Optimizing Visual-Token Allocation in Long-Video MLLMs
September 4, 2026
An empirical study shows that query-selected frame selection outperforms uniform sampling by 6.9 points on the LongVideoBench hour-long bin. The researchers found that Orthogonal Matching Pursuit, a sparse-approximation algorithm, performs competitively with modern purpose-built selectors.
HOW THIS AFFECTS YOU
●
builderYou can improve long-video model performance by implementing query-aware frame selection instead of uniform sampling.
●
researcherUse sparse-approximation algorithms as high-performance baselines for video token selection.