Sangoh Lee

01about

Ph.D. student at POSTECH GSAI, working with Prof. Wook-Shin Han in the Data Systems Lab. Research focuses on Physical AI, Vision-Language-Action models, and World Models. I believe physical intelligence will be the next major leap in AI.

Previously: Natural Language Processing, Knowledge Graph Question Answering, and AI for Databases — publications at ACM SIGMOD 2024 and EMNLP 2025.

02interests

Physical AI Vision-Language-Action Models World Models Embodied Agents

03news

Aug 2026NewPreprint on Intention Distillation (INDI) for VLA action decoders is now on arXiv. [arXiv] [project]
May 2026VAP paper is accepted to ICML 2026 🎉
Jan 2026Received the BK21 Best Paper Award (Semi-Grand Prize) from POSTECH GSAI for SAFE.
Dec 2025Preprint on Visual Attentive Prompting (VAP) for personalizing VLAs is now on arXiv. [arXiv] [project]
Nov 2025SAFE accepted to EMNLP 2025 main track (Suzhou, China).
Jun 2024ASM and ASM in Action accepted to ACM SIGMOD 2024.
Jul 2023Awarded the SBS Cultural Foundation Scholarship (≈75,000 USD).
May 2023Awarded the Postechian Fellowship for the highest grades entering POSTECH GSAI.
Feb 2023Started Ph.D. at POSTECH GSAI in the Data Systems Lab.

04publications

arXiv 2026 · preprint

Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

Sangoh Lee, Sangwoo Mo, Wook-Shin Han

Intention Distillation (INDI) distills behavior-level intent from a frozen teacher into the VLA action decoder, so the policy acts on what the demonstration was trying to achieve — not just the motor commands it contained.

arXiv pdf project

ICML 2026 · regular track · Seoul, South Korea

Bring My Cup! Personalizing Vision-Language-Action Models with Visual Attentive Prompting

Sangoh Lee, Sangwoo Mo, Wook-Shin Han

A training-free perceptual adapter that gives frozen VLAs top-down selective attention — personalized object manipulation from a few reference images.

arXiv pdf project code

EMNLP 2025 · main track · Suzhou, China

SAFE: Schema-Driven Approximate Distance Join for Efficient Knowledge Graph Querying

Sangoh Lee, Sungho Park, Wook-Shin Han

A schema-driven framework for robust query graph generation and efficient KG retrieval, combining an Approximate Distance Join algorithm with compact schema graphs.

pdf

SIGMOD 2024 · research track · Santiago, Chile

ASM: Harmonizing Autoregressive Model, Sampling, and Multi-dimensional Statistics Merging for Cardinality Estimation

Kyoungmin Kim, Sangoh Lee, Injung Kim, Wook-Shin Han

A learned cardinality estimator combining autoregressive models, sampling-based join merging, and multi-dimensional statistics merging.

pdf

SIGMOD 2024 · demonstration · Santiago, Chile

ASM in Action: Fast and Practical Learned Cardinality Estimation

Sangoh Lee, Kyoungmin Kim, Wook-Shin Han

A demonstration of ASM's internal estimation process and plan-space exploration, showcasing end-to-end execution-time advantages.

pdf demo video

2026 · 2

arXiv

Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models
Sangoh Lee, Sangwoo Mo, Wook-Shin Han

arXiv project
ICML

Bring My Cup! Personalizing Vision-Language-Action Models with Visual Attentive Prompting
Sangoh Lee, Sangwoo Mo, Wook-Shin Han

arXiv project code

2025 · 1

EMNLP

SAFE: Schema-Driven Approximate Distance Join for Efficient Knowledge Graph Querying
Sangoh Lee, Sungho Park, Wook-Shin Han

pdf

2024 · 2

SIGMOD

ASM: Harmonizing Autoregressive Model, Sampling, and Multi-dimensional Statistics Merging for Cardinality Estimation
Kyoungmin Kim, Sangoh Lee, Injung Kim, Wook-Shin Han

pdf
SIGMOD demo

ASM in Action: Fast and Practical Learned Cardinality Estimation
Sangoh Lee, Kyoungmin Kim, Wook-Shin Han

pdf video

05ongoing

🪞

Ego–World Factorization for Robot World Models

Robot world models treat the future as one entangled prediction problem, yet much of a robot's own future is already determined by its controller and kinematics. This work asks what a robot should predict about itself before predicting the world.

🧭

Active Spatial Intelligence in Embodied MLLM Agents

Embodied agents can act to gather the views they need, yet they often commit to an early belief and stop exploring. This work examines what limits active perception in MLLM agents and how much of that gap can be closed.

🧪

Dissociating Kinematics, Force, and Control in Video World Models for Manipulation

A systematic audit of whether the Physics Emergence Zone in video world models (e.g., DreamDojo, DreamZero, V-JEPA 2) encodes manipulation-relevant contact dynamics — friction, mass, and force inference.

06notes & posts

Core robotics Rigid Body Motion: Rigid Body Transformation Oct 16, 2025

07cvpdf ↗

Education

2023 –Ph.D. in AIPOSTECH GSAI · Wook-Shin Han · GPA 4.27/4.30
2019 – 23B.S. in EEPOSTECH · Salutatorian (2/256) · GPA 4.25/4.30
2017 – 19Daegu Il Science HSDG1S

Experience

'22 – '23Research InternPOSTECH Data Systems Lab · sampling-based cardinality estimation

Teaching

2025 –TA · Deep Learning (AIGS538)
2023TA · Database Systems (CSED421)
'20 – '22Tutor · Calculus, Linear Algebra

08honors & awards

2026BK21 Best Paper Award (Semi-Grand Prize), POSTECH GSAI
2023SBS Cultural Foundation Scholarship (≈ 75,000 USD)
2023Postechian Fellowship (highest grades entering POSTECH GSAI)
2023Kwanjeong Scholarship (declined in favor of SBS)
2023Awards of Proud Postechian Member with the Highest Honor
2023Newcomb Lim Ki-Hong Design Project Best Award (POSTECH EE)
2022KEC Science Education Foundation Scholarship
2022Awards of Korea Military Academy Principal (ROND program)
2021National Science and Technology Scholarship