Researchers have developed a new framework called MAVEN-T for real-time multi-agent trajectory prediction in autonomous driving systems. The framework uses a high-capacity teacher model to predict the future movements of multiple agents and then distills this information into a smaller student model that can run on limited on-board computation. This approach improves accuracy while reducing inference costs, making it suitable for real-time deployment. Experiments show that the student model achieves significant parameter compression and inference acceleration compared to existing methods. --- Why it matters: This matters to engineers working on autonomous driving systems because it provides a more efficient way to predict the movements of multiple agents in real-time, which is crucial for collision checking, behavior planning, and control. Source: https://arxiv.org/abs/2604.10169