HiThink Research
Institute for AI Industry Research (AIR), Tsinghua UniversityThe Hong Kong University of Science and Technology (Guangzhou)

SuperGlance: Driving at a Glance with a Vision-Language Model

Jieyuan Pei1,2,3,*Xin Yan4,*Mingwei Xu5Haiyi Zhang3Siyu Zou2Zhangyi Hu2,3Zhuoxuan Li6Yuguang Yang7Rongjunchen Zhang2,†Cheng Chi6Hui Xiong3,✉Yan Wang1,✉
1Institute for AI Industry Research (AIR), Tsinghua University2HiThink Research3The Hong Kong University of Science and Technology (Guangzhou)4Agency for Science, Technology and Research (A*STAR), Singapore5University of Washington6Renmin University of China7Beihang University
*Equal contribution†Project leader✉Corresponding authors

Checkpoints: SuperGlance-9B (final) · SuperGlance-9B-Stage1 · SuperGlance-9B-Stage2

SuperGlance and benchmark results
SuperGlance plans in one forward pass, 6.2 times faster than Qwen-Drive-1.0, and leads it on ten of eleven QA and all driving benchmarks; ring bars start above zero and longer ADE bars mean lower error.

Video

Abstract

Autonomous driving is turning to vision-language-action models (VLAs). They answer and plan by decoding, one step at a time. Each step is another forward pass, so a plan can take over a second of GPU time. Embodied AI often pairs such a slow, reasoning System 2 with a fast System 1 that acts. Human drivers plan at a glance, with System 1. Faster designs usually put System 1 outside the vision-language model (VLM), in a planner without its abilities. Yet before a VLM writes its first token, it has already read the scene into one hidden state. We show that, once trained, this hidden state is enough to drive and to answer questions about what the model sees. SuperGlance reads it to turn a decoder-only VLM into an encoder-style driving policy. Its linear head scores a fixed set of 5,273 candidate trajectories. With the set fixed, rewards are computed offline, so supervised targets can be reward-weighted and reinforcement learning maximizes the exact expected reward. Before and during policy training, multiple-choice training keeps the VLM’s visual abilities in one pass, so the policy still answers questions about driving scenes. Training questions are built against shortcuts. Each must be answered right in two option orders. Against drift, a KL term keeps each stage near its starting model. One pass offers VLAs a more plausible path to System 1. SuperGlance answers visual multiple-choice questions as well as its base VLM does by decoding. It plans in 84 ms of GPU time, six times faster than Qwen-Drive-1.0, and at least as well on NAVSIM, PAI-AV and WOD-E2E. In AlpaSim closed loop, km per at-fault incident holds for both planners up to a 300-ms delay and drops beyond. With GPU time charged as delay, SuperGlance’s lead over Qwen-Drive-1.0 grows from 2.8 to 3.9 times. Delayed by 1.2 s, our model falls below Qwen-Drive-1.0 at its own 518 ms.

Method

SuperGlance turns a decoder-only vision-language model, Qwen3.5-9B, into an encoder-style driving policy that plans in one forward pass. A linear Policy Head on the last hidden state at the answer position, [ANS], scores a fixed set of 5,273 candidate trajectories, and the language-model head still answers multiple-choice questions. As the set is fixed, every candidate’s reward is computed offline, so supervised targets can be reward-weighted and reinforcement learning maximizes the exact expected reward. Multiple-choice training before and during policy training keeps the VLM’s visual abilities in one pass.

Three ways a vision-language model produces a plan
Three ways a vision-language model produces a plan. Only (c) has no loop: SuperGlance reads the last hidden state h once, at [ANS], and its Policy Head gives a probability to each of 5,273 fixed candidates. Each bottom strip shows one plan’s measured GPU time to scale; in (a), our VLM writing text runs off its axis.

Results

Time per plan and distance in closed loop
Time per plan and distance in closed loop. One pass plans in 84 ms of GPU time, a 10-step flow expert in 518 ms and the same VLM writing text in 5,558 ms (top); each time (dotted) acts as the next 100-ms lag in AlpaSim (bottom). Both planners keep their distance per at-fault incident within 8% up to a 300-ms lag; at 1.2 s ours falls to 0.259 km, below Qwen-Drive-1.0 at its own 518 ms. At 5 s every plan arrives after its last pose (hollow): both planners drive more slowly and vehicles behind hit them, which is not at fault and ends the scenario as a pass.
MethodRLColl.↓Off-road↓Prog.↑km / incident↑Pass↑
Each plan’s GPU time as delay
Qwen-Drive-1.0, 518 ms32.6/18.119.367.90.229/0.31736.6
 + reasoning, 1.2 s26.7/21.036.351.80.145/0.1589.3
SuperGlance, 84 ms✓10.9/6.87.588.50.970/1.24872.2
No delay (simulator default)
Qwen-Drive-1.032.8/14.415.471.10.284/0.46158.7
Qwen-Drive-1.0✓43.6/15.613.157.30.194/0.38353.6
SuperGlance11.2/6.08.887.90.886/1.19972.1
SuperGlance✓10.6/5.97.688.60.975/1.31873.5
AlpaSim closed loop. With each plan’s GPU time as its delay, our RL model drives 3.9 times as far per at-fault incident as Qwen-Drive-1.0 (95% interval 3.3 to 4.7) and passes 72.2% of scenarios, against at most 36.6%. Qwen-Drive-1.0 plans with reasoning except in its 518 ms row. Coll. and km / incident: all / at-fault; Coll., off-road, progress and pass in %. Shading runs from the worst to the best value of each column within a group.
Qwen3.5-9BSuperGlance after stage
BenchmarkletterdecodedQwen-Drive-1.0123
Vision
CV-Bench86.3686.4785.2987.6087.6687.40
BLINK79.3776.6277.6084.6881.3479.37
3DSRBench45.3957.0946.8160.2760.5058.29
Language
MMLU75.9185.2170.6979.7979.6378.61
MMLU-Pro50.4173.9344.3559.0357.9856.85
ARC-C93.7796.1691.3895.3995.1495.31
HellaSwag88.6789.4981.8391.0091.0490.15
WinoGrande72.5385.6467.4076.4077.4376.40
BoolQ88.6288.2987.6589.7289.4889.42
TruthfulQA74.4276.2572.9567.0766.2265.61
Driving
DriveLM60.8160.9565.0673.5584.3385.51
Question answering. Accuracy (%) after each training stage, against our base model and Qwen-Drive-1.0’s 4B VLM, all read as one option letter except the decoded column (Qwen3.5-9B’s written replies, thinking off). After stage 3, our letter is level with or above those replies on every vision benchmark and 24.56 points above them on DriveLM, the only benchmark in our training data. Shading runs from the worst to the best score of each row.

BibTeX

@article{pei2026superglance,
  title   = {{SuperGlance}: Driving at a Glance with a Vision-Language Model},
  author  = {Pei, Jieyuan and Yan, Xin and Xu, Mingwei and Zhang, Haiyi and Zou, Siyu and Hu, Zhangyi and Li, Zhuoxuan and Yang, Yuguang and Zhang, Rongjunchen and Chi, Cheng and Xiong, Hui and Wang, Yan},
  journal = {arXiv preprint arXiv:2611.19138},
  year    = {2026}
}