Reliable AI from foundations to deployment.

Foundation Models · Privacy & Unlearning

Preventing Degradation in Sequential Machine Unlearning

연속적인 정보 삭제가 누적될 때 발생하는 모델 성능 저하를 분석하고 완화합니다.

Research Question How can models keep forgetting without progressively losing their capabilities?

Reliable Computing

Accelerating Variational Quantum Algorithms with Quantum Natural Gradient

QNG의 측정·피드백 비용을 줄여 변분 양자 알고리즘의 최적화를 가속합니다.

Research Question How can we reduce measurement and feedback latency in quantum natural gradient optimization?

Quantum Computing · Efficient Optimization

Foundation Models

Hallucination Detection in Multi-Agent Systems

여러 AI 에이전트가 상호작용할 때 발생하고 전파되는 hallucination을 탐지합니다.

Research Question How do hallucinations emerge and propagate among interacting AI agents?

Multi-Agent Systems · Reliability

Physical World · Robustness

Detecting Shortcut Bias in World Models for Autonomous Driving

자율주행 월드 모델이 실제 세계의 구조가 아닌 데이터의 shortcut에 의존하는 현상을 탐지합니다.

Research Question Are world models learning the physical world — or shortcuts in the data?

Foundation Models · Robustness

Mitigating Spurious Correlations in Protein Foundation Models

단백질 모델이 사용하는 비생물학적 spurious signal과 shortcut을 탐지하고 완화합니다.

Research Question Do protein models rely on non-biological shortcuts, and how can we identify and mitigate them?

AI for Science

Foundation Models

Efficient Pruning for Diffusion Language Models

Diffusion language model의 구조적 특성을 활용해 성능을 유지하면서 모델을 효율적으로 경량화합니다.

Research Question How should diffusion language models be pruned while preserving generation quality?

Diffusion Language Models · Model Efficiency

Foundation Models · Alignment & Explainability

Mechanistic Interpretability for High-Quality LLM Data Generation

Large Language Model의 학습 데이터를 생성할 때 mechanistic interpretability를 활용해 데이터의 품질을 평가하고 향상하는 방법을 연구합니다.

Research Question Can mechanistic interpretability guide the generation of higher-quality data for large language models?

Mechanistic Interpretability · Data Generation

Foundation Models · Robustness

Visual Hallucination Reduction in Vision-Language Models

Vision-language model이 생성한 각 주장이 충분하고 일관된 시각적 근거로 뒷받침되는지 검증해 visual hallucination을 줄입니다.

Research Question How can vision-language models verify visual evidence before accepting generated claims?

Vision-Language Models · Visual Hallucination

Reliable Computing

Data-Informed Quantum Kernels for Time-Series Forecasting

시계열 데이터의 구조적 특성을 inductive bias로 반영한 양자 커널을 설계합니다.

Research Question Can data-informed inductive biases make quantum kernels better suited for structured time-series forecasting?

Quantum Machine Learning · Time-Series Forecasting · Inductive Bias

Reliable Computing

Verifiable Distributed Computing Against Colluding Malicious Workers

분산 계산 환경에서 일부 참여자들이 서로 공모해 잘못된 결과를 내더라도 계산 결과를 검증할 수 있는 방법을 연구합니다.

Research Question How can distributed computation remain verifiable when malicious workers collude?

Verifiable Computing · Distributed Computing · Secure AI Computation