RosettaSim Achieves State-of-the-Art Long-Term Traffic Simulation.

Interactive traffic simulation is a crucial component for developing and testing autonomous driving systems, acting as a vital world model. A significant challenge in creating long-horizon simulations is accurately modeling sustained multi-agent interactions, especially when agents dynamically enter and exit the scene, leading to fluctuating token cardinality. This research proposes that the solution lies in combining the architectural inductive biases and statistical priors of large-scale sequence models, such as Large Language Models (LLMs). Probing experiments revealed that LLMs, even small, heavily frozen ones, can rapidly adapt to traffic modeling due to the transferability of attention mechanisms and distributional consistency between motion tokens and natural language. Building on this insight, the researchers introduce RosettaSim, a unified framework that projects scene topology, agent states, and spawning intents into a structured autoregressive stream with variable length. This approach achieves both high short-term accuracy and stable long-horizon simulation fidelity. Furthermore, to address the challenge of evaluating extended rollouts where one-to-one agent correspondence fades, they introduce Retrieval-based Traffic Evaluation (RTE). RTE retrieves semantically similar real-world scenarios as context-aware reference anchors for evaluation. Experiments on the Waymo Open Sim Agent Challenge (WOSAC) demonstrate that RosettaSim achieves state-of-the-art performance in both short- and long-term simulation, and RTE shows a stronger correlation with standard metrics than existing approaches.