Yuhua Li

STD is proposed, a hierarchical token pruning framework that adapts token selection mechanisms to the functional role of each network stage, and introduces a Stability-Adaptive Trigger in deep layers to execute pruning only during semantically stable phases.

Bridging the gap between domains with tailored noise and selective inpainting can revolutionize few-shot object detection performance.

Counterintuitively, pushing *away* certain image tokens from their corresponding text embeddings boosts CLIP's few-shot cross-domain performance.

Fine-tuning vision-language models for cross-domain few-shot learning makes them *worse* at distinguishing between classes due to an exacerbated "attention sink" problem, but a simple token re-weighting scheme can fix it.

Object detectors in new visual domains suffer from "astigmatism," but mimicking the human eye's foveal vision can bring them into focus.