Qwen-Drive-1.0-RL achieved 91.4 points on the public NAVSIM v1.1 leaderboard, but that number did not come from a vehicle outputting a single trajectory. For each scenario, the system first generated six candidates, and the evaluator then selected the highest-scoring one under the PDMS rules. The result is real, but its meaning needs to be narrowed.

What this is

Qwen-Drive is a planning model from Alibaba’s Tongyi Qianwen designed for autonomous driving scenarios. NAVSIM is a benchmark that evaluates a vehicle’s planning trajectories using simulation data, while PDMS combines multiple metrics to assess trajectory quality comprehensively.

On the same leaderboard, Qwen-Drive scored 90.7 by default with a single trajectory after supervised fine-tuning (SFT, learning from human demonstrations) and reinforcement learning (RL, further optimization through reward signals). Enabling six candidates and post-hoc selection raised the score to 91.4, a difference of 0.7 points.

What matters here is not simply whether the model “can drive,” but the distinction between two capabilities: whether it can generate a potentially good trajectory, and whether it can select one at the time of driving using only the information then available. An evaluator that knows the ground truth and picks the highest-scoring result is engaging in post-hoc selection. This measures the upper bound of the candidate set; it does not prove that the vehicle has acquired such judgment.

Industry view

From a research perspective, the six-candidate score remains valuable. The gap between default output and selected output can help teams separate generation quality from selection quality, while providing a target for training a selector. Alibaba’s team improved the score from 87.8 to 91.4 through a combination of training methods, inference conditions, and selection budget; the gain cannot be attributed entirely to “reasoning capability.”

We do not support using the 91.4 score directly to define a production integration plan. The public leaderboard does not deliver a selector that reads only contemporaneous observations and chooses the best trajectory under real-road conditions, nor does it provide statistical significance. In the source code, different metrics may also evaluate different candidate trajectories. More importantly, the number of scenarios included in the aggregate, missing caches, and the time grid can all affect reproducibility. The article also explicitly notes that its authors only reviewed the paper, documentation, and code; they did not run the model, simulator, or vehicle. We therefore cannot conclude from this analysis that the official results are wrong.

Our judgment is that Qwen-Drive’s progress should be understood as “strong candidate generation capability,” not “real-time trajectory selection solved.” Moving from a high benchmark score to deployment on real vehicles still requires a deployable and verifiable selection mechanism.

Impact on regular people

For enterprise IT: When evaluating intelligent-driving or Agent-class systems, enterprises should require vendors to distinguish default output, candidate-set upper bounds, and real selection capability. They should also require disclosure of effective sample sizes and information boundaries, so that offline best-of-N results are not mistaken for online performance.

For individual professionals: This distinction can also be used to evaluate other AI products. Generating multiple answers and having the backend choose one is not the same, either commercially or in terms of cost structure, as a system that produces a usable answer on its first attempt.

For the consumer market: If similar methods reach passenger vehicles, what consumers ultimately experience will still depend on whether the vehicle can make the correct choice in time. Increasing the number of candidates also means greater compute and energy requirements; a score of 91.4 does not automatically translate into a more reliable driving experience.