Yuqian Fu

Distilling multi-pass top-down keypoint detection into a single full-image forward pass breaks the latency bottleneck of generalist vision models while securing 53.1 PQ on COCO panoptic segmentation.

This work identifies and formalizes this failure mode, which is term destructive resource preemption: obtaining the resources required for a requested task by terminating, overwriting, evicting, or degrading an incumbent task.

StepGuard slashes attack success rates by over 77% while maintaining nearly all utility, setting a new standard for safety in LLM-based agents.

Over 1,500 submissions revealed stark differences in model performance across diverse domains, highlighting the challenges of generalizing egocentric video understanding.