Kangwook Ko

Much of the reported progress in multi-object tracking stems from inconsistent baselines and hyperparameter tuning rather than algorithmic breakthroughs—and this controlled empirical audit reveals which modular components actually move the needle.

Forgetting individual information in MLLMs can be achieved without compromising model performance, thanks to a novel method that localizes updates to specific layers.

Adapting visual features directly in a low-rank subspace can drastically improve VLMs' adversarial robustness without sacrificing performance.

Few-shot tuning can distort embeddings, but Subspace Tuning stabilizes adaptation, leading to robust generalization even for unseen classes.