Qwen-Drive's 91.4 NAVSIM score is an oracle best-of-6 result
A Juejin AI technical audit finds that Qwen-Drive-1.0-RL's reported 91.4 PDMS score on NAVSIM v1.1 navtest comes from selecting the highest-scored trajectory among six candidates per scene using evaluation-time PDMS, while the default single trajectory scores 90.7. The analysis says this oracle result describes the potential of the candidate set rather than deployable selection ability, and it also examines evaluation denominators, time grids, data splits, and closed-loop evidence. The audit relied on official documentation, papers, and source code without running the model, simulator, scorer, or vehicle tests.
The distinction between an oracle benchmark result and a deployable selector is central to interpreting Qwen-Drive's reported performance and planning any deployment evaluation.
Qwen -Drive 的 91.4 分,究竟选中了什么?
一张成绩表里,Qwen-Drive-1.0-RL 的 NAVSIM 分数既是 90.7,也可以是 91.4。区别藏在后者旁边的 best-of-6:每个场景生成六条候选,再用该场景的 PDMS 评测分数选最高的一条。91.4 如实描述了这次评测,却没有证明车辆在当时可用的信息下,已经知道该选择哪条轨迹。1
这不是要抹去多候选结果。它能告诉研究者,候选 集合 里还有多少潜力;默认输出与择优输出的距离,也能帮助分辨生成和选择两个问题。但拿这个数字制定接入方案时,必须补上选择所用的信息、评分场景的分母、时间网格,以及结果来自哪种交互环境。否则数字没有抄错,工程判断仍可能做错。
下面沿官方评测文档和固定源码,把一个最高分还原成可比较的实验。本文引用的 模型 成绩均为官方报告;本次只核对论文、文档与代码,并做静态检查和显式时间算术,没有运行权重、数据集、评分器、仿真或车辆。
#Qwen-Drive-1.0-RL#Qwen-Drive#NAVSIM v1.1#PDMS#AlpaSim#PAI-AV-NuRec