John Seon Keun Yi

I am a Computer Science Ph.D student at Boston University. I am fortunate to be advised by Prof. Dokyun "DK" Lee.

I received Masters and Bachelors degrees in Computer Science from Georgia Institute of Technology, where I worked on data-efficient learning and methodologies for human-robot interaction.

Email: jskyi {at} bu {dot} edu

CV  /  Google Scholar  /  Github

My research interest mainly focuses on the reliability and interpretability of foundation models. I am interested in using causal reasoning and mechanistic interpretability to promote the safe, auditable deployment of language models in business and society.

Fine-tuning framework to internalize multi-agent debate inside a single language model. Additionally, we identify and steer agent subspaces inside the internalized model.

An evaluation benchmark for physical commonsense reasoning covering 100+ languages and cultures.

Continuously learning to drive from an infinite flow of YouTube driving videos.

A comprehensive survey on Masked Autoencoders(MAE) in vision and other fields.

We use simple self-supervised pretext tasks and a loss-based sampler to sample both representative and difficult data from an unlabeled pool.

We use semantic scene graphs to disambiguate referring expressions in an interactive object grounding scenario. This is effectively useful in scenes with multiple identical objects.

We use 2D inpainting methods to complete occlusions and imperfections in 3D building point cloud scans.

This page is adapted from this template which is created by Jon Barron.