Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features

Sining Ang; Yan Wang; Yuguang Yang

REVIEW 4 major objections 5 minor 19 references

Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A quality-gated planner can exit a video diffusion model early and still drive well.

desk verdict A well-executed systems paper with a genuinely new depth-vs-noise diagnostic and a useful adaptive multi-exit planner; the headline numbers are plausible, but a missing noise-seed variance analysis leaves the deployed result partly pinned to an uncontrolled random draw. read the letter →

arxiv 2608.06008 v1 pith:EHVU2BMV submitted 2026-08-06 cs.RO

Sining Ang , Yuguang Yang , Yan Wang This is my paper

classification cs.RO

keywords early-exitplanningvideodiffusionmodelsworld-actionautonomousdrivingtrajectoryqualityscoringadaptiveinferenceDiTfeaturesNAVSIM

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks how much of a video diffusion model a driving policy really needs to run. It finds that varying the video-noise level barely changes planning scores, while the depth of the transformer matters a lot: strong trajectories can be decoded from middle layers, and the best middle layer beats the full-depth exit. Based on that, it builds Adaptive-WAM, which attaches trajectory decoders to six intermediate blocks of a Wan2.2-5B video model and uses a lightweight quality scorer to stop as soon as a decoded trajectory looks good. The deployed planner skips the iterative denoising loop and video decoding entirely, averaging 170 ms on an A100 while reaching 90.8 PDMS on NAVSIM v1, 89.9 EPDMS on NAVSIM v2, and 0.88 m average L2 error with 0.08% collisions on nuScenes zero-shot.

What carries the argument

The central object is a multi-exit world-action model built on the Wan2.2-5B video DiT: six independent ReCogDrive-style five-step trajectory diffusion heads are attached to blocks 5, 9, 15, 18, 22, and 30, all reading from the same conditional forward at fixed video-noise index 17. A fine-tuned DINOv2-Small scorer with six component heads predicts NAVSIM sub-scores for each decoded trajectory, and a controller stops at the first exit whose predicted quality $\hat{q}_j$ reaches a threshold $\eta$, otherwise reuse of cached hidden states continues to the next exit. This avoids the unconditional classifier-free-guidance branch, the iterative DDIM denoising loop, and VAE video decoding, reducing end-to-end planning to roughly 170 ms.

What would settle it

Run the adaptive planner on NAVSIM v1 with the same seed but ten different random draws for the future-latent noise at index 17 and compute the PDMS spread; if the spread exceeds the 0.17-point improvement of adaptive routing over the fixed block-15 exit, the reported gain could be a noise artifact. Additionally, sweep the noise index across all 40 steps, not just the five tested indices, and check whether the 0.13-0.15 point plateau persists.

Watch

Extended reading notes

Core claim

The central discovery is that a single conditional forward through a video DiT, at a fixed noise index and without any denoising, yields intermediate hidden states from which a five-step trajectory diffuser can decode planning-quality trajectories. Across the five tested noise indices, planning scores vary by at most 0.15 points, whereas across the six exit depths they vary by more than six points; block 15 is the strongest fixed exit, yet no depth dominates every scene. Adaptive-WAM therefore reads one trajectory per attempted exit, keeps the best candidate, and continues to deeper blocks only when a lightweight DINOv2-based scorer predicts the quality threshold is not met. With this routing, it improves PDMS from 90.62 (best fixed exit) to 90.79 while cutting latency from 190 ms to 170 ms, and a fixed block-22 variant with 64 proposals reaches 92.6 PDMS.

Load-bearing premise

The load-bearing premise is that one conditional Wan forward, with the future-latent slice seeded from scheduler noise at index 17 and no denoising or noise averaging, yields hidden features that stably support both trajectory decoding and quality scoring.

Editorial extensions

If this is right

If correct, large video world models can be used for real-time closed-loop driving without ever rendering future frames, since the planner consumes only the current image and cached text context.
The quality-gated routing means compute is spent per scene: scenes that an intermediate layer already solves cost about 112 ms at threshold 80, while harder scenes continue deeper, producing a smooth performance–latency frontier controlled by the threshold.
The representation transfers across datasets without fine-tuning, suggesting that intermediate video-DiT features encode traffic dynamics general enough to support zero-shot planning on nuScenes.
The fixed block-22 64-proposal variant shows that a single intermediate exit with proposal coverage can exceed the headline numbers of full-video world-action models, further supporting the depth-vs-noise finding.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper does not test variance over scheduler-noise draws for the future-latent slice; a natural extension is to measure PDMS spread across multiple noise seeds at index 17 to confirm the result is not an artifact of one random draw.
The tie-aware quality scorer, which predicts component scores rather than a strict total order, could transfer to other domains with saturated and frequently tied rewards, such as motion planning or robotic manipulation from video models.
A testable consequence of the early-exit claim is that a single DiT prefix could support multiple driving sub-tasks with per-task exit decisions, not just per-scene routing, further reducing the amount of world-model computation each decision requires.
The 78x latency reduction (13.22 s to 170 ms) is measured on an A100; the structural advantage of one conditional forward versus 80 DiT forwards should persist on less powerful hardware, but the exact speedup will vary.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The thing to know about this paper: it cleanly separates two axes people usually conflate—video noise level and DiT depth—and shows depth matters much more than the five tested noise levels. That controlled measurement, plus the quality-gated multi-exit planner built on it, is new. The paper is also honest about what is auxiliary: the 92.6 PDMS number comes from a separate fixed-exit 64-proposal model trained with pseudo-expert targets, and the authors keep it labeled as such. The adaptive single-trajectory system at 90.8 PDMS and 170 ms on an A100 is the real contribution, and it is well supported by a threshold sweep, scorer diagnostics, latency decomposition, and ten-seed layer-wise statistics.

The soft spots are real but mostly fixable. The biggest one, which the stress-test note gets right, is that the deployed hidden state in Eq. (1) uses a single scheduler-noise draw for the future latent at s*=17, and there is no variance over noise seeds. The paper's noise-level experiment varies five scheduler indices, not random realizations, so it does not establish that the intermediate features are stable to the draw. Since training feeds real noised future frames while deployment feeds pure noise, this is a genuine train/deploy mismatch. It is an empirical gap rather than an established failure, but it is load-bearing: if the random future slice shifts the decoded trajectory or the scorer's quality estimate, then the routing decision and the 170 ms/90.8 PDMS numbers are properties of one seed. A revision should report mean and variance over at least a few noise seeds for the deployed path.

Second, the main PDMS and latency numbers have no error bars, and the threshold eta, the noise index, and the exit set are all validation-selected. That is standard practice in this benchmark-driven subfield, but it does mean the reported adaptive gain over fixed B15 (0.17 PDMS) is close to the noise floor. Third, the scorer is trained on evaluator component targets and the DiffGRPO reward is the NAVSIM planner score, so the headline metric is partly calibrated to itself. The authors are transparent about this, and the threshold sweep shows the adaptive benefit is modest; the main claim is latency reduction, which survives the circularity concern. Fourth, code is promised but not yet released, so the 92.6 auxiliary result is not independently checkable yet.

Who is this for? Anyone working on world-action models, efficient diffusion inference, or driving planning with generative priors. It is a serious empirical contribution that deserves referee time. My recommendation: send it to peer review, and make the noise-seed variance analysis and error bars a required revision condition. The central architecture and depth-dependence finding are solid enough that this should be published after those gaps are closed.

Referee Report

4 major / 5 minor

Summary. The paper proposes Adaptive-WAM, a multi-exit trajectory planner built on a Wan2.2-5B video-diffusion backbone. It attaches independent ReCogDrive-style trajectory diffusion heads to six selected DiT blocks, trains a lightweight DINOv2-Small scorer to predict NAVSIM metric components, and at inference decodes one trajectory per attempted exit, terminating when the best accumulated predicted quality exceeds a validation-selected threshold η. The deployment path is a single conditional Wan forward at a fixed video-noise index s*=17, with no future-video denoising and no VAE video decoding. The paper reports 90.8 PDMS on NAVSIM v1 and 89.9 EPDMS on v2 for the adaptive single-trajectory planner, 92.6 PDMS for an auxiliary fixed-exit 64-proposal variant, zero-shot nuScenes L2 error of 0.88 m with 0.08% collision rate, and 170 ms end-to-end latency versus 190 ms for the fixed block-15 planner and 320 ms for the full-depth planner.

Significance. If the results hold, the paper makes a useful contribution: a controlled study separating video-noise level from DiT depth, a practical early-exit mechanism for large world-action models, and a transparent efficiency evaluation. The paper is detailed and commendable in its ablations: backbone-adaptation comparisons, scorer-backbone diagnostics, a threshold sweep, tie-aware scorer failure analysis, a latency decomposition, and ten-seed aggregation for the layer-wise analyses. The auxiliary 64-proposal result is clearly separated from the adaptive single-trajectory claim. However, several load-bearing empirical decisions need strengthening before the central claims are fully supported: the headline numbers lack error bars, the deployed single-forward representation has no reported noise-realization variance, the threshold and related hyperparameters are validation-selected without a clearly defined split, and the adaptive latency accounting needs an exit-distribution breakdown.

major comments (4)

[Section D.1, Eq. (1), Appendix F] The deployed hidden state h_l = F_{1:l}(I,d(o);s*) is computed from a single conditional Wan forward whose future-latent slice is initialized from scheduler noise at index 17. The motivating noise-level experiment (Table 13) varies five scheduler indices but not noise realizations, so it demonstrates insensitivity to index, not to the random draw. Because the training-time forward uses real future frames noised to index 17 while deployment replaces those frames with a random future slice, the train/deploy distribution shift is unquantified. Please report the mean and standard deviation of PDMS and latency over at least 10 noise draws for the fixed exits and the adaptive policy, or average the latent over multiple draws; this is needed to support the central claim that intermediate representations are stable enough for quality-guided routing.
[Table 5 and Section 5] The headline adaptive improvement over the strongest fixed single-trajectory exit is 90.79 versus 90.62 PDMS, a difference of 0.17 points. No error bars or seed-level statistics are reported for Table 5, even though Section 3 reports ten-seed aggregations for layer-wise scores. Given the small difference, the claim that adaptive routing improves over every fixed single-trajectory exit needs confidence intervals or multi-seed results; otherwise the reported gain may be within run-to-run variation.
[Section 4, Eq. (7), Section C.1] Threshold η, noise index s*=17, and exit set E are validation-selected, but Section C.1 states that NAVSIM does not define an official train/validation partition for the converted corpus. It is therefore unclear how the validation set used for threshold selection is constructed and whether the reported test numbers are affected by selection on the same data. Please specify the split discipline for η selection and report the threshold sweep on a separate holdout or via cross-validation.
[Section K, Table 5] For the adaptive policy at η=90, 94.1% of scenes terminate within the first three exits (B5, B9, B15), and each attempted exit adds a trajectory-head evaluation and a scorer evaluation. It is not obvious how this can average 170 ms while the fixed B15 planner, which evaluates only one head and no scorer, averages 190 ms, unless a large fraction of adaptive runs terminate at B5 or B9. Please report the exit-index distribution and a per-component latency breakdown (VAE encoding, Wan prefix by exit, per-exit trajectory head, per-exit scorer) to reconcile the efficiency claim.

minor comments (5)

[Section 4, Eq. (3)] The component-vector notation in Eq. (3) is garbled: "ˆrℓ =σ(a ℓ) = (dNC, [DAC, [DDC, [TTC, cEP, \Comf)" contains mismatched brackets and a stray backslash. Please rewrite it as a clean tuple of the six predicted components.
[Section 3 and Table 5] Section 3 says "Reported PDMS aggregates validation-best checkpoints over ten seeds," but Tables 2, 3, and 5 present single numbers without variance. Please clarify whether these main-table numbers are also ten-seed aggregates and, if so, include error bars or a variance column.
[Table 4 and Appendix L] The text says DriveVA is "slightly better" on the nuScenes averages, but the exact gaps (0.04 m L2 and 0.02% collision rate) are not stated in the main text; please state them and note explicitly that the efficiency comparison is not quantified in the same table.
[Figure 5 caption] The caption mentions green and red trajectory colors, but the figure appears not to encode these colors legibly in the manuscript; please ensure the visual encoding is clear or adjust the caption.
[Appendix E] The auxiliary 64-proposal result uses CLOVER-derived pseudo-expert targets scored with the true NAVSIM evaluator and training-time future occupancy. This privileged supervision is disclosed in the appendix but should be stated more prominently in the main text so that the 92.6 PDMS result is not mistaken for a purely self-supervised or fully comparable single-trajectory result.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are measured against external official evaluators, and the learned scorer and validation-selected threshold are transparent routing components, not the reported metrics themselves.

full rationale

The paper's central derivation chain is empirically self-contained rather than circular. The noise-level and depth findings (Section 3, Tables 1 and 13) are direct measurements on the official NAVSIM navtest protocol with ten seeds; the paper explicitly limits the claim to the five tested noise indices and does not assert invariance to all schedulers or noise draws. The adaptive planner's reported PDMS (90.8), EPDMS (89.9), and nuScenes L2/collision numbers are computed by the official NAVSIM and nuScenes evaluators, not by the learned trajectory-quality scorer. The scorer's predicted Q is used only for routing and candidate selection, and its reliability is separately diagnosed on a held-out candidate pool (Section H). Training the scorer on evaluator-provided component targets is transparent supervised learning of a proxy; it does not reduce the reported evaluation to the scorer's own output. The threshold η is selected on validation, which is standard model selection rather than a fitted parameter being renamed as a prediction. The only self-citation is CLOVER (Ang et al. 2026) for pseudo-expert target construction in the auxiliary fixed-B22 64-proposal model (Appendix E), but that protocol is described in detail in the appendix and does not bear on the main early-exit or single-trajectory adaptive claims; it is therefore not load-bearing circularity. The absence of variance over the scheduler-noise draw at s*=17 is a legitimate empirical robustness gap, but it is not a self-referential reduction: no equation in the paper defines the reported result in terms of the noise draw. No step was found where Eq. X equals Eq. Y by construction or where a fitted value is reported as the prediction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper's central claims are empirical, not derived. The system depends on benchmark-tuned hyperparameters (video noise index, exit set, quality threshold, RL reward weights, pseudo-expert targets) and on domain assumptions about NAVSIM and Wan2.2 transfer. No new physical or theoretical entities are postulated; the quality scorer and exit heads are trained artifacts evaluated inside the paper.

free parameters (5)

Quality threshold eta = 90
Selected on NAVSIM validation (Table 5). It sets the desired quality-computation operating point; the headline 170 ms and 90.79 PDMS numbers are measured at this threshold.
Video noise index s* = 17
Fixed after testing five indices on NAVSIM; robustness across tested indices is shown, but all reported results use this single index.
Exit block set E = {5, 9, 15, 18, 22, 30}
Hand-chosen readout depths from the 30-block Wan2.2 DiT; layer-wise analysis uses these blocks.
DiffGRPO reward weights = 10/5/2/0 for progress/TTC/comfort/driving direction
Manual planner-RL weights, adopted from ReCogDrive; affects trajectory-head refinement.
Pseudo-expert target settings (auxiliary 64-proposal model) = 180 candidates scored, 50 retained, threshold 0.8, coverage weight 0.5
Used by the fixed B22 64-proposal model to achieve 92.6 PDMS via evaluator-scored candidate targets; not used by the adaptive single-trajectory planner.

assumptions (6)

domain assumption Wan2.2-TI2V-5B video-diffusion pretraining provides transferable spatiotemporal representations for driving.
Section 4 initializes the visual dynamics backbone from Wan2.2; the method's value depends on this transfer.
domain assumption NAVSIM non-reactive PDMS/EPDMS is a valid proxy for driving-planning quality.
All headline numbers are computed on this benchmark; generalization to real driving is assumed.
ad hoc to paper A single conditional Wan forward at fixed noise index with a noisy future latent gives a stable planning representation.
Section D.1: the future part of the latent is initialized from scheduler noise; no averaging over noise draws is reported. If false, results depend on an uncontrolled random draw.
domain assumption The quality scorer's predicted PDMS components generalize from the diagnostic pool to unseen scenes.
Section H reports diagnostics on 12,146 scenes but provides no formal guarantee; routing depends on this.
domain assumption Five-step ReCogDrive trajectory diffusion heads with DiffGRPO refinement are appropriate trajectory decoders.
Borrowed from ReCogDrive; all exits use the same protocol so depth comparisons are controlled.
domain assumption Validation-best checkpoint selection over ten seeds is a fair model-selection rule.
Section 3 reports layer-wise statistics aggregated over ten seeds; main results use the best validation checkpoints.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features." pith.science (2026). https://pith.science/paper/EHVU2BMV

@misc{pith2026260806008,
  author       = {Pith},
  title        = {Pith review of: Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHVU2BMV}},
  note         = {Machine review of arXiv:2608.06008}
}

read the original abstract

Large video diffusion models provide rich spatiotemporal priors for autonomous driving, but existing world-action models often inherit the cost of iterative future-video generation even though deployment only requires an ego trajectory. We ask a more basic question: how much of a video diffusion model must be executed to make a reliable driving decision? Through a controlled study of video denoising timesteps and Diffusion Transformer (DiT) depth, we find that planning performance is largely insensitive to the tested video-noise levels, whereas strong trajectories can already be decoded from intermediate layers. Based on this observation, we introduce Adaptive-WAM, a quality-aware multi-exit planner built on a Wan2.2-5B backbone. Trajectory diffusion heads are attached to selected DiT blocks, and a lightweight trajectory-quality scorer terminates inference once the best trajectory decoded so far satisfies a quality threshold; otherwise, computation continues from the cached hidden state to a deeper exit. The deployed planner therefore avoids the iterative classifier-free denoising loop and VAE decoding required for future-video synthesis, while dynamically allocating backbone depth according to trajectory quality. On NAVSIM, the adaptive single-trajectory planner achieves 90.8 PDMS; a separate fixed-exit variant reaches 92.6 PDMS with 64 proposals. It further obtains 89.9 EPDMS on NAVSIM v2, yielding the best reported results among the compared front-view video world-model planners. Without target-domain fine-tuning, Adaptive-WAM transfers to nuScenes with 0.88 m average L2 error and a 0.08\% collision rate. On an A100, adaptive routing improves PDMS from 90.62 to 90.79 while averaging 170 ms end-to-end planning latency, approximately 10\% below the 190 ms fixed block-15 planner and 47\% below the 320 ms fixed full-depth planner. Code will be released.

Figures

Figures reproduced from arXiv: 2608.06008 by the authors.

**Figure 2.** Post-RL Jaccard overlap of high-quality scene sets. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Pairwise large-advantage counts before and after planner RL. Panels report the mean or standard deviation over ten [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Overview of Adaptive-WAM. Wan2.2 retains video supervision while six intermediate blocks feed independent [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Selected layer-wise trajectory overlays. Trajectory colors encode scores: green marks high-scoring trajectories, whereas [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 5 canonical work pages

[5]

Hu, Y.; Yang, J.; Chen, L.; Li, K.; Sima, C.; Zhu, X.; Chai, S.;Du,S.;Lin,T.;Wang,W.;etal.2023b

DrivingWorld: Constructing world model for autonomous driving via video GPT.arXiv preprint arXiv:2412.19505. Hu, Y.; Yang, J.; Chen, L.; Li, K.; Sima, C.; Zhu, X.; Chai, S.;Du,S.;Lin,T.;Wang,W.;etal.2023b. Planning-oriented autonomous driving. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 17853– 17862. Huang, G.; C...

arXiv
[7]

InEuropean Conference on Computer Vision, 469–485

DrivingDiffusion: Layout-guided multi-view driving scenarios video genera- tionwithlatentdiffusionmodel. InEuropean Conference on Computer Vision, 469–485. Springer. Li, Y.; Fan, L.; He, J.; Wang, Y.; Chen, Y.; Zhang, Z.; and Tan, T. 2025b. Enhancing end-to-end autonomous driving with latent world model. InInternational Conference on Learning Representati...

arXiv 2025
[8]

Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang, S

Driveva: Video action models are zero-shot drivers.arXiv preprint arXiv:2604.04198. Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang, S

arXiv
[9]

Min, C.; Zhao, D.; Xiao, L.; Zhao, J.; Xu, X.; Zhu, Z.; Jin, L.; Li, J.; Guo, Y.; Xing, J.; et al

Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control.arXiv preprint arXiv:2603.10448. Min, C.; Zhao, D.; Xiao, L.; Zhao, J.; Xu, X.; Zhu, Z.; Jin, L.; Li, J.; Guo, Y.; Xing, J.; et al

arXiv
[10]

Peebles,W.;andXie,S.2023.Scalablediffusionmodelswith transformers

Dinov2: Learning robust visual fea- tures without supervision.Transactions on Machine Learn- ing Research Journal. Peebles,W.;andXie,S.2023.Scalablediffusionmodelswith transformers. InProceedings of the IEEE/CVF international conference on computer vision, 4195–4205. Shao,Z.;Wang,P.;Zhu,Q.;Xu,R.;Song,J.;Bi,X.;Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al

work page 2023
[11]

Song,J.;Meng,C.;andErmon,S.2020

Deepseekmath: Pushing the limits of mathematical reasoning in open lan- guage models.arXiv preprint arXiv:2402.03300. Song,J.;Meng,C.;andErmon,S.2020. Denoisingdiffusion implicit models.arXiv preprint arXiv:2010.02502. Teerapittayanon, S.; McDanel, B.; and Kung, H.-T

arXiv 2020
[13]

InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 15449– 15458

Para-drive: Parallelized architecture for real-time au- tonomousdriving. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 15449– 15458. Xia, T.; Li, Y.; Zhou, L.; Yao, J.; Xiong, K.; Sun, H.; Wang, B.;Ma,K.;Chen,G.;Ye,H.;etal.2025. Drivelaw:Unifying planningandvideogenerationinalatentdrivingworld.arXiv preprint arXiv:...

arXiv 2025
[14]

InProceedings of the Computer Vision and Pattern Recognition Conference, 1602–1611

Goalflow: Goal-driven flow matching for multimodal trajectories generation in end-to- end autonomous driving. InProceedings of the Computer Vision and Pattern Recognition Conference, 1602–1611. Yang, J.; Chitta, K.; Gao, S.; Chen, L.; Shao, Y.; Jia, X.; Li, H.;Geiger,A.;Yue,X.;andChen,L.2026. Resim:Reliable world simulationfor autonomous driving.Advances ...

work page 2026

Show all 19 references

[15]

Zeng, S.; Chang, X.; Xie, M.; Liu, X.; Bai, Y.; Pan, Z.; Xu, M.; and Wei, X

Fast-wam: Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666. Zeng, S.; Chang, X.; Xie, M.; Liu, X.; Bai, Y.; Pan, Z.; Xu, M.; and Wei, X

arXiv
[16]

Advances in Neural Information Processing Systems, 38: 67299–67318

Futuresightdrive: Thinking visually with spatio-temporal cot for autonomous driving. Advances in Neural Information Processing Systems, 38: 67299–67318. Zhang,K.;Tang,Z.;Hu,X.;Pan,X.;Guo,X.;Liu,Y.;Huang, J.;Yuan,L.;Zhang,Q.;Long,X.-X.;etal.2025. Epona:Au- toregressive diffusio...

2025 arXiv
[17]

No observed future latent, unconditional classifier-free-guidance branch, or video VAE decoding is used by the planning path

The future part of the latent is initialized from scheduler noise at that index; the current image provides the conditioned latent slice. No observed future latent, unconditional classifier-free-guidance branch, or video VAE decoding is used by the planning path. Hidden states...

2020
[18]

Non-LoRA Wan parameters and the VAE are frozen. Wan LoRA item Setting Target modulesq, k, v, o Rank / alpha / dropout 32 / 64 / 0.05 Optimizer AdamW Base learning rate2×10 −4 Per-GPU batch / GPUs 5 / 4 Gradient accumulation 4 Effective batch 80 Training length 80 epochs Precis...

2026
[90]

Exit by B15

Off-diagonal overlaps range from 0.69 to 0.82. Thus, the exits solve substantially overlapping but non-identical scene sets: an early exit is sufficient for most scenes, while deeperexitscanstillrecovercasesnotsolvedbytheglobally strongest intermediate block. G.3 Directional L...

2024
[2017]

Tong,W.;Sima,C.;Wang,T.;Chen,L.;Wu,S.;Deng,H.;Gu, Y.;Lu,L.;Luo,P.;Lin,D.;etal.2023

Branchynet:Fastinferenceviaearlyexitingfromdeepneural networks.arXiv preprint arXiv:1709.01686. Tong,W.;Sima,C.;Wang,T.;Chen,L.;Wu,S.;Deng,H.;Gu, Y.;Lu,L.;Luo,P.;Lin,D.;etal.2023. Sceneasoccupancy. InProceedings of the IEEE/CVF International Conference on Computer Vision, 8406...

2023 arXiv
[2018]

InInternational con- ference on learning representations

Multi-scale dense networks for resource efficient image classification. InInternational con- ference on learning representations. Jiang, B.; Chen, S.; Xu, Q.; Liao, B.; Chen, J.; Zhou, H.; Zhang,Q.;Liu,W.;Huang,C.;andWang,X.2023.Vad:Vec- torized scene representation for effici...

2023 arXiv
[2020]

InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621–11631

nuscenes: A multimodal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621–11631. Chen, S.; Jiang, B.; Gao, H.; Liao, B.; Xu, Q.; Zhang, Q.; Huang,C.;Liu,W.; andWang,X.2024. Vadv2:End-to-end vectorized auto...

2024 arXiv
[2024]

Gao, R.; Chen, K.; Xie, E.; Hong, L.; Li, Z.; Yeung, D.-Y.; and Xu, Q

Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking.Advances in Neural Information Processing Systems, 37: 28706–28719. Gao, R.; Chen, K.; Xie, E.; Hong, L.; Li, Z.; Yeung, D.-Y.; and Xu, Q. 2024a. Magicdrive: Street view generation with diverse3dge...

2024 arXiv
[2025]

Caesar, H.; Bankiti, V.; Lang, A

Vavim and vavam: Autonomous driving through video generative modeling.arXiv preprint arXiv:2502.15672. Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O

arXiv
[2026]

Bartoccioni,F.;Ramzi,E.;Besnier,V.;Venkataramanan,S.; Vu, T.-H.; Xu, Y.; Chambon, L.; Gidaris, S.; Odabas, S.; Hurych, D.; et al

CLOVER: Closed-loop value estimation and ranking for end-to-end autonomous driving planning.arXiv preprint arXiv:2605.15120. Bartoccioni,F.;Ramzi,E.;Besnier,V.;Venkataramanan,S.; Vu, T.-H.; Xu, Y.; Chambon, L.; Gidaris, S.; Odabas, S.; Hurych, D.; et al

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [5]

Hu, Y.; Yang, J.; Chen, L.; Li, K.; Sima, C.; Zhu, X.; Chai, S.;Du,S.;Lin,T.;Wang,W.;etal.2023b

DrivingWorld: Constructing world model for autonomous driving via video GPT.arXiv preprint arXiv:2412.19505. Hu, Y.; Yang, J.; Chen, L.; Li, K.; Sima, C.; Zhu, X.; Chai, S.;Du,S.;Lin,T.;Wang,W.;etal.2023b. Planning-oriented autonomous driving. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 17853– 17862. Huang, G.; C...

arXiv

[2] [7]

InEuropean Conference on Computer Vision, 469–485

DrivingDiffusion: Layout-guided multi-view driving scenarios video genera- tionwithlatentdiffusionmodel. InEuropean Conference on Computer Vision, 469–485. Springer. Li, Y.; Fan, L.; He, J.; Wang, Y.; Chen, Y.; Zhang, Z.; and Tan, T. 2025b. Enhancing end-to-end autonomous driving with latent world model. InInternational Conference on Learning Representati...

arXiv 2025

[3] [8]

Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang, S

Driveva: Video action models are zero-shot drivers.arXiv preprint arXiv:2604.04198. Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang, S

arXiv

[4] [9]

Min, C.; Zhao, D.; Xiao, L.; Zhao, J.; Xu, X.; Zhu, Z.; Jin, L.; Li, J.; Guo, Y.; Xing, J.; et al

Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control.arXiv preprint arXiv:2603.10448. Min, C.; Zhao, D.; Xiao, L.; Zhao, J.; Xu, X.; Zhu, Z.; Jin, L.; Li, J.; Guo, Y.; Xing, J.; et al

arXiv

[5] [10]

Peebles,W.;andXie,S.2023.Scalablediffusionmodelswith transformers

Dinov2: Learning robust visual fea- tures without supervision.Transactions on Machine Learn- ing Research Journal. Peebles,W.;andXie,S.2023.Scalablediffusionmodelswith transformers. InProceedings of the IEEE/CVF international conference on computer vision, 4195–4205. Shao,Z.;Wang,P.;Zhu,Q.;Xu,R.;Song,J.;Bi,X.;Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al

work page 2023

[6] [11]

Song,J.;Meng,C.;andErmon,S.2020

Deepseekmath: Pushing the limits of mathematical reasoning in open lan- guage models.arXiv preprint arXiv:2402.03300. Song,J.;Meng,C.;andErmon,S.2020. Denoisingdiffusion implicit models.arXiv preprint arXiv:2010.02502. Teerapittayanon, S.; McDanel, B.; and Kung, H.-T

arXiv 2020

[7] [13]

InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 15449– 15458

Para-drive: Parallelized architecture for real-time au- tonomousdriving. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 15449– 15458. Xia, T.; Li, Y.; Zhou, L.; Yao, J.; Xiong, K.; Sun, H.; Wang, B.;Ma,K.;Chen,G.;Ye,H.;etal.2025. Drivelaw:Unifying planningandvideogenerationinalatentdrivingworld.arXiv preprint arXiv:...

arXiv 2025

[8] [14]

InProceedings of the Computer Vision and Pattern Recognition Conference, 1602–1611

Goalflow: Goal-driven flow matching for multimodal trajectories generation in end-to- end autonomous driving. InProceedings of the Computer Vision and Pattern Recognition Conference, 1602–1611. Yang, J.; Chitta, K.; Gao, S.; Chen, L.; Shao, Y.; Jia, X.; Li, H.;Geiger,A.;Yue,X.;andChen,L.2026. Resim:Reliable world simulationfor autonomous driving.Advances ...

work page 2026