Strada-LLM: Graph LLM for traffic prediction

The integration of large language models (LLMs) into automated driving systems has enhanced vehicle trajectory prediction and lane-change intentions by transforming complex driving contexts into understandable representations. Despite their promising capabilities, the robustness of these LLM-based models in safety-critical environments remains unexamined. This study conducts a vulnerability analysis of LLM-enabled vehicle trajectory prediction, proposing a differential evolution attack that targets a single kinematic feature of surrounding vehicles.

Graph neural networks (GNN) have demonstrated significant potential across various fields, including materials science and social sciences. This study investigates the impact of batching algorithms on training time and model performance for GNNs, focusing on static and dynamic batching methods. Results indicate that optimizing the batching algorithm can yield speedups of up to 2.7 times, depending on factors such as data type, model architecture, batch size, hardware, and training steps.