SimWAM scores 91.5 on NAVSIM, then discards its video model

SimWAM scored 91.5 on NAVSIM's PDMS driving-planning benchmark by training a full video-prediction model alongside its action planner. Then it deletes the video half entirely, per the paper on arXiv. The result beats existing WAM planners on the same benchmark. It also runs at lower latency, because the surviving half never renders a video frame to make a decision.

For production driving stacks, every millisecond of inference latency is a hardware cost and a safety constraint, paid on every frame the car sees. A planner that matches the benchmark score without a video branch to run is one fewer expensive step between camera input and a driving decision.

The trick is architectural, not pruning after the fact. SimWAM co-trains the video expert and the action expert with joint flow matching. An isolated attention mask keeps the action side from ever looking at future frames while it learns. Once training ends, the model doesn't need the video branch to plan, so it gets cut. The resulting planner also transferred zero-shot to nuScenes, a separate driving dataset, without retraining.

If SimWAM's result holds, most world-model driving planners are burning inference latency on video prediction that only ever helped during training. Worth watching whether other WAM systems can strip their video branch this cleanly, or whether SimWAM's masking trick is hard to copy.

Code and trained weights shipped with the paper, so the latency and zero-shot numbers are checkable directly, not just claimed.