AI News

GeoGR compresses multi-view 3D vision-language inputs with voxel consolidation and group-wise LLM skipping

ABC3 min read
Research