Home
 |
Kwang-Sung Jun Associate Professsor Graduate School of Artificial Intelligence Department of Computer Science and Engineering Pohang University of Science and Technology kwangsungjun@postech.ac.kr RIST 4-dong 4410-ho, 67 Cheongam-ro, Pohang 37673 South Korea Google scholar CV |
I am currently actively recruiting self-motivated students for MS or PhD. I am also recruiting unpaid undergraduate interns, preferably onsite.
Shoot me an email if you are interested!
Intro
Broadly, I am interested in interactive machine learning (IML), which includes reinforcement learning, bandits, Bayesian optimization, and active learning (see below for some introduction of these). I also develop novel and practical confidence bounds that often becomes key tools for constructing efficient IML algorithms.
Recently, I have been looking into IML problems arising from LLMs, including alignment, post-training, reasoning, and test-time scaling.
I also work closely with Krafton AI on post training of LLMs and with Meta Platforms on efficient A/B testing systems. Previously, I was an assistant professor at the University of Arizona.
News
Apr’26-Jul’26: I AISTATS, 1 ICLR, 2 ICML, 1 COLT, 1 TMLR
Jan’26: I have moved to POSTECH, which is one of the top research universities in South Korea!
Jul’25: 1 paper in ICML and 1 paper in COLT.
Jan’25: 2 papers accepted to AISTATS! (i) Linear extension of the MED algorithm (or Maillard sampling) that is randomized, off-policy-friendly, and nearly optimal. (ii) Estimating the value of the maximum mean among K arms (rather than estimating who achieves it) that can surpass the performance of the oracle who knows the best arm and reports its sample mean!
Nov’24: I gave a talk at UW-Madison SILO on 'Confidence Sequences via Online Learning’ (video link).
Oct’24: 2 papers accepted to NeurIPS: (i) led by Yao, bandit experiments with instrumental variables, so you can experiment when you can't experiment! (ii) led by Junghyun, unified confidence sets for generalized linear models and finally removing the norm dependence in logistic bandits with novel regret analysis!
May’24: 2 papers accepted to ICML: (i) adapting to the unknown noise level in linear bandits (ii) low-rank matrix recovery with better guarantees than the nuclear-norm regularization, how we can perform design of experiments for better subspace recovery, and how to solve low-rank bandits with them!
May’24: 1 paper accepted to COLT on PAC-Bayes bounds with a different divergence that is better than KL!
Interactive machine learning
Interactive machine learning (IML) is an umbrella term for ML problems where the agent's decision is an action that affects which data/feedback s/he gets to receive as opposed to prediction that does not. For example, recommending a product is an action since we get click feedback for the recommended product, not for the unrecommended products (and we don't know what would have happened if we recommended another item). Forecasting tomorrow's weather is a prediction since the weather tomorrow is not affected by which forecast we have made.
If you are getting feedback (i.e., correct/incorrect) on the output from GenAI models (e.g., a response from an LLM), then the right approach is to view the generated outputs as 'action’ and leverage tools from interactive machine learning!
Service
Senior Area Chair: ICML, AAAI
Area chair: COLT, ALT, NeurIPS
Program committee: AISTATS