- 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
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 inputs | Hz | EPDMS (%) ↑ | ADE (m) ↓ | FDE (m) ↓ |
|---|---|---|---|---|
| Original | 4 | 79.138±0.308 | 0.382±0.006 | 0.893±0.005 |
| Original | 6 | 80.079±0.570 | 0.391±0.024 | 0.898±0.030 |
| Original | 8 | 81.558±0.078 | 0.371±0.001 | 0.876±0.006 |
| Original | 10 | 81.343±0.059 | 0.373±0.008 | 0.879±0.007 |
| Image noise | 4 | 79.9585±0.7249 | 0.3743±0.0029 | 0.8794±0.0081 |
| Image noise | 6 | 80.7704±0.2253 | 0.3721±0.0006 | 0.8782±0.0018 |
| Image noise | 8 | 80.9530±0.4114 | 0.3739±0.0016 | 0.8821±0.0015 |
| Image noise | 10 | 81.3080±0.0492 | 0.3722±0.0021 | 0.8789±0.0057 |
| Image + state noise | 4 | 78.6778±0.1964 | 0.6546±0.0128 | 1.3413±0.0213 |
| Image + state noise | 6 | 78.7034±0.1069 | 0.6673±0.0088 | 1.3628±0.0151 |
| Image + state noise | 8 | 79.2125±0.1257 | 0.6924±0.0066 | 1.4010±0.0120 |
| Image + state noise | 10 | 78.8919±0.2594 | 0.7219±0.0006 | 1.4552±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) ↗.
| Model | Evaluated configuration | Epochs in Table 4 |
|---|---|---|
| Drive-JEPA [2] | Pose-only. Released V-JEPA [3] ViT-L encoder, newly initialized planner, L1 ego-trajectory loss. | 20 |
| E2EDriver | Original 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.
Thank you for your feedback
Thank you for your time and thoughtful feedback.
Response to Reviewer Aow2 (1/2)
AW1. We have added a within-family E2EDriver [1] capacity sweep in Table 8 (Lines 873–890) ↗.
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.
AW2. We have added iteration-matched controls for AutoVLA [2] and DriveVLA-W0 [3] in Table 5 (Lines 487–496) ↗.
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.