CVPR 2026 Accepts ByteDance Seed's SpatialTree: A New Framework for MLLM Spatial Intelligence

A joint research team from Zhejiang University, ByteDance Seed, and Beijing University of Railways has introduced the SpatialTree framework, which has been accepted at CVPR 2026. This framework aims to redefine how multimodal language models handle spatial intelligence, addressing ongoing challenges in understanding distances, volumes, and multi-faceted relationships. SpatialTree organizes capabilities into four layers: perception, mental planning, mental simulation, and agent efficiency. The SpatialTree-Bench dataset encompasses 27 sub-capabilities for spatial assessment.