Research focus
Information-theoretic foundations
We focus on the theoretical understanding of diverse deep learning problems, including generative models, privacy and safety, and network compression.
Yonsei University · Department of Artificial Intelligence
AI • ISL
Artificial Intelligence and Information Systems Laboratory (AI-ISL) is a deep learning research lab focused on information-theoretic approaches to machine learning problems.
Join Us
Become part of AI-ISL
Join our lab to work on foundational research in discrete diffusion language models, machine unlearning, safety, and LLM quantization.
Core themes
Current priorities
- Advancing the theoretical understanding of discrete diffusion language models
- Developing machine unlearning methods for large language models
- Building safety alignment and agent-level controls for LLM behavior
- Applying LLM quantization to make large models efficient and deployable
News
Lab updates
- CVPR 2026 Findings
Our group has 1 paper accepted:
Memorization In Stable Diffusion Is Unexpectedly Driven by CLIP Embeddings. - ICLR 2026
Our group has 3 papers accepted:
A2D: Any-Order, Any-Step Safety Alignment for Diffusion Language ModelsRainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMsRethinking Benign Relearning: Syntax as the Hidden Driver of Unlearning Failures. - SaTML 2026
Our group has 1 paper accepted:
Differentially Private Adaptation of Diffusion Models via Noisy Aggregated Embeddings. - AAAI 2026
Our group has 1 paper accepted:
An Information Theoretic Evaluation Metric For Strong Unlearning. - NeurIPS 2025
Our group has 2 papers accepted:
SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early AlignmentInformation-Theoretic Discrete Diffusion. - EMNLP 2025 Main
Our group has 2 papers accepted:
R-TOFU: Unlearning in Large Reasoning ModelsSEPS: A Separability Measure for Robust Unlearning in LLMs. - ICML 2025 Workshop on Tiny Titans
Our group has 1 oral presentation:
Preserve then Quantize: Dominant-Subspace Guided Low-Rank Reconstruction. - ACL 2025 Findings
Our group has 1 paper accepted:
Assigning Distinct Roles to Quantized and Low-Rank Matrices Toward Optimal Weight Decomposition. - ICML 2025 Spotlight
Our group has 1 paper accepted:
Understanding and Mitigating Memorization in Generative Models via Sharpness of Probability Landscape. - AAAI 2025 Workshop on Privacy-Preserving Artificial Intelligence
Our group has 1 paper accepted:
Understanding Memorization In Generative Models Through A Geometric Framework. - NeurIPS 2024 Workshop on Statistical Foundations of LLMs
Our group has 1 paper accepted:
Adversarial Sample-Based Approach for Tighter Privacy Auditing in Final Model-Only Scenarios. - ICML 2024
Our group has 1 paper accepted:
Improved Communication-Privacy Trade-offs in L2 Mean Estimation under Streaming Differential Privacy. - Lab news
ISL moved to the Department of Artificial Intelligence at Yonsei University. - NeurIPS 2023 Workshop on Diffusion Models
Our group has 1 paper accepted:
LoRA can Replace Time and Class Embeddings in Diffusion Probabilistic Models. - NeurIPS 2023
Our group has 2 papers accepted:
Censored Sampling of Diffusion Models Using 3 Minutes of Human FeedbackExact Optimality of Communication-Privacy-Utility Tradeoffs in Distributed Mean Estimation. - ICCV 2023 Workshop on Low-Bit Quantized Neural Networks
Our group has 1 oral presentation: Fully Quantized Always-on Face Detector Considering Mobile Image Sensors. - ICML 2023 Workshop on Federated Learning
Our group has 1 paper accepted:
Exact Optimality of Communication-Privacy-Utility Tradeoffs in Distributed Mean Estimation. - ECCV 2022
Our group has 1 paper accepted:
Prune Your Model Before Distill It. - ICML 2022
Our group has 1 paper accepted:
Neural Tangent Kernel Analysis of Deep Narrow Neural Networks. - AISTATS 2022
Our group has 1 paper accepted:
An Information-Theoretic Justification for Model Pruning. - ICML 2021
Our group has 1 paper accepted:
WGAN with an Infinitely Wide Generator Has No Spurious Stationary Points.