Reproduce “Graph Mamba Operator: A Latent Simulator for Interacting Particle Systems” 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
Karn Tiwari, Niladri Dutta, N M Anoop Krishnan, Prathosh A P
GraMO is a latent simulator that integrates graph-based spatial interactions with long-range temporal state-space modeling within a unified recurrent framework for interacting particle systems. This approach enhances long-horizon prediction accuracy, evidenced by an average 67.3% reduction in Final Mean Squared Error on MoCap data and improved performance across N-body and robotics benchmarks.
Key insights
- The framework learns an input-conditioned neural operator with dynamics that are linear in the lifted latent state, where the transition operator itself dynamically adapts to the current graph state.
- Stability and long-range information propagation are ensured through a normalized adjacency matrix and HiPPO-initialized temporal operators, preventing error accumulation over extended prediction horizons.
- GraMO induces an adaptive graph ARMA process in the latent space, efficiently capturing multi-hop spatial dependencies and long-term memory with coefficients that dynamically adjust based on input.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.