acceptodds
Under review as a conference paper at ICLR 2027

CEDAR: Concise, Evidence-grounded Driving Action Recommendation via Cascade Reasoning and Policy Refinement

Abstract

Large vision-language models have advanced autonomous driving toward interpretable scene understanding, yet existing models tend to produce verbose responses without explicitly surfacing the visual evidence behind their decisions, limiting their practicality for real-time driving assistance. We formulate the task of concise, evidence-grounded driving recommendation generation, where actionable decisions are explicitly backed by decision-relevant visual evidence and contextual reasoning. To this end, we construct CODA-GRD (Grounding, Reasoning & Decision), a unified Evidence-Reason-Suggestion benchmark providing structured supervision for visual grounding, reasoning, and action recommendation. Built upon it, we propose CEDAR, trained via a progressive pipeline: Entity Presence Grounding Pretraining aligns visual entities with structured evidence; Chain-of-Insight Cascade Reasoning (CCR) propagates context from Reason to Evidence to Suggestion under a reason-aware contrastive objective; and Driving Policy Refinement applies branch-wise Group Relative Policy Optimization with task-specific rewards. Together, these stages enable CEDAR to generate concise recommendations explicitly grounded in visual evidence and reasoning. Experiments on CODA-GRD and three external benchmarks, OmniDrive, LingoQA, and DriveCoT, show consistent improvements over strong baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Autonomous Driving, Vision-Language Models, Evidence-Grounded Reasoning, Visual Grounding, Reinforcement Learning

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_11VSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to generate short driving suggestions together with reasons and localized visual evidence. The authors build CODA-GRD by reorganizing CODA-LM annotations and propose CEDAR based on Florence-2. The method combines grounding pretraining, a latent Reason–Evidence–Suggestion cascade, reason-aware contrastive learning, and branch-wise GRPO. Experiments on CODA-GRD and three additional driving datasets show improvements on the reported metrics over several baselines.

Critical strengths
  1. The task is practically useful. Separating the suggestion, reason, and object locations makes the output easier to inspect and provides a clear interface for studying driving recommendations.
  2. On CODA-GRD, the overall score improves from 26.25 to 29.82, with improvements also reported on LingoQA and DriveCoT. These results provide useful evidence that the combined design improves the model on the evaluated tasks.
Critical weaknesses
  1. My main concern is that the evaluation does not directly validate the claimed evidence-grounded and concise recommendations. The text-overlap metrics, including ROUGE-L used as the training reward, do not directly assess whether the generated evidence and reasons support the suggested action. Table 8 evaluates annotation quality, while Figures 4–5 mainly show annotation examples involving construction vehicles and similar, direct suggestions. The paper does not provide comparative model outputs on cases requiring evidence selection among multiple objects or reasoning about spatial relations. It remains unclear whether the gains reflect these abilities or mainly common associations between object categories and suggestions. Quantitative evaluation of conciseness is also missing.

  2. The method introduces several modules and training stages, but the benefit of some components remains unclear. Although the paper provides several ablations, the cumulative component study does not fully isolate their contribution in the final system. RCA improves Overall only from 28.22 to 28.24 in Table 4(a), while the Evidence average decreases from 73.43 to 72.42 after adding the pretraining tasks in Table 4(e). These results need further explanation and estimates of variation across runs. A simpler Florence-2 configuration using matched data and a comparable training budget, including GRPO, would also help determine whether the additional components are needed.

Questions for authors
  1. Do you have results for a simpler Florence-2 baseline using the same training data and a comparable training budget, including GRPO?

  2. Could you provide comparisons on cases requiring evidence selection among multiple objects or spatial reasoning, with human assessment of decision correctness and evidence support?

  3. Could you clarify how Reason and Suggestion annotations are extracted, and how the external datasets are prepared and split?

  4. Do the external results use models trained only on CODA-GRD or models fine-tuned on each target dataset?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 11VS(1/2)

We thank the reviewer for recognizing the practical value of separating driving suggestions, reasoning, and object locations, as well as the consistent improvements on CODA-GRD, LingoQA, and DriveCoT. We address the concerns as follows:

Q1: Evaluation of Evidence Grounding

Cross-component evaluation. We agree that text-overlap metrics alone cannot establish whether Evidence and Reason support Suggestion. Following your comments, we therefore add CLIP-based Evidence–Reason (E–R) and Reason–Suggestion (R–S) similarity, together with human assessment of Evidence–Suggestion (E–S) support on 200 samples:

ModelE–R ↑R–S ↑E–S(Human) ↑
Gemini-2.5-Flash0.2280.7856.0%
Qwen3-VL-235B0.2350.7664.0%
MiniCPM-V-2.60.2290.8066.0%
Qwen2-VL-2B0.2310.8070.0%
Baseline0.2270.7668.5%
Baseline+CCR0.2360.8170.5%
Baseline+CCR+RCA0.2410.8071.0%
CEDAR-0.89B0.2420.8175.0%

Compared with the Baseline, CEDAR improves E–R from 0.227 to 0.242, R–S from 0.76 to 0.81, and E–S human agreement from 68.5% to 75.0%. These improvements provide complementary evidence of stronger cross-component consistency, while CEDAR remains competitive with external models.

These evaluation results have been added to Table 3 and Sec. 4.1 (Page 8, Lines 414–424) of the revised manuscript.

Q2: Model outputs on complex Driving Scenarios

Following your comments, we have added four qualitative examples covering simple and complex scenes with multiple traffic entities, including pedestrians, construction vehicles, barriers, and surrounding cars (Appendix 6.4, Fig. 5 (Page 17)). Unlike Figs. 6–7, these examples present model-generated Evidence, Reason, and Suggestion, illustrating their relationships in different driving scenarios.

Q3: Conciseness of Generated Driving Suggestions

The original CODA-LM driving_suggestion averages 91.65 words, and OmniDrive action averages 89.70 words, whereas CODA-GRD Suggestion averages 20.16 words. Moreover, CEDAR-generated Suggestions average 20.84 words. Our objective is to make driving actions explicit and concise while preserving supporting explanations separately in Reason and Evidence.

The length comparison has been added in Sec. 4.1 (Page 7, Lines 376-377) .

Q4: Annotation Construction and Dataset Preparation

(i) Annotation extraction. We adopt rule-guided text decomposition using causal discourse cues and contextual information to separate action-oriented Suggestions and explanatory Reasons from original CODA-LM driving suggestions, followed by manual screening and refinement.

(ii) External datasets. We clarify the preparation and splits in Appendix 6.1(External Datasets, Page 14, Lines 702–710):

  • OmniDrive: Extract Suggestion–Reason pairs aligned with front-view images; 20,000 training and 8,130 validation samples from the original training set, and 6,019 test samples from the original validation set.
  • LingoQA: Extract pairs from causally structured answers, using an 8:1:1 training/validation/test split.
  • DriveCoT: Preserve original splits and retain valid front-view samples with extracted Suggestion–Reason pairs.
  • DriveLM: Used exclusively for auxiliary grounding pretraining, with a 9:1 training/validation split.

External benchmarks evaluate Suggestion and Reason without structured Evidence supervision.

AuthorsOpenReview ↗

Response to Reviewer 11VS(2/2)

Q5: Component Contributions and Training Stages

(i) Independent ablations. To address your concern about the limited improvement from RCA (in Table 5(a)), we further evaluate Baseline+CCR,Baseline+RCA, Baseline+Pre, and Baseline+RL independently.

MethodEvidenceSuggestionReasonOverall
Baseline72.6420.4413.5126.25
Baseline+CCR73.4421.3217.0328.22
Baseline+RCA72.9720.8917.1628.01
Baseline+Pre71.9620.5213.6426.23
Baseline+RL75.8221.1114.3627.42

Although RCA provides only a modest additional gain when combined with CCR (28.22→28.24), applying RCA independently improves Overall from 26.25 to 28.01, particularly benefiting Reason (13.51→17.16). This shows that RCA contributes to reasoning performance, although its benefits may partially overlap with CCR. Meanwhile, RL independently improves Evidence from 72.64 to 75.82. Pretraining alone does not improve Overall (26.25→26.23), as its grounding-oriented objectives differ from the final generation objectives. This is expected, since pretraining is designed to provide grounding support rather than directly optimize downstream performance.

The results are included in Table 5(a)(Page 9) of the revised manuscript .

(ii) Contribution of Grounding Pretraining. We acknowledge the reviewer's observation that Evidence decreases from 73.43 to 72.42 after adding ECA and EPV (Pretraining in Table 5(e)). However, grounding pretraining is designed to strengthen entity-level visual grounding and support the joint learning of Evidence, Reason, and Suggestion, rather than optimize Evidence in isolation. With pretraining, Suggestion improves from 21.16 to 21.54, Reason from 17.24 to 18.33, and Overall from 28.24 to 28.68. These results suggest that pretraining contributes to a better balance across the three tasks, despite the modest decrease in Evidence performance.

To further examine its contribution under RL-based refinement, we compare CEDAR with variants that remove either the entire pretraining stage or EPV alone:

MethodEvidenceSuggestionReasonOverall
CEDAR w/o pre74.4722.5818.6629.59
CEDAR w/o EPV76.1922.0718.6629.67
CEDAR75.7522.6118.6729.82

Removing the entire pretraining stage reduces Evidence from 75.75 to 74.47 and Overall from 29.82 to 29.59, further supporting the contribution of grounding pretraining to the complete framework. Removing EPV alone, in contrast, increases Evidence to 76.19 but decreases Suggestion from 22.61 to 22.07 and Overall from 29.82 to 29.67. This indicates that EPV introduces a trade-off in Evidence performance while contributing to better Suggestion generation and overall performance. Taken together, these results suggest that grounding pretraining contributes to a better balance across the three tasks, despite the modest decrease in Evidence.

These results are reported in Table 5(a) (Page 9), and Table 10 (Page 16) of the revised manuscript.

(iii) Comparable training budget. To further address your concern about whether the improvements come mainly from additional training stages, we evaluate a simpler Florence-2 configuration incorporating Pretraining and GRPO, without CCR and RCA. The comparison is presented below:

MethodEvidenceSuggestionReasonOverall
Baseline + Pre + GRPO70.5621.1614.0626.44
CEDAR75.7522.6118.6729.82

Although the simpler configuration also includes Pretraining and GRPO, its Overall score reaches 26.44, compared with 29.82 for CEDAR. CEDAR achieves higher performance across Evidence, Suggestion, and Reason.

The results are included in Table 11 of the revised manuscript (Page 16).

Q6: External Benchmark Training Protocol

The external results use dataset-specific fine-tuning, rather than zero-shot evaluation of a CODA-GRD-only model. CEDAR and comparison models are fine-tuned on the respective target datasets and evaluated on their test sets.

Compared with the task-specific Florence-2 Baseline, CEDAR improves Overall from 46.93 to 51.21 on LingoQA, 40.36 to 43.38 on DriveCoT, and 21.90 to 22.56 on OmniDrive (Tables 2 and 4 and Appendix Table 7).

Dear Reviewer 11VS, thank you sincerely for the time and effort in reviewing. As the discussion deadline approaches in just under two days, we would like to kindly ask whether you have any additional questions or if there are any further experiments you would like us to consider. We are truly happy to provide additional theoretical analysis or experimental results as soon as possible to help clarify any points of concern. Your feedback is truly valuable to our work, and we would be grateful if we could receive your recognition.

Reviewer_xiQPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper makes two main contributions: 1. Build a new benchmark CODA-GRD from an existing benchmark CODA-LM. The curated new benchmark explicitly constains evidence grounding, reasoning and explicit decision process. The authors didn't build this benchmark from scratch, but instead reorganize another CODA-LM benchmark. Therefore, the benchmark angle contribution isn't that obvious. 2. Based on the newly curated benchmark, the authors introduced a new training recipe tailored to improve the performance. The new training recipe consists of three stages and the overall framework is called Chain-of-Insight Cascade Reasoning. Experiment with the new training recipe show significant performance gain comparing with other baselines.

Critical strengths

From my understanding, the most critical strength lies in the proposed progressive training recipe that combines grounding pretraining, contrastive reasoning supervision, as well as branch-wise policy refinement. This training recipe improves the CODA-GRD average from 26.25 to 29.82 (13.6% relative performance gain), such training recipe also brings improvements across three external benchmarks. These consistent results provide strong empirical support for the effectiveness of the overall training strategy.

Critical weaknesses

The main concern to me lies in the practical motivation for the proposed task and benchmark remains insufficiently substantiated. While explicit evidence and reasoning could facilitate interpretability, the paper does not demonstrate their benefits for driving decisions, safety, or human usability.

Furthermore, CODA-GRD largely restructures existing CODA-LM annotations without establishing what practically important capability its evaluation uniquely captures. Consequently, improvements under the proposed supervision and metrics demonstrate successful optimization of the formulated task, while leaving its necessity and practical usefulness unresolved.

Questions for authors

See the weakness section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer xiQP(1/2)

We thank the reviewer for recognizing the practical value of separating driving suggestions, reasoning, and object locations, as well as the effectiveness of our progressive training strategy and its improvements across multiple benchmarks. Regarding the concerns raised, we provide the following clarifications:

Q1: Practical Motivation and Usability of CEDAR

Our primary objective is to generate concise, actionable driving suggestions while retaining supporting visual evidence and contextual reasoning. We would like to address your concern by answering the following questions:

(i) Why concise suggestions matter? As discussed in the Introduction, existing driving VLMs often combine actionable instructions with lengthy scene descriptions and causal explanations, <u>making the key driving action less immediately accessible</u>. Our statistical analysis shows that the original CODA-LM driving_suggestion field contains 91.65 words on average, while the OmniDrive action field averages 89.70 words. In contrast, CODA-GRD extracts concise Suggestions averaging 20.16 words. Our goal is not merely to shorten responses, but to make the recommended driving action explicit while preserving its supporting information separately. These word-count statistics have been added to Sec. 4.1 (Page 7, Lines 376–377) of the revised manuscript.

(ii) What are the roles of Evidence and Reason? Evidence and Reason are supporting outputs. Evidence identifies decision-relevant traffic objects and their locations, while Reason explains the context behind the recommendation, making the suggestion convincing. These components provide complementary supervision during training and interpretable information for inspection and verification. Their role is to strengthen the decision representation without requiring lengthy explanations in every user-facing recommendation.

To further examine their contribution, we compare the Baseline trained with Suggestion alone (Baseline + S) against joint Evidence–Reason–Suggestion supervision (Baseline + ERS SFT).

MethodB-1B-4MR-LCAVG
Baseline + S26.7810.4232.7530.360.3220.13
Baseline + ERS SFT27.0010.9032.9131.020.3820.44

AVG denotes the average of the five Suggestion metrics.

Joint ERS supervision improves Suggestion AVG from 20.13 to 20.44, with gains across all five metrics, indicating that Evidence and Reason provide useful complementary supervision for Suggestion generation.

(iii) Latent integration and flexible output. As described in Section 3.2.2 (Eq. 7), CCR progressively integrates reasoning-aware and evidence-aware features into the Decision Unit. Reason, Evidence, and Suggestion are generated from interconnected task-specific representations derived from a shared multimodal feature space and decoded separately through a shared Language Decoder. Thus, Suggestion benefits from integrated reasoning and evidence representations without requiring their corresponding texts to be decoded first. This design allows the framework to prioritize concise driving recommendations while retaining the ability to provide supporting explanations when needed.

AuthorsOpenReview ↗

Response to Reviewer xiQP(2/2)

Q2: Distinct Value and Evaluation of CODA-GRD

CODA-GRD's contribution lies in reformulating the original annotations into structured supervision for concise, evidence-grounded driving recommendation generation.

(i) A clearer task formulation. In CODA-LM, the driving_suggestion field combines driving actions with contextual and causal explanations. As illustrated in Figure 1 and Section 3.1, CODA-GRD separates actionable instructions into Suggestion, preserves explanatory information as Reason, and constructs spatially grounded Evidence from regional annotations. This enables the model to learn concise recommendations while retaining complementary grounding and reasoning supervision.

(ii) Separate learning objectives and evaluation. CODA-GRD enables Evidence, Reason, and Suggestion to be supervised and evaluated individually. Evidence is evaluated using F1 and IoU, while Reason and Suggestion are assessed using BLEU, METEOR, ROUGE-L, and CIDEr. This provides a more fine-grained assessment of visual grounding, contextual reasoning, and driving recommendation quality than a single combined response score.

(iii) Empirical effectiveness. As reported in Table 5(a), our progressive training strategy improves all three components:

ModelEvidence (Avg.)Suggestion (Avg.)Reason (Avg.)Overall
Baseline72.6420.4413.5126.25
CEDAR75.7522.6118.6729.82

Furthermore, CEDAR improves the Overall score on LingoQA from 46.93 to 51.21 and DriveCoT from 40.36 to 43.38 (Tables 2 and 4). These results demonstrate that the proposed training strategy benefits recommendation and reasoning generation beyond CODA-GRD.

In summary, CEDAR aims to provide concise, actionable, and safety-oriented driving recommendations, with Evidence and Reason serving as complementary supervision and interpretable support. CODA-GRD enables these capabilities to be learned and evaluated separately, while CEDAR integrates them to produce driving suggestions that clearly communicate recommended actions and their supporting context. This design prioritizes decision clarity and safety-relevant information, making the recommendations more suitable for driving assistance.

Dear Reviewer xiQP, thank you sincerely for the time and effort in reviewing. As the discussion deadline approaches in just under two days, we would like to kindly ask whether you have any additional questions or if there are any further experiments you would like us to consider. We are truly happy to provide additional theoretical analysis or experimental results as soon as possible to help clarify any points of concern. Your feedback is truly valuable to our work, and we would be grateful if we could receive your recognition.

Dear Reviewer xiQP,

Thank you again for your valuable time and constructive review. With the discussion phase nearing its end, we wanted to gently check if our rebuttal has resolved your concerns, or if any further clarifications or experiments would be helpful.

We are fully available to provide additional details promptly. Thank you very much for your thoughtful consideration of our work.

Best regards,

The Authors

Reviewer_kXm3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a useful problem in autonomous driving. A VLM should explain not only what to do, but also why, and connect its explanation to visual evidence. The paper makes two main contributions: a data preparation and training pipeline, and a cascade architecture that connects reasoning, visual evidence, and driving suggestions.

Critical strengths

Providing visual evidence alongside driving suggestions makes the outputs easier to understand and verify. The proposed framework brings these tasks together, and the relatively small model achieves competitive results.

Critical weaknesses
  1. Qwen2.5-VL-3B is used both to supplement missing bounding-box annotations and as a baseline. CEDAR performs better, which is possible, but the role of these generated labels is not sufficiently clear. In particular, if the test set also contains Qwen-generated labels, the evaluation may partly measure agreement with these labels rather than independent grounding accuracy.
  2. The paper uses contrastive alignment and cascade reasoning to strengthen the connection between Reason, Evidence, and Suggestion. However, the gains in the relevant ablations are small. I would like to see concrete examples or visual analyses showing whether these components actually make the three outputs more consistent and better connected.
Questions for authors
  1. What proportion of the training and test annotations is generated by Qwen2.5-VL-3B, and how are these labels independently verified?
  2. What mechanism ensures that the generated suggestion is supported by the reported visual evidence? How does the method handle contradictions between its Evidence, Reason, and Suggestion outputs?
4 comments
AuthorsOpenReview ↗

Response to Reviewer kXm3(1/2)

We thank the reviewer for recognizing the value of providing visual evidence alongside driving suggestions, the unified design of CEDAR, and its competitive performance with a lightweight model. Regarding the concerns raised, we provide the following clarifications:

Q1: Role and Reliability of Qwen2.5-VL-3B Annotations

(i) Baseline clarification. Our primary baseline is Florence-2-large-ft (0.77B parameters), not Qwen2.5-VL-3B. As detailed in Appendix 6.1 (Florence-2 Baseline), Florence-2 Baseline uses the same training data and supervision format as CEDAR, while excluding Entity Presence Grounding Pretraining, CCR, RCA, and Driving Policy Refinement. This controlled comparison isolates the contributions of our proposed components.

(ii) The Role of Qwen2.5-VL-3B. Qwen2.5-VL-3B serves only as an auxiliary annotation tool to supplement missing bounding boxes for decision-relevant pedestrians described in the original CODA-LM annotations. It does not generate Reason or Suggestion labels.

(iii) Annotation proportion. Qwen-assisted annotations account for 5.54% of the training annotations and 5.76% of the test annotations. The remaining annotations originate from CODA-LM.

(iv) Annotation Verification and Reliability. To ensure annotation reliability, we apply GPT-4o-assisted verification: GPT-4o identifies and filters potentially incorrect or inconsistent Qwen-generated annotations, after which the retained annotations undergo manual quality assessment. As reported in Appendix Table 9, we evaluate localization reliability on 300 images, achieving 96.00% (576/600) , and evaluate Reason consistency on 200 images, obtaining 95.75% (383/400) . These manual assessments provide additional evidence of annotation reliability beyond the Qwen-generated outputs.

Q2: Qualitative Analysis of Cross-Component Connections

To address this concern, we provide two additional qualitative comparisons of Baseline, Baseline+CCR, and Baseline+CCR+RCA in Fig. 4 of Appendix 6.4 (Qualitative Analysis), illustrating the improvements in Evidence, Reason, and Suggestion generation. These examples demonstrate how CCR and RCA strengthen the connections among visual evidence, contextual reasoning, and actionable driving suggestions.

Q3: How Is Suggestion Supported by Visual Evidence?

Our CEDAR establishes the connection between the suggestion and visual evidence through structured supervision and cascade reasoning, rather than an explicit rule-based verification mechanism.

(i) Structured supervision. CODA-GRD annotates Evidence, Reason, and Suggestion jointly for the same driving scene, so the three fields are inherently coupled by construction: the Reason is grounded in the Evidence, and the Suggestion follows from the Reason. Training on these annotations therefore teaches the model to keep the three fields consistent with one another.

(ii) Cascade reasoning. CCR module progressively injects reasoning-aware and evidence-aware representations into the Decision Unit, grounding Suggestion generation in both visual evidence and contextual information.

(iii) Additional consistency evaluation. Following your comments, we evaluate cross-component relationships using CLIP-based similarity and human assessment on 200 samples for E–S.

ModelE–R ↑R–S ↑E–S(Human) ↑
Gemini-2.5-Flash0.2280.7856.0%
Qwen3-VL-235B0.2350.7664.0%
MiniCPM-V-2.60.2290.8066.0%
Qwen2-VL-2B0.2310.8070.0%
Baseline0.2270.7668.5%
Baseline+CCR0.2360.8170.5%
Baseline+CCR+RCA0.2410.8071.0%
CEDAR(0.89B parameters)0.2420.8175.0%

E–R measures CLIP similarity between Evidence-grounded image regions and Reason, while R–S measures Reason–Suggestion semantic similarity. For E–S, since suggestion often does not mention the specific visual objects, we manually assess whether generated suggestions are supported by the reported visual evidence.

Compared with the Baseline, CEDAR improves E–R from 0.227 to 0.242, R–S from 0.76 to 0.81, and E–S human agreement from 68.5% to 75.0%. CEDAR remains competitive with external models. These results provide complementary evidence of improved cross-component connections.

The results are included in Table 3 and Sec. 4.1 (Page 8, Lines 414–424) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer kXm3(2/2)

Q4: Handling Contradictions Among Evidence, Reason, and Suggestion

CEDAR does not employ an explicit contradiction detection or correction module. Instead, it promotes cross-component coherence through structured supervision and latent feature refinement.

Specifically, CCR integrates Reason and Evidence representations into the Decision Unit, while RCA strengthens visual–Reason alignment. Driving Policy Refinement further optimizes generation through task-specific rewards. Our primary objective is to generate concise, actionable driving suggestions supported by relevant visual evidence and contextual reasoning, rather than guarantee strict logical consistency across all generated statements.

The additional evaluations in Q3 show that CEDAR achieves the highest E–R similarity and E–S human agreement among the evaluated variants, together with improved R–S similarity over the Baseline. These findings support the effectiveness of our design, while explicit contradiction detection remains a direction for future work.

We additionally explored explicit consistency rewards during Driving Policy Refinement. Specifically, E–S consistency uses CLIP-based image–text matching between the evidence-grounded visual regions and the generated Suggestion as a reward signal: RESlocal=sim⁡CLIP(IE,S)R_{\mathrm{ES}}^{\mathrm{local}} = \operatorname{sim}{\mathrm{CLIP}}(I_E,S). Similarly, E–R consistency uses CLIP-based image–text matching between the same evidence-grounded visual regions and the generated Reason as a reward signal: RERlocal=sim⁡CLIP(IE,R)R_{\mathrm{ER}}^{\mathrm{local}} = \operatorname{sim}_{\mathrm{CLIP}}(I_E,R).

MethodEvidenceSuggestionReasonOverall
+ E–S Consistency76.2821.2317.4528.83
+ E–R Consistency78.2921.4215.8528.58
CEDAR75.7522.6118.6729.82

Neither consistency reward improves overall performance. This suggests that explicitly emphasizing semantic consistency does not necessarily improve the overall task, while the existing structured supervision and latent refinement already provide useful consistency constraints. We acknowledge the importance of cross-component consistency and will further investigate explicit contradiction detection and correction in future work.

Dear Reviewer kXm3, thank you sincerely for the time and effort in reviewing. As the discussion deadline approaches in just under two days, we would like to kindly ask whether you have any additional questions or if there are any further experiments you would like us to consider. We are truly happy to provide additional theoretical analysis or experimental results as soon as possible to help clarify any points of concern. Your feedback is truly valuable to our work, and we would be grateful if we could receive your recognition.

Reviewer_kXm3OpenReview ↗

Maintain my score

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers: Summary of Revisions

We sincerely thank all reviewers for their constructive feedback, which has helped us improve the clarity and rigor of our work. We have addressed the concerns in our individual responses and uploaded the revised manuscript.

The main revisions include:

  • Clarifying the role of Qwen2.5-VL-3B in supplementary bounding-box annotation, reporting the annotation proportions, and providing human assessments of annotation reliability.
  • Strengthening the motivation for concise driving recommendations through word-count comparisons with CODA-LM and OmniDrive, and clarifying the supporting roles of Evidence and Reason.
  • Adding cross-component consistency evaluations based on CLIP similarity and human assessment, together with comparisons against additional vision-language models.
  • Expanding component ablations with independent CCR, RCA, Pretraining, and RL configurations, as well as combined settings to examine their respective contributions.
  • Providing further analyses of grounding pretraining, including ECA and EPV ablations and comparisons with a baseline trained using Pretraining and GRPO.
  • Extending qualitative comparisons across Baseline, Baseline+CCR, and Baseline+CCR+RCA, and adding challenging driving scenarios involving multiple traffic participants and obstacles.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.