Shaoyi Du

Forget specialized counting models – Count Anything can now count *anything* you describe in an image, from cells to crops to cars, across diverse visual domains.

LLMs can now natively understand hypergraphs, unlocking their ability to reason about complex, multi-way relationships that traditional graph-based methods miss.

Domain-specific knowledge hypergraphs can now be extracted with significantly improved quality by dynamically learning and applying extraction skills, outperforming static few-shot learning.

Weight regularization, often overlooked in parameter-efficient continual learning, can still significantly improve the stability-plasticity trade-off, even when using low-rank adapters.