Suppressing the Shortcut Hidden Behind the Gains of VideoLLM Fine-Tuning
Youngwoo Shin, Jongsuk Kim†, Junmo Kim†
Suppressing the shortcut that fine-tuning builds in VideoLLMs by intervening on the update through projection.
Ph.D. Candidate · Electrical Engineering, KAIST
I am a Ph.D. student at the Statistical Inference & Information Theory Laboratory (SIIT) at KAIST, advised by Prof. Junmo Kim.
I study how multimodal large language models represent and reason over time, and how to interpret the generation process to align better with intention.
I am currently looking for research internship opportunities. Feel free to reach out!
Youngwoo Shin, Jongsuk Kim†, Junmo Kim†
Suppressing the shortcut that fine-tuning builds in VideoLLMs by intervening on the update through projection.
Youngwoo Shin, Yusung Ro, Minseo Kim, Junmo Kim†
NeurIPS 2026
Reinforcing fading temporal signals in VideoLLMs at inference time, without additional training.
Youngwoo Shin*, Jiwan Hur*, Junmo Kim†
ICLR 2026
Training-free guidance for visual autoregressive (VAR) models, derived from an information-theoretic view of next-scale generation.
Changho Choi*, Youngwoo Shin*, Gyojin Han, Dong-Jae Lee, Junmo Kim†
ACM Multimedia 2025
The first benchmark and baseline MLLM for spatio-temporal understanding of 4D LiDAR.
* equal contribution · † corresponding author