Reproduce “Hierarchical Spatiotemporal World Modeling with Content-Adaptive Spatial Addressing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tiantian Li, Junfeng Lu, Yifei Wang, Zhaoyang Zhang, Lei Wang, Tianyi Qian
HiSTFormer predicts future first-person images by combining recent temporal context with spatially indexed retrieval of older experiences. Temporal layers process recent observations and actions; spatial layers retrieve from a bounded cache using path-integrated addresses adjusted by content-dependent offsets. On 9×9 MemoryMaze, it achieves LPIPS 0.2920 ± 0.0037 for 500-frame prediction, versus CoME’s 0.3169 ± 0.0020, assuming reliable ground-truth headings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.