Janghoon Cho Ph.D.

Staff Engineer, Qualcomm AI Research
Seoul, South Korea

I work on efficient multimodal AI, with a focus on visual token compression, long-video understanding, and multimodal retrieval. My goal is to help make multimodal AI run efficiently, both in the cloud and on devices with limited computational resources.

My background spans submodular optimization and audio and speech processing. At Qualcomm, I have also led speaker verification development toward deployment on mobile devices. I received my Ph.D. in Electrical Engineering from KAIST, advised by Chang D. Yoo.

Email Google Scholar CV

Recent Work

FLoC: Facility Location-Based Efficient Visual Token Compression for Long Video Understanding

ICLR 2026 First author

A training-free method that selects a compact, representative set of visual tokens for long-video understanding. FLoC uses submodular optimization and works across video-language models without depending on the user query.

Paper PDF

Preserving the details that matter in long video understanding

visual token compression ratio 1/32

MLVU · Needle QA · Qwen2-VL 7B

Green boxes: tokens selected by FLoC

Question

What is the woman wearing during the summer sunset?

FLoC predictionA hat and sunglasses✓ Correct answer

Compare methods on this example

Predictions reported in Figure 8 for the same video and question, at a compression ratio of 1/32
Method	Prediction	Result
FLoC	A hat and sunglasses	Correct
DivPrune	A dress and heels	Incorrect
TS-LLaVA	A dress and heels	Incorrect
Spectral Clustering	A dress and heels	Incorrect
K-means	A hat and sunglasses	Correct
K-medoids	A dress and heels	Incorrect

One qualitative example from Cho et al., FLoC, Figure 8. Original image regions, cropped and rearranged; CC BY 4.0. View original figure.

Experience & Education

2018 – Present: Qualcomm AI ResearchStaff Engineer, Korea
Efficient multimodal AI and audio/speech research. Speaker verification model and training pipeline development in collaboration with commercialization teams.
2017 – 2018: SK hynixSenior Engineer, Solution Algorithm Team
2018: Ph.D., Electrical Engineering, KAISTAdvisor: Chang D. Yoo
Dissertation: Summarizing Distribution: Submodular Probability Density Cover.
2011 / 2009: M.S. / B.S., Electrical Engineering, KAIST

Publications

Google Scholar

Beyond research

I like to unwind through physical activity. I enjoy basketball and continue to play today. During university, I played on the university basketball team and was named MVP at an intramural basketball tournament.

More recently, I took up running and completed a 10 km race in 48 minutes.

Contact

For research collaboration and professional opportunities:
janghoon.cho86@gmail.com

U.S. immigration: EB-2 NIW petition (I-140) approved; documentarily qualified (DQ).