With advancements in LLMs, users increasingly rely on AI to understand complex information, yet face unique challenges based on their prior knowledge. In this talk, I will present two projects designed to address these challenges by aligning AI support with each user's knowledge state, fostering deeper understanding and engagement with content.

거대 언어 모델을 이용하는 코드 자동 완성, 결함 수정 등 코드에 적용하는 다양한 연구가 발전하고 있지만, 언어 모델을 활용하는 방법은 프롬프트 엔지니어링이나 추가 학습(Fine-tuning) 방법 등으로 제한적입니다. 이 발표에서는 정적 분석 데이터와 강화 학습을 이용해서 언어 모델이 더 안전한 코드를 생성하도록 학습하는 아이디어를 제시하고, 실험적인 결과를 공유드리겠습니다.

From vibrant eating broadcasts (Mukbang) to recipe videos on YouTube, multimedia food content brings enjoyment and a sense of connection to our daily lives. However, for individuals with eating disorders, such food-related media can backfire: tempting food videos may inadvertently trigger disordered behaviors like binge eating. In this talk, I will discuss these challenges and introduce FoodCensor, an intervention designed to promote mindful engagement with digital food content and support a healthier online experience.

"Data-Efficient Unsupervised Interpolation Without Any Intermediate Frame for 4D Medical Images" 논문(CVPR'24 Accepted)의 main method 소개 및 저널 extension을 위한 motivation 및 additional contribution에 대해 설명하겠습니다.

This talk introduces a proactive self-tracking system designed for mental health research, utilizing a multimodal smart speaker to deliver surveys based on user context transitions. By adapting survey timing and modality to the user's immediate environment, the system aims to increase engagement and compliance, enhancing data quality. A four-week field study revealed user preferences for touch interactions over voice commands, with interaction choice influenced by ongoing activities. Key findings offer valuable design insights for creating effective, context-aware mental health tools for home environments.

NeurlPS'24 Dataset and Benchmark에서 Spotlight 논문으로 선정된 데이터 기반의 LLM Hallucination 평가기법을 소개하고, 데이터베이스와 LLM 같은 이종 분야 결합의 시너지 효과에 대하 직접적인 경험을 공유할 예정입니다.

State-of-the-art multilingual LLMs have shown their vulnerabilities in non-English, code-switching attacks, revealing a spurious correlation between language resources and safety alignment. This talk briefly introduces code-switching curriculum learning, an effective and efficient training strategy for the language transfer of LLMs. The code-switching curriculum learning advances the cross-lingual alignment of LLMs, which enables mitigation of the aforementioned spurious correlation.

학부에서 배움에 정진하던 시기를 지나 대학원에서 창조적 연구 활동을 하고 계신 학생분들께 제가 같은 과정을 거쳐오며 생각한 바를 공유해 드리고, 연구 주제를 결정할 때 고려할 사항들과 긴 연구의 여정을 지탱해 줄 학문의 즐거움에 대해 함께 생각해 보는 시간을 가졌으면 합니다.