Seungju Han
Visiting @ Allen Institute for AII am currently a visiting student researcher at Allen Institute for AI (AI2), advised by Yejin Choi. At AI2, I have also worked with Nouha Dziri, Youngjae Yu, and Jack Hessel. Previously, I received my B.S. in Electrical and Computer Engineering from Seoul National University.
I am interested in developing simple, intuitive, and scalable learning algorithms to enhance the capabilities of large language models. In recent years, I have been investigating the limitations of current LLMs, including:
- Safety of LLMs: I developed tools to detect the harms and refusals in user-LLM interactions (WildGuard) and a novel red-teaming framework to identify and mitigate the risks of LLMs without sacrificing other capabilities (WildTeaming).
- Visual-grounding of LLMs: I studied scalable ways to build data for training and evaluating vision-language models, for example, employing internet-scale (20 million) videos to teach vision-grounded dialogue (CHAMPAGNE) and creating challenging vision-language benchmarks that current models struggle to solve, such as understanding social norms, humor, and visual arguments.
I also worked on research and engineering at a startup (Hyperconnect; acquired by Match Group for $1.7B), focusing on problems related to open-domain conversation (diversity of responses, in-context learning, retrieval augmentation, distillation) and text-to-speech (few-shot learning) to build social chatbot products, aiming to solve people’s loneliness. Additionally, I tackled product-related ML problems like long-tail classifications.
Email: wade3han at snu.ac.kr