acceptodds
Withdrawn from ICLR 2027

Benchmarking Evidence Sufficiency in Vision-Language Models for Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision–language models (VLMs) for autonomous driving are typically evaluated by answer correctness, even when the supplied observations may not contain sufficient evidence to justify an answer. We introduce DriveAnswer, a benchmark of 20,151 questions across seven reasoning families, in which each question is evaluated under matched strong- and weak-evidence configurations. Each pair preserves the underlying scene fact and semantic answer while changing whether the observations support answering, making abstention the appropriate response under weak evidence. Across five zero-shot VLMs, strong-condition accuracy substantially overstates evidence-sensitive performance: the best baseline reaches 64.69% strong-evidence accuracy but only 41.96% matched-pair success. A controlled adaptation study separates answer, sufficiency, grounding, and matched-pair supervision. Answer-only fine-tuning nearly eliminates abstention, whereas strong-weak sufficiency supervision restores selective behavior; grounding supervision produces a separate localization gain. On the held-out test set, DriveAnswer-Gemma reaches 81.58% matched-pair success, while explicit grounding and sufficiency supervision raises joint localization-presence ([email protected]) from 53.43% to 77.16% and joint answer-grounding accuracy ([email protected]) to 61.98%. These results show that semantic correctness, evidence sufficiency, and visual grounding are distinct capabilities that should be evaluated jointly.

Primary area:
datasets and benchmarks
Keywords:
autonomous driving benchmark, VLM benchmark, VLM autonomous driving dataset, evidence-sufficiency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oQXjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DriveAnswer, a driving VQA benchmark for evaluating whether vision-language models can determine when the available visual evidence is sufficient to answer a question. Built from nuScenes, it contains 20,151 grounded questions across seven spatial, temporal, cross-view, and decision-related families. For each question, the authors construct matched strong- and weak-evidence configurations that preserve the underlying scene fact and question while retaining or removing required camera views or temporal observations. Models are expected to answer under strong evidence and abstain under weak evidence, while also localizing the supporting visual regions. The paper evaluates five VLMs using answer accuracy, abstention, matched-pair success, and grounding metrics, and conducts staged adaptation experiments on Gemma and Qwen using answer–abstention, grounding, and matched-pair supervision. The results show that semantic correctness, evidence-sensitive abstention, and visual grounding can differ substantially, while direct supervision improves the latter two capabilities.

Critical strengths
  1. The paper introduces an interesting setting that evaluates not only whether a model answers correctly, but also whether the available visual evidence is sufficient to support an answer.

  2. The matched strong–weak evidence pairs and grounding annotations provide a useful benchmark for jointly evaluating answering, abstention, and evidence localization. The dataset construction also involves substantial effort.

Critical weaknesses
  1. Limited task scope. The main benchmark focuses on retrospective perception and scene understanding. It does not address evidence sufficiency for future prediction or planning, where sufficient evidence is harder to define.

  2. Unvalidated benefit to actual planning. The paper evaluates only answer/abstention and grounding performance. It does not test whether training with DriveAnswer improves the trajectory-planning ability of an autonomous driving VLA. For example, the authors could fine-tune the same driving model with and without the proposed QA supervision and compare ADE/FDE, collision rate, or closed-loop performance. Therefore, the practical value of evidence-aware QA for autonomous driving remains unclear.

  3. Incremental contribution. Abstention under insufficient evidence has been widely studied in reliable VQA and selective prediction. This work mainly applies the idea to nuScenes through matched evidence pairs and grounding annotations, without demonstrating cross-dataset generalization or benefits to actual driving systems.

Questions for authors

See weaknesses.

2 comments

We sincerely thank the reviewer for the thoughtful feedback. We appreciate the opportunity to elaborate on DriveAnswer's scope, its relevance to autonomous driving, and its relationship to prior work.

1. Scope and extension toward prediction and planning

We appreciate the suggestion to consider future prediction and planning, where uncertainty and evidence sufficiency become especially consequential. DriveAnswer takes a complementary first step by examining whether a driving VLM's conclusions are supported by the observations available to it. Even at the observational level, this requires reasoning across multiple camera views, temporal context, actor relationships, motion states, and map information. Through question-conditioned strong–weak pairs, DriveAnswer provides a controlled way to examine how models respond when evidence needed for a particular conclusion becomes unavailable.

This complements work such as DriveLM (Sima et al., ECCV 2024), which connects graph-structured visual question answering with perception, prediction, and planning. We see observational sufficiency as a foundation for further investigating evidence-aware prediction and planning, and will clarify this scope and future direction in the camera-ready version.

2. Practical significance for autonomous driving

We thank the reviewer's suggestion for evaluating downstream planning performance through ADE/FDE, collision rates, and closed-loop testing. We agree that demonstrating the benefits of evidence-aware supervision for planning is a valuable next step, and will clarify this direction in the revised manuscript.

We would also like to emphasize DriveAnswer's practical relevance through scene understanding and observational sufficiency as complementary foundations for reliable planning. DriveBench (Xie et al., ICCV 2025) identifies a critical reliability concern: driving VLMs can produce plausible responses based on general knowledge or textual cues without adequate visual grounding. Such unsupported interpretations could influence downstream planning decisions, even when conventional answer-accuracy metrics fail to reveal the underlying problem.

DriveAnswer directly addresses this evaluation gap. By distinguishing semantic correctness from observational support, its matched strong–weak protocol and grounding metrics, including Joint@t and JAGS, provide a diagnostic assessment of whether driving-scene conclusions are both correct and supported by the available evidence, an important diagnostic step before downstream planning. This perspective is also consistent with Drive4C (Sohn et al., CVPRW 2025), which emphasizes that aggregate driving scores do not necessarily reveal the underlying reasons for model failures and therefore evaluates scenario understanding separately from driving performance.

We therefore see DriveAnswer's diagnostic contribution as valuable independently, with future planning-oriented studies as a strong next step. Our experiments demonstrate that models can learn the benchmark's targeted capabilities; they are not intended as claims of improved trajectory planning. We will make this distinction explicit in the revision.

3. Relationship to selective prediction and existing driving benchmarks

We agree with the reviewer’s observation that abstention and selective prediction are well-established research directions. However, we see DriveAnswer as building on this foundation while addressing a complementary question: not only whether a model's answer is likely to be correct, but whether the observations provided actually support that answer.

For example, Reliable VQA (Whitehead et al., ECCV 2022) studies selective prediction by estimating the likelihood of answer correctness. UNK-VQA (Guo et al., TPAMI 2024) more directly investigates unanswerability through image and question perturbations. DriveAnswer builds on these ideas in the autonomous-driving setting by holding the question and underlying scene fact fixed while varying the evidence available to support the answer.

Through matched strong–weak pairs and observation-level grounding, we can distinguish a semantically correct but unsupported answer from one justified by the supplied observations. This evidence-centered distinction complements confidence-based selective prediction, particularly in multi-view and temporal driving scenes where the information required for an answer may be distributed across observations.

Our existing Table 1 compares DriveAnswer with representative driving benchmarks. We will expand the related-work discussion to clarify these connections and better communicate the contribution of question-conditioned observational sufficiency evaluation. We once again thank the reviewer for the feedback and look forward to future discussion.

We include the reference for the papers we mentioned below.

References:

Sima, C., Renz, K., Chitta, K., Chen, L., Zhang, H., Xie, C., ... & Li, H. (2024, September). DriveLM: Driving with Graph Visual Question Answering. In European Conference on Computer Vision (pp. 256-274). Cham: Springer Nature Switzerland.

Xie, S., Kong, L., Dong, Y., Sima, C., Zhang, W., Chen, Q. A., ... & Pan, L. (2025, October). Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data, and Metric Perspectives. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 6285-6297). IEEE.

Sohn, T. S., Dillitzer, M., Bach, J., Corso, J. J., Brühl, T., Schwager, R., ... & Sax, E. (2025, June). Drive4C: A Closed-Loop Benchmark on what Foundation Models Really Need to be Capable of for Language-Guided Autonomous Driving. In *2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) *(pp. 3859-3869). IEEE.

Whitehead, S., Petryk, S., Shakib, V., Gonzalez, J., Darrell, T., Rohrbach, A., & Rohrbach, M. (2022, October). Reliable Visual Question Answering: Abstain Rather Than Answer Incorrectly. In European Conference on Computer Vision (pp. 148-166). Cham: Springer Nature Switzerland.

Guo, Y., Jiao, F., Shen, Z., Nie, L., & Kankanhalli, M. (2024). UNK-VQA: A Dataset and a Probe Into the Abstention Ability of Multi-Modal Large Models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12), 10284-10296.

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

DriveAnswer contains 20,151 questions from nuScenes, each shown under a strong and a weak set of original camera frames. The question and the annotated scene fact stay fixed. The strong set keeps every observation the authors mark as required and expects the semantic answer. The weak set removes or withholds at least one required observation and expects ABSTAIN. Any evidence unit labeled PARTIAL, AMBIGUOUS, or UNAVAILABLE makes the configuration non-sufficient, and every non-sufficient configuration has the same target, ABSTAIN. Grounding is obtained in a separate call and scored by IoU against regions projected from the annotations.

Critical strengths

The construction keeps the question, the queried entities, and the semantic reference fixed while changing which original frames are supplied, and it states the checks required for that invariance. The authors also separate rule compliance from human judgment: a stratified review of 120 questions agrees that the weak configuration is insufficient in 117/120 cases, and the paper explicitly says that the automatic 100% rates are not independent estimates of perceptual agreement.

Critical weaknesses

The abstention rule is stricter than the paper can justify. Any missing required observation makes the whole question ABSTAIN, so a correct answer on a weak input is scored as a model error. The paper never checks whether the remaining frames still support that answer. The strict target therefore keeps the labels tidy, but it does not show that those inputs are actually unanswerable.

The manuscript never explains why a model can still state the correct fact when its own rule marks the evidence as insufficient. That explanation is needed to justify calling the input unanswerable. Setting the target to ABSTAIN does not provide it.

Questions for authors
  1. On weak inputs, what fraction of non-abstain outputs exactly match the semantic reference? For those matches, do the remaining frames still visually support that reference?

  2. If the abstention label is removed from training, or if entire families of weak interventions are held out, does the model still abstain on the corresponding test inputs?

4 comments

We thank the reviewer for the thoughtful questions concerning observational sufficiency, semantic correctness, and abstention supervision. We conducted additional analyses to address these questions directly.

1. Semantic agreement on non-abstention cases and weak configuration evidence insufficiency audit (W1, Q1)

Semantic agreement. We examined whether non-abstaining predictions on weak inputs match the underlying scene truth. On the six-family test set, 7/12, 11/18, and 10/17 such predictions match the semantic reference for Gemma Stages B, C, and D, respectively. These subsets are small because the adapted models abstain on nearly all weak inputs.

Additional human validation. To directly assess whether weak observations contain sufficient evidence, we conducted a blinded visual audit of 100 weak-condition examples using complete imagery and questions stratified by family. These examples include all the 47 non-abstaining examples provided by Gemma Stages B, C, and D. The human annotator assessed the provided weak observations without seeing model predictions, semantic references, or sampling categories.

The results were: overall 93/100 (93%) judged insufficient; seven uncertain, and zero judged sufficient. By auditing 47 weak-condition cases where Gemma Stages B, C, and D answered instead of abstaining, we asked whether the provided observations contained sufficient evidence to justify any answer. 45 were judged insufficient and two uncertain; none was judged sufficient. Thus, correct semantic predictions made by finetuned models in weak configurations do not imply that the remaining visual inputs support those predictions.

These findings complement the 117/120 agreement reported in our original stratified human validation, while directly examining the semantic-match cases highlighted by the reviewer. We will incorporate the additional audit protocol, sample construction, and qualitative examples in the revision.

2. Abstention on unseen data (Q2)

We additionally evaluated a 57-pair manually screened cross-scene pilot. Weak observations were substituted from very similar scenes while matching camera configuration and image count, controlling the characteristic missing-view and duplicated-frame cues, while creating a more challenging task for deciding between answering and abstaining. Gemma Stages C and D abstained on 63.2% and 61.4% of these weak inputs, respectively. This is important because both models continue to abstain on the majority of weak observations on new data.

The experiment provides complementary evidence that the learned abstention behavior extends beyond recognizing the original intervention structure. We will include this stress test alongside our evaluation to make the robustness analysis more comprehensive.

We appreciate the reviewer’s questions and will incorporate these additional results to clarify the benchmark’s validity and the role of sufficiency supervision.

We thank the reviewer again for their thoughtful feedback on abstention supervision. We appreciate the opportunity to further investigate this question and would like to share additional ablation results that complement our earlier response.

2. Abstention without the label, and on held-out interventions (Q2)

To better isolate the contributions of each training objective, we trained an additional variant, A2, using strong-condition answer and grounding supervision but without explicit weak-condition abstention targets. Together with our existing baselines and training stages, this provides the following comparison on Gemma-4-12B:

No sufficiency targetSufficiency target
No groundingZero-shot: 67.31 / 69.47 / 44.35 / 21.55/ 48.6%Stage B: 81.52 / 99.58 / 81.20 / 42.79/ 52.7%
GroundingA2: 83.85 / 64.81 / 49.19 / 38.41/ 78.1%Stage C: 82.58 / 99.36 / 82.05 / 62.37/ 76.0%

Values: strong accuracy / weak abstention / pair success / [email protected] (%)/ Grounded Share (%). Grounded share is calculated as [email protected] divided by pair success.

Abstention rate alone does not fully capture evidence-sufficiency behavior. Zero-shot Gemma abstains on 69.47% of weak inputs but achieves only 44.35% pair success, reflecting its lower accuracy on strong inputs containing sufficient evidence. A2, despite receiving no explicit weak-abstention supervision, improves strong accuracy to 83.85% and pair success to 49.19%, while retaining 64.81% weak abstention. This illustrates through strong-answer sufficiency and grounding, the model can more reliably reason about answering-abstaining compared to the zero-shot Gemma model.

The two training objectives contribute complementary capabilities. Sufficiency supervision increases pair success from about 44–49% to about 81–82%, both with and without grounding. However, pair success alone does not establish whether the model correctly localizes the required evidence.

To examine this distinction, we calculate the proportion of successful QA pairs that also satisfy all required grounding conditions, given by [email protected] divided by pair success. This grounded share is approximately 49–53% without grounding supervision (zero-shot: 48.6%; Stage B: 52.7%) and 76–78% with grounding supervision (A2: 78.1%; Stage C: 76.0%). These results suggest that a model can correctly answer and abstain on a question pair without consistently localizing all required supporting observations, differentiating between a sound answer and a justified decision. Stage B illustrates this distinction: despite achieving 81.20% pair success, its [email protected] is 42.79%. With both objectives, Stage C achieves 82.05% pair success and 62.37% [email protected]. We observe a similar improvement in [email protected] for Qwen3-VL-8B, from 39.01% to 62.26%.

We believe these additional results help clarify the complementary roles of answer/abstention supervision and grounding supervision. They also reinforce the motivation for DriveAnswer's multi-faceted evaluation: correct answers and abstention decisions are important, but evaluating whether those decisions are supported by the required visual evidence provides an additional perspective. We recognize that these aggregate results do not, by themselves, rule out reliance on structural cues. We therefore view them as complementary to the structure-controlled evaluation discussed in our earlier response.

We sincerely appreciate the reviewer's questions and we look forward to further discussion.

Thank you for the response from the author. The current cross-scenario experiment reports how often the model fails to respond when there is insufficient evidence. Additionally, I am quite curious about comparing A2 and Stage C on the same set of 57 pairs of samples. How many pairs of these models can answer correctly when there is sufficient evidence and refuse to answer when there is insufficient evidence?

We sincerely thank the reviewer for the follow-up and for suggesting this direct comparison.

We evaluated A2 and Stage C on the same 57 manually selected and validated matched strong/weak pairs, with identical questions, matched camera configurations, and observation counts. Matched-pair success requires both a correct answer on the strong configuration and abstention on the corresponding weak configuration.

MetricA2Stage C
Strong accuracy64.9% (37/57)70.2% (40/57)
Weak abstention47.3% (27/57)63.2% (36/57)
Matched-pair success29.8% (17/57)38.6% (22/57)

A pairwise comparison shows that Stage C succeeds on six pairs where A2 fails, which contain mostly multi-grounding tasks, whereas A2 succeeds on only one pair where Stage C fails, which is a relative position question. Both stages fail the majority of the temporal evidence cases. These results suggest that explicit sufficiency supervision contributes to improved paired answering and abstention under matched observation structures, beyond strong-answer and grounding supervision. We appreciate the reviewer's suggestion and look forward to further discussion.

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

This paper introduces DriveAnswer, a nuScenes-based autonomous driving VQA benchmark designed to evaluate whether vision-language models (VLMs) can determine if available visual evidence is sufficient to justify an answer. Across 20,151 questions spanning seven reasoning families, each question is paired with a strong configuration (requiring the correct semantic answer) and a weak configuration (where critical evidence is withheld, requiring ABSTAIN), alongside visual grounding annotations. Zero-shot evaluation of five open-source VLMs reveals poor evidence sensitivity, whereas staged fine-tuning on Gemma-4-12B and Qwen3-VL-8B with sufficiency and grounding supervision boosts Matched-Pair Success to ~82% and joint answer-grounding success ([email protected]) to ~63%

Critical strengths
  1. Evaluating evidence sufficiency alongside answer accuracy and visual grounding addresses a realistic safety gap in autonomous driving under partial observability.

  2. Preserving the underlying scene fact and question while modifying only visual evidence availability isolates whether a model conditions its answer or abstention on actual observations.

Critical weaknesses
  1. In Appendix H (Figures 7 and 8), weak configurations contain repeated identical frames, whereas strong configurations provide distinct consecutive frames. Moreover, 39.69% of weak configurations completely remove target visibility. Fine-tuned models might achieve 98–99% Weak Abstention Accuracy (Table 4) simply by detecting duplicate frames or missing views rather than reasoning about evidence sufficiency, while Strong Accuracy on families like Map and Temporal Relation remains modest (~50–58% in Tables 21 and 23). The no-image shortcut control in Appendix E only tests zero-shot Gemma-4-12B and does not rule out shortcut exploitation by fine-tuned models.

  2. Appendix A.5 states that Cross-view and Temporal Recovery examples are retained only when added observations restore missing evidence, implying the strong-condition answer is always affirmative. Consequently, Strong Accuracy and UAR sum to 100% across all zero-shot models (Table 19), and fine-tuned models reach 99–100% Pair Success (Tables 21, 23), reducing these tasks to binary abstention checks. Furthermore, for questions like "Can the silver sedan's identity be recovered from another camera's view?", when no supporting view is provided in the weak condition, the true answer relative to the admitted input should arguably be "No" rather than ABSTAIN.

  3. In Table 2(b), InternVL3.5-8B, Molmo2-ER, and LLaVA-OneVision-7B achieve 75.59%–83.97% Presence accuracy but near-zero [email protected] (0.05%–0.35%). Because VLMs use different coordinate formats and grounding conventions, the lack of details on model-specific prompting and box parsing makes it unclear whether these results reflect genuine localization failure or parsing mismatches.

Questions for authors
  1. How exactly are weak configurations constructed regarding frame repetition, frame counts, and camera views across families? Can the authors report a simple baseline (or fine-tuned no-image/frame-hash control) that predicts ABSTAIN solely from input structure or duplicate-frame detection?

  2. Could the authors confirm the ground-truth answer distribution for Cross-view and Temporal Recovery, and clarify why withholding a recovery view makes the answer ABSTAIN rather than "No"?

  3. How were bounding boxes prompted, extracted, and normalized for InternVL3.5-8B, Molmo2-ER, and LLaVA-OneVision-7B, and did the authors verify that their near-zero [email protected] is not caused by coordinate-format mismatches?

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

Related papers

Discussion (0)

Sign in to comment.