Pith. sign in

REVIEW 3 major objections 6 minor 119 references

WAM-Diff2: Hierarchical AR-to-Diffusion Distillation for Highly Efficient Autonomous Driving VLA

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A three-stage hierarchical distillation converts a pre-trained autoregressive driving VLA into a block-causal masked diffusion VLA that matches planning quality while decoding 2.8x faster.

desk verdict Solid distillation recipe and real planning speedup, but multi-task parity claim is contradicted by the paper's own Table 1. read the letter →

arxiv 2608.01035 v1 pith:YOK64JBA submitted 2026-08-02 cs.RO cs.AIcs.CV

Zhihao Zhu , Hanlin Shang , Mingwang Xu , Feipeng Cai , Zhuolin He , Yaoyi Li , Jianhua Han , Hang Xu
show 1 more author
Siyu Zhu
This is my paper
classification cs.ROcs.AIcs.CV
keywords vision-language-actionautonomousdrivingdiscretediffusiondistillationblock-causalattentionexposurebiastrajectoryplanningefficientinference
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

This paper tries to bridge two disconnected paradigms in autonomous driving: autoregressive vision-language-action (VLA) generalists, which reason broadly but decode slowly and suffer from exposure bias, and diffusion policies, which decode in parallel but are usually narrow and single-task. The authors propose WAM-Diff2, a discrete diffusion VLA that inherits the semantic knowledge of a pre-trained autoregressive teacher through three staged distillation phases. If the conversion works as claimed, practitioners could keep the multi-task intelligence of large autoregressive models while gaining a 2.8x decoding speedup, rising to 15.1x with system-level optimizations, without sacrificing planning quality.

What carries the argument

The central mechanism is the block-causal attention mask combined with a three-stage distillation curriculum. Tokens within the same decoding block attend to each other bidirectionally while inter-block dependencies remain causal; progressively increasing the block size from B=1 (pure autoregressive) to B=32 (near-bidirectional) interpolates between the paradigms, and Jensen-Shannon divergence distillation from a smaller-block diffusion teacher and then from an 8B diffusion teacher aligns the student's token-level distributions.

What would settle it

Train the B=32 diffusion student directly from the AR teacher without the progressive block-size curriculum on the same data and measure NAVSIM PDMS; the paper's claim of a smooth curriculum-driven transition predicts a large drop (about 84.1 PDMS) without the stages and recovery to about 88.3 PDMS with them, so observing parity without the curriculum would falsify it.

Watch

Extended reading notes

Core claim

WAM-Diff2 shows that an autoregressive VLA can be translated into a block-causal masked diffusion VLA through a three-stage hierarchical distillation: (1) progressive block-wise adaptation, which expands the decoding block size B from 1 to 32, with each variant initialized from its predecessor to relax causal attention into bidirectional in-block refinement; (2) block-wise distillation, where a stable small-block diffusion teacher distills larger-block students using a symmetric Jensen-Shannon divergence over noisy intermediate states; and (3) model-wise cross-scale distillation, where an 8B diffusion teacher transfers reasoning to a 2B student. On NAVSIM v1, the B=32 student scores 87.44 PD

Load-bearing premise

Stage I's curriculum assumes that progressively enlarging the decoding block size, with each variant initialized from the previous one, smoothly interpolates between causal and bidirectional attention so the model retains the AR teacher's semantic knowledge; if that interpolation fails, the later distillation stages only teach the student to imitate a degraded attention pattern.

Editorial extensions

If this is right

  • A single 2B diffusion VLA checkpoint can match its autoregressive foundation: 87.44 vs 88.14 PDMS on NAVSIM v1, with minor tradeoffs on DriveBench, LingoQA, and COCO.
  • The 2.8x algorithmic decoding speedup compounds with FlashInfer and CUDA Graphs to a cumulative 15.1x latency reduction (22.7 ms/token to 1.5 ms/token) with negligible planning degradation.
  • Bidirectional iterative refinement mitigates exposure bias: average per-waypoint L2 error drops 5.8%, and the error reduction widens from 0.002 at the first waypoint to 0.082 at the eighth.
  • Block size B acts as a runtime-configurable latency/quality knob: going from B=4 to B=32 raises vanilla throughput from 68.3 to 124.8 TPS while keeping planning metrics within about 0.5 PDMS.
  • The three-stage recipe yields a unified multi-task model without task-specific heads, covering planning, perception, and scene understanding from a single checkpoint.

Reading between the lines

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

  • If the recipe generalizes, any pre-trained autoregressive VLA (robotics, multimodal chat, world models) could be re-targeted to masked diffusion, making the AR-to-diffusion translation a reusable efficiency layer rather than a driving-specific fix.
  • The reported top-K overlap ratio between teacher and student could serve as a cheap, model-agnostic predictor of cross-paradigm distillation success, allowing practitioners to decide when to attempt the conversion before committing compute.
  • Because the error-reduction curve widens monotonically with horizon, the closed-loop benefit of bidirectional refinement should grow with route length; a natural test is measuring infraction rates on longer Bench2Drive routes.
  • Stage III being optional implies a family of latency/quality operating points (fast 2B without cross-scale distillation versus reasoning-heavy 2B with it), a Pareto frontier the paper does not explicitly chart.
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.

Referee Report

3 major / 6 minor

Summary. WAM-Diff2 proposes to convert a pre-trained autoregressive VLA (Qwen3-VL) into a block-causal masked-diffusion VLA through three distillation stages: progressive block-size expansion (B=1→32), block-wise JSD distillation from smaller-block diffusion teachers, and cross-scale distillation from an 8B diffusion teacher. The paper reports multi-task results on DriveBench, LingoQA, COCO, NAVSIM v1/v2, and Bench2Drive, and claims strict performance parity with the autoregressive foundation while obtaining a 2.8× decoding speedup (up to 15.1× with FlashInfer and CUDA Graphs). The core evidence is the B=32 2B model: 87.44 PDMS vs. 88.14 for the AR baseline on NAVSIM, and 124.8 TPS vs. 44.5 TPS. The 2.8× speedup and planning-level parity are supported by the reported numbers, but the broader multi-task parity claim is contradicted by the same table on understanding and perception benchmarks.

Significance. If the strong claim were correct—that an autoregressive generalist can be turned into a parallel diffusion model without multi-task loss—this would be a practically important contribution to efficient autonomous-driving VLA deployment. The paper has genuine strengths: it evaluates a single unified checkpoint across multiple benchmarks, includes systematic ablations of the three stages (Table 4), compares distillation objectives (Table 5), and quantifies exposure-bias mitigation with paired L2-error reductions (Figure 7). However, the paper's own numbers do not support the headline 'performance parity' claim as stated. The core methodology is still defensible if reframed as achieving planning parity with small, explicitly quantified degradations on non-planning tasks, and the 2.8× algorithmic speedup is credible. The larger 15.1× claim is weakened by cross-hardware comparisons. With corrected claims and additional measurement details, the paper would have clear value.

major comments (3)
  1. [Figure 4 and §4.2] The retention percentages in Figure 4 ('100.5% retained', '100.0% retained') are computed against the original Qwen3-VL-2B model, not against the AR teacher that the method is claimed to preserve (the Ours-2B B=1 row in Table 1). Against the actual AR teacher, the B=32 model retains only about 95.3% on DriveBench, 96.2% on LingoQA, and 92.6% on COCO. This is an internal arithmetic inconsistency in the presentation: the claimed preservation target is the fine-tuned AR baseline, not the base Qwen3-VL. The figure and accompanying text must be corrected and the baseline clearly identified.
  2. [Appendix A.3, Table 1, §3.3] The claimed 15.1× cumulative speedup mixes hardware stacks. The 2.8× algorithmic speedup (124.8 TPS vs. 44.5 TPS) is computed on the Ascend platform, consistent with Table 1. The FlashInfer and CUDA Graph numbers (673.4 TPS, 1.5 ms/token) are, according to Appendix A.3, evaluated only on a CUDA-compatible platform. Comparing the 1.5 ms/token CUDA result to the 22.7 ms/token Ascend baseline conflates algorithmic gains with platform differences. The paper should either report the AR baseline on the same CUDA stack or clearly label the 15.1× as a cross-platform upper bound, not a directly measured cumulative speedup.
  3. [§3.2 Stage III, Table 1] The 8B diffusion teacher used in Stage III is never evaluated. Table 1 lists Qwen3-VL-8B, but that is the raw autoregressive model, not the '8B block-32 diffusion model' produced by the same three-stage pipeline and used for distillation. Because Stage III is credited with recovering multi-task intelligence lost during compression, the paper should report the 8B diffusion teacher's scores on the same benchmarks under the same protocol. Without this, the reader cannot tell whether the teacher is actually a strong diffusion model or whether the cross-scale distillation is propagating a degraded teacher. The ablation in Table 7 only compares AR teacher vs. diffusion teacher at the same 2B scale, not the 8B teacher's absolute capability.
minor comments (6)
  1. [Table 1] No error bars, confidence intervals, or multiple-seed results are reported. Given that some of the claimed 'parity' differences are 0.5–0.8 PDMS, the absence of variance information makes it hard to distinguish true parity from noise.
  2. [Eq. (5) and §3.2] Stage II is described as on-policy distillation, but the loss in Eq. (5) is written over 'identical corrupted states' without specifying whether the noisy states are sampled from the student's own generative process or from a fixed corruption schedule. The on-policy aspect should be made explicit, or the terminology should be softened.
  3. [Table 5] The difference between forward KL (88.2) and JSD (88.3) is 0.1 PDMS. The text says JSD 'achieves peak performance' and 'preserving both mode-seeking precision and broad semantic coverage'; this is an overinterpretation of a negligible difference unless statistical significance is shown.
  4. [Appendix A.4] The motion planning prompt template contains three '<image>' placeholders while the text says 'front-view image'; clarify whether this is one image or three camera views, and whether the same prompt is used for NAVSIM and Bench2Drive.
  5. [Figure 3] The denoising-step curves are useful, but the figure does not state which block size B, which teacher, or which evaluation protocol produced these curves. Without this context, the reader cannot reproduce or interpret the saturation effect.
  6. [§4.1] The paper would benefit from a statement about code/model release, as the method involves several nontrivial training stages and the current description does not include enough detail (e.g., exact mask-ratio schedule, remasking confidence threshold) for exact reproduction.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: central claims are measured against external benchmarks; only a minor non-load-bearing self-citation exists.

full rationale

The derivation chain is empirically grounded at every stage. Stage I initializes each larger-block student from its smaller-block predecessor (Eq. 4) and is evaluated by external benchmarks; Stage II distills from a smaller-block diffusion teacher and is evaluated on NAVSIM; Stage III distills from an 8B diffusion teacher and is again evaluated on external multi-task benchmarks. No fitted constant is renamed as a prediction: the claimed 2.8x speedup and 15.1x latency reduction are measured timings, and the planning scores come from NAVSIM/Bench2Drive. The single same-group citation is Xu et al. 2025a (WAM-Diff), used only in the related-work enumeration of diffusion trajectory models; it is not the justification for the paper's central premise. The top-K overlap statistic (Eq. 7) and the teacher ablation (Tables 6-7) are internal empirical observations, not imported uniqueness theorems. The skeptic's point that Table 1 shows B=32 is 2.4-2.9 points below the AR baseline on DriveBench, LingoQA, and COCO is an internal consistency/correctness issue rather than circularity, because those numbers are externally measured, not derived from the method's own inputs. Therefore no circular step is established.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The central claims rest on external benchmarks, so the circularity burden is low. The main non-standard premises are that semantic knowledge survives the attention-mask relaxation, that JSD distillation preserves multimodality, that the self-generated 8B diffusion teacher is a strong teacher, and that discrete tokens are sufficient for trajectories. The only free parameters that materially affect the headline results are the block size B and the inference-time denoising schedule T, neither fully specified.

free parameters (3)
  • Inference-time denoising steps T = not stated; Figure 3 suggests saturation near 8-16
    Throughput and quality both depend on T; without the exact T, Table 1 latency and accuracy numbers cannot be reproduced.
  • Block size B at deployment = 32 in final model
    The speed/accuracy trade-off is controlled by B; paper explores B=4,8,16,32 but gives no selection criterion beyond a Pareto table.
  • Remasking confidence threshold and mask ratio schedule
    Parallel remasking fixes a subset of tokens with the highest prediction confidence, but the threshold and per-step mask ratio are not given, so the exact decoding procedure is underspecified.
assumptions (5)
  • standard math Absorbing-state masked discrete diffusion forward process with transition matrix Q_t and closed-form marginals is valid.
    Invoked in Section 3.1 Eq. (2); standard result from Austin et al. 2021 and related discrete diffusion literature.
  • domain assumption Semantic knowledge of the pre-trained AR VLA survives the block-causal attention relaxation and can be re-expressed as a denoising model.
    Core premise of Stage I and II; asserted in Section 3.2 and validated only by the paper's own ablations.
  • domain assumption JSD between token distributions preserves multi-modal trajectory structure and avoids mode collapse.
    Used as the distillation loss in Eq. (5) and (8); only empirical support is Table 5.
  • ad hoc to paper The 8B diffusion teacher, produced by the same pipeline, is a strong enough teacher to improve the 2B student.
    Stage III assumes this; Table 7 measures the teacher only indirectly through the student's PDMS, never directly.
  • domain assumption Discrete tokenization of continuous waypoints and bounding boxes is sufficiently precise for the target tasks.
    The unified tokenizer maps continuous outputs to discrete tokens; the paper's own Limitations section admits quantization artifacts.
invented entities (2)
  • Block-causal attention pattern
    purpose: Allows bidirectional parallel token refinement within decoding blocks while keeping causal order across blocks, bridging AR and diffusion attention.
    Introduced in Section 3.2 and 3.3; its effectiveness is demonstrated only through this paper's internal experiments, with no external validation or formal analysis.
  • 8B diffusion teacher
    purpose: Source of high-level semantic knowledge in Stage III cross-scale distillation; its token predictions guide the 2B student.
    This checkpoint is created by the authors' own pipeline and is never benchmarked directly; the paper reports neither its NAVSIM nor DriveBench scores.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WAM-Diff2: Hierarchical AR-to-Diffusion Distillation for Highly Efficient Autonomous Driving VLA." pith.science (2026). https://pith.science/paper/YOK64JBA

@misc{pith2026260801035,
  author       = {Pith},
  title        = {Pith review of: WAM-Diff2: Hierarchical AR-to-Diffusion Distillation for Highly Efficient Autonomous Driving VLA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOK64JBA}},
  note         = {Machine review of arXiv:2608.01035}
}
read the original abstract

Vision-Language-Action (VLA) models have emerged as a prominent paradigm for end-to-end autonomous driving; however, their efficient deployment is severely constrained by high computational latency and exposure bias arising from sequential autoregressive decoding. Conversely, while specialized diffusion policies enable low-latency, parallel execution, training them from scratch typically yields narrow, single-task architectures that lack holistic visual-linguistic reasoning. Successfully transforming pre-trained autoregressive generalists into parallel diffusion models could combine multi-task cognitive intelligence with execution efficiency, yet this transition presents a formidable architectural challenge due to mismatched attention patterns (causal versus bidirectional) and divergent optimization objectives. To bridge this divide, we introduce WAM-Diff2, a multi-task discrete diffusion VLA framework powered by a three-stage hierarchical distillation strategy. By structuring the architectural shift through progressive block-wise adaptation, block-wise distillation, and model-wise cross-scale distillation, WAM-Diff2 preserves the underlying semantic foundations of the base model while accelerating inference. Extensive evaluations across driving understanding, perception, and planning benchmarks demonstrate that WAM-Diff2 effectively mitigates exposure bias and achieves performance parity with autoregressive baselines. Crucially, the autoregressive-to-diffusion transition yields a 2.8x decoding speedup, which scales to an ultimate 15.1x acceleration when combined with system-level optimizations including FlashInfer and CUDA Graphs.

Figures

Figures reproduced from arXiv: 2608.01035 by the authors.

Figure 1
Figure 1. Efficiency, multi-task performance, and planning robustness of WAM-Diff2. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architectural overview of WAM-Diff2 and the three-stage hierarchical distillation framework. (a) Multi-view images, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Effect of denoising steps. PDMS improves as the number of denoising steps increases and gradually saturates at [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Performance–efficiency trade-offs of WAM-Diff2 across multiple benchmarks. WAM-Diff2 consistently preserves [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Qualitative results of WAM-Diff2 on Bench2Drive. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 10
Figure 10. Figure 10: Failure cases visualization on NAVSIM. decoding throughput rises from 68.3 to 124.8 TPS with negli￾gible planning degradation (∆ ≤ 0.5 PDMS), demonstrating flexible, runtime-configurable control over execution latency. Efficiency Analysis. The throughput acceleration …
Figure 8
Figure 8. Figure 8: Visualization of understanding results [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Visualization of perception results. Speed–Accuracy Pareto Frontier. We analyze the infer￾ence dynamics of WAM-Diff2 across varying refinement budgets and decoding block sizes (B). As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 11
Figure 11. Figure 11: Failure cases visualization on Bench2Drive. [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Representative extreme-weather driving scenarios [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Representative driving scenarios in Bench2Drive [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

119 extracted references · 12 canonical work pages

  1. [1]

    arXiv preprint arXiv:2410.23262 , year=

    Emma: End-to-end multimodal model for autonomous driving , author=. arXiv preprint arXiv:2410.23262 , year=

    arXiv

  2. [2]

    arXiv preprint arXiv:2503.19755 , year=

    Orion: A holistic end-to-end autonomous driving framework by vision-language instructed action generation , author=. arXiv preprint arXiv:2503.19755 , year=

    arXiv

  3. [3]

    arXiv preprint arXiv:2506.13757 , year=

    AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning , author=. arXiv preprint arXiv:2506.13757 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2503.23463 , year=

    Opendrivevla: Towards end-to-end autonomous driving with large vision language action model , author=. arXiv preprint arXiv:2503.23463 , year=

    arXiv

  5. [5]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  6. [6]

    The Thirteenth International Conference on Learning Representations , year=

    Diffusion-Based Planning for Autonomous Driving with Flexible Guidance , author=. The Thirteenth International Conference on Learning Representations , year=

  7. [7]

    arXiv preprint arXiv:2512.11872 , year=

    WAM-Diff: A Masked Diffusion VLA Framework with MoE and Online Reinforcement Learning for Autonomous Driving , author=. arXiv preprint arXiv:2512.11872 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2502.09992 , year=

    Large Language Diffusion Models , author=. arXiv preprint arXiv:2502.09992 , year=

    arXiv

Show all 119 references
  1. [9]

    arXiv preprint arXiv:2505.16933 , year=

    LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning , author=. arXiv preprint arXiv:2505.16933 , year=

    arXiv

  2. [10]

    IEEE Robotics and Automation Letters , volume=

    Drivegpt4: Interpretable end-to-end autonomous driving via large language model , author=. IEEE Robotics and Automation Letters , volume=. 2024 , publisher=

    2024

  3. [11]

    arXiv preprint arXiv:2402.12289 , year=

    Drivevlm: The convergence of autonomous driving and large vision-language models , author=. arXiv preprint arXiv:2402.12289 , year=

    arXiv

  4. [12]

    European conference on computer vision , pages=

    Drivelm: Driving with graph visual question answering , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  5. [13]

    arXiv preprint arXiv:2511.19221 , year=

    Percept-WAM: Perception-enhanced world-awareness-action model for robust end-to-end autonomous driving , author=. arXiv preprint arXiv:2511.19221 , year=

  6. [14]

    arXiv preprint arXiv:2505.16278 , year=

    DriveMoE: Mixture-of-experts for vision-language-action model in end-to-end autonomous driving , author=. arXiv preprint arXiv:2505.16278 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:2604.02190 , year=

    UniDriveVLA: Unifying Understanding, Perception, and Action Planning for Autonomous Driving , author=. arXiv preprint arXiv:2604.02190 , year=

  8. [16]

    Advances in neural information processing systems , volume=

    Structured denoising diffusion models in discrete state-spaces , author=. Advances in neural information processing systems , volume=

  9. [17]

    Advances in Neural Information Processing Systems , volume=

    A continuous time framework for discrete denoising models , author=. Advances in Neural Information Processing Systems , volume=

  10. [18]

    arXiv preprint arXiv:2310.16834 , year=

    Discrete diffusion modeling by estimating the ratios of the data distribution , author=. arXiv preprint arXiv:2310.16834 , year=

    arXiv

  11. [19]

    Advances in Neural Information Processing Systems , volume=

    Discrete flow matching , author=. Advances in Neural Information Processing Systems , volume=

  12. [20]

    arXiv preprint arXiv:2508.15487 , year=

    Dream 7b: Diffusion large language models , author=. arXiv preprint arXiv:2508.15487 , year=

    arXiv

  13. [21]

    arXiv preprint arXiv:2505.15809 , year=

    Mmada: Multimodal large diffusion language models , author=. arXiv preprint arXiv:2505.15809 , year=

    arXiv

  14. [22]

    arXiv preprint arXiv:2508.12603 , year=

    Vilad: A large vision language diffusion framework for end-to-end autonomous driving , author=. arXiv preprint arXiv:2508.12603 , year=

    arXiv

  15. [23]

    arXiv preprint arXiv:2509.20109 , year=

    Discrete diffusion for reflective vision-language-action models in autonomous driving , author=. arXiv preprint arXiv:2509.20109 , year=

  16. [24]

    arXiv preprint arXiv:2602.14577 , year=

    DriveFine: Refining-Augmented Masked Diffusion VLA for Precise and Robust Driving , author=. arXiv preprint arXiv:2602.14577 , year=

  17. [25]

    arXiv preprint arXiv:2306.08543 , year=

    Minillm: Knowledge distillation of large language models , author=. arXiv preprint arXiv:2306.08543 , year=

    arXiv

  18. [26]

    The twelfth international conference on learning representations , year=

    On-policy distillation of language models: Learning from self-generated mistakes , author=. The twelfth international conference on learning representations , year=

  19. [27]

    arXiv preprint arXiv:2604.13016 , year=

    Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe , author=. arXiv preprint arXiv:2604.13016 , year=

    arXiv

  20. [28]

    arXiv preprint arXiv:2604.07944 , year=

    On-Policy Distillation of Language Models for Autonomous Vehicle Motion Planning , author=. arXiv preprint arXiv:2604.07944 , year=

    arXiv

  21. [29]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data and Metric Perspectives , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  22. [30]

    European Conference on Computer Vision , pages=

    Lingoqa: Visual question answering for autonomous driving , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  23. [31]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    nuscenes: A multimodal dataset for autonomous driving , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  24. [32]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  25. [33]

    Advances in Neural Information Processing Systems , volume=

    Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking , author=. Advances in Neural Information Processing Systems , volume=

  26. [34]

    Advances in Neural Information Processing Systems , volume=

    Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving , author=. Advances in Neural Information Processing Systems , volume=

  27. [35]

    Advances in Neural Information Processing Systems , volume=

    Simple and effective masked diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  28. [36]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Maskgit: Masked generative image transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [37]

    Advances in neural information processing systems , volume=

    Scheduled sampling for sequence prediction with recurrent neural networks , author=. Advances in neural information processing systems , volume=

  30. [38]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    A survey on vision-language-action models for autonomous driving , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  31. [39]

    arXiv preprint arXiv:2512.16760 , year=

    Vision-language-action models for autonomous driving: Past, present, and future , author=. arXiv preprint arXiv:2512.16760 , year=

  32. [40]

    European Conference on Computer Vision , pages=

    Dolphins: Multimodal language model for driving , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  33. [41]

    European Conference on Computer Vision , pages=

    Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  34. [42]

    Proceedings of the computer vision and pattern recognition conference , pages=

    Omnidrive: A holistic vision-language dataset for autonomous driving with counterfactual reasoning , author=. Proceedings of the computer vision and pattern recognition conference , pages=

  35. [43]

    arXiv preprint arXiv:2310.01415 , year=

    Gpt-driver: Learning to drive with gpt , author=. arXiv preprint arXiv:2310.01415 , year=

    arXiv

  36. [44]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Lmdrive: Closed-loop end-to-end driving with large language models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  37. [45]

    arXiv preprint arXiv:2410.22313 , year=

    Senna: Bridging large vision-language models and end-to-end autonomous driving , author=. arXiv preprint arXiv:2410.22313 , year=

    arXiv

  38. [46]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Drivegpt4-v2: Harnessing large language model capabilities for enhanced closed-loop autonomous driving , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  39. [47]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Drivinggpt: Unifying driving world modeling and planning with multi-modal autoregressive transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  40. [48]

    arXiv preprint arXiv:2510.12796 , year=

    DriveVLA-W0: World models amplify data scaling law in autonomous driving , author=. arXiv preprint arXiv:2510.12796 , year=

    arXiv

  41. [49]

    Advances in neural information processing systems , volume=

    Argmax flows and multinomial diffusion: Learning categorical distributions , author=. Advances in neural information processing systems , volume=

  42. [50]

    Advances in neural information processing systems , volume=

    Diffusion-lm improves controllable text generation , author=. Advances in neural information processing systems , volume=

  43. [51]

    arXiv preprint arXiv:2210.08933 , year=

    Diffuseq: Sequence to sequence text generation with diffusion models , author=. arXiv preprint arXiv:2210.08933 , year=

    arXiv

  44. [52]

    Advances in neural information processing systems , volume=

    Simplified and generalized masked diffusion for discrete data , author=. Advances in neural information processing systems , volume=

  45. [53]

    arXiv preprint arXiv:2406.03736 , year=

    Your absorbing discrete diffusion secretly models the conditional distributions of clean data , author=. arXiv preprint arXiv:2406.03736 , year=

    arXiv

  46. [54]

    arXiv preprint arXiv:2503.09573 , year=

    Block diffusion: Interpolating between autoregressive and diffusion language models , author=. arXiv preprint arXiv:2503.09573 , year=

    arXiv

  47. [55]

    arXiv preprint arXiv:2503.04482 , year=

    Generalized interpolating discrete diffusion , author=. arXiv preprint arXiv:2503.04482 , year=

    arXiv

  48. [56]

    arXiv preprint arXiv:2501.15564 , year=

    Diffusion-based planning for autonomous driving with flexible guidance , author=. arXiv preprint arXiv:2501.15564 , year=

    arXiv

  49. [57]

    arXiv preprint arXiv:2602.20577 , year=

    Efficient and Explainable End-to-End Autonomous Driving via Masked Vision-Language-Action Diffusion , author=. arXiv preprint arXiv:2602.20577 , year=

  50. [58]

    arXiv preprint arXiv:2508.20072 , year=

    Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies , author=. arXiv preprint arXiv:2508.20072 , year=

    arXiv

  51. [59]

    arXiv preprint arXiv:2509.06932 , year=

    Llada-vla: Vision language diffusion action models , author=. arXiv preprint arXiv:2509.06932 , year=

    arXiv

  52. [60]

    arXiv preprint arXiv:2509.25681 , year=

    dvla: Diffusion vision-language-action model with multimodal chain-of-thought , author=. arXiv preprint arXiv:2509.25681 , year=

  53. [61]

    arXiv preprint arXiv:2603.25406 , year=

    MMaDA-VLA: Large Diffusion Vision-Language-Action Model with Unified Multi-Modal Instruction and Generation , author=. arXiv preprint arXiv:2603.25406 , year=

  54. [62]

    arXiv preprint arXiv:2603.26320 , year=

    DFM-VLA: Iterative Action Refinement for Robot Manipulation via Discrete Flow Matching , author=. arXiv preprint arXiv:2603.26320 , year=

    arXiv

  55. [63]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    arXiv

  56. [64]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the knowledge in a neural network , author=. arXiv preprint arXiv:1503.02531 , year=

    arXiv

  57. [65]

    Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    Sequence-level knowledge distillation , author=. Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    2016

  58. [66]

    arXiv preprint arXiv:1910.01108 , year=

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter , author=. arXiv preprint arXiv:1910.01108 , year=

    1910 arXiv

  59. [67]

    arXiv preprint arXiv:1511.06732 , year=

    Sequence level training with recurrent neural networks , author=. arXiv preprint arXiv:1511.06732 , year=

    arXiv

  60. [68]

    arXiv preprint arXiv:2604.19710 , year=

    SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery Samples for Vision-Language-Action Model , author=. arXiv preprint arXiv:2604.19710 , year=

    arXiv

  61. [69]

    arXiv preprint arXiv:2507.04049 , year=

    DIVER: Reinforced Diffusion Breaks Imitation Bottlenecks in End-to-End Autonomous Driving , author=. arXiv preprint arXiv:2507.04049 , year=

    arXiv

  62. [70]

    arXiv preprint arXiv:2512.07745 , year=

    DiffusionDriveV2: Reinforcement learning-constrained truncated diffusion modeling in end-to-end autonomous driving , author=. arXiv preprint arXiv:2512.07745 , year=

  63. [71]

    arXiv preprint arXiv:2603.06049 , year=

    Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models , author=. arXiv preprint arXiv:2603.06049 , year=

  64. [72]

    arXiv preprint arXiv:2604.12656 , year=

    FeaXDrive: Feasibility-aware Trajectory-Centric Diffusion Planning for End-to-End Autonomous Driving , author=. arXiv preprint arXiv:2604.12656 , year=

    arXiv

  65. [73]

    arXiv preprint arXiv:2408.03326 , year=

    Llava-onevision: Easy visual task transfer , author=. arXiv preprint arXiv:2408.03326 , year=

    arXiv

  66. [74]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    arXiv

  67. [75]

    arXiv preprint arXiv:2412.05271 , year=

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling , author=. arXiv preprint arXiv:2412.05271 , year=

    arXiv

  68. [76]

    Advances in Neural Information Processing Systems , volume=

    Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline , author=. Advances in Neural Information Processing Systems , volume=

  69. [77]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Think twice before driving: Towards scalable decoders for end-to-end autonomous driving , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  70. [78]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  71. [79]

    arXiv preprint arXiv:2305.10430 , year=

    Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes , author=. arXiv preprint arXiv:2305.10430 , year=

    arXiv

  72. [80]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Drivesuprim: Towards precise trajectory selection for end-to-end planning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  73. [81]

    arXiv preprint arXiv:2601.05083 , year=

    Driving on Registers , author=. arXiv preprint arXiv:2601.05083 , year=

  74. [82]

    arXiv preprint arXiv:2406.06978 , year=

    Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation , author=. arXiv preprint arXiv:2406.06978 , year=

    arXiv

  75. [83]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Planning-oriented autonomous driving , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  76. [84]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Vad: Vectorized scene representation for efficient autonomous driving , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  77. [85]

    arXiv preprint arXiv:2503.07656 , year=

    Drivetransformer: Unified transformer for scalable end-to-end autonomous driving , author=. arXiv preprint arXiv:2503.07656 , year=

    arXiv

  78. [86]

    arXiv preprint arXiv:2506.08052 , year=

    Recogdrive: A reinforced cognitive framework for end-to-end autonomous driving , author=. arXiv preprint arXiv:2506.08052 , year=

    arXiv

  79. [87]

    arXiv preprint arXiv:2602.18887 , year=

    SafeDrive: Fine-Grained Safety Reasoning for End-to-End Driving in a Sparse World , author=. arXiv preprint arXiv:2602.18887 , year=

  80. [88]

    arXiv preprint arXiv:2503.12170 , year=

    Diffad: A unified diffusion modeling approach for autonomous driving , author=. arXiv preprint arXiv:2503.12170 , year=

    arXiv

  81. [89]

    arXiv preprint arXiv:2604.16514 , year=

    BARD: Bridging AutoRegressive and Diffusion Vision-Language Models Via Highly Efficient Progressive Block Merging and Stage-Wise Distillation , author=. arXiv preprint arXiv:2604.16514 , year=

    arXiv

  82. [90]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    arXiv

  83. [91]

    arXiv preprint arXiv:2502.14786 , year=

    Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features , author=. arXiv preprint arXiv:2502.14786 , year=

    arXiv

  84. [92]

    2025 , eprint=

    Qwen2.5-VL Technical Report , author=. 2025 , eprint=

    2025

  85. [93]

    2025 , eprint=

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models , author=. 2025 , eprint=

    2025

  86. [94]

    2025 , eprint=

    Percept-WAM: Perception-Enhanced World-Awareness-Action Model for Robust End-to-End Autonomous Driving , author=. 2025 , eprint=

    2025

  87. [95]

    2023 , eprint=

    Planning-oriented Autonomous Driving , author=. 2023 , eprint=

    2023

  88. [96]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Para-drive: Parallelized architecture for real-time autonomous driving , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  89. [97]

    2022 , eprint=

    TransFuser: Imitation with Transformer-Based Sensor Fusion for Autonomous Driving , author=. 2022 , eprint=

    2022

  90. [98]

    arXiv preprint arXiv:2408.03601 , year=

    Drama: An efficient end-to-end motion planner for autonomous driving with mamba , author=. arXiv preprint arXiv:2408.03601 , year=

    arXiv

  91. [99]

    2026 , eprint=

    VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning , author=. 2026 , eprint=

    2026

  92. [100]

    arXiv preprint arXiv:2302.13971 , year=

    Llama: Open and efficient foundation language models , author=. arXiv preprint arXiv:2302.13971 , year=

    arXiv

  93. [101]

    IEEE Robotics and Automation Letters , volume=

    Artemis: Autoregressive end-to-end trajectory planning with mixture of experts for autonomous driving , author=. IEEE Robotics and Automation Letters , volume=. 2025 , publisher=

    2025

  94. [102]

    arXiv preprint arXiv:2509.13769 , year=

    Adathinkdrive: Adaptive thinking via reinforcement learning for autonomous driving , author=. arXiv preprint arXiv:2509.13769 , year=

  95. [103]

    Advances in Neural Information Processing Systems , volume=

    Generalized jensen-shannon divergence loss for learning with noisy labels , author=. Advances in Neural Information Processing Systems , volume=

  96. [104]

    arXiv preprint arXiv:2410.14655 , year=

    Bridging the training-inference gap in llms by leveraging self-generated tokens , author=. arXiv preprint arXiv:2410.14655 , year=

    arXiv

  97. [105]

    arXiv preprint arXiv:2403.06963 , year=

    The pitfalls of next-token prediction , author=. arXiv preprint arXiv:2403.06963 , year=

    arXiv

  98. [106]

    arXiv preprint arXiv:2004.12817 , year=

    LightPAFF: A two-stage distillation framework for pre-training and fine-tuning , author=. arXiv preprint arXiv:2004.12817 , year=

    2004 arXiv

  99. [107]

    Neurocomputing , volume=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  100. [108]

    arXiv preprint arXiv:2503.19779 , year=

    Pygraph: Robust compiler support for cuda graphs in pytorch , author=. arXiv preprint arXiv:2503.19779 , year=

  101. [109]

    Proceedings of Machine Learning and Systems , volume=

    Flashinfer: Efficient and customizable attention engine for llm inference serving , author=. Proceedings of Machine Learning and Systems , volume=

  102. [110]

    2025 , eprint=

    DiffusionDriveV2: Reinforcement Learning-Constrained Truncated Diffusion Modeling in End-to-End Autonomous Driving , author=. 2025 , eprint=

    2025

  103. [111]

    2024 , eprint=

    Benchmarking Object Detectors with COCO: A New Path Forward , author=. 2024 , eprint=

    2024

  104. [112]

    arXiv preprint arXiv:2512.13636 , year=

    MindDrive: A Vision-Language-Action Model for Autonomous Driving via Online Reinforcement Learning , author=. arXiv preprint arXiv:2512.13636 , year=

    arXiv

  105. [113]

    arXiv preprint arXiv:2606.12396 , year=

    VLGA: Vision-Language-Geometry-Action Models for Autonomous Driving , author=. arXiv preprint arXiv:2606.12396 , year=

    arXiv

  106. [114]

    arXiv preprint arXiv:2605.31271 , year=

    DriveMA: Driving Vision-Language-Action Models with verifiable Meta-Actions , author=. arXiv preprint arXiv:2605.31271 , year=

    arXiv

  107. [115]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Nord: A data-efficient vision-language-action model that drives without reasoning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  108. [116]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Drivemoe: Mixture-of-experts for vision-language-action model in end-to-end autonomous driving , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  109. [117]

    arXiv preprint arXiv:2601.05640 , year=

    SGDrive: Scene-to-Goal Hierarchical World Cognition for Autonomous Driving , author=. arXiv preprint arXiv:2601.05640 , year=

  110. [118]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Resad: Normalized residual trajectory modeling for end-to-end autonomous driving , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  111. [119]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Simlingo: Vision-only closed-loop autonomous driving with language-action alignment , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.