At the intersection, look ahead before you turn: previewing a short future route cuts Porto distance JSD to 1.01

Road-network trajectory generation shifts from simulating 2D locations to modeling behavior on the road network, where each step is an intersection turn decision. Diffusion-style generators treat a trajectory as a whole sequence of 2D locations and denoise it from noise, often overlooking the next-transition choice at intersections and producing trajectories that drift off the road network. On Porto, Beijing and San Francisco, using the same HOSER graphs and splits, matched origin-destination conditions, replicate counts and generation budgets, the Markov-history variant reaches the lowest Porto distance-distribution JSD at 1.01 while the full-history variant is 1.53, both below HOSER's 1.57; on Beijing and San Francisco the full-history variant reaches 1.68 and 1.39, below HOSER's 2.04 and 1.82.

At each step the model first predicts a finite-horizon continuation of the route, then uses that preview to score every legal outgoing transition. A conventional stepwise generator models only the one-step conditional distribution of the next state, so future information is not explicitly represented as context that shapes the current decision. The preview comes from a future decoder, and the transition network normalizes only over topologically legal candidates, with the STOP event admissible only at the destination; ablating the candidate-conditioned continuation loss raises Porto distance-distribution JSD from 1.53 to 2.55.

After each executed movement the model re-anticipates from the updated history instead of running the preview open-loop. Executing a predicted route open-loop would leave later steps without a chance to correct against the actual path taken. At generation time the fixed future window is rebuilt from the realized history at every decision and only the selected movement is executed; on San Francisco the Markov-history variant reaches the destination in 94.98% of rollouts and selects STOP successfully in 94.92%.

The report covers origin-destination conditioned generation on Porto, Beijing and San Francisco using the public HOSER road graphs and official splits, evaluated with matched test conditions, replicate counts and generation budgets on distance and radius JSD, Hausdorff distance and EDR. It is aimed at researchers and engineers working on road-network trajectory generation, graph-constrained sequential decision making and urban mobility modeling. The source states no next-step experiment.

The anticipation horizon bounds only the explicitly represented future, not the total generated trajectory length; candidate-conditioned continuation predictions have no factual counterfactual continuation for unexecuted candidates during training and are compared against the same observed suffix. The San Francisco Markov-history variant has distance-distribution JSD of 2.68, above the full-history variant's 1.39, so the effect of history length varies by city. Generation is also bounded by a maximum transition budget, and exhausting it is recorded as a timeout rather than a model-selected termination.