Tag
DRIVE proposes a dual-level skill modeling framework that separates reasoning knowledge from interaction knowledge for web agents under continual learning, achieving a 52.8% task success rate on WebArena, outperforming the skill-free baseline by 7.3 percentage points.