A research paper out of arXiv proposes a new architecture for autonomous driving that scores 77.28% on a closed-loop benchmark — the best reported result on that test.
Current vision-language-action (VLA) planners for self-driving take one of two approaches: predict trajectories through a planning head that only loosely ties language reasoning to motion, or generate full trajectories coordinate-by-coordinate in long token sequences. Both have problems. The first produces weak semantic-to-action alignment; the second is slow, error-prone, and buries meaning in low-information coordinate tokens. AnchorVLA sidesteps both by inserting a middle layer — trajectory-pattern anchors — that represent whole driving maneuvers rather than individual waypoints. The language model reasons over those compact anchor tokens, then a residual flow module fills in the fine-grained trajectory within that anchor's motion space.
The anchor idea matters because it preserves what large language models are actually good at — high-level commonsense reasoning about traffic and intent — without forcing them to micromanage geometry. That separation of concerns is the same principle behind hierarchical planning in robotics broadly, and applying it here produces measurable gains: a Success Rate of 77.28 and a Driving Score of 89.92 on Bench2Drive, a closed-loop benchmark that runs simulated end-to-end driving rather than just scoring static predictions.
Closed-loop benchmarks are harder to game than open-loop ones, so the result carries more weight than typical leaderboard numbers — though lab-to-road transfer remains the field's unsolved problem, and a benchmark score is not a safety certification.