KAIST
Graduate research
Language-model internals, simulated populations and deliberation, and satellite-based measurement
Chaemin Jang
M.S. Student · KAIST School of Computing
Advised by Jihee Kim and Dongman Lee. I study language models from three sides: what happens inside them, how well they stand in for people, and what they can measure where data is scarce.
NOWUPDATED OCTOBER 2026
- How knowledge taught only as statements becomes retrievable in every format
- What changes an LLM agent's mind in multi-agent deliberation
- How the order of answer options biases LLM survey responses
- How well LLMs predict which demographic groups differ in surveys, and by how much
- Label-free crop mapping for North Korea, with Stanford
NEWS10 ITEMS · SCROLL
Presenting at NeurIPS 2026 workshops in Paris and Atlanta
From Simulated Citizens to Simulated Deliberation: Challenges in Representation and Interaction
Our project goes on stage at SBS D-Forum 2026
Presented at EMNLP 2026
Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
Presented at COLM 2026
In-Context Examples Suppress Scientific Knowledge Recall in LLMs
Accepted at NeurIPS 2026 SocialAgent and UserSim workshops
From Simulated Citizens to Simulated Deliberation: Challenges in Representation and Interaction
Accepted to EMNLP 2026
Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
Accepted to COLM 2026
In-Context Examples Suppress Scientific Knowledge Recall in LLMs
Looking inside the model
How models store, retrieve, and use what they learn.
The model playing people
How models reproduce populations, opinions, and human behavior.
The model reading the ground
How models measure places that lack data.
PUBLICATIONS5 PAPERS
2026
-
Preprint
Isolating a Language Model’s Judgment with a Label-Free Subspace
Chaemin Jang, Donggyu Lee, Jeesoo Kim, Dongman Lee, Jihee Kim
Keeps its judgment after it can no longer express it.
-
Preprint
Learning a Fact Is Not Learning How to Retrieve It
Chaemin Jang, Jihee Kim, Dongman Lee
Learns a fact, but retrieves it only the way it was taught to be asked.
-
Preprint
From Simulated Citizens to Simulated Deliberation: Challenges in Representation and Interaction
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
Deliberates convincingly, but not the way a real public does.
-
EMNLP Findings
Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
Chaemin Jang, Dongman Lee, Jihee Kim
Describes a distribution it cannot sample from.
-
COLM
In-Context Examples Suppress Scientific Knowledge Recall in LLMs
Chaemin Jang, Woojin Park, Hyeok Yun, Dongman Lee, Jihee Kim
Drops the formula it knows once examples appear.
PROJECTS3 FUNDED
-
SIMULATION
Human–AI Deliberation for SBS D-Forum 2026
SBS · with KAIST KIXLAB · 2026
Simulated 10,000 citizens deliberating in 1,000 rooms on polarizing issues, for SBS D-Forum 2026.
-
SIMULATION
AI-Agent-Based Social Simulation: Validation and Alignment with Human Behavior
Ministry of Science and ICT · AI-Science Hub · 2026
Building and validating an LLM-agent social simulation framework against human deliberation data.
-
SATELLITE &
LOW-RESOURCEMapping North Korean Agriculture from Satellite Imagery
Ministry of Unification · with Stanford University · 2026
10 m crop maps of the Hwanghae provinces for 2019–2025, built without field labels and checked against official statistics.
EXPERIENCE
NSUSLAB
AI Engineer
LLM-based translation and customer-inquiry tools
KAIST & IBS
Undergraduate research
Graph learning and multimodal imaging
OTDA
Founder
Designer fashion-sharing platform