Memory Forcing: Spatio-Temporal Memory for Consistent Scene Generation on Minecraft
TLDR
Memory Forcing uses spatial memory and hybrid training to improve long-term consistency in autoregressive video diffusion for Minecraft world modeling.
Reasoning
The paper introduces a novel framework combining spatial memory with training strategies to balance exploration and revisit consistency, showing strong empirical results. However, the abstract lacks details on baseline comparisons and quantitative metrics, and the approach is limited to Minecraft.
Read-first score
Read-first score 58.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.
Field roles
Rank sensitivity
Stability: volatile; rank range: 367.
Keyword Scores
Deep Analysis
Innovations
- Hybrid Training: exposes distinct gameplay regimes to guide model reliance on temporal vs spatial memory
- Chained Forward Training: extends autoregressive training with model rollouts to encourage spatial memory reliance
- Point-to-Frame Retrieval: efficiently retrieves history by mapping currently visible points to source frames
- Incremental 3D Reconstruction: maintains and updates an explicit 3D cache
Methodology
Memory Forcing pairs training protocols with a geometry-indexed spatial memory. It uses Hybrid Training to distinguish exploration and revisit regimes, Chained Forward Training for autoregressive rollouts, Point-to-Frame Retrieval for efficient history retrieval, and Incremental 3D Reconstruction for explicit 3D caching.
Key Results
Memory Forcing achieves superior long-term spatial consistency and generative quality across diverse environments while maintaining computational efficiency for extended sequences.