PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Reliable Vision-Language Models for Autonomous Driving

Manasa Mariam Mammen, Priyanka Mary Mammen, Zafer Kayatas, Stefan Wagner

arXiv:2610.01531v1cs.AIcs.CV

TL;DR

Visual degradation can affect both VLM predictions and confidence in safety-critical autonomous driving. The paper evaluates five models across four driving QA datasets under camera corruptions and tests inference-time VEA. Corruption effects and VEA gains vary across conditions, and improved accuracy does not always mean more reliable confidence.

  • Problem

    Visual corruption may affect VLM accuracy and confidence in driving, but reliability across driving tasks and visual input settings remains unclear.

    Source: §1 · p. 2

  • Method

    The study evaluates five VLMs on four driving QA benchmarks under four camera corruptions and tests inference-time VEA.

    Source: §1 · p. 2

  • Results

    Corruption effects vary, and VEA improves performance for some model–dataset combinations but not consistently; accuracy gains do not always accompany more reliable confidence.

    Source: §4.2 · p. 6, §5 · p. 7

  • Takeaways & Limitations

    Performance, calibration, and uncertainty discrimination do not always move together, so task accuracy alone does not capture confidence reliability.

    Source: §4.1 · p. 5, §5 · p. 7

    VEA re-weights spatial evidence but does not add temporal motion information, limiting its direct ability to address reasoning that depends strongly on temporal cues.

    Source: §4.2 · p. 6

Abstract

from arXiv · show

Vision-Language models (VLMs) are increasingly being explored in autonomous driving for tasks such as scene understanding, driving reasoning, decision-making, and end-to-end driving. As their role becomes more prominent, ensuring their robustness and reliability is increasingly important. In real-world conditions, visual inputs may be degraded by sensor imperfections and environmental conditions, potentially affecting both model predictions and their associated confidence. Such degradation is especially concerning in autonomous driving, where safety-critical decisions require models to make accurate predictions and recognize when their predictions may be unreliable. In this work, we evaluate five VLMs (Qwen3.5-9B, Gemma4-E4B, LLaVA-OneVision-7B, DriveFusion/DriveFusionQA-4B, and NVIDIA Alpamayo-1.5-10B) across four driving-related QA datasets with different visual input settings, including single-frame, multi-view, multi-frame, and monocular inputs. Our results show that the effects of visual corruption vary across models, datasets, and input settings, with changes in accuracy and confidence reliability and also differing across conditions. We then apply Visual Evidence Augmentation ($\mathrm{V}{\scriptstyle \mathrm{EA}}$), a recent inference-time method to examine whether it can improve model reliability under degraded visual conditions. We find that $\mathrm{V}{\scriptstyle \mathrm{EA}}$ improves performance for some models and datasets, although the gains are not consistent across all settings.

1 Introduction

The introduction frames VLM reliability under degraded camera inputs as an open issue for safety-critical autonomous driving. It evaluates visual corruptions and inference-time enhancement across driving tasks and input settings.

1 Introduction

  • End-to-end autonomous driving maps multimodal sensory inputs directly to driving actions, but current approaches still face gaps in long-tail and safety-critical situations.The introduction identifies sparse supervision and high-level reasoning as needs in these situations.Source: §1 · p. 1
  • Real-world visual degradation can affect both prediction accuracy and confidence, while VLM reliability in driving remains unclear.The introduction highlights rare, difficult cases and safety-critical decisions as reasons to examine this reliability question.Source: §1 · p. 2
  • The study evaluates five recent driving-related VLMs on four QA benchmarks spanning different reasoning tasks and visual input formats.The models are limited to 10B parameters or fewer, and the benchmarks include single images, consecutive frames, and multi-view camera inputs.Source: §1 · p. 2
  • The evaluation asks how glare, fog, motion blur, and lens occlusion affect accuracy and calibration, and whether inference-time VEA improves robustness.The corruptions represent illumination, weather, motion-related, and partial-visibility degradations.Source: §1 · p. 2

2 Related Work

Prior work examines VLM robustness to visual corruption and proposes inference-time methods to improve visual grounding. However, uncertainty estimation and calibration remain less explored in robustness benchmarks.

2 Related Work

  • DriveBench [29] and USB assess VLM robustness under diverse visual corruptions, with USB also evaluating decision accuracy and reasoning quality.DriveBench examines multiple driving tasks, while USB applies perturbations to unstructured driving situations.Source: §2 · pp. 2–3
  • AWDB studies winter-driving video, and ObsDriveBench [32] evaluates multimodal reasoning in real-world adverse conditions.The related work notes conditions including fog, rain, snow, and low light.Source: §2 · p. 3
  • Robustness benchmarks generally emphasize performance degradation, while uncertainty estimation and calibration remain largely underexplored.Inference-time visual-grounding methods include decoding changes, prompting or intermediate reasoning, attention-based input refinement, and internal-representation approaches [33] [34] [35] [36] [37].Source: §2 · p. 3

3 Experimental Setup

The experimental setup evaluates driving VLMs on four datasets using benchmark-specific scoring and measures of accuracy, uncertainty discrimination, and calibration. It also applies four controlled camera corruptions.

3 Experimental Setup

  • The setup section introduces the datasets, models, evaluation metrics, and visual corruptions used in the experiments.These components define the evaluation protocol described in the subsequent setup subsections.Source: §3 · p. 3

Datasets

  • The study evaluates four driving QA datasets, including open-ended, multiple-choice, spatial-reasoning, and monocular 3D tasks.It uses 1,000 samples each from DrivingVQA and STRIDE-QA Bench, 1,117 from NuScenes-QA-mini, and all 50 OSR examples.Source: Datasets · p. 3

Models

  • Five general-purpose and driving-specific models are evaluated, including Alpamayo-1.5-10B through its visual question-answering interface.The set also includes Qwen3.5-9B, Gemma4-E4B, LLaVA-OneVision-7B, and DriveFusionQA-4B.Source: Models · pp. 3–4

Dataset-specific evaluation

  • Scoring follows each benchmark: exact-match accuracy for DrivingVQA and OSR, GPT-4 judging for NuScenes-QA-mini, and localization thresholds for STRIDE-QA Bench.STRIDE-QA Bench counts predictions correct within ±25% distance error and ±10° heading error.Source: Dataset-specific evaluation · p. 4

Metrics

  • Accuracy, VAUQ-based AUROC, ECE, and Brier score measure task performance, uncertainty discrimination, and calibration.VAUQ combines predictive entropy with an image-information score; lower Brier scores indicate better calibration.Source: Metrics · p. 4

3.3 Visual Corruptions

  • Four corruptions—glare, fog, motion blur, and lens occlusion—are applied independently with fixed-seed randomization and per-image severity sampling.Glare uses a randomly centered Gaussian pattern; fog uses a smoothed noise mask, motion blur a horizontal linear kernel, and lens occlusion blurred elliptical regions.Source: §3.3 · p. 4

4 Results

The section compares clean and corrupted visual inputs across driving tasks, then assesses whether VEA improves performance and confidence reliability under degradation.

4.1 Robustness Analysis

  • On DrivingVQA, accuracy changes only slightly for most models, but confidence metrics reveal differences not reflected in accuracy alone.DriveFusionQA-4B stays near chance AUROC_VAUQ despite relatively low ECE and Brier; Gemma4-E4B’s AUROC_VAUQ drops more than its accuracy under some corruptions.Source: §4.1 · p. 5, Table 1 · p. 5
  • Corruption has a stronger effect on NuScenes-QA-mini, where DriveFusionQA-4B’s slight accuracy gains coincide with declining AUROC_VAUQ.Qwen3.5-9B, Gemma4-E4B, and LLaVA-OV-7B retain above-chance AUROC_VAUQ under several corruptions, but their signal is less reliable than in clean inputs.Source: §4.1 · p. 5, Table 2 · p. 6
  • No model performs consistently well across accuracy, calibration, and AUROC_VAUQ under visual corruption.Some models are better calibrated but discriminate weakly between correct and incorrect predictions, while others discriminate more strongly but calibrate poorly.Source: §4.1 · p. 5

4.2 Visual Evidence Enhancement

  • Across four datasets, VEA sometimes preserves accuracy gains under corruption, but its effects vary by model and task.The reported accuracy changes are not consistent across all model–dataset combinations.Source: §4.2 · pp. 6–7, Table 3 · p. 6
  • VEA reweights spatial evidence but does not add temporal motion information, limiting its direct help with temporal-cue-dependent STRIDE-QA reasoning.Alpamayo-1.5-10B consistently loses accuracy on STRIDE-QA Bench despite benefiting strongly on DrivingVQA and NuScenes-QA-mini and mostly improving on OSR.Source: §4.2 · p. 6
  • VEA’s accuracy gains do not always accompany more reliable confidence, with different models showing distinct calibration and discrimination changes.Alpamayo-1.5-10B gains accuracy while AUROC_VAUQ, ECE, and Brier worsen; DriveFusionQA-4B gains little accuracy while its reliability metrics generally improve.Source: §4.2 · p. 7, Figure 2 · p. 7

5 Conclusion

Across clean and corrupted visual conditions, model performance and confidence reliability did not always move together. VEA improved performance for some model–dataset combinations, but its effects were task- and model-dependent.

5 Conclusion

  • Performance and confidence reliability did not always change in the same way across the evaluated models and conditions.No model performed consistently well across accuracy, calibration, and AUROC.Source: §5 · p. 7
  • VEA improved performance for some model–dataset combinations, with many gains remaining under corruption.Its effects depended on the task and model.Source: §5 · p. 7
  • Improved accuracy with VEA did not always come with more reliable confidence.Source: §5 · p. 7

A Supplementary Results

Supplementary results show that corruption effects and VEA’s reliability changes vary across tasks and models. Accuracy, confidence discrimination, and calibration can disagree, and OSR’s small sample makes its condition-level differences noisy.

STRIDE-QA Bench

  • STRIDE-QA Bench has the lowest overall accuracy among the four datasets, reflecting the difficulty of its spatiotemporal reasoning setting.Corruption changes are mixed rather than clear gains or losses from a single metric.Source: A Supplementary Results · p. 10, Table 4 · p. 11
  • On STRIDE-QA Bench, LLaVA-OV-7B remains below chance in AUROCVAUQ across all conditions.Its confidence is therefore not reliably aligned with visually grounded correctness in this setting.Source: A Supplementary Results · p. 10, Table 4 · p. 11

OSR

Supplementary VEA results

  • VEA does not produce a uniform reliability change across NuScenes-QA-mini, STRIDE-QA Bench, and OSR.Accuracy or calibration gains are not always matched by better AUROCVAUQ.Source: A Supplementary Results · p. 10

B VEA Computation Time

VEA adds evidence-extraction time, generally below one second on three datasets but higher on STRIDE-QA Bench. The overhead is model- and task-dependent and matters for real-time use.

B VEA Computation Time

  • VEA’s added evidence-extraction time is generally below one second on DrivingVQA, NuScenes-QA-mini, and OSR.Figure 6 reports average added time per sample.Source: B VEA Computation Time · p. 11
  • On STRIDE-QA Bench, added time reaches about 1.3 seconds for Qwen3.5-9B and 3 seconds for DriveFusionQA-4B.Source: B VEA Computation Time · p. 11
  • VEA’s overhead varies by model and task and is added on top of original inference time, making it relevant to strict real-time deployment.The passage identifies reducing computation cost as necessary for such applications.Source: B VEA Computation Time · p. 11

C Implementation Details

The implementation materials identify model-specific VEA steering layers and summarize hardware, generation, VAUQ, and evaluation settings. The supplementary figures cover VEA effects on three datasets and its added evidence-extraction time.

C Implementation Details

Related papers

Loading 2610.01531v1…