腾讯混元、清华、南洋理工联手，「以小博大」破解空间智能算力与记忆断裂难题 | ECCV 2026

TL;DR - Spatial-TTT is a 2B-parameter model that uses test-time training to encode streaming video into dynamically updated “fast weights,” providing long-term 3D spatial memory with linear compute growth. It reportedly outperforms much larger models on spatial benchmarks while processing videos up to 120 minutes without out-of-memory failures.

• Its decoder interleaves three TTT layers with one full-attention layer, preserving global semantic reasoning while compressing long-range spatial history.
• Each TTT layer combines fast-weight memory with sliding-window attention; lightweight 3D convolutions inject local geometric and temporal structure into visual features.
• Training uses dense descriptions of roughly 16,000 indoor scans followed by about 3 million spatial question-answering examples, with window annealing to shift long-range memory responsibilities toward TTT.
• Spatial-TTT-2B scores 64.4 on VSI-Bench and 76.2% on MindCube-Tiny; fixed-size sliding-window and pending-update caches prevent memory usage from growing with the full video history.