Sung-Ju Lee

Prof. at KAIST.

[Taeckyung Lee @taeckyung_lee]

Excited to present our poster at the ICML 2026 AI for Math Workshop tomorrow!

Dense Supervision but Sparse Guidance: A Token-Level Analysis of On-Policy Self-Distillation

We asked a simple question: What does token-level supervision in on-policy self-distillation (e.g., SDPO) actually teach the model?

Our key finding: the highest-KL tokens are usually not the direct cause of failure. They often correspond to comments or other non-executable text, yet a single teacher-preferred token can still redirect the model toward a completely different, and often correct, solution trajectory.

This suggests that self-distillation works less by local token correction and more by trajectory steering, offering a new perspective on why token-level objectives can outperform uniformly distributed sequence-level rewards.

If you're interested in RLVR, self-distillation, or LLM reasoning, come say hi at the poster!

Thanks to my amazing collaborators Jeonghye Kim, @hyung_jun_yoon, and @wewantsj .

#ICML2026 #LLM #RLVR #SelfDistillation #MachineLearning

[Yewon Kim @haiyewon]

A Design Space for Live Music Agents 🎷🎹🥁 #CHI2026

What does it take for AI to truly jam with you? We surveyed 184 live music agents across AI, HCI, and Computer Music fields to map the design space, and where it's headed.

🗓️ Talk: Fri Apr 17, 12:15PM · P1 Room 132
📄 Paper: https://t.co/vNfdqrdV23
🔗 Interactive demo: https://t.co/XD1lP5pGCB

[Sung-Ju Lee @wewantsj]

@punwitter Love you, bro!

[Hyungjun Yoon @hyung_jun_yoon]

Pleased to share that our paper, “SelfReplay: Adapting Self-Supervised Sensory Models via Adaptive Meta-Task Replay,” is accepted to #SenSys2025!

Huge thanks to my collaborators, Jaehyun, Biniyam, @GaoleDai4, Prof. Mo Li, @Taesik_MobileAI, @kimin_le2, and @wewantsj! (1/2)

[Adiba Orzikulova @AdibaOrz]

⌚️Wearables that can collaborate via bio-signals?

📜Check out our latest #SenSys2025 paper, BioQ, to discover how biological cues enhance context-aware multi-device collaboration.

Huge thanks to @dvasile7, Ian Tang, @raswak, @wewantsj, @ChulhongM for making this possible!

[Taeckyung Lee @taeckyung_lee]

We're happy to announce that our work on test-time adaptation with binary feedback has been accepted to #ICML2025!

We will discuss how we can utilize human feedback for TTA.

Special thanks to collaborators
@sornswgn @JunsuKim97 @jinwoos0417 @Taesik_MobileAI @wewantsj !!

[Ryuhaerang Choi @CHI2025 @haerang_hci]

I’ll present our #CHI2025 paper, “Private Yet Social: How LLM Chatbots Support and Challenge Eating Disorder Recovery” in LLM for Health session.

🕤9:36 AM Wed 30. Apr
📍Annex Hall F206
📃https://t.co/8AkdzL7aoo
Grateful to Taehan, @subin_hci, Prof. Jennifer G Kim, and @wewantsj

[Subin Park @subin_hci]

Excited to share our #CHI2025 LBW paper, ""I know my data doesn’t leave my phone, but still feel like being wiretapped": Understanding (Mis)Perceptions of On-Device AI Vishing Detection Apps" in Yokohama!! 😆😆

🔗 Project link: https://t.co/IlcKESNuNE

[Yewon Kim @haiyewon]

What role can AI play in unlocking musical creativity?

At #CHI2025 (🥇 Best Paper Award 🥇), we present Amuse: a songwriting tool for musicians that turns photos, sounds, and stories into chords 🎶

📄 https://t.co/ScJX1AGccv
🎧 https://t.co/FtkINAXdEK

[Seungjoo Lee @seungjoo99]

🚀 (FL)² enables 99.9% label-free federated learning!
I am on my way to #NeurIPS2024

📍 Catch us at #NeurIPS
- Date: Wednesday, Dec 11
- Time: 11 a.m. PST — 2 p.m. PST
- Location: East Exhibit Hall A-C #3603

Paper: https://t.co/LZ0eB74Xhd

[Sung-Ju Lee @wewantsj]

@NoSyu Thanks for the invitation! Had a blast. The pic looks like I’m speaking in front of nobody! Haha

[JinYeong Bak @NoSyu]

Thank you for an intersting talk and for sharing your research work and insights with SKKU students, Prof. Sung-Ju Lee
https://t.co/w26P1ukFjg
"Embracing NLP for Mobile Computing Innovations"

[Hyungjun Yoon @hyung_jun_yoon]

👋ChatGPT can recognize your hand gestures from your smartwatch!
I’m on my way to #EMNLP2024!
Excited to present our By My Eyes: Grounding LLMs with Sensor Data via Visual Prompting 🎉

Find me at Poster Session E (Riverfront) on 11/13, 16:00-17:30. Let’s chat!
📽️Demo:
(1/2)

[Seungjoo Lee @seungjoo99]

🚀 Thrilled to share that "(FL)²: Overcoming Few Labels in Federated Semi-Supervised Learning" has been accepted at #NeurIPS2024 @NeurIPSConf 🎉

Stay tuned for camera-ready version and code release 👀

Huge thanks to the amazing team: Long, @jaeminshin31, and @wewantsj 👏

[Hyungjun Yoon @hyung_jun_yoon]

🎉 Thrilled to share our paper “By My Eyes: Grounding Multimodal Large Language Models with Sensor Data via Visual Prompting” is accepted at #EMNLP2024 Main! 🙌 Huge thanks to Biniyam, Prof. @Taesik_MobileAI, @kimin_le2, and @wewantsj!

📄 Preprint: https://t.co/J4UyJVCPrp (1/2)

[Sung-Ju Lee @wewantsj]

@punwitter @HPE_labs @kaistpr Thanks for hosting! Loved the energy of your lab!!!

[Puneet Sharma @punwitter]

It was wonderful to have ex @HPE_Labs colleague, now Professor at KAIST (@kaistpr) S.J. Lee, visit and give a talk on "Channeling Mobile Computing Innovations into Everyday Health"
Thanks for visiting @wewantsj and telling us about fab research your NMSL is doing.

[Taeckyung Lee @taeckyung_lee]

Curious about *estimating test-time adaptation (TTA) accuracy* without labeled data or source data access? Check out our #CVPR2024 paper on AETTA, a novel method for TTA accuracy estimation.
📄 Paper: https://t.co/Ck3S4wNEVn
💻 Code: https://t.co/jw1lx71l0V

(1/3)

[Ryuhaerang Choi @CHI2025 @haerang_hci]

Our research on FoodCensor was featured on Daejeon MBC, a Korean national TV broadcaster (in Korean)!
🎥 https://t.co/Zo3o0Akkt7

Hope this could bring more attention and support to the population with eating disorders❤️

Thanks again, @subin_hci @h4nsujin @wewantsj

[Adiba Orzikulova @AdibaOrz]

📱Struggling with smartphone addiction? Discover how AI can help with our latest #CHI2024 paper, Time2Stop ⏳2️⃣✋

📜Paper: https://t.co/vIECoeiMQh
📽️Teaser: https://t.co/MBKAY6ZhGd
📆Wednesday 2PM HST, 🗺️ 316B.