Dongho Kim
I'm an M.S. student in Statistics & Data Science at Northwestern University, advised by Prof. Han Liu in the MAGICS Lab. I also collaborate with the NSF–Simons SkAI Institute and am applying to PhD programs for Fall 2027.
My research focuses on reliable AI: making the reliability of AI systems verifiable and measurable rather than assumed. As AI becomes both an instrument for scientific and social inquiry and an actor in social and economic systems, I pursue three complementary directions: agentic systems with built-in verification (SPINE, co-first author), evaluation of foundation models on challenging scientific data (StarEmbed, ICML 2026), and verified agentic workflows for computational social science (Sci2Pol-Agent, ongoing).
Research
Verification-first language-model agents
Agentic systems that ground outputs in evidence, use specialized verification roles, and require observable validation before completion. (SPINE; Sci2Pol-Agent)
Statistical evaluation under uncertainty
Evaluation designs that separate workflow effects from model effects and quantify uncertainty in AI-generated outputs. (Sci2Pol-Agent)
Representation learning for irregular scientific data
Studying when pretrained representations transfer under irregular sampling, heteroskedastic noise, and distribution shift. (StarEmbed, ICML 2026)
News
- Presented StarEmbed at ICML 2026 and received the Outstanding Paper Award after an oral presentation at the AI4Physics Workshop.
- Joined the NSF–Simons SkAI Institute collaboration on foundation models for astronomical time series.
- Started my M.S. at Northwestern University; joined MAGICS Lab (Prof. Han Liu).
Publications & Preprints
SPINE: Bridging the Cyber-Physical Gap with Agentic AI
M. Ham*, D. Kim*, C. Lee*, J. Wang, J. Zhao, G. Ye, S. Park, M. J. Kim, Y. Zhang, H. Liu (*equal contribution)
Preprint2026
StarEmbed: Benchmarking Time Series Foundation Models on Astronomical Observations of Variable Stars
W. Li*, H.-Y. Chen*, N. Rehemtulla*, V. Shah, D. Kim, D. Wu, Q. Lin, A. Miller, H. Liu (*equal contribution)
Ongoing
Sci2Pol-Agent: Verification-First Agentic Harness for Science-to-Policy Writing
A training-free, multi-role agent workflow that routes drafts through evidence extraction, support checking, targeted revision, and format control. A same-backbone evaluation isolates workflow effects and finds gains concentrated in claim verification. Current work treats verifier judgments as noisy measurements of output accuracy.
Ongoing · 2026