acceptodds
Under review as a conference paper at ICLR 2027

DriveScout: Sharpening MLLM Perception in Autonomous Driving with Adaptive and Robust Visual Focusing

Abstract

Multimodal Large Language Models (MLLMs) are increasingly used in autonomous driving for perception, reasoning, and interpretable scene understanding. Yet safety-critical small-scale targets, such as pedestrians, traffic signs, and distant vehicles, remain difficult to recognize, particularly when motion blur weakens their already limited visual evidence. Think-with-Images MLLMs can acquire additional evidence by cropping and enlarging local regions, but blur may destabilize the selected region, while indiscriminate image operations add inference cost. To study this problem, we introduce SCOUT, a paired driving VQA benchmark built from driving scenes in nuScenes, where each clear observation is matched with motion-degraded variants for small-target understanding. We further propose DriveScout, a reinforcement learning framework that formulates visual focusing as selective evidence acquisition. DriveScout combines Target-Grounded Spatial Credit with Necessity-Aware Focus Optimization, while a Progressive Blur Curriculum supports learning as the visual evidence becomes increasingly degraded. Experiments on SCOUT and two additional driving benchmarks show that DriveScout improves robustness to blur while reducing redundant visual operations.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Models, Autonomous Driving, Tool-Augmented Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nnpVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DriveScout is an RL‑based framework that enables adaptive visual cropping‑and‑zooming for multimodal large language models to recognize small safety‑critical driving targets under motion blur. The authors also construct SCOUT, a paired VQA benchmark with clear and motion‑degraded driving samples.

Critical strengths

Problem identification is valuable: It points out the under‑studied problem of tiny safety‑relevant object understanding under motion blur for driving multimodal models.

Critical weaknesses
  1. Simulation‑only degradation limitation: All motion blur is synthetic image‑space corruption, not real camera motion blur captured in physical driving. Synthetic blur cannot faithfully reproduce real rolling‑shutter, sensor noise and complex object motion, so the simulated training distribution has large domain gap with real‑world driving inputs.
  2. Task scope mismatch for autonomous driving: The whole system only addresses VQA perception tasks. It is never integrated into downstream driving planning or closed‑loop simulation. We cannot verify whether better VQA performance actually translates to better driving decisions.
  3. Behavior instability risk: Ablation shows strong accuracy‑efficiency trade‑off. Tuning penalty for redundant cropping can sharply drop overall recognition accuracy, which suggests brittle hyper‑parameter dependence without robust out‑of‑box performance.
Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Reponse to W1, W2 and W3

Response to W1: Generalization to Naturally Blurred Driving Scenes

We thank the reviewer for highlighting the domain gap between synthetic and real-world motion blur. DriveScout transfers directly to naturally blurred driving observations in a zero-shot setting, as demonstrated by our evaluation on real driving frames from BDD100K MOT. To assess generalization beyond synthetic degradation, we conduct an additional zero-shot evaluation on BDD100K MOT [1]. Specifically, we select naturally blurred frames from real driving videos and construct 990 manually verified VQA samples, without applying any synthetic blur.

On these samples, DriveScout achieves 72.22% accuracy, outperforming the strongest baseline, Base (SFT+GRPO) + Restormer [2], by 5.45 percentage points. Compared with Base, DriveScout also reduces the crop rate from 26.46% to 8.99% and the average inference latency from 4.430 s to 2.647 s. SCOUT provides matched clear and motion-degraded observations across blur severities and directions, enabling controlled analysis of small-target recognition and visual focusing. The zero-shot BDD100K results complement this controlled evaluation with naturally occurring blur. Further experimental details and results, including the small-target subset evaluation, are provided in Appendix A.6 of the revised manuscript.

MethodAcc. (%) ↑Crop Rate (%) ↓Avg. Output Tokens ↓Avg. Latency (s) ↓Δ Latency (%) ↓Avg. Input Tokens ↓
Base (SFT+GRPO)64.2426.46152.784.4300.001723.93
Base + Restormer66.7724.44141.215.114+15.441710.56
Base + DeblurGAN-v2 [3]65.5625.56146.334.620+4.291716.02
DriveScout72.228.9971.402.647−40.251621.94

Δ Latency is computed as (method latency − Base latency) / Base latency × 100%; negative values indicate lower latency.

Response to W2: Scope and Relevance to Downstream Driving Decisions

We thank the reviewer for this valuable suggestion. We agree that downstream planning and closed-loop evaluation are important for assessing complete autonomous driving systems. We would like to clarify that our work focuses on perception-level visual understanding, rather than end-to-end planning and control.

To our knowledge, existing open-loop and closed-loop driving benchmarks have placed limited emphasis on small-target understanding under motion blur, making its potential contribution to downstream driving decisions difficult to assess directly. We believe that accurately understanding distant vehicle orientations, pedestrian motion, and traffic signs can provide valuable environmental cues for driving decisions. Furthermore, DriveScout explores adaptive acquisition of necessary local visual evidence, offering a promising direction toward more robust and efficient driving perception.

Response to W3: Hyperparameter Sensitivity and Behavioral Stability

We thank the reviewer for raising this concern. We acknowledge that an excessively strong redundant-focus penalty may reduce recognition accuracy, reflecting an accuracy–efficiency trade-off rather than necessarily indicating instability under small hyperparameter changes.

To further investigate this, we conduct a finer-grained hyperparameter analysis around the default penalty of 0.08. When the penalty varies from 0.06 to 0.10, accuracy remains within 53.9%–54.7%, with a maximum difference of only 0.8 percentage points. Moreover, the crop rate remains within 27.9%–29.1% for penalty values between 0.07 and 0.10.

These results demonstrate that DriveScout maintains relatively stable recognition performance and focusing behavior around its default configuration, rather than relying on a narrowly tuned penalty value.

Tool PenaltyCrop Rate (%) ↓Accuracy (%) ↑
0.0633.354.3
0.0727.953.9
0.08 (Ours)28.554.2
0.0928.054.4
0.1029.154.7

The corresponding revisions are included in the revised manuscript and marked in orange for changes specific to your comments or purple for changes shared across reviewers.

References

[1] Yu, F., et al. "BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning." CVPR, 2020.

[2] Zamir, S. W., et al. "Restormer: Efficient Transformer for High-Resolution Image Restoration." CVPR, 2022.

[3] Kupyn, O., et al. "DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better." ICCV, 2019.

We sincerely thank the reviewer for recognizing the importance of studying small safety-critical objects under motion blur in autonomous driving. We hope that our rebuttal and additional experiments have adequately addressed the main concerns raised in the review. We would greatly appreciate it if the reviewer could consider our responses and share any further feedback. We sincerely thank the reviewer again for the constructive feedback and suggestions that have helped us improve our paper.

Reviewer_Ru4wOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates small-target understanding in autonomous driving using multimodal large language models, with particular emphasis on motion-degraded observations. The authors identify two challenges: motion blur can cause inaccurate visual focusing, while unnecessary cropping operations introduce additional inference overhead. To investigate these issues, they introduce SCOUT, a paired VQA benchmark containing clear and motion-blurred driving scenes. The proposed DriveScout framework employs reinforcement learning with three components: Target-Grounded Spatial Credit (TGSC) for crop localization, Necessity-Aware Focus Optimization (NAFO) for selective visual operations, and a Progressive Blur Curriculum (PBC) for robust training. Experiments on SCOUT, SURDS, and DriveBench demonstrate improvements in recognition accuracy and reductions in visual operation frequency.

Critical strengths
  1. The paper focuses on a practically relevant challenge involving small-target perception under degraded driving conditions.
  2. SCOUT provides paired clear and blurred observations, enabling controlled evaluation of perception robustness.
  3. The proposed approach achieves promising recognition performance while substantially reducing the frequency of cropping operations.
Critical weaknesses
  1. Does the method actually alleviate crop drift? A central motivation is that motion blur destabilizes crop localization. However, the main experiments focus on final answering accuracy, with little quantitative analysis of crop localization quality across blur severities. The reported gains therefore do not clearly establish that the proposed method improves localization robustness.

  2. Fewer crops do not necessarily imply more appropriate focusing decisions. Although NAFO substantially reduces the overall crop rate, the paper does not examine whether the remaining operations are concentrated on samples that genuinely benefit from additional visual inspection. This makes it difficult to assess the claimed necessity-aware behavior.

  3. Actual computational savings are left unmeasured. Crop rate and generated-token counts provide useful indicators, but they do not fully characterize the additional visual encoding and model inference costs associated with focusing. Without latency or computational measurements, the practical efficiency benefits remain uncertain.

  4. The robustness conclusions are tied to synthetic degradation. Most evaluations employ parameterized motion blur, which may not capture the complexity of naturally occurring blur in driving scenes. The generalization of the proposed strategy to realistic motion degradation is consequently unclear.

Questions for authors
  1. A breakdown of crop localization accuracy across different blur levels would help determine whether TGSC effectively addresses the crop drift described in the Introduction.
  2. What proportion of focusing operations occurs on samples where direct prediction fails? An analysis contrasting clear and blurred inputs, as well as easy and difficult cases, would better characterize NAFO's behavior.
  3. Please clarify the practical inference overhead introduced by visual focusing, ideally with end-to-end latency measurements. Additional evaluation on naturally blurred driving observations would also strengthen the robustness claims.
6 comments
AuthorsOpenReview ↗

Response to W1/Q1 and W2/Q2 (Part1/2)

Response to W1: Crop Localization Under Motion Blur

We thank the reviewer for the suggestion. Both natural-inference and forced-cropping evaluations show that TGSC improves crop localization under motion blur. To quantify crop localization across blur levels, we report [email protected], Coverage, and IoU. [email protected] measures the percentage of samples recorded as cropped with normalized center error d=((c^x−cx)/w)2+((c^y−cy)/h)2≤0.2d=\sqrt{((\hat{c}_x-c_x)/w)^2+((\hat{c}_y-c_y)/h)^2}\leq 0.2, where (cx,cy)(c_x,c_y) and (w,h)(w,h) are the center and dimensions of the human-annotated reference box. Coverage measures the fraction of the reference box covered by the predicted crop, and IoU measures their spatial overlap.

(1) Natural inference. We compare PBC and PBC+TGSC on matched samples where both models independently choose to crop. Across all blur levels, TGSC improves [email protected], Coverage, and IoU by an average of 4.17, 2.89, and 2.94 percentage points (pp), respectively.

(2) Forced cropping. We use the same prefix to require cropping by DriveScout w/o TGSC and the full DriveScout, reducing the influence of different focusing decisions. TGSC improves the three metrics across all blur levels by an average of 4.43, 3.23, and 3.73 pp, respectively.

The forced-cropping evaluation reduces the influence of different decisions to invoke cropping. Its consistent gains strengthen the evidence that TGSC improves the spatial quality of acquired visual evidence under blur. Detailed results are provided in Appendix A.7 (Tables 13–14) of the revised manuscript.

Table 1. Crop localization under natural inference. Results are computed on the same samples cropped by both models at each blur level. All localization metrics are in percent (%).

Blur LevelMethod[email protected] ↑Coverage ↑IoU ↑
CleanPBC67.7881.2845.77
PBC+TGSC70.3182.4847.39
LightPBC61.0277.6844.31
PBC+TGSC66.4880.1646.54
MediumPBC57.2073.2238.36
PBC+TGSC63.0178.2143.80
HeavyPBC48.9468.8432.91
PBC+TGSC51.8371.7235.36

Table 2. Crop localization under forced cropping. Both models receive the same cropping-enforcing prefix. All localization metrics are in percent (%).

Blur LevelMethod[email protected] ↑Coverage ↑IoU ↑
CleanDriveScout w/o TGSC69.3082.6645.12
DriveScout73.5083.4347.39
LightDriveScout w/o TGSC62.3076.1242.29
DriveScout68.2080.7246.30
MediumDriveScout w/o TGSC58.9073.7137.97
DriveScout64.0077.5943.05
HeavyDriveScout w/o TGSC45.9067.7331.35
DriveScout48.4071.4134.91

Response to W2: Necessity-Aware Visual Focusing

NAFO increases the share of beneficial crops and reduces redundant crops, improving the usefulness of visual focusing beyond simply lowering crop frequency. To examine whether NAFO distinguishes beneficial from redundant crops, we evaluate the same SCOUT test samples under mixed-blur conditions using normal inference and forced direct answering without cropping. Among samples where a model crops during normal inference, we define the Beneficial Crop Ratio as the fraction for which direct answering fails but crop-assisted answering succeeds, and the Redundant Crop Ratio as the fraction for which both answers are correct. Both ratios use the total number of cropped samples as the denominator.

Adding NAFO to Base (SFT+GRPO) increases the Beneficial Crop Ratio from 28.45% to 51.69% and reduces the Redundant Crop Ratio from 25.13% to 17.91%. Likewise, adding NAFO to DriveScout w/o NAFO increases the beneficial ratio from 31.75% to 67.37% and reduces the redundant ratio from 30.22% to 17.19%. Together, the increased beneficial-crop ratio and reduced redundant-crop ratio support NAFO’s role in directing visual computation toward useful local evidence. The corresponding results appear in Section 5.3.1 (Table 5), with the direct-view protocol and NAFO details in Appendices A.2 and A.4.

Table 3. Necessity analysis under mixed-blur conditions. Both ratios are percentages of samples cropped during normal inference.

MethodBeneficial Crop Ratio ↑Redundant Crop Ratio ↓
Base (SFT+GRPO)28.4525.13
Base + NAFO51.6917.91
DriveScout w/o NAFO31.7530.22
DriveScout67.3717.19
AuthorsOpenReview ↗

Response to W3/Q3 and W4/Q3 (Part2/2)

Response to W3: Practical Inference Efficiency

To assess practical efficiency beyond generated-token counts, we additionally measure cumulative input tokens, output tokens, crop rate, and end-to-end inference latency. Compared with Base (SFT+GRPO), DriveScout reduces average input tokens by 19.93%, output tokens by 29.59%, and total latency from 4.811 s to 4.054 s (15.73%), while reducing crop rate from 57.30% to 28.50%.

The latency reduction is primarily driven by the lower crop rate, which reduces the frequency of second-round visual encoding and inference. These efficiency gains are accompanied by an improvement of 5.5 percentage points in blurred accuracy over Base on SCOUT, demonstrating a joint benefit in recognition performance and inference cost. The detailed stage-wise latency breakdown in Appendix A.8 (Tables 15–16) provides further evidence.

Removing NAFO increases latency from 4.054 s to 4.844 s while lowering blurred accuracy from 54.2% to 53.7%. Removing TGSC lowers accuracy to 50.4% and increases latency to 4.304 s. Without PBC, latency is slightly lower at 4.014 s, but accuracy drops to 52.6%, illustrating the accuracy–efficiency trade-off. The accuracy results are reported in Table 6 of the manuscript; the w/o rows denote DriveScout ablations.

Table 4. Inference efficiency on SCOUT. Input tokens are cumulative across inference rounds; latency is measured end to end.

MethodCrop Rate (%) ↓Avg. Input Tokens ↓Avg. Output Tokens ↓Total Latency (s) ↓
Base (SFT+GRPO)57.303950.61132.814.811
w/o PBC27.003133.3892.074.014
w/o TGSC37.703421.55108.134.304
w/o NAFO58.903978.20136.644.844
DriveScout28.503163.2693.514.054

Response to W4: Generalization to Naturally Blurred Driving Scenes

DriveScout transfers directly to naturally blurred driving observations in a zero-shot setting, as demonstrated by our evaluation on real driving frames from BDD100K MOT. To evaluate robustness beyond synthetic degradation, we conduct an additional zero-shot evaluation on BDD100K MOT [1]. We select naturally blurred frames from real driving videos and construct 990 manually verified vehicle- and pedestrian-related VQA samples, including 533 small-target samples whose target area does not exceed 1% of the image. No synthetic blur is added to these test frames.

DriveScout achieves 72.22% accuracy on the full set, outperforming the strongest baseline, Base (SFT+GRPO) + Restormer [2], by 5.45 percentage points. On the small-target subset, DriveScout achieves 67.54% accuracy, exceeding the same baseline by 3.75 percentage points. SCOUT provides matched clear and motion-degraded observations across blur severities and directions, enabling controlled analysis of small-target recognition and visual focusing. The zero-shot BDD100K results complement this controlled evaluation with naturally occurring blur. The gains on the small-target subset further support the usefulness of the learned focusing policy in the small-target setting motivating SCOUT. Further evaluation details and complete results, including the small-target subset, are provided in Section 5.2.4 and Appendix A.6 (Tables 11–12) of the revised manuscript.

Table 5. Zero-shot evaluation on naturally blurred BDD100K MOT [1] frames (990 VQA samples). No synthetic blur is added. Input tokens are cumulative across inference rounds.

MethodAcc. (%) ↑Crop Rate (%) ↓Avg. Input Tokens ↓Avg. Output Tokens ↓Total Latency (s) ↓Δ Latency (%) ↓
Base (SFT+GRPO)64.2426.461723.93152.784.4300.00
Base + Restormer [2]66.7724.441710.56141.215.114+15.44
Base + DeblurGAN-v2 [3]65.5625.561716.02146.334.620+4.29
DriveScout72.228.991621.9471.402.647−40.25

Δ Latency is computed as (method latency − Base latency) / Base latency × 100%; negative values indicate lower latency.

The corresponding revisions are included in the revised manuscript and marked in green for changes specific to your comments or purple for changes shared across reviewers.

References

[1] Yu, F., et al. "BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning." CVPR, 2020.

[2] Zamir, S. W., et al. "Restormer: Efficient Transformer for High-Resolution Image Restoration." CVPR, 2022.

[3] Kupyn, O., et al. "DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better." ICCV, 2019.

Thank the author for the detailed responses and additional experiments. However, the real-world evaluation is limited to one dataset, and localization robustness under heavy blur remains insufficiently validated. I therefore consider the work insufficiently developed and maintain my score.

Reviewer_8joPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, author aim to do QA for small and distant objects on the street for driving situation. They introduce SCOUT. it is a nuScenes-based benchmark dataset. each clear image there has a synthetic blur levels. They also fine-tune Theme with several module on the proposed SCOUT. In specific, first, answering question without cropping. There is penalty if QA pass since it means the cropping is not necessary (NAFO). Second, they propose TGSC to score the overlay between cropping box and GT (TGSC). Third, they apply blur as data augmentation (PBC). Overall, they add several operations to the QA dataset collection so it is more suitable for the problem settings.

Critical strengths

(+) the dataset collection for those challenging cases can be helpful for the community. And the dataset itself can bring performance improvement to the thyme already. The proposed module further improve the performance.

(+) the propose of judging necessity of cropping for QA is straightforward, simple, and effective. Author break the problems into two challenges and addresses them with three simple modules. the output token gets obviously lower

Critical weaknesses

(-) it will be helpful if author can bring a naive solution (non-VLM based) to the comparison as well to further show the necessity of VLM in such problem settings. For example, rule-based, deblur before QA, etc

(-) does author test the generalization on other datasets or other real-world datasets? For example, the blur in the training data is pure synthetic and may not be realistic enough for real-world cases as you can imagine there could be really complex spatially varying blur in real-world cases.

(-) the efficiency table only shows the generated token, however it will be more helpful if input tokens and latency can be reported as well.

(-) I think a figure to show the accuracy and crop rate together can be helpful in someway as it may show the distribution statistic of the dataset and capability of methods. for example, thyme crop rate is 100% and lower accuracy. It is quite like a point plot

Questions for authors

I have listed all my concern in the weakness

4 comments
AuthorsOpenReview ↗

Response to W1 and W2 (Part1/2)

Response to W1: Comparison with Simple Baselines.

We thank the reviewer for the suggestion. SCOUT requires identifying language-referred targets and understanding their semantic attributes, which conventional object detectors cannot directly address without additional task-specific components. To examine a simple alternative, we evaluate two deblurring-before-QA baselines by applying Restormer [1] or DeblurGAN-v2 [2] before inference with Base (SFT+GRPO).

On SCOUT, Restormer and DeblurGAN-v2 provide some improvement under blur, but DriveScout outperforms them by 1.7 and 3.0 percentage points, respectively. Restormer and DeblurGAN-v2 incur additional deblurring time of 0.81 s and 0.38 s, respectively. Their end-to-end latencies are 5.53 s and 5.21 s, compared with 4.05 s for DriveScout. DriveScout achieves higher accuracy through adaptive visual focusing without requiring a separate deblurring stage. The comparison highlights the value of optimizing visual evidence acquisition for the QA objective, including both crop localization and the necessity of local inspection.

MethodBlur Acc. (%) ↑Gain vs Base (pp) ↑Total Latency (s) ↓
Base (SFT+GRPO)48.7—4.81
Base + Restormer52.5+3.85.53
Base + DeblurGAN-v251.2+2.55.21
DriveScout54.2+5.54.05

Gain is relative to Base. Total latency includes deblurring, when used, and all inference rounds.

Response to W2: Generalization to Real-World Motion Blur.

We thank the reviewer for raising this point. DriveScout transfers directly to naturally blurred driving observations in a zero-shot setting, as demonstrated by our evaluation on real driving frames from BDD100K MOT[3]. To assess generalization beyond synthetic blur, we conduct an additional zero-shot evaluation on BDD100K MOT using naturally blurred frames from real driving videos. Following the SCOUT test-set construction procedure, we select vehicle and pedestrian targets, exclude ambiguous cases, and construct VQA samples. Answers are manually annotated, and the question–answer pairs are manually verified. The resulting test set contains 990 samples, including 533 small-target samples whose targets occupy at most 1% of the image area.

On the full 990-sample set, DriveScout achieves 72.22% accuracy, outperforming the strongest baseline, Base (SFT+GRPO) + Restormer, by 5.45 percentage points. Compared with Base, DriveScout reduces crop frequency by 66.02% and end-to-end latency by 40.25%. The zero-shot BDD100K results complement the controlled evaluation on SCOUT with naturally occurring blur. On these real driving frames, DriveScout achieves higher accuracy with fewer crops and lower end-to-end latency than Base. Further evaluation details, including results on the 533-sample small-target subset, are provided in Appendix A.6 (Tables 11–12) of the revised manuscript.

MethodAcc. (%) ↑Crop Rate (%) ↓Avg. Input Tokens ↓Avg. Output Tokens ↓Latency (s) ↓Δ Latency (%) ↓
Base (SFT+GRPO)64.2426.461723.93152.784.4300.00
Base + Restormer66.7724.441710.56141.215.114+15.44
Base + DeblurGAN-v265.5625.561716.02146.334.620+4.29
DriveScout72.228.991621.9471.402.647−40.25

Δ Latency is computed as (method latency − Base latency) / Base latency × 100%; negative values indicate lower latency.

AuthorsOpenReview ↗

Response to W3 and W4 (Part2/2)

Response to W3: Additional Inference Efficiency Evaluation.

We thank the reviewer for the suggestion. We additionally report average input tokens, output tokens, and end-to-end inference latency on SCOUT to complement the output-token analysis in the paper.

Compared with Base (SFT+GRPO), DriveScout reduces average input tokens by 19.93%, output tokens by 29.59%, and end-to-end inference latency from 4.811 s to 4.054 s (15.73%). The latency reduction primarily stems from a lower crop rate (57.3% to 28.5%), which reduces the frequency of second-round inference and its associated overhead, while first-round inference time remains similar. These efficiency gains are accompanied by an improvement of 5.5 percentage points in blurred accuracy over Base on SCOUT, demonstrating a joint benefit in recognition performance and inference cost. A detailed stage-wise latency breakdown is provided in Appendix A.8 (Tables 15–16) of the revised manuscript, which provides further evidence.

Removing NAFO increases latency from 4.054 s to 4.844 s while lowering blurred accuracy from 54.2% to 53.7%. Removing TGSC lowers accuracy to 50.4% and increases latency to 4.304 s. Without PBC, latency is slightly lower at 4.014 s, but accuracy drops to 52.6%, illustrating the accuracy–efficiency trade-off. The accuracy results are reported in Table 6 of the manuscript; the w/o rows denote DriveScout ablations.

MethodAvg. Input Tokens ↓Avg. Output Tokens ↓Latency (s) ↓Crop Rate (%) ↓
Base (SFT+GRPO)3950.61132.814.81157.3
w/o PBC3133.3892.074.01427.0
w/o TGSC3421.55108.134.30437.7
w/o NAFO3978.20136.644.84458.9
DriveScout3163.2693.514.05428.5

Response to W4: Accuracy–Crop Rate Visualization.

We thank the reviewer for the suggestion. We have added an Accuracy–Crop Rate scatter plot on SCOUT in Figure 4 of the revised manuscript, comparing Thyme, Base (SFT+GRPO), the ablations, and DriveScout.

Among the plotted methods, DriveScout achieves the highest accuracy (54.2%) with a crop rate of 28.5%. Compared with Base, it improves accuracy by 5.5 percentage points while reducing the crop rate by 28.8 percentage points. Although DriveScout w/o PBC has a slightly lower crop rate (27.0%), its accuracy is 1.6 percentage points below DriveScout. The plot therefore highlights DriveScout's favorable accuracy–crop-rate trade-off.

The corresponding revisions are included in the revised manuscript and marked in blue for changes specific to your comments or purple for changes shared across reviewers.

References

[1] Zamir, S. W., et al. "Restormer: Efficient Transformer for High-Resolution Image Restoration." CVPR, 2022.

[2] Kupyn, O., et al. "DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better." ICCV, 2019.

[3] Yu, F., et al. "BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning." CVPR, 2020.

thanks for the rebuttal. experiments address all my concerns. therefore, I lean to keep my current rate

Thank you for your positive feedback and for taking the time to review our rebuttal and additional experiments. We are glad that our responses have addressed all your concerns. We sincerely appreciate your thoughtful review and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the Area Chair

We thank the three reviewers and the Area Chair for their careful reading and constructive feedback, which have helped us improve the paper. Following their suggestions, we have expanded the evaluation with additional baselines, analyses of visual focusing, practical efficiency measurements, and evaluation on naturally blurred driving images. We have also examined hyperparameter sensitivity and clarified the scope and limitations of our study. We have uploaded a revised PDF.

Reviewer 8joP, thank you for your suggestions on strengthening the baseline comparisons and evaluating real-world generalization and inference efficiency more comprehensively. We have added deblurring-before-QA comparisons and zero-shot evaluation on naturally blurred BDD100K frames, expanded the efficiency measurements to include input tokens and end-to-end latency, and added an accuracy–crop rate visualization.

Reviewer Ru4w, we appreciate your emphasis on direct evidence of crop localization, focusing necessity, and practical computational savings. We have added crop-localization analyses across blur levels, examined the usefulness of crop-assisted inference, and reported cumulative token usage and end-to-end latency. We have also evaluated zero-shot transfer to naturally blurred BDD100K frames.

Reviewer nnpV, thank you for highlighting the synthetic-to-real gap, downstream relevance, and hyperparameter sensitivity. We have added real-world blur evaluation and a finer-grained sensitivity analysis around the default penalty. We have also clarified the perception-level scope of our study and its limitations regarding downstream planning and closed-loop control.

The Summary of Changes on the first page of the revised manuscript provides clickable links to the relevant revisions in the main text and appendix, with additions coloured by reviewer and shared changes marked in a separate colour. We would be glad to discuss any point further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.