MJev-Spatial: Unified Classification for Spatial Intelligence
October 14, 2026
MJev-Spatial replaces autoregressive text decoding for spatial tasks with a unified classification framework. It uses pretrained query and key projections to score candidates like 2D points or scalars without adding new parameters.
HOW THIS AFFECTS YOU
●
researcherYou can achieve faster, more efficient spatial reasoning in VLMs by avoiding expensive text parsing.