Original title: Long-Horizon Consistent and Interaction-Aware World Models for Multi-Style End-to-End Driving
Yuxuan Han, Kunyuan Wu, Liyunong Yang, Zilu Wang +3 · September 2026
Self-driving cars can learn by imagining future scenarios in a simulated mental model, but these imagined futures often fall apart over long timeframes, miss interactions with other vehicles, and struggle with different driving styles. To fix this, researchers created StyleDrive. It keeps imagined futures steady by linking past states with new ones using a gated attention mechanism to stop errors from building up. It also separates details that matter directly to the car from irrelevant traffic noise, making choices clearer. Finally, it uses a technique called Group Relative Policy Optimization to compare whole paths instead of grading every single step, allowing the car to adapt to various driving styles without retraining. Tested on the Bench2Drive benchmark, StyleDrive achieved an 88.44 driving score and a 66.82 success rate, and successfully drove a real automated vehicle.
Imagine you are planning a week-long road trip. Each day, you check the weather for the next day, but the forecast for day three changes depending on what you did on day two. It's hard to plan because early mistakes throw off later predictions. In this paper, the world model is like that weather forecast for driving. The new system makes sure that predictions stay accurate over many steps, and it separates your car's actions from other drivers' moves, so planning is easier.
Self-driving cars must make safe decisions over long drives. This system helps them anticipate future events more accurately, which could reduce accidents and allow vehicles to adapt to different driving styles.
The abstract does not specify limitations, but it notes evaluation on the Bench2Drive closed-loop driving benchmark and deployment on a real automated guided vehicle platform, suggesting that broader real-world applicability may require further testing.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Long-Horizon Consistent and Interaction-Aware World Models for Multi-Style End-to-End Driving — Yuxuan Han, Kunyuan Wu, Liyunong Yang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.