Can self-driving cars plan fast enough while keeping options open?

End-to-end autonomous driving needs both diverse maneuvers and sub-100ms inference. CLEAR swaps diffusion's slow iterative denoising for a single-step latent sampling, then uses a fine-tuned language model to read the scene and pick between preset diversity levels. On NAVSIM, it scores 93.7 PDMS—matching rich behavior generation without the real-time penalty.