Jaehyuk Jang

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.

Text dominance in Audio LLMs can be mitigated through a novel back-patching technique that enhances audio representations, challenging the status quo of multimodal processing.

Integrating trainable prompts into the audio encoder can significantly boost few-shot learning performance in Audio-Language Models, outperforming traditional text-only approaches.

Explicitly aligning audio and video streams in a multimodal Transformer boosts emotion recognition, showing that ignoring frame-rate differences hurts performance.

Forget full-network finetuning: adapting only a low-dimensional decoder subspace unlocks state-of-the-art zero-shot depth completion with significantly improved efficiency.