🏢 UC Santa Cruz

CRATE-a: A new white-box vision transformer architecture achieves 85.1% ImageNet accuracy by strategically scaling model size and datasets, outperforming prior white-box models and preserving interpre…

VLMs struggle with insufficient visual info for Q&A; this work introduces a novel Directional Guidance task and a data augmentation framework, significantly improving VLM performance by teaching them …

ECO prompts enable efficient LLM unlearning by corrupting prompts flagged for forgetting, achieving promising results across various LLMs and tasks with minimal side effects.

FairnessWithoutHarm achieves fairer ML models without sacrificing accuracy by using an influence-guided active sampling method that doesn’t require sensitive training data.

Spiking Autonomous Driving (SAD) is the first unified SNN for autonomous driving, achieving competitive performance in perception, prediction, and planning while significantly reducing energy consumpt…