TL;DR
LaM-SLidE introduces a novel framework for modeling the stochastic trajectories of spatial dynamical systems (like molecules or crowds) directly within a compressed, fixed-size latent space. By introducing assignable identifier representations (IDs), the model reaps the efficiency and scalability benefits of latent generative models while maintaining perfect traceability of individual entities over time. The result is a highly scalable architecture that outperforms state-of-the-art graph neural networks and geometric diffusion models while requiring up to 100x fewer function evaluations.
What Problem Does This Paper Address?
Modeling how spatial dynamical systems evolve over time—whether it's the chaotic dance of atoms in a molecule or the collective behavior of pedestrians in a crowd—is a fundamental challenge in science and engineering. Traditionally, researchers use Graph Neural Networks (GNNs) to model these systems by treating entities as nodes and spatial proximities as edges. However, GNNs struggle to scale because their computational complexity grows with the number of entities and interactions.
On the flip side, latent space generative models (like Latent Diffusion Models used in image and video generation) are incredibly efficient because they compress data into a lower-dimensional space before modeling its evolution. Unfortunately, standard latent models cannot easily be applied to physical systems because they use a fixed input/output structure and completely destroy the "traceability" of individual entities—making it impossible to track which specific atom or person went where.
Methods
LaM-SLidE (Latent Space Modeling of Spatial Dynamical Systems via Linked Entities) bridges this gap using a two-stage architecture:
- Stage 1: Entity-Preserving Autoencoder: The encoder takes $N$ input tokens—each representing an entity's position, features, and a unique, randomly assigned identifier (ID) from a fixed pool. Using cross-attention, it maps these variable entities into a constant number ($L$) of learned latent tokens. Crucially, the decoder can reconstruct the exact physical properties of each entity by using the assigned IDs as queries to retrieve information from the latent space.
- Stage 2: Latent Flow Matching: Once the system is compressed into a fixed-size latent representation, a flow-based generative model (built on the stochastic interpolants framework) simulates the system's time evolution purely in latent space. The architecture uses transformer blocks that alternate between spatial and temporal dimensions to predict future latent states.
Datasets
The authors evaluated LaM-SLidE across several distinct domains to test its generalizability:
- Pedestrian Trajectory Forecasting: The ETH-UCY dataset, capturing real-world human crowd movement across five different scenes.
- Basketball Player Forecasting: The SportVU NBA movement dataset, modeling complex adversarial and cooperative player/ball dynamics.
- N-Body System Dynamics: Simulated particle environments including Charged Particles, Spring Dynamics, and Gravitational Systems.
- Small Molecule Dynamics: The MD17 dataset, containing molecular dynamics trajectories for eight small molecules (ranging from 9 to 21 atoms).
- Protein Dynamics: A Tetrapeptide (4AA) dataset used to evaluate long-term prediction horizons (up to 10,000 consecutive frames) on unseen molecules.
Key Insights and Ideas
- IDs as Retrieval Mechanisms: Instead of explicitly modeling physical bonds or constraints, LaM-SLidE relies on the attention mechanism to treat entity IDs as keys/queries. The decoder acts as an associative memory, consistently mapping the same identifier to the same entity across different conformations.
- Decoupling Complexity from Entity Count: Because the generative modeling happens entirely within a fixed number of latent tokens ($L$), the computational cost of predicting future trajectories does not scale directly with the number of particles or agents in the scene.
- Eliminating Inductive Biases Beneficially: Unlike many competing models that hardcode spatial equivariance or velocity features, LaM-SLidE intentionally omits these constraints. Instead, it relies on data augmentation (random rotations/translations), allowing it to scale gracefully with model size and parameter count.
Key Results
- State-of-the-Art Accuracy: LaM-SLidE achieved the lowest Average Discrepancy Error (ADE) and Final Discrepancy Error (FDE) on the MD17 molecular dynamics benchmark, outperforming all equivariant GNN baselines.
- Massive Efficiency Gains: Compared to GeoTDM (the second-best geometric diffusion model), LaM-SLidE requires 10x to 100x fewer Number of Function Evaluations (NFEs) to generate trajectories, drastically reducing inference time.
- Cross-Molecule Generalization: The architecture successfully trained a single unified model on all eight molecules in the MD17 dataset simultaneously, outperforming baselines that were specialized and trained on individual molecules.
- Favorable Scaling: Experiments on the tetrapeptide dataset demonstrated that performance consistently improves as the model scales up (tested from 4M up to 28M parameters).
Limitations and Future Work
The authors noted a distinct trade-off between the number of latent space vectors ($L$) used to encode system states and the performance of the generative model. While compressing the system into fewer latents improves generative efficiency, it places a higher burden on the autoencoder's reconstruction accuracy. Additionally, while LaM-SLidE shows remarkable cross-domain generalization, the authors emphasize that it is not yet a universal Graph Foundation Model (GFM), noting that future work should explore emergent properties when training on even larger, multi-domain datasets.
Summary updated