Discrete diffusion takes the wheel: MD-Drive reaches 88.9 PDMS on NAVSIM with 153K training samples
Synopsis
Autonomous-driving vision-language models have mostly generated trajectories token by token, letting errors accumulate along the sequence. MD-Drive discretizes trajectory coordinates into tokens on a standard vision-language architecture and denoises the whole sequence with masked discrete diffusion, adding trajectory discretization in supervised fine-tuning, diversity filtering in reinforcement learning, and a decoding strategy that switches direction with scene risk at inference. On the NAVSIM closed-loop benchmark, MD-Drive reaches 88.9 PDMS, 6.5 and 3.8 points above the same-architecture autoregressive baselines DrivingGPT and FSDrive, using 153K training samples.
Interpretation
An autonomous-driving vision-language model can stop generating trajectories token by token and instead plan the whole trajectory at once as a sequence of tokens to be denoised. Earlier autoregressive vision-language models generate step by step through next-token prediction, so inference depends on sequential generation and easily accumulates errors, limiting long-horizon planning in complex scenes. On the NAVSIM closed-loop benchmark, the discrete-diffusion MD-Drive reaches 88.9 PDMS, against 82.4 for the same-architecture autoregressive baseline DrivingGPT and 85.1 for FSDrive.
Each 2D waypoint coordinate is mapped into a 1D codebook, the full trajectory becomes a flattened token sequence, and a masked diffusion model recovers it by iterative denoising. Supervised fine-tuning replaces text serialization of continuous coordinates with trajectory discretization to suppress noise from redundant tokens; reinforcement learning selects a checkpoint balancing pass@1 and pass@8 accuracy and diversity as its starting point and keeps only mid and hard samples. On NAVSIM, the continuous baseline without discretization scores 68.5 PDMS, discretization raises it to 78.2, and the setting with vocabulary size 0.25 and quantization error 0.125 reaches 81.4.
Perspective
The results cover the NAVSIM closed-loop benchmark and the nuScenes and DriveCoT open-loop datasets, with DreamVL as the base model trained in bf16 on four NVIDIA H800 GPUs; this is the demonstrated evaluation scope, and the material reports no real-vehicle or broader road-scene validation.
The material notes that driving-behavior annotations in DriveCoT are themselves ambiguous, since several actions may be valid in one scene, and partly attributes the weighted-F1 gap between MD-Drive and DriveCoT-Agent to this; checkpoint choice also depends on multi-seed pass@8 stability, and a checkpoint with lower average PDMS performed better after GRPO.
