AI through 0 and 1" (2024)
- "Textbook is all you need" (2025)
Teaching Assistant (MATLAB)
Sogang University Department of Mathematics | 1.5 Years
AI Technology Journalism
The Sogang Herald & IMDS Newsletter
Analyzed global AI trends and communicated complex technologies to the public. Bridging the gap between technical research and broader impact.
AI Expo 2024
Analyzed the latest trends in LLM applications and Generative AI solutions across industries.
World IT Show 2024
Investigated on-device AI innovations and the integration of AI in consumer electronics.
Google Cloud Summit 2024
Reported on Google's enterprise AI strategies and cloud infrastructure advancements.
Why AIM Intelligence?
Transforming "Formal Methods" into practical "AI Safety" guarantees.
Safety via Formalism
I don't just find vulnerabilities; I structuralize them. Applying mathematical rigor to explain why alignment failures occur (e.g., Alignment Faking).
Global Communication
Experience in presenting at global venues (ISIS 2025). Ready to collaborate with global AI safety researchers and communicate complex ideas effectively in English.
8-Week Impact
Proposed Goal: Research methodology to induce Alignment Faking via Red Teaming. Delivering not just reports, but reproducible evaluation pipelines.