We introduce Latent-WAM++, a representation-centered World Action Model (WAM) for end-to-end autonomous driving. Existing WAMs are primarily trained to predict future states or final actions, leaving the intermediate world and action representations that connect prediction to planning only indirectly constrained. Latent-WAM++ matches the supervision of each representation to its role in planning. World Representation Construction combines scene-ego future prediction, SIGReg-based distribution stabilization, and frozen geometric priors to learn a temporally predictive, non-degenerate, and geometry-aware world representation without an EMA target encoder. Trajectory MeanFlow Planner directly predicts multimodal trajectory proposals while being optimized with an Improved MeanFlow velocity-space objective, enabling one-step sampling. Visual-Action Representation Alignment further aligns the resulting pre-output action features with a stop-gradient target derived from current visual semantics. Latent-WAM++ achieves state-of-the-art performance on NAVSIM v2 and under zero-shot evaluation on HUGSIM, reaching 90.1 EPDMS on NAVSIM v2 and 47.0 average RC and 30.1 average HD-Score on HUGSIM. Training the planner end to end costs 190 GPU-hours for Latent-WAM++ versus 1,600 for Latent-WAM, a reduction of 88.1%.
Figures & tables
Figure 1: NAVSIM v2 performance and the cost of training planners end to end under identical hardware and data-loading settings. Our planner improves EPDMS by 0.8 over Latent-WAM ( 90.1 vs. 89.3 ) and costs 88.1% less to train end to end ( 190 vs. 1,600 GPU-hours).
Figure 2: Overview of Latent-WAM++. (a) WoRC constructs the world representation through scene prediction, distribution regularization, and additional geometric context. (b) TrajMF generates multimodal trajectory proposals in one step and aggregates them into the final trajectory. (c) VARA uses stop-gradient visual features to encourage semantic consistency between the trajectory-predictive action representation and the driving scene.
Method
NC ↑
DAC ↑
DDC ↑
TLC ↑
EP ↑
TTC ↑
LK ↑
HC ↑
EC ↑
EPDMS ↑
TransFuser ( Chitta et al. 2023 )
96.9
89.9
97.8
99.7
87.1
95.4
92.7
98.3
87.2
76.7
DriveSuprim ( Yao et al. 2026 )
97.5
96.5
99.4
99.6
88.4
96.6
95.5
98.3
77.0
83.1
ReCogDrive ( Li et al. 2026 )
98.3
95.2
99.5
99.8
87.1
97.5
96.6
98.3
86.5
83.6
DiffusionDrive ( Liao et al. 2025 )
98.2
95.9
99.4
99.8
87.5
97.3
96.8
98.3
87.7
84.5
World4Drive ( Zheng et al. 2025 )
97.8
96.3
99.4
99.8
88.3
97.1
97.7
98.0
53.9
84.8
DriveVLA-W0 ( Li et al. 2025b )
98.5
99.1
98.0
99.7
86.4
98.1
93.2
97.9
58.9
86.1
Table 1: Comparison on NAVSIM v2 ( Dauner et al. 2024 ; Cao et al. 2025 ) . Best and second-best results are bold and underlined.
RC ↑
HD-Score ↑
Method
E
M
H
X
Avg.
E
M
H
X
Avg.
UniAD ( Hu et al. 2023 )
58.6
41.2
40.4
26.0
40.6
48.7
29.5
27.3
14.3
28.9
VAD ( Jiang et al. 2023 )
38.7
27.0
25.5
23.0
27.9
24.3
9.9
10.4
8.2
12.3
LTF ( Chitta et al. 2023 )
68.4
40.7
36.9
25.5
41.4
52.8
24.6
19.8
8.1
24.8
GTRS-Dense † ( Li et al. 2025c )
64.2
50.0
20.7
22.3
38.0
55.5
39.0
11.7
14.3
28.6
MeanFuser ( Wang et al. 2026a )
76.7
41.1
37.9
28.4
43.9
67.1
24.3
21.5
11.0
28.4
Table 2: Zero-shot closed-loop evaluation on HUGSIM ( Zhou et al. 2026 ) across easy (E), medium (M), hard (H), and extreme (X) settings. † : GTRS-Dense is evaluated on 345 scenarios.
Figure 3: Cross-scene visual-token statistics without stabilization, with EMA, and with SIGReg. (a) Mean feature standard deviation per token identity; (b) Empirical CDF of same-token cosine similarities across scenes. SIGReg produces the largest standard deviation and lowest similarity.
Figure 4: NAVSIM v2 trajectory comparison. Yellow and green denote model and human-driver trajectories. Latent-WAM++ follows the reference more consistently than World4Drive and Latent-WAM in the displayed scenes.
Figure 5: Action representations w/o and w/ VARA. Colors are K=8 clusters from frozen DINOv2 CLS features.
ID
DWM
Stabilizer
Occ
Geometry Distill.
Decoder
VARA
EPDMS ↑
Δ EPDMS
1
×
×
×
×
MLP
×
88.2
–
2
✓
×
×
×
MLP
×
88.0
−0.2
3
✓
EMA \mbox+86.6M
×
×
MLP
×
88.6
+0.4
4
✓
SIGReg
×
×
MLP
×
88.8
+0.6
5
✓
SIGReg
✓
×
MLP
×
89.1
+0.9
6
✓
SIGReg
✓
×
TrajMF
×
89.5
+1.3
Table 3: Unified ablation of world-representation learning, action decoding, and geometric supervision. IDs index configurations; the EMA subscript reports additional parameters, and Geometry Distill. denotes feature alignment to the frozen geometry teacher. Δ EPDMS is measured relative to configuration 1 (88.2).
Decoder
Pred.
Loss
EPDMS ↑
Iter. ↓
MLP
x
x
89.1
16,160
iMF
u
v
89.4
16,160
TrajMF
x
v
89.5
8,080
Table 4: Action-decoder ablation under the same training recipe with all other modules fixed. iMF and TrajMF differ only in their network prediction targets: iMF predicts u , whereas TrajMF predicts x . Iter. denotes the training iteration attaining the reported EPDMS.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Training split
Test split
Semantic class
Voxel count
Share (%)
Voxel count
Share (%)
background
41,531,592,449
92.6545
5,164,837,360
94.8239
vehicle
2,899,332,761
6.4682
255,186,420
4.6851
pedestrian
208,170,168
0.4644
12,563,558
0.2307
generic_object
143,724,597
0.3206
8,735,890
0.1604
traffic_cone
18,276,533
0.0408
1,810,800
0.0332
Appendix
Table S1: Semantic-class distribution of occupied NAVSIM-Occ voxels. Percentages are computed within each split; free denotes unoccupied space.
Figure S1: Projected NAVSIM-Occ semantic occupancy annotations for the left, front, and right camera views; colors identify classes.
Figure S2: Qualitative trajectory comparison among World4Drive ( 8 ) , Latent-WAM, and our method. Yellow curves denote model-planned trajectories, whereas green curves denote human-driver trajectories.
Figure S3: HUGSIM trajectory-planning visualizations on nuScenes and KITTI-360. Yellow trajectory points denote the trajectories planned by our model.
Figure S4: HUGSIM trajectory-planning visualizations on Waymo and PandaSet. Yellow trajectory points denote the trajectories planned by our model.