VectorWorld: Efficient Streaming World Model via Diffusion Flow on Vector Graphs

Closed-loop evaluation of autonomous-driving policies requires interactive simulation beyond log replay. Existing generative world models suffer three gaps: history-incompatible initialization, sampling latency exceeding real-time budgets, and compounding kinematic infeasibility. We propose VectorWorld, a streaming vector-graph world model that incrementally generates ego-centric lane--agent tiles during rollout. VectorWorld couples a motion-aware gated VAE for history-compatible initialization, an edge-gated relational DiT with interval-conditioned MeanFlow and JVP-based large-step supervision for solver-free outpainting, and $\Delta$Sim, a physics-aligned NPC policy with hybrid discrete--continuous actions and differentiable kinematic logit shaping. On Waymo Open Motion and nuPlan, VectorWorld improves map fidelity, initialization validity, and density calibration, enabling stable real-time $1\mathrm{km}+$ closed-loop rollouts.

Testing self-driving cars only by replaying recorded drives is limited because it cannot show what would happen if the car made a different decision or if traffic continued beyond the recorded clip. VectorWorld is a simulator that builds a driving world as the test vehicle moves. It creates new road and traffic scenes ahead of the vehicle, gives newly created cars short motion histories so they behave naturally from the start, and keeps other vehicles' movements physically reasonable. Because the generation is fast, the simulator can support long closed-loop tests in which a driving policy repeatedly observes, acts, and receives new surroundings. In experiments on public driving datasets, VectorWorld produced more consistent road layouts, fewer invalid starting scenes, and more stable long-distance simulations. This makes it useful for stress-testing and training autonomous-driving policies in a controlled environment before using more expensive or risky real-world evaluation.