My research looks at language models from three sides. From the inside: what a model has learned, how that knowledge is stored and retrieved, and what stops it from being used. From the outside: how faithfully models can stand in for people, as survey respondents, personas, and participants in public deliberation. And in the field: how models can measure places where ground-truth data is scarce.
01 Looking inside the model
Language models often hold more than they show. I study what is stored inside a model, how it is retrieved, and what gets in the way of using it, whether that is the phrasing of a request, the examples in a prompt, or the instruction itself. The work combines controlled training setups with probing and steering of hidden states.
Chaemin Jang, Donggyu Lee, Jeesoo Kim, Dongman Lee, Jihee Kim
When instructions change, the differences in a model’s hidden states concentrate in a low-dimensional expression subspace that can be found without labels. Removing it isolates the model’s judgment, which stays reliable under new instructions: with a negated question, AUROC rises from 0.504 to 0.825 for entailment and from 0.480 to 0.891 for safety.
Chaemin Jang, Jihee Kim, Dongman Lee
We separate learning a fact from learning to retrieve it. Two models learn the same new facts as plain statements; the one that earlier saw facts in five request forms can retrieve the new facts through other phrasings. Steering the hidden state just before the answer can turn retrieval on or off, across facts and relations.
Chaemin Jang, Woojin Park, Hyeok Yun, Dongman Lee, Jihee Kim
In-context examples make models rely less on scientific formulas they can already recall, shifting them toward fitting patterns in the examples, even when the examples come from that same formula. The shift holds across 60 latent-structure tasks in five domains, whether accuracy falls, holds, or appears to rise.
02 The model playing people
Language models are increasingly used to stand in for people. I test when that works by measuring simulated respondents and simulated deliberation against real survey and deliberative-poll data, and I look for methods that bring simulations closer to the populations they claim to represent.
Chaemin Jang, Junsik Min, Jaewoo Choi, Donggyu Lee, Haiin Lee, Junyoung Park, Namhee Kim, Hyunwoo Kim, Jungwon Kim, Juho Kim, Nuri Kim, Jihee Kim
Accepted at SocialAgent and UserSim workshops, NeurIPS 2026
Simulating a deliberating public needs agents whose opinions reflect a population and change through discussion. Across six LLMs, national surveys, controlled experiments and the 2017 Shin-Kori deliberative poll, agents argue convincingly but represent opinion too narrowly, and change positions even when they cannot see each other.
Chaemin Jang, Dongman Lee, Jihee Kim
Silicon sampling treats each model call as an independent survey respondent, but instruction-tuned models collapse to one answer: the same persona gives the same answer on more than half of a public-opinion benchmark. Asking the model to describe the distribution in a single call more than halves error against human survey data, and Prompt-Perturbed Argyle cuts error by 21% when per-persona answers are needed.
SBS · with KAIST KIXLAB
Simulated 10,000 citizens deliberating in 1,000 rooms on polarizing issues, for SBS D-Forum 2026.
Ministry of Science and ICT · AI-Science Hub
Building and validating an LLM-agent social simulation framework against human deliberation data.
03 The model reading the ground
Many places that matter for policy have little reliable data. I build methods that read satellite imagery without field labels, combining existing classifiers and self-training, and validate the resulting maps against official statistics.
Ministry of Unification · with Stanford University
10 m crop maps of the Hwanghae provinces for 2019–2025, built without field labels and checked against official statistics.