AnalysisAI ModelsAugust 29, 2026

Parallel Tube Decoding speeds up spatio-temporal video grounding

Paper proposes Parallel Tube Decoding, which removes autoregressive dependencies so spatial and temporal video grounding happen simultaneously, cutting latency while improving accuracy. It also introduces Decoupled Block Attention and localization-aware policy optimization.

1 source

More stories today

Open the live feed