Research
Our research follows one thread, Corruption-to-Reconstruction (C2R), across three pillars: making vision-language foundation models efficient, taking them from perception to action in the physical world, and keeping them trustworthy and secure. This agenda is supported by the NRF Outstanding Early-Career Researcher Program (2026-2029).
Efficient Learning of Vision-Language Foundation Models Core
Salience- and token-aware masking cuts pre-training time by up to 50% without losing representation quality. Corruption applied at the right place lets VLMs see fine-grained detail (ECC) and makes vision-language encoders robust (Aether). We also select the image-text pairs that actually matter, and compress large vision and CLIP models for on-device and AI-semiconductor settings.
1-A. Training: corruption-based pre-training
1-B. Data: filtering and selection
1-C. Deployment: efficient inference and adaptation
From Perception to Action: toward Embodied and Physical AI
An embodied agent needs two things: acting from pixels and perceiving the 3D world. We have worked on sample-efficient, environment-agnostic visual reinforcement learning, and on monocular depth and stereo matching that keeps adapting in changing real-world conditions. We are now combining these with our efficient VLMs to build vision-language-action (VLA) and video models for robots and autonomous driving (new direction, launched in 2026).
2-A. Acting from pixels: visual reinforcement learning
2-B. 3D perception for driving and robots
2-C. Vision-Language-Action models (ongoing)
Ongoing project, launched in 2026
Trustworthy and Secure AI
As models become more capable, we study how to keep them verifiable: detecting AI-generated images of people through social-gaze consistency (with Prof. James M. Rehg, UIUC), auditing deepfake detectors, analyzing attack surfaces of AI agents, and measuring when LLM verifiers fail together. This line is funded by the National Security Research Institute, the IITP AI-agent safety program with KAIST, and the ITRC AI Security Research Center.
3-A. Authenticity: deepfake and AI-generated image detection (Social AI)
3-B. Security of AI agents and LLMs
3-C. Robustness
AI for Science and Industry
AI-Semiconductor
[P12] Deeper Is Not Better for Quantized CLIP, with the COSS Next-Generation Semiconductor project (see Funding below).
Audio-visual music understanding
Two industry-academia R&D projects with Spartan SW (Phase I and Phase II).
Funding and Collaborators
Since Sep. 2025: 7 research grants as Principal Investigator and participation in 3 national programs.
As Principal Investigator
| Project | Funding agency | Period |
|---|---|---|
| NRF Outstanding Early-Career Researcher Program: “Enhancing Training Efficiency and Performance of Large-Scale AI Models via the Corruption-to-Reconstruction (C2R) Framework” | National Research Foundation of Korea (NRF) | Mar. 2026 - Feb. 2029 |
| Analysis and Roadmap of Agentic AI-based Autonomous Security Technologies for the National and Public Sector | National Security Research Institute (NSR) | Apr. 2026 - Oct. 2026 |
| Efficient Representation Learning and System Optimization with Large-Scale Models for AI-Semiconductor Convergence | COSS Next-Generation Semiconductor (MOE / NRF) | May 2026 - Dec. 2026 |
| Temporally Aligned Audio-Visual-Text Multimodal Foundation Model for Short-Form Music Content | Industry-Academia R&D (Spartan SW), Phase II | Fall 2026 |
| Audio-Visual Multimodal Deep Learning Models for Music Content Analysis | Industry-Academia R&D (Spartan SW), Phase I | Spring 2026 |
| Pre-training with a General-Purpose Decoder | Soongsil University Research Fund for New Faculty | Sep. 2025 - Aug. 2026 |
| Curriculum Development: four new AI courses for the RISE Program and the AI Bootcamp Program | Soongsil University (RISE / AI Bootcamp) | 2025 - 2026 |
As Participating Faculty in National Programs
| Project | Funding agency / Partners | Period |
|---|---|---|
| Autonomous Control and Proactive Defense Mechanisms for Enhancing the Reliability and Safety of AI Agents | IITP (MSIT). With KAIST (lead), Seoul National University, Coontec | Apr. 2026 - Dec. 2029 |
| AI Security Research Center (ITRC), Director: Prof. Souhwan Jung | IITP-ITRC (MSIT) | 2025 - Present |
| AI-Centered University Program (“AI Omniversity”): co-author of the winning proposal and participating faculty | MSIT / IITP | 2026 - 2033 |