andrew kim / research

papers

semantic multiplicity creates strategic flexibility in streaming ai agents 2026

co-author, one of seven, marked equal contribution. economics for machine learning workshop, neurips 2026. pdf.

when an ai agent streams an action and one action can be written more than one way, it can wait for the other side's reaction before it commits. we work out when that's possible and test it on real models.

same text, different transitions: action insufficiency in lean provers 2026

co-author, one of four. mathnlp 2026. pdf.

lean proof states that print the same can still react differently to the same tactic. we found 103 verified cases in mathlib, and a fix for proof search that costs 8% more runtime.

honors

usaco platinum 2024-25
the top division of the usa computing olympiad

national economics challenge, national runner-up 2026
team captain, david ricardo division. we placed second in the country

georgia economics challenge, state champion 2026

john locke global essay competition, high commendation 2026
economics category. the cost of having no exit, and the certificate

international invitational math optimization challenge, top ten 2024-25

cash app teen advisory council 2024
selected for facemood

roles

student researcher, mantis ai at mit csail 2026
ml workflows for high dimensional and time series data, with embeddings, evaluation, and visual tools for looking at datasets

ex roblox developer 2025-26
one of several developers on archived, a game that had made over five figures by the time i left

captain, lambert economics team 2026

legislative fellow, georgia general assembly 2026
worked out of representative au's office on health research and events, drafted newsletters, and sat in on town halls and constituent meetings

director of events, junior economics club 2026
ran tours of synovus and invesco, an open house, and an equifax case study event