Vision-Language-Action (VLA) models have recently achieved notable progress in end-to-end autonomous driving by integrating perception, reasoning, and control within a unified multimodal framework. However, they often lack explicit modeling of temporal dynamics and global world consistency, which limits their foresight and safety. In contrast, world models can simulate plausible future scenes but generally struggle to reason about or evaluate the imagined future they generate. In this work, we present VLA-World, a simple yet effective VLA world model that unifies predictive imagination with reflective reasoning to improve driving foresight. VLA-World first uses an action-derived feasible trajectory to guide the generation of the next-frame image, capturing rich spatial and temporal cues that describe how the surrounding environment evolves. The model then reasons over this self-generated future imagined frame to refine the predicted trajectory, achieving higher performance and better interpretability. To support this pipeline, we curate nuScenes-GR-20K, a generative reasoning dataset derived from nuScenes, and employ a three-stage training strategy that includes pretraining, supervised fine-tuning, and reinforcement learning. Extensive experiments demonstrate that VLA-World consistently surpasses state-of-the-art VLA and world-model baselines on both planning and future-generation benchmarks. Project page: https://vlaworld.github.io
Figures & tables
Figure 2: Comparison of the (a) VLA, (b) World Model, and (c) the proposed VLA-World paradigm.
Figure 3: The illustration of the three-stage training and inference pipeline of VLA-World. Our training pipeline consists of three key stages: (a) visual pretrain to activate generation capability, (b) supervised fine-tuning to seed conceptual knowledge to the model, and (c) reinforcement learning with GRPO to explore higher performance like humans.
Method
ST-P3 metrics
UniAD metrics
LLM
L2 (m) ↓
Collision (%) ↓
L2 (m) ↓
Collision (%) ↓
1s
2s
3s
Avg.
1s
2s
3s
Avg.
1s
2s
3s
Avg.
1s
2s
3s
Avg.
Non-Autoregressive methods
ST-P3* [ECCV22] [ 23 ]
1.33
2.11
2.90
2.11
0.23
0.62
1.27
0.71
-
-
-
-
-
-
-
-
-
VAD [ICCV23] [ 29 ]
0.69
1.22
1.83
1.25
0.06
0.68
2.52
1.09
-
-
-
-
-
-
-
-
-
VAD* [ICCV23] [ 29 ]
0.17
0.34
0.60
0.37
0.04
0.27
0.67
0.33
-
-
-
-
-
-
-
-
-
Table 1: End-to-end trajectory planning results on nuScenes [ 10 ] . We evaluate L2 error and collision rate following the respective protocols of ST-P3 [ 23 ] and UniAD [ 25 ] . The ∗ indicates the use of additional ego-state information. Results for VAD [ 29 ] and UniAD [ 25 ] are taken from BEV-Planner [ 38 ] .
Method
DriveGAN
DriveDreamer
Drive-WM
GenAD
GEM
Doe-1
FSDrive
VLA-World
[CVPR21 [ 31 ] ]
[ECCV24 [ 59 ] ]
[CVPR24 [ 61 ] ]
[CVPR24 [ 69 ] ]
[CVPR25 [ 21 ] ]
[arxiv24 [ 80 ] ]
[NeurIPS25 [ 74 ] ]
Type
GAN
Diffusion
Diffusion
Diffusion
Diffusion
Autoregressive
Autoregressive
Autoregressive
Resolution
256 × 256
128 × 192
192 × 384
256 × 448
576 × 1024
384 × 672
128 × 192
128 × 192
FID ↓
73.4
52.6
15.8
15.4
10.5
15.9
10.1
9.8
Table 2: Comparison of future frame generation results with different generative models evaluated by FID ( ↓ ) on the nuScenes [ 10 ] dataset.
Figure 4: Visualization of our VLA-World compared with the SOTA FSDrive [ 74 ] . The top row shows the current scene, and the bottom row shows the generated scene for the next 0.5 seconds. The red trajectory is the model prediction, and the green trajectory is the ground truth. Zoom in for a better view.
Method
Lateral ( F1 )↑
Longitudinal ( F1 )↑
forward
left
right
keep
acc.
dec.
stop
Qwen2-VL-2B
62.43
22.75
28.65
40.70
50.23
49.21
41.04
Qwen2-VL-2B †
92.60
61.78
66.52
56.42
74.32
76.10
74.85
VLA-World
95.88
74.22
75.06
60.98
81.42
80.04
81.24
Table 3: Action prediction performance on the nuScenes dataset measured by F1 score (%). † denotes models trained on nuScenes.
Method
L2 Error (m) ↓
1s
2s
3s
Avg.
(a)
w/o. P.T.
0.35
0.56
0.81
0.57
w/o. SFT
0.35
0.79
1.40
0.85
w/o. RL
0.43
0.70
1.01
0.71
(b)
w/o. Perception
0.42
0.73
1.09
0.75
w/o. Generation
0.41
0.67
0.96
0.68
Table 4: Ablation studies of trajectory planning L2 errors (ST-P3) on nuScenes to validate each proposed component.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Data sample of (a) pretraining stage, (b) supervised fine-tuning, and (c) reinforcement learning.
Figure 6: Comparison between our VLA-World and the state-of-the-art FSDrive [ 74 ] on generating the future frame at next 0.5 seconds.
Res.
L2 Error (m) ↓
1s
2s
3s
Avg.
36000
0.03
0.14
0.98
0.38
52884
0.11
0.27
0.52
0.30
Appendix
Table 5: Evaluation of trajectory planning L2 errors (ST-P3) on nuScenes with varying input view resolutions.
Figure 7: Comparison of 3-second future trajectory predictions generated by our VLA-World and the state-of-the-art FSDrive [ 74 ] . Zoom in for a better view.
Method
L2 Error (m) ↓
1s
2s
3s
Avg.
Qwen2-VL-2B
0.11
0.27
0.52
0.30
Qwen2.5-VL-3B
0.05
0.08
0.76
0.29
Qwen2-VL-7B
0.03
0.03
0.47
0.18
Appendix
Table 6: Evaluation of trajectory planning L2 errors (ST-P3) on nuScenes with varying model sizes.
Method
L2 Error (m) ↓
1s
2s
3s
Avg.
w/o. Mixed
0.27
0.47
0.73
0.49
Qwen2-VL-2B
0.11
0.27
0.52
0.30
Appendix
Table 7: Evaluation of trajectory planning L2 errors (ST-P3) on nuScenes with training strategy.