Reproduce “Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory” 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
Ying Yang, Guiyu Zhang, Lianghua Huang, Chang Nie, Chenyang Si, Haofan Wang, Shaoshuai Shi, Li Jiang
SMI organizes video-world memory by spatial region, using a fine-tuned MLLM to remove redundant chunks, retrieve memories for the next action, and reject unreliable additions. On HY1.5, it reaches a Camera-Constrained Scene Consistency score of 67.0200 versus 51.6053 for Base, while removing 83.6842% of memory items. Its latency is 1.38 times Base, and it does not improve every metric.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.