Xinyu Fu

Achieving state-of-the-art multimodal performance with only 208.62 million unique images and a theoretical training cost of just $400K challenges the notion that larger datasets and budgets are always necessary for success.

SoftSkill achieves up to 42.1 points improvement on LiveMath by transforming lengthy Markdown skills into a few powerful virtual tokens.

Lightweight GUI agents can achieve surprising task completion rates by offloading planning to a pre-computed, app-specific knowledge graph.