OSWAM cuts driving world-model inference to one step at 40ms

OSWAM, a new world action model for autonomous driving and embodied control, generates driving actions in a single diffusion step and runs with compiled inference latency as low as 40ms on an NVIDIA RTX 5090, according to the arXiv preprint. The authors report roughly a 3x speedup over direct-policy WAM baselines, which they say currently need "tens of denoising steps to achieve competitive action quality."

The trick is initialization. Instead of starting the diffusion process from random noise, the paper constructs what it calls "imagination tokens to initialize the diffusion process, providing physical priors and action-conditioned representations," so the model does not have to iteratively denoise a coherent action sequence out of nothing. The authors pair this with "privileged information distillation with future conditioning" during training. The claim is that OSWAM "can efficiently and effectively generate high quality actions in just one diffusion step."

The abstract publishes no per-benchmark scores, no breakdown by direct-policy versus Joint-WAM comparator, and no evaluation outside the RTX 5090 setup. The quality claim is phrased as "comparable or even better results in both autonomous driving and embodied control" rather than specific deltas against named baselines.