Yan Ma

Curvature estimates for graphical solutions to curvature quotient equations are proven without convexity or $Γ_{k+1}$-admissibility, unlocking analysis for a broader class of geometric problems.

Pretraining loss actively misleads video data curation: broad data coverage systematically beats aggressive quality filtering across downstream benchmarks in both Wan 2.1 and V-JEPA 2.1.

Retaining critical information in robot locomotion can yield a 15% performance boost while simplifying the model's architecture.

AnchorScore reveals a reliable method to identify which classes MLLMs struggle to annotate, potentially saving researchers significant time and resources in evaluation.