Reproduce “LT-Mem: Volatility-Aware Spatio-Temporal Memory for Lifelong Scene Understanding” 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
Yumin Lee, Hyoseok Ju, Giseop Kim
LT-Mem maintains object histories, current states, and behavior statistics across changing places. It links observations across sessions, classifies events, and adapts updates to object volatility. On combined Lab-S and Lab-L tasks, it scores 0.910, 0.820, and 0.600 on Event F1, QA-Event, and QA-Freq, versus VLM-Batch’s 0.790, 0.680, and 0.333, using 438K versus 7,114K reported tokens. Evidence is limited to controlled ten-session tests.
Key insights
- Establishing persistent cross-session object identity is a foundational requirement for all forms of longitudinal temporal reasoning, as demonstrated by the severe performance drop without it.
- Adaptive memory update decisions, conditioned on an object's learned volatility, significantly improve the accuracy of change detection and event logging by accommodating diverse object dynamics.
- A structured memory architecture that separates current states, event histories, and long-term statistics enables efficient and scalable complex temporal queries over extended periods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.