CoverPrune Preserves 3D VLM Geometric Structure via Optimal Transport
August 14, 2026
CoverPrune reduces 3D VLM inference bottlenecks by reframing token pruning as an Optimal Transport problem. Unlike diversity-based methods that select outliers, this training-free approach maximizes visual evidence coverage to maintain multi-view consistency and geometric structures.
HOW THIS AFFECTS YOU
●
builderYou can reduce 3D VLM inference latency without losing the spatial reasoning capabilities required for geometric tasks.
●
researcherThis moves pruning objectives from simple diversity maximization to evidence-based coverage in 3D spaces.