Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory
SMI uses an understanding model to manage long-range spatial memory for long-video world models.
The paper says longer memory sequences make spatial context harder to manage in action-conditioned video prediction. SMI handles that with four operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering. The authors report improvements across baselines, benchmarks, and world-model backbones in memory sparsity, generation stability, and spatial consistency. Source: HF Daily Papers' note.
The paper says longer memory sequences make spatial context harder to manage in action-conditioned video prediction. SMI handles that with four operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering. The authors report improvements across baselines, benchmarks, and world-model backbones in memory sparsity, generation stability, and spatial consistency. Source: HF Daily Papers' note.
score 4