LatentSightDrive: Progressive Foresight Internalization for Autonomous Driving

Future reasoning provides valuable foresight for autonomous driving, but not all future evidence is equally relevant to the current ego-trajectory decision. The challenge is to select useful future evidence and internalize it into the planner, so that its benefits no longer depend on the complete future-reasoning process at inference. We propose LatentSightDrive (LSDrive), a framework for Progressive Foresight Internalization following a Select-and-Internalize paradigm. LSDrive maps external future evidence and planner representations into a common foresight-slot space and applies Hierarchical Foresight Selection to guide their alignment. Scene-Adaptive Guidance allocates supervision according to planning difficulty and representation stability, while Planning-Relevant Foresight Scoring estimates each external slot's contribution to ego-trajectory prediction. The selected evidence shapes the planner's latent representation through weighted alignment, with training progressing from uniform bootstrapping to selective refinement. At inference, only the direct planner is retained. Experiments on NAVSIM and nuScenes demonstrate improvements over uniform transfer across latent-world and visual-CoT settings. Compared with explicit visual-CoT planning on nuScenes, LSDrive reduces average displacement error and collision rate by 9.3% and 31.7%, respectively, while reducing per-sample generation time by 31.7%.