Traffic forecasting drops step-by-step sampling: CoReFlow emits a full forecast in one pass and wins 11 of 12 point-metric cases across four road networks
Synopsis
Probabilistic traffic forecasting has relied on multi-step diffusion sampling, which is slow and often trails deterministic baselines on point accuracy. CoReFlow learns the flow map, the solution operator of the probability-flow ODE, directly and relaxes trajectory consistency through the conditional mean, so a single forward pass produces the full forecast. On four real road-network datasets, PEMS03, PEMS04, PEMS07 and PEMS08, it ranks first in 11 of 12 deterministic metric cases and takes all 14 reported comparisons against generative baselines on PEMS07 and PEMS08.
Interpretation
Probabilistic traffic forecasting can now produce a complete forecast in a single forward pass instead of repeatedly calling the network along a diffusion trajectory. Earlier diffusion and flow-matching generative models captured uncertainty through multi-step iterative sampling, which is slow and often falls short of the strongest deterministic baselines on point accuracy. On four real road-network datasets, PEMS03, PEMS04, PEMS07 and PEMS08, CoReFlow ranks first in 11 of 12 deterministic metric cases; the sole exception is RMSE on PEMS03, where HyperD ranks first.
The training objective no longer demands whole-trajectory self-consistency; it restricts consistency through the conditional mean, stabilizing optimization when the conditional distribution is close to unimodal. Existing flow-map training objectives such as self-distillation and MeanFlow Identity preserve trajectory consistency to keep multimodal structure, and bring gradient instability and early-training error accumulation. On PEMS07 and PEMS08, swapping in MeanFlow Identity, improved MeanFlow Identity or the self-distillation objective raises prediction error, and replacing node-level adaptive weighting with batch-level weighting consistently degrades accuracy.
In generative comparisons, CoReFlow attains the lowest RMSE, MAE and MAPE under both one-sample and ten-sample aggregation, and the best CRPS on both datasets. Generative baselines DiffSTG, USTD and SpecSTG show higher errors on the same two road networks. CoReFlow is best in all 14 reported comparisons on PEMS07 and PEMS08; inference needs one network function evaluation per trajectory, whereas SpecSTG needs 600.
Perspective
The demonstrated setting is four highway flow datasets with one hour of input and one hour of forecast at five-minute resolution; the authors explicitly limit the conclusion to regimes where the conditional future is close to unimodal and plan to extend the relaxed flow map to richer multimodality and longer horizons.
The authors state that the objective trades a small approximation bias for stable supervision, with the largest benefit when the conditional future is close to unimodal; evaluation covers only highway flow data over a one-hour horizon, and other domains may exhibit stronger multimodality.