Kwang-Sung (Kwang) Jun (@kwangsungjun) profile picture

associate professor at POSTECH. ML theory and recently LLMs! i am a fan of regret minimalism, both in life and research.

tucson az
Joined June 2009
Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)Photo by Kwang-Sung (Kwang) Jun (@kwangsungjun)
#aistats2027 #alt2027 #iclr2027 I recently found many cases where scores are bipolar. One review says the paper is readable and the other review says the paper is completely unreadable. My hypothesis: probably, the former relied on AI to read a paper, and the latter didn't.
3
1
34
5,554
Upon inspecting the paper, it is not well written -- AI slob that alters/rephrases precisely-defined technical terms and blurs the meaning and assumes prior knowledge of the topic and 'declarative' sentences instead of procedural or guiding.
1
2
616
For pre-AI standard, it should be rejected. For now, I don't have the answer. However, will the future of papers be written via AI and read via AI and maybe no one cares whether or not it is readable without AI?
3
543
Kwang-Sung (Kwang) Jun retweeted
👋 KRAFTON's Raon is entering the open-weight scene! Meet Raon-1 mini, our first model, trained from scratch. On par with leading models in its size class, while running faster and using less memory. Available under the Apache 2.0 license. huggingface.co/KRAFTON/Raon-…
48
89
21
379
32,749
I met my PhD advisor at #icml2026 who now writes ai-free music!
2
43
3,758
Kwang-Sung (Kwang) Jun retweeted
Our new paper at #ICML2026 studies why simple on-policy preference learning (e.g., online DPO) can outperform their offline counterparts! Joint work with Jihun Yun @jasondeanlee @kwangsungjun Presenting tomorrow (July 8) 5pm, Hall A #4514! arxiv.org/abs/2601.08421 1/5
1
7
41
3,413
#ICML2026 poster presentation today at 2:30pm: #4512 "Fixed Budget is No Harder Than Fixed Confidence in Best-Arm Identification up to Logarithmic Factors" Joint work with Houssam's team at Meta and my team from U of Arizona with co-first authors Kapilan and Yinan
2
1
19
1,479
Now, one can see that the existence of our algorithm and guarantee implies that FB is no harder than FC, ignoring logarithmic factors..!
1
88
due to traveling constraints, i will present the poster instead of the co-first authors. i hope to see you at the poster!
71
🚀Exciting news! Our project "Orchestrating Large Language Models via Contextual Partial Monitoring" has been selected for the NVIDIA Academic Grant Program! Many thanks to to NVIDIA the hardware support and I look forward to making breakthroughs! #NVIDIAGrant @NVIDIAAIDev
1
7
735
I am heading to JSM2025, Sunday-Tuesday. #JSM2025 I will give a talk at "Advances in Adaptive Experimentation and Anytime-Valid Inference" session on Sunday. Let's meet up if you like to discuss anything!
4
735
We propose three new algorithms under our key modeling assumption that the preference is just a way to get some info about the target language model \pi^*. (3/n)
1
278
3. We also propose a REBEL-like objective, but potentially better and natural, called preference distillation. The theoretical guarantee is the same as DPO. (6/n)
1
192
All in all, we provide learning-theoretic justification of the RLHF objective and provide theoretical analyses, all under the viewpoint that alignment is a DISTRIBUTION LEARNING from pairwise feedback. Please consider coming to our poster and have a fun discussion! (7/7)
176
Juno Kim and I will be presenting a poster at the MoFA workshop in #ICML2025. Time: today 12-1:30pm, 4-5pm. Title: "Alignment as Distribution Learning: Your Preference Model is Explicitly a Language Model" (1/n)
1
1
308
We never understood how "the RLHF objective" can be justified. The alignment problem itself never involves rewards! So, we explored under what assumptions we can develop new methods or justify existing ones, all from the DISTRIBUTION LEARNING viewpoint. (2/n)
1
123