acceptodds
Under review as a conference paper at ICLR 2027

More Camera Frames, Better Driving? Within-Scene Information Scaling in Autonomous Driving

Abstract

Autonomous driving requires models to learn vehicle motion, agent interactions, and scene evolution from driving data. Existing scaling efforts focus on increasing model capacity and expanding driving data coverage, largely treating driving scenes as the unit of training information. However, how the information exposed within each scene should scale with model capacity remains underexplored. We study within-scene information scaling, using temporal sampling frequency as a controlled probe to vary camera-frame density within fixed driving scenes. We formulate a minimal information–capacity model that captures the interaction among bounded driving-relevant information, increasing driving-irrelevant exposure, and finite model capacity. The model characterizes the capacity-dependent scaling behavior of within-scene information and predicts that the preferred temporal sampling frequency increases with model capacity under its assumptions. We conduct controlled frequency sweeps across compact end-to-end models, large vision-language-action (VLA) models, and world-action models on Waymo, nuScenes, and PAVE and in NAVSIM-based settings, with additional iteration-matched comparisons. AutoVLA and DriveVLA-W0 attain their lowest reported nuScenes trajectory errors at the highest tested frequency, yet further reductions above 4 Hz are much smaller for DriveVLA-W0. Drive-JEPA achieves strong driving performance on NAVTEST already at 2 Hz. These findings distinguish how much a model improves with denser temporal exposure from how well it already performs under sparse temporal exposure. They motivate evaluating driving performance alongside frequency-response curves and jointly developing model architectures, learning objectives, and within-scene data construction to utilize driving-relevant information that affects the driving decision while limiting the influence of driving-irrelevant exposure.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Information Utilization, Model Capacity, Temporal Redundancy, Information Scaling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_Aow2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies within-scene information scaling in autonomous driving: how increasing the density of training timestamps within fixed driving scenes affects performance while keeping each model’s input construction and prediction target unchanged. It proposes a phenomenological information–capacity model that predicts a higher optimal sampling frequency with greater model capacity under its assumptions. Empirical validation includes frequency sweeps across compact end-to-end, vision-language-action, and world-action models on Waymo, nuScenes, PAVE, and NAVSIM, supplemented by iteration-matched controls for compact models and adaptive-sampling comparisons. Results show model- and dataset-dependent improvements, plateaus, or degradation as sampling becomes denser, motivating evaluation of full frequency-response curves alongside absolute driving performance.

Critical strengths
  • This paper is well-written and easy to follow.
  • The topic is quite worth studying, and the paper is informative.
  • Some experiments are sufficient.
Critical weaknesses
  • The central capacity-dependent scaling claim is not adequately validated. Section 3 derives an increasing optimal sampling frequency with model capacity under a phenomenological information–capacity model. However, its information and capacity quantities are neither measured nor calibrated against the experiments. The empirical comparisons change architecture, pretraining, inputs, and supervision alongside parameter count, without varying capacity within a controlled model family. Consequently, the results establish model-dependent frequency responses but do not isolate capacity as their cause. This weakens the empirical support for the paper’s central theoretical contribution.

  • The optimization-budget confound remains unresolved for the large-model results. The main open-loop sweeps fix epochs, so denser sampling also produces more optimization updates (Section 4). The iteration-matched controls in Table 5 cover only the compact models, leaving AutoVLA’s gains without an equivalent control. Its apparent preference for denser sampling could therefore partly reflect a need for additional optimization rather than a greater ability to exploit within-scene information. Since the contrast between compact models and AutoVLA is important evidence for the scaling interpretation, this missing control materially limits the conclusions.

  • The proposed mechanism of irrelevant exposure competing for capacity is assumed rather than demonstrated. The formulation groups redundant content and task-irrelevant visual variation into a linearly increasing exposure term and imposes a penalty through Equation 2. Yet repeated task-relevant examples need not have the same effect as distracting information, and the experiments do not separate these factors. Performance saturation or degradation alone cannot distinguish the proposed mechanism from alternatives involving sample weighting or optimization. Without targeted interventions or measurements, the theory remains a plausible explanation rather than an experimentally supported account of why denser sampling helps or hurts.

  • Important experimental details are insufficiently specified for reproducibility and interpretation. Appendix C does not identify the exact DriveVLA-W0 variant used or fully specify the Drive-JEPA training configuration. The adaptive 4 Hz setting lacks an operational sampling rule, while Appendix D acknowledges the need for scene identifiers, timestamp manifests, and model-specific validity rules without providing those details in the manuscript. Moreover, the AutoVLA configuration omits the original method’s CoT and reinforcement fine-tuning, limiting the scope of conclusions about that method. These omissions are consequential because the paper’s contribution depends primarily on controlled empirical comparisons: readers need to reconstruct the sampled training sets and model configurations to assess whether the reported differences arise from sampling frequency.

Questions for authors

Please refer to weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Aow2 (1/2)

W1. The central capacity-dependent scaling claim is not adequately validated. […] The empirical comparisons change architecture, pretraining, inputs, and supervision alongside parameter count, without varying capacity within a controlled model family.

AW1. We have added a within-family E2EDriver [1] capacity sweep in Table 8 (Lines 873–890) ↗.

Capacity (width)HzADE (m) ↓FDE (m) ↓EPDMS (%) ↑
6420.5281±0.01391.1749±0.003769.0919±0.1539
6440.4674±0.01561.0369±0.003773.4165±0.1795
6460.4434±0.00220.9847±0.001375.8087±0.0989
6480.4529±0.00080.9975±0.002175.0123±0.1051
64100.4454±0.00100.9893±0.001677.7761±0.1348
51220.395±0.0140.919±0.01376.888±0.298
51240.382±0.0060.893±0.00579.138±0.308
51260.391±0.0240.898±0.03080.079±0.570
51280.371±0.0010.876±0.00681.558±0.078
512100.373±0.0080.879±0.00781.343±0.059
1,02420.4533±0.00001.1282±0.000566.5062±0.0445
1,02440.4404±0.00021.0967±0.003367.8559±0.0049
1,02460.4320±0.00081.0764±0.002869.0045±0.0989
1,02480.4258±0.00191.0575±0.004469.6615±0.1948
1,024100.4223±0.00051.0476±0.001270.0125±0.1161

Within the E2EDriver family, increasing model capacity raises the preferred temporal sampling frequency for ADE/FDE (Lines 361–368 ↗). This directly supports the capacity-dependent prediction in Section 3 ↗ and Appendix B ↗.

Table 7 (Lines 865–870) ↗ lists widths and parameter counts. Only width and its adapter change. All variants share the pretrained backbone, inputs, objective and training protocol (Lines 893–897 ↗). Values are NAVTEST means ± sample SD over three seeds.


W2. The optimization-budget confound remains unresolved for the large-model results. […] Its apparent preference for denser sampling could therefore partly reflect a need for additional optimization rather than a greater ability to exploit within-scene information.

AW2. We have added iteration-matched controls for AutoVLA [2] and DriveVLA-W0 [3] in Table 5 (Lines 487–496) ↗.

ModelHzEpochsEPDMS (%) ↑ADE (m) ↓FDE (m) ↓
AutoVLA22577.881±0.0550.515959±0.0005751.166427±0.001166
AutoVLA10581.135±0.0850.485±0.0001.103±0.001
DriveVLA-W022086.787±0.0510.6198±0.00121.2016±0.0009
DriveVLA-W010487.366±0.0930.6194±0.0031.2053±0.005

The iteration-matched controls show that denser temporal sampling exposes additional driving-relevant information for AutoVLA. More optimization updates on the same sparse anchors cannot fully recover that benefit. DriveVLA-W0 recovers much of its EPDMS gap through longer training on sparse anchors.

AutoVLA uses answer-only SFT (batch 32) and DriveVLA-W0 uses Emu3Pi0 (batch 96). Values are NAVTEST means ± sample SD over three seeds.

We completed the 4, 6 and 8 Hz runs for both models after submission and before the rebuttal period. We also retrained DriveVLA-W0 at 10 Hz before the rebuttal period following data loss. This update is documented in Lines 430–433 ↗.

[1] E2EDriver. Mufti. GitHub repository (accessed 2026).

[2] AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning. Zhou et al. NeurIPS'25.

[3] DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving. Li et al. ICLR'26.

AuthorsOpenReview ↗

Response to Reviewer Aow2 (2/2)

W3. The proposed mechanism of irrelevant exposure competing for capacity is assumed rather than demonstrated. […] Yet repeated task-relevant examples need not have the same effect as distracting information, and the experiments do not separate these factors.

AW3. We have added a targeted driving-irrelevant exposure intervention at fixed E2EDriver [1] capacity in Table 10 (Lines 983–998) ↗.

Training inputsHzEPDMS (%) ↑ADE (m) ↓FDE (m) ↓
Original479.138±0.3080.382±0.0060.893±0.005
Original680.079±0.5700.391±0.0240.898±0.030
Original881.558±0.0780.371±0.0010.876±0.006
Original1081.343±0.0590.373±0.0080.879±0.007
Image noise479.9585±0.72490.3743±0.00290.8794±0.0081
Image noise680.7704±0.22530.3721±0.00060.8782±0.0018
Image noise880.9530±0.41140.3739±0.00160.8821±0.0015
Image noise1081.3080±0.04920.3722±0.00210.8789±0.0057
Image + state noise478.6778±0.19640.6546±0.01281.3413±0.0213
Image + state noise678.7034±0.10690.6673±0.00881.3628±0.0151
Image + state noise879.2125±0.12570.6924±0.00661.4010±0.0120
Image + state noise1078.8919±0.25940.7219±0.00061.4552±0.0039

At fixed capacity, additional driving-irrelevant exposure reduces the benefit of denser temporal sampling for trajectory prediction. This supports finite-capacity competition (Section 3 ↗). For DriveVLA-W0, additional training on the same sparse, task-relevant anchors improves driving performance (Table 5 ↗).

All conditions use width 512, unchanged targets and clean NAVTEST evaluation (mean ± sample SD over three seeds).


W4. Appendix C does not identify the exact DriveVLA-W0 variant used or fully specify the Drive-JEPA training configuration. The adaptive 4 Hz setting lacks an operational sampling rule […] Moreover, the AutoVLA configuration omits the original method’s CoT and reinforcement fine-tuning, limiting the scope of conclusions about that method.

AW4(1/3). We have detailed the models in Appendix C (Lines 929–933 and 939–943) ↗ and their training budgets in Table 11 (Lines 1027–1033) ↗.

ModelEvaluated configurationEpochs in Table 4
Drive-JEPA [2]Pose-only. Released V-JEPA [3] ViT-L encoder, newly initialized planner, L1 ego-trajectory loss.20
E2EDriverOriginal width 512, trainable SqueezeNet 1.1 backbone.10
AutoVLA [4]Qwen2.5-VL-3B-Instruct [5], answer-only SFT. CoT supervision and reinforcement fine-tuning are omitted.5
DriveVLA-W0 [6]Emu3Pi0. NuPlan-pretrained Emu3 with the flow-matching action expert and trainable VLM backbone.4

DriveVLA-W0 uses downstream action prediction. Drive-JEPA retains score selection and uses pose-only supervision because NAVSIM-10Hz lacks PDM scores. Table 4 ↗ uses final-epoch checkpoints.


AW4(2/3). We have specified the adaptive 4 Hz sampling rule in Appendix D (Lines 1086–1092) ↗.

About half the anchors preserve uniform coverage. We select the rest using eight normalized measures of the past 1.5 s of ego motion, relaxing 0.125 s spacing only to fill the quota. A4 yields 180,304 AutoVLA and 214,856 DriveVLA-W0 samples (Lines 1096–1097 ↗).

Lines 1094–1099 ↗ list the sampling scripts and retained scene/log IDs, image IDs, timestamps, validity records, sample lists and normalization values.


AW4(3/3). We have specified AutoVLA's answer-only SFT (Lines 922–923 ↗) without CoT supervision or reinforcement fine-tuning (Lines 1037–1039 ↗).

Our results concern this setup.

[1] E2EDriver. Mufti. GitHub repository (accessed 2026).

[2] Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to-End Driving. Wang et al. arXiv'26, 2601.22032.

[3] Revisiting Feature Prediction for Learning Visual Representations from Video. Bardes et al. TMLR'24.

[4] AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning. Zhou et al. NeurIPS'25.

[5] Qwen2.5-VL Technical Report. Bai et al. arXiv'25, 2502.13923.

[6] DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving. Li et al. ICLR'26.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you for your time and thoughtful feedback.

Reviewer_C5q1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates within-scene information scaling in autonomous driving, focusing on how temporal camera-frame sampling frequency affects driving performance across different model architectures and capacities. The authors propose a simplified information-capacity model that assumes driving-relevant information saturates with increasing sampling frequency, while driving-irrelevant exposure continues to grow. Based on this formulation, they argue that the preferred sampling frequency increases with model capacity. Experiments are conducted across Waymo, nuScenes, PAVE, and NAVSIM using compact end-to-end models, VLA models, and world-action models. The results reveal model-dependent frequency-response patterns, with some models benefiting from denser temporal sampling while others achieve comparable performance at lower frequencies.

Critical strengths
  1. The paper keeps the underlying driving scenes fixed within each frequency sweep, enabling comparisons across different training data densities.
  2. The experiments cover multiple datasets and model architectures, providing a relatively broad empirical comparison.
Critical weaknesses
  1. Insufficient validation of the central claim. The paper predicts that the preferred sampling frequency increases with model capacity, but the empirical comparisons involve models with different architectures, input representations, and training objectives. Without varying model capacity within a controlled architecture and training setup, the experiments do not isolate the effect of capacity on the preferred sampling frequency.

  2. Weak connection between theory and experiments. The theoretical analysis assumes that driving-relevant information saturates with sampling frequency, while driving-irrelevant exposure grows approximately linearly. However, these quantities are phenomenological and are not empirically measured. The observed frequency-response curves therefore do not directly validate the proposed information–capacity mechanism.

  3. Incomplete experimental controls and ambiguous interpretation of sampling frequency. Increasing sampling frequency changes both training-anchor density and the number of optimization updates. Although iteration-matched controls are provided for compact models, equivalent controls are missing for the larger models. Moreover, for models whose per-sample visual input contains only a current frame or a fixed temporal context, denser training anchors do not necessarily increase the temporal information available in each input. This complicates the interpretation of sampling frequency as a consistent measure of within-scene information scaling.

Questions for authors

See Weaknesses.

7 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer C5q1 (1/2)

W1. Insufficient validation of the central claim. […] Without varying model capacity within a controlled architecture and training setup, the experiments do not isolate the effect of capacity on the preferred sampling frequency.

AW1(1/2). We have added a within-family E2EDriver [1] capacity sweep in Table 8 (Lines 873–890) ↗.

Capacity (width)HzADE (m) ↓FDE (m) ↓EPDMS (%) ↑
6420.5281±0.01391.1749±0.003769.0919±0.1539
6440.4674±0.01561.0369±0.003773.4165±0.1795
6460.4434±0.00220.9847±0.001375.8087±0.0989
6480.4529±0.00080.9975±0.002175.0123±0.1051
64100.4454±0.00100.9893±0.001677.7761±0.1348
51220.395±0.0140.919±0.01376.888±0.298
51240.382±0.0060.893±0.00579.138±0.308
51260.391±0.0240.898±0.03080.079±0.570
51280.371±0.0010.876±0.00681.558±0.078
512100.373±0.0080.879±0.00781.343±0.059
1,02420.4533±0.00001.1282±0.000566.5062±0.0445
1,02440.4404±0.00021.0967±0.003367.8559±0.0049
1,02460.4320±0.00081.0764±0.002869.0045±0.0989
1,02480.4258±0.00191.0575±0.004469.6615±0.1948
1,024100.4223±0.00051.0476±0.001270.0125±0.1161

Within the E2EDriver family, increasing model capacity raises the preferred temporal sampling frequency for ADE/FDE, as our model predicts.

Table 7 (Lines 865–870) ↗ lists widths and parameter counts. Only width and its adapter change. All variants share the pretrained backbone, inputs, objective and training protocol (Lines 893–897 ↗). Values are NAVTEST means ± sample SD over three seeds.


AW1(2/2). We have added a structured-input HiVT [2] sweep in Table 9 (Lines 973–980) ↗.

HzADE (m) ↓FDE (m) ↓
20.5847±0.00191.6629±0.0018
40.5363±0.00211.5507±0.0024
60.5185±0.00111.4914±0.0019
80.5305±0.00081.5345±0.0036
100.5018±0.00141.4571±0.0021

HiVT shows that even a compact model can benefit from denser temporal sampling when structured inputs directly provide driving-relevant information.

HiVT-128 has 2.56M trainable parameters and uses agent histories, vectorized lanes, traffic signals and an oracle route command on a curated Waymo Open Motion Dataset subset. Values are three-second errors (mean ± reported SD).

[1] E2EDriver. Mufti. GitHub repository (accessed 2026).

[2] HiVT: Hierarchical Vector Transformer for Multi-Agent Motion Prediction. Zhou et al. CVPR'22.

AuthorsOpenReview ↗

Rebuttal to Reviewer C5q1 (2/2)

W2. Weak connection between theory and experiments. […] The observed frequency-response curves therefore do not directly validate the proposed information–capacity mechanism.

AW2. We have added a targeted driving-irrelevant exposure intervention at fixed E2EDriver [1] capacity in Table 10 (Lines 983–998) ↗.

Training inputsHzEPDMS (%) ↑ADE (m) ↓FDE (m) ↓
Original479.138±0.3080.382±0.0060.893±0.005
Original680.079±0.5700.391±0.0240.898±0.030
Original881.558±0.0780.371±0.0010.876±0.006
Original1081.343±0.0590.373±0.0080.879±0.007
Image noise479.9585±0.72490.3743±0.00290.8794±0.0081
Image noise680.7704±0.22530.3721±0.00060.8782±0.0018
Image noise880.9530±0.41140.3739±0.00160.8821±0.0015
Image noise1081.3080±0.04920.3722±0.00210.8789±0.0057
Image + state noise478.6778±0.19640.6546±0.01281.3413±0.0213
Image + state noise678.7034±0.10690.6673±0.00881.3628±0.0151
Image + state noise879.2125±0.12570.6924±0.00661.4010±0.0120
Image + state noise1078.8919±0.25940.7219±0.00061.4552±0.0039

At fixed model capacity, additional driving-irrelevant exposure lowers the preferred temporal sampling frequency for ADE/FDE. Increasing model capacity shifts that frequency upward (Table 8 ↗). Together, these interventions support finite-capacity competition in our model (Section 3 ↗).

We add noise to camera images alone or to both images and ego-state history. All conditions use width 512, ten epochs per frequency, unchanged targets and clean NAVTEST evaluation. Values are means ± sample SD over three seeds.


W3. Incomplete experimental controls and ambiguous interpretation of sampling frequency. […] Although iteration-matched controls are provided for compact models, equivalent controls are missing for the larger models. Moreover, for models whose per-sample visual input contains only a current frame or a fixed temporal context, denser training anchors do not necessarily increase the temporal information available in each input.

AW3(1/2). We have added iteration-matched controls for AutoVLA [2] and DriveVLA-W0 [3] in Table 5 (Lines 487–496) ↗.

ModelHzEpochsEPDMS (%) ↑ADE (m) ↓FDE (m) ↓
AutoVLA22577.881±0.0550.515959±0.0005751.166427±0.001166
AutoVLA10581.135±0.0850.485±0.0001.103±0.001
DriveVLA-W022086.787±0.0510.6198±0.00121.2016±0.0009
DriveVLA-W010487.366±0.0930.6194±0.0031.2053±0.005

The iteration-matched controls show that denser temporal sampling exposes additional driving-relevant information for AutoVLA. More optimization updates on the same sparse anchors cannot fully recover that benefit. DriveVLA-W0 recovers much of its EPDMS gap through longer training on sparse anchors.

AutoVLA uses answer-only SFT (batch 32) and DriveVLA-W0 uses Emu3Pi0 (batch 96). Within each model, the initialization protocol, input format and optimizer stay fixed. Values are NAVTEST means ± sample SD over three seeds.

We completed the 4, 6 and 8 Hz runs for both models after submission and before the rebuttal period. We also retrained DriveVLA-W0 at 10 Hz before the rebuttal period following data loss. See Lines 430–433 ↗.


AW3(2/2). Thank you for highlighting this distinction. We vary training-anchor density within fixed driving scenes. Each model’s per-input temporal context and prediction target stay fixed (Section 4 (Lines 296–300) ↗). Denser anchors expose additional scene timestamps and their supervision during training (Appendix D (Lines 1070–1076) ↗).

AutoVLA benefits from those additional timestamps at matched updates with the same inference inputs. We considered adding historical frames, which would change the inference-time input requirements and could increase latency.

[1] E2EDriver. Mufti. GitHub repository (accessed 2026).

[2] AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning. Zhou et al. NeurIPS'25.

[3] DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving. Li et al. ICLR'26.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you for your time and thoughtful feedback.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Author Rebuttal by Authors

We sincerely thank all the reviewers for their time and their thoughtful comments and questions.

  • Our research topic is quite worth studying, and the paper is informative, well-written and easy to follow (Aow2). We study a useful data construction question, and our frequency sweeps provide practical evidence(Wq13). Our experiments cover multiple datasets and model architectures, providing a relatively broad empirical comparison (C5q1).
  • We have made the following contributions.
    • We have proposed a minimal information–capacity model for within-scene information scaling in Section 3 and Appendix B (C5q1, Aow2).
    • We have added a within-family E2EDriver capacity sweep in Tables 7 and 8 (C5q1, Wq13, Aow2).
    • We have added a targeted driving-irrelevant exposure intervention at fixed capacity in Table 10 (C5q1, Aow2).
    • We have added iteration-matched controls for AutoVLA and DriveVLA-W0 in Table 5 (C5q1, Wq13, Aow2).
  • Reporting both trajectory errors and NAVSIM scores is valuable(Wq13).

We have done our best to address the reviewers’ questions. We believe their comments and the resulting revisions have made the paper stronger, and we thank all the reviewers for their help.

AuthorsOpenReview ↗

Response to Reviewer Wq13 (1/2)

W1. The experiments do not isolate model capacity as the reason for the different frequency responses. […] Meanwhile, the relevant information, irrelevant exposure, and effective capacity in the theoretical model are not measured.

AW1. We have varied E2EDriver [1] width within one architecture in Table 8 (Lines 873–890) ↗ and added a targeted driving-irrelevant exposure intervention at fixed width in Table 10 (Lines 983–998) ↗, reproduced below.

Training conditionTraining frequencyEPDMS (%) ↑3-s ADE (m) ↓3-s FDE (m) ↓
Original training4 Hz79.138±0.3080.382±0.0060.893±0.005
Original training6 Hz80.079±0.5700.391±0.0240.898±0.030
Original training8 Hz81.558±0.0780.371±0.0010.876±0.006
Original training10 Hz81.343±0.0590.373±0.0080.879±0.007
Image-only noise4 Hz79.9585±0.72490.3743±0.00290.8794±0.0081
Image-only noise6 Hz80.7704±0.22530.3721±0.00060.8782±0.0018
Image-only noise8 Hz80.9530±0.41140.3739±0.00160.8821±0.0015
Image-only noise10 Hz81.3080±0.04920.3722±0.00210.8789±0.0057
Image + state noise4 Hz78.6778±0.19640.6546±0.01281.3413±0.0213
Image + state noise6 Hz78.7034±0.10690.6673±0.00881.3628±0.0151
Image + state noise8 Hz79.2125±0.12570.6924±0.00661.4010±0.0120
Image + state noise10 Hz78.8919±0.25940.7219±0.00061.4552±0.0039

Across the three E2EDriver variants, we keep the pretrained backbone, inputs, objective and training protocol fixed. Only the internal width and its adapter change. Increasing width raises the preferred temporal sampling frequency for ADE/FDE. At fixed width, adding image noise lowers it even with intact ego-state history. These complementary interventions test the capacity-dependent prediction and support finite-capacity competition in Section 3 ↗.

We held E2EDriver width fixed at 512 and evaluated all conditions on clean NAVTEST inputs. Values are means ± sample SD over three seeds. Bold marks the best mean for each metric within each condition.


W2. Training budget remains a confounding factor in the large-model results. The main open-loop sweeps fix the number of epochs, so higher sampling frequencies also receive more optimization updates.

AW2. We have added NAVTEST iteration-matched controls for AutoVLA [2] and DriveVLA-W0 [3] in Table 5 (Lines 487–496) ↗.

ModelHzEpochsEPDMS (%) ↑ADE (m) ↓FDE (m) ↓
AutoVLA22577.881±0.0550.515959±0.0005751.166427±0.001166
AutoVLA10581.135±0.0850.485±0.0001.103±0.001
DriveVLA-W022086.787±0.0510.6198±0.00121.2016±0.0009
DriveVLA-W010487.366±0.0930.6194±0.0031.2053±0.005

Iteration-matched AutoVLA controls show that denser temporal sampling exposes additional driving-relevant information that additional optimization on sparse data cannot fully recover. Comparing the fixed-epoch baseline in Table 4 ↗ with Table 5 shows that DriveVLA-W0 recovers much of its EPDMS gap through longer training on the same sparse anchors.

The main open-loop sweeps fix epochs, so denser sampling increases both training-anchor density and optimization updates (Lines 301–303 ↗). Table 5 separates these effects on NAVTEST.

AutoVLA uses answer-only SFT (batch 32) and DriveVLA-W0 uses Emu3Pi0 (batch 96). Values are NAVTEST means ± sample SD over three seeds.

[1] E2EDriver. Mufti. GitHub repository (accessed 2026).

[2] AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning. Zhou et al. NeurIPS'25.

[3] DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving. Li et al. ICLR'26.

AuthorsOpenReview ↗

Response to Reviewer Wq13 (2/2)

Q1. Do you have a capacity sweep within the same model family, with the architecture design, pretraining, and objective otherwise controlled? This would help determine whether the observed frequency response is actually driven by capacity.

AQ1. Yes. We have added a within-family E2EDriver [1] capacity sweep in Table 8 (Lines 873–890) ↗.

Capacity (width)HzADE (m) ↓FDE (m) ↓EPDMS (%) ↑
6420.5281±0.01391.1749±0.003769.0919±0.1539
6440.4674±0.01561.0369±0.003773.4165±0.1795
6460.4434±0.00220.9847±0.001375.8087±0.0989
6480.4529±0.00080.9975±0.002175.0123±0.1051
64100.4454±0.00100.9893±0.001677.7761±0.1348
51220.395±0.0140.919±0.01376.888±0.298
51240.382±0.0060.893±0.00579.138±0.308
51260.391±0.0240.898±0.03080.079±0.570
51280.371±0.0010.876±0.00681.558±0.078
512100.373±0.0080.879±0.00781.343±0.059
1,02420.4533±0.00001.1282±0.000566.5062±0.0445
1,02440.4404±0.00021.0967±0.003367.8559±0.0049
1,02460.4320±0.00081.0764±0.002869.0045±0.0989
1,02480.4258±0.00191.0575±0.004469.6615±0.1948
1,024100.4223±0.00051.0476±0.001270.0125±0.1161

Within the E2EDriver family, increasing model capacity raises the preferred temporal sampling frequency for ADE/FDE, as our model predicts.

Table 7 (Lines 865–870) ↗ lists widths of 64, 512 and 1,024, corresponding to 0.97M, 12.20M and 46.50M parameters. Only width and its adapter change. All variants share the pretrained backbone, inputs, objective and training protocol (Lines 893–897 ↗). Values are NAVTEST means ± sample SD over three seeds.


Q2. Do the AutoVLA and DriveVLA-W0 trends remain when the number of optimization updates is matched across sampling frequencies? Please also clarify the training-budget policy used for Table 4.

AQ2. We have added iteration-matched controls for AutoVLA [2] and DriveVLA-W0 [3] in Table 5 (Lines 487–496) ↗.

ModelHzEpochsEPDMS (%) ↑ADE (m) ↓FDE (m) ↓
AutoVLA22577.881±0.0550.515959±0.0005751.166427±0.001166
AutoVLA10581.135±0.0850.485±0.0001.103±0.001
DriveVLA-W022086.787±0.0510.6198±0.00121.2016±0.0009
DriveVLA-W010487.366±0.0930.6194±0.0031.2053±0.005

No, the EPDMS gains do not remain unchanged. Both models still score higher at 10 Hz than at 2 Hz after matching updates. Iteration-matched AutoVLA controls show that denser temporal sampling exposes additional driving-relevant information that additional optimization on sparse data cannot fully recover. DriveVLA-W0 recovers much of its EPDMS gap through longer training on the same sparse anchors (Table 4 ↗). Values are NAVTEST means ± sample SD over three seeds.

Table 4 fixes the epoch count within each model and reports the final-epoch checkpoint at every frequency (Section 4 (Lines 301–305) ↗). Table 11 (Lines 1027–1033) ↗ gives the epoch budgets below.

NAVTEST configurationEpochs per frequency
Drive-JEPA (pose-only)20
E2EDriver10
AutoVLA (answer-only SFT)5
DriveVLA-W0 (Emu3Pi0)4

Higher sampling frequencies therefore receive more updates in Table 4. Both large models achieve their best mean EPDMS at the densest uniform sampling frequency tested. After submission and before the rebuttal period, we completed the 4, 6 and 8 Hz runs for both models and retrained DriveVLA-W0 at 10 Hz after data loss (Lines 430–433 ↗).

[1] E2EDriver. Mufti. GitHub repository (accessed 2026).

[2] AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning. Zhou et al. NeurIPS'25.

[3] DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving. Li et al. ICLR'26.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.