ProgressiveDrive: Reconciling fidelity and consistency in world models for autonomous driving

World-model-based autonomous driving benefits from latent imagination, but BEV raster decoding must preserve decision-relevant spatial structures and temporal consistency during long-horizon rollouts. Existing compact world models commonly use efficient one-step decoders, which may provide insufficient BEV-level geometric and occupancy fidelity, while frame-wise reconstruction alone can leave cross-frame inconsistency unresolved. To address this problem, we propose ProgressiveDrive, a compact BEV world-model framework that improves decoder-side spatial fidelity and temporal stability. It introduces a progressive latent refiner that replaces one-shot decoding with a short deterministic refinement process, trained with BEV reconstruction and auxiliary denoising objectives to recover clearer lane, drivable-area, and occupancy structures without diffusion sampling at deployment. A GRU-based temporal adapter further aggregates recent latent states into sequence-aware conditioning to improve rollout consistency without modifying the RSSM dynamics core. On dense-traffic CARLA Navigation tasks with online-randomized routes, ProgressiveDrive achieves a success rate of 74.30%, reduces the collision rate to 10.51%, and lowers the out-of-lane rate to 8.25%. It also maintains strong BEV reconstruction fidelity and obtains the lowest Flicker and Jitter among the compared variants, showing that BEV-level reconstruction and temporal conditioning can improve closed-loop driving.