OmniTraffic: A Controllable Generation Pipeline and Benchmark for Spatio-Temporal Traffic Reasoning

OmniTraffic is a newly introduced controllable generation pipeline and benchmark designed for spatio-temporal traffic reasoning, featuring 12 reconstructed 3D traffic environments and surveillance footage from two countries. The benchmark includes a three-level task hierarchy for scene perception, multi-view and temporal reasoning, and decision support, generating 8 million visual question-answering samples and a human-verified test set of 3,000 instances. Evaluation of eleven frontier multimodal large language models (MLLMs) reveals significant performance gaps, particularly in topology-grounded and spatio-temporal reasoning tasks, highlighting the potential of simulation-generated data for improving traffic-specific multimodal reasoning capabilities.