acceptodds
Under review as a conference paper at ICLR 2027

SpanVLA: Learning from Negative-Recovery Samples with Fast Action Bridging for Vision-Language-Action Model

Abstract

Vision-Language-Action (VLA) models bring rich world knowledge and reasoning capabilities to autonomous driving. However, most driving VLAs are trained only on positive expert demonstrations, which teach models what to do but provide limited supervision on which behaviors to avoid and how to recover from these failures, leading to limited robustness. We introduce SpanVLA, a robust and efficient driving VLA, together with nuReasoning-NR, the first real-world driving dataset and benchmark for learning from negative and recovery samples. An extension of nuReasoning, it comprises 36K reasoning-intensive driving scenarios, including 3K suboptimal trajectories and 3K expert recovery trajectories. To exploit their asymmetric supervision, we introduce Negative–Recovery Reinforcement Fine-Tuning (NR-RFT), which penalizes undesirable behaviors and rewards expert corrective actions. Because challenging scenarios produce rollout groups dominated by similar suboptimal or even infeasible trajectories, it further incorporates an advantage correction to preserve informative optimization signals and encourage exploration. For efficient action generation, SpanVLA retains autoregressive reasoning and introduces a lightweight flow-matching action bridge conditioned on sparse-layer KV caches and initialized from the historical trajectory. Extensive experiments on NAVSIM v1/v2, nuReasoning, and the nuReasoning-NR benchmark demonstrate strong and robust planning performance and efficient action generation. Qualitative results across diverse challenging scenarios further highlight its planning robustness and recovery capability. The dataset and benchmark code will be released to support future research.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Model, Autonomous Driving

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TUsFOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents SpanVLA, a vision-language-action framework for autonomous driving that learns from negative and recovery examples while enabling efficient trajectory generation. First, it introduces nuReasoning-NR, a dataset and benchmark containing positive driving demonstrations, suboptimal driving behaviors, and subsequent expert recovery trajectories. Second, it proposes Negative–Recovery Reinforcement Fine-Tuning (NR-RFT), which penalizes trajectories resembling negative examples and rewards alignment with expert recovery trajectories, together with an advantage correction for difficult rollout groups. Third, its Efficient Action Bridge uses a lightweight flow-matching action expert conditioned on sparse VLM KV caches and initialized from historical trajectories to generate continuous future trajectories. Experiments on NAVSIM, nuReasoning, and nuReasoning-NR evaluate planning performance, the benefits of negative–recovery supervision, and action-generation efficiency.

Critical strengths
  1. The introduction of nuReasoning-NR is a valuable contribution. By including real-world suboptimal driving behaviors and subsequent expert recovery trajectories alongside positive demonstrations, the dataset supports studying both what a driving policy should avoid and how it should recover from difficult states.

  2. The paper provides a coherent SFT and GRPO-based training recipe for exploiting these different forms of supervision. NR-RFT distinguishes undesirable trajectories from expert corrections through appropriate penalties and rewards. The experiments support its practical value: NR training improves SpanVLA from 60.6 to 69.6 PDMS and AutoVLA from 59.1 to 64.1 on nuReasoning-NR, while the ablations show additional gains from recovery samples beyond negative supervision alone.

Critical weaknesses
  1. The ablations support the overall framework but do not fully isolate several key contributions. For the Efficient Action Bridge, Table 4 compares different systems with different reasoning lengths, while Table 5 changes both bridge size and historical initialization in the full-versus-sparse caching comparison. These evaluations are also limited to NAVSIM v1. For NR-RFT, Figure 4 omits a positive–recovery baseline, and the reward ablation removes the negative penalty and recovery reward together, leaving the additional value of negative supervision once recovery data are available unclear. Matched action-generation comparisons and more targeted data/reward ablations would strengthen the attribution of the reported gains to the proposed components.
  2. The dataset construction is not described in sufficient operational detail. Appendix B provides the overall composition, scenario coverage, examples, and reasoning-annotation pipeline, but does not clearly specify how negative events are selected, how segment boundaries are determined, or what qualifies as a successful expert recovery. More detailed scenario-level train/test statistics would also help characterize the benchmark. Since nuReasoning-NR is a central contribution, these details are important for assessing annotation consistency, coverage, and reproducibility.
Questions for authors
  1. Dataset construction: What operational criteria were used to select negative events and identify successful expert recoveries? Please clarify how segment boundaries were determined, how these judgments were checked, and how scenario-level separation between training and evaluation was enforced. A breakdown of scenario categories across the two splits would also help characterize coverage.

  2. Negative versus recovery supervision: How much does negative supervision contribute when recovery examples are already available? Could you provide a positive–recovery comparison or separately ablate the negative penalty and recovery reward under matched training budgets? The existing ablations establish the overall benefit, but do not isolate this conditional contribution.

  3. Efficient Action Bridge: Can you provide a comparison between autoregressive action generation and the proposed bridge using the same VLM checkpoint and reasoning configuration, reporting planning quality, latency, and additional parameter counts? For the full-versus-sparse cache comparison, please clarify the effect of matching historical initialization and bridge capacity. Evidence on nuReasoning-NR would further help assess the bridge in the negative/recovery setting.

2 comments

Weakness 1, Question 2 and 3: Evidence from the Action-Bridge and NR-RFT Ablations

We thank the reviewer for carefully examining the ablation studies. We clarify the existing evidence and provide an additional positive–recovery comparison below.

Choice of evaluation benchmark. NAVSIM v1 provides a large, standardized evaluation set of approximately 12K non-trivial driving scenarios, making it a useful setting for assessing planning quality and generation efficiency across diverse situations. Our Table 4 additionally reports reasoning-length statistics averaged over 10K samples. This scale motivates using NAVSIM v1 for the main bridge ablations.

Negative versus recovery supervision. Negative examples identify behaviors to discourage, while recovery demonstrations provide corrective references. Figure 4 reports 60.6 PDMS for positive-only training, 64.2 for positive–negative training, and 69.6 for the complete positive–negative–recovery recipe.

We additionally evaluated a positive–recovery baseline, which achieves 65.4 PDMS. This comparison uses the same initialization, training-step and rollout budgets, downstream bridge-training procedure, and evaluation and selection protocol. With recovery supervision already present, adding negative supervision improves PDMS by 4.2 points, from 65.4 to 69.6.

Training configurationPDMS
Positive only60.6
Positive + negative64.2
Positive + recovery65.4
Positive + negative + recovery69.6

Reward ablations examine both the joint contribution and parameter sensitivity. In Figure S4, with positive, negative, and recovery training data present, disabling both reference-based reward terms yields 87.6 PDMS, compared with 90.8 when both are enabled. Weight settings of 2, 5, and 8 yield 89.1, 90.8, and 88.4, respectively; proximity thresholds of 0.8, 1.7, and 2.5 m yield 89.7, 90.8, and 89.2.

Action-generation comparison. Table 4 compares autoregressive decoding and continuous action generation, and we add SpanVLA with autoregressive decoding here. Actually, SpanVLA and AutoVLA use the same Qwen2.5-VL VLM backbone. For ten-waypoint trajectories, the reported results are:

ConfigurationPDMSReasoning tokensAction-generation time
AutoVLA, autoregressive89.1230.40 s
SpanVLA, autoregressive89.8150.40 s
SpanVLA, L1 head85.1150.02 s
SpanVLA, flow matching90.8150.08 s

The flow-matching configuration achieves 1.0 points higher PDMS and 5× faster action generation than SpanVLA with autoregressive action decoding. Action-generation latency is measured separately from reasoning latency. These results demonstrate the quality–efficiency trade-off of the evaluated systems.

Bridge-design ablations. Table 5 provides further comparisons:

ConfigurationPDMSAction-generation time
Full caching88.10.18 s
Last layer79.30.01 s
Interval 482.20.05 s
Interval 2, without historical initialization86.40.08 s
Interval 2, with historical initialization90.80.08 s

In particular, with interval-2 conditioning fixed, historical initialization improves PDMS by 4.4 points without increasing the measured generation time. The full-cache baseline also includes the historical initialization, and the interval-2 baseline gets better results.

Weakness 2 and Question 1: Dataset Construction and Split Protocol

We thank the reviewer for carefully reading our paper and identifying the operational details that need clarification. We hope the following explanation makes the collection, annotation, and split procedures clearer.

Negative-event selection. The negative and recovery samples are collected during real-world testing of a deployed SOTA driving policy, with a driving expert onboard. Experts intervene when they judge the policy’s behavior to be unsafe or potentially dangerous, including situations involving unsafe traffic interactions or potential traffic-rule violations. We retrieve the corresponding driving-log segments using failure and takeover tags recorded during testing. These events therefore reflect expert-identified problems encountered during actual driving.

Segment boundaries and recovery demonstrations. Sample anchors are selected manually relative to the expert’s takeover time, without a fixed temporal offset across events. Each sample contains its own 1.5-second observation history and a five-second future reference trajectory sampled at 2 Hz. Negative and recovery segments correspond to different timestamps within the same event; they are temporally associated demonstrations, rather than alternative trajectories starting from an identical state. Recovery samples contain trajectories executed by driving experts after takeover.

Manual quality checks. Following collection, we manually reviewed the candidate segments. This review is separate from the reasoning-annotation audit in Appendix B.3, which concerns positive CoT annotations.

Training–evaluation separation and scenario coverage. We partition the dataset by driving log, keeping each event’s negative segments, corresponding recovery segments, and neighboring windows in the same split. Thus, segments from the same driving log are not shared between training and evaluation. The evaluation subset contains 300 negative and 300 recovery samples.

Figure 1 summarizes the dataset’s scenario coverage, including cut-in and lead braking, stop signs, vulnerable road users, construction zones, lane bias, ego lane changes, and other scenarios. Both splits originate from the same collection pipeline.

Reviewer_zm1ROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper identifies that existing vision‑language‑action driving models mainly learn from positive expert demonstrations and lack supervision from negative failure trajectories and corresponding recovery behaviours. The authors construct nuReasoning‑NR, a new real‑world dataset containing negative, positive and recovery driving samples. They propose SpanVLA together with Negative‑Recovery Reinforcement Fine‑Tuning (NR‑RFT). NR‑RFT applies asymmetric behavioural rewards to penalize undesirable behaviours and reward expert recovery corrections. A lightweight sparse‑KV‑cache action‑bridging module is further designed to improve inference efficiency for continuous trajectory generation. Evaluations on NAVSIM benchmarks and the proposed nuReasoning‑NR dataset show improved overall planning performance and stronger robustness on challenging negative‑recovery scenarios, with comprehensive ablation studies validating each core component.

Critical strengths
  1. The paper builds nuReasoning‑NR, a real‑world driving dataset explicitly containing negative failure trajectories and expert recovery samples. It fills a gap in existing driving datasets which mostly focus only on successful demonstrations, and enables systematic research on learning from driving failures and recovery behaviours.
  2. The proposed NR‑RFT in Section 5 introduces asymmetric reward design for negative‑recovery samples. It adapts GRPO with advantage correction to handle problematic near‑identical roll‑out groups, offering a practical solution for stable post‑training under mixed positive‑negative driving supervision. The lightweight action‑bridging module with sparse KV‑cache conditioning effectively reduces generation latency while preserving planning quality.
  3. The paper delivers high‑quality figures with strong visual clarity. Diagrams for model architecture, dataset examples and ablation plots are well‑organised, with properly labelled sub‑panels and distinct colour coding. Visualizations effectively help readers understand negative‑recovery scenarios, module workflow and experimental contrasts without heavy reliance on table numbers.
  4. Extensive evaluations are conducted across NAVSIM v1/v2 and the custom nuReasoning‑NR benchmark. Systematic ablations dissect the contribution of negative‑recovery supervision, advantage correction components and different bridging configurations. Both quantitative metrics and rich qualitative case studies support the claimed improvements.
Critical weaknesses
  1. Several initiatives in the reward design of Section 4.3 do not hold up under scrutiny. The negative penalty is a clipped distance to the one recorded negative trajectory, so it vanishes beyond the threshold δ, and a rollout that is bad in a different way receives no penalty at all, which means the reward never states what good driving is in these scenarios, only that one specific trajectory should be avoided. Removing ego progress from the driving reward for negative samples deletes the only term that rewards forward motion, leaving pure risk avoidance. Finally, when all rollouts in a group fail, assigning the same virtual reward to every rollout produces no within-group ranking signal. Please clarify how NR-RFT learns to distinguish less harmful failures from more severe failures in such groups, or whether these groups are excluded or handled separately. These choices jointly explain why the appendix shows NR training dropping ego progress on navhard by roughly 20 points while the EPDMS gain comes almost entirely from the safety terms, i.e., the policy just becomes more conservative. That outcome is the opposite of the paper's stated motivation, which presents negative samples as a cure for hesitant behavior, and the main text never discusses it.
  2. The best-of-6 evaluation may introduce an evaluation-protocol issue. The paper appears to retain the candidate with the highest PDMS or EPDMS, which is also the metric being reported. Please clarify whether ground-truth evaluation scores are used to select the candidate. If so, this would amount to oracle access to the test metric and would make the comparison with other methods unfair. If an internal model score is used instead, the paper should define this score and explain whether it was available during training. In addition, the non-best-of-6 results appear to be very close to the competing methods, with differences of only a few tenths of a point. These margins may be within normal run-to-run variation. The state-of-the-art claim should therefore be qualified unless the authors provide statistically reliable comparisons.
  3. Several details needed for reproduction are missing or inconsistent. The reward weights λ_N and λ_R are never specified numerically (the appendix only says a "moderate setting" performed best), and δ is only implied through an ablation. The SFT loss definition is ambiguous. The paper assigns a weight of 0 to slow-thinking samples and a weight of 1 when CoT is activated, but it is unclear whether these refer to the same subset of samples. If they are the same, the stated formulation would appear to remove the language-model loss from samples containing reasoning. Please clarify the exact masking rule and provide a concrete example. L_Action also sums over a range that is a subset of L_LM, so the two losses double-count action tokens.
Questions for authors
  1. Citation error in Table S1: DriveVLA-W0 (?) .
  2. On navhard, NR training reduces ego progress by roughly 20 points while improving the safety terms. Why does training on negative samples make the policy more conservative, and how is this consistent with the claim that such samples reduce hesitant behavior? Is the same trade-off present on nuReasoning-NR, and if so, why is it not reported?
8 comments

Weakness 1: Reward Design and the Safety–Progress Trade-off

We thank the reviewer for their careful reading and for highlighting the need to clarify the reward design and the observed driving behavior.

First, we do see the benefit of training with negative and recovery data in navhard and nuReasoning-NR benchmarks.

The negative penalty supplements a broader driving objective. Its purpose is to discourage proximity to a recorded undesirable trajectory. Clipping ensures that increasing this distance beyond the threshold provides no further discouraging reward from that term. If we have no threshold δ for the additional negative reward, the best trajectory would be the one that is farthest away from the negative sample, where the vehicle will choose to go to the right-turn lane in a left-turn scenario. Other unsafe rollouts remain subject to the driving reward, while positive and expert-recovery examples provide supervision for desirable behavior. Thus, avoiding the recorded negative trajectory is only one part of the learning signal, and an amount of negative, positive, and recovery data provides scaling and comprehensive supervision signal for the policy training.

The virtual anchor supplies a negative advantage, not a failure-severity ranking. It modifies the baseline rather than assigning an identical virtual reward to every rollout. For negative or recovery queries whose final rollout rewards are all zero, we retain the group without resampling. The anchor yields Ai=−1/(G+1)A_i=-1/(G+1) for each rollout. It therefore provides a common negative advantage but does not distinguish severity among equally scored failures. The core challenge of the all-zero group is that it provides no advantage from the mean 0 and std 0, and the additional virtual samples provide an additional anchor to identify that zero-reward behavior is bad and provide the learning signal during training. Most rollouts with zero reward are trajectories with collisions, and all of them are more harmful failures. The failure level would be distinguished by various non-zero rollouts in a large amount of negative data. We will clarify that more clearly.

Safety and progress must be considered together in challenging scenarios. High ego progress alone does not establish successful driving: a trajectory can make substantial progress while still resulting in a collision. On navhard, the NR configuration improves three key safety-related scores consistently: no-at-fault collision increases from 86.9 to 89.8, driving-direction compliance from 87.1 to 94.8, and time-to-collision from 82.7 to 88.2. These gains cover collision avoidance, rule compliance, and collision-risk anticipation.

DrivoR [1] also mentioned that high ego progress leads to worse EPDMS in navhard, so they suggest that prioritizing ego progress too heavily can hurt EPDMS scores, especially in the second stage of navhard, where perturbations to the ego vehicle’s position and heading call for a cautious return to a safe path, and aggressive forward motion increases the risk of collisions. Our qualitative observations also show that the policy becomes more conservative to be consistent with the safety-oriented training objective.

The progress reduction is not universal across our evaluations. On nuReasoning-NR, adding NR supervision improves EP from 58.6 to 66.9, PDMS from 60.6 to 69.6, and comfort from 92.5 to 93.2; all remaining reported metrics improve or remain unchanged. These results demonstrate that negative and recovery supervision can support improvements in both safety and progress. Our motivation is to correct context-dependent undesirable behavior, including unnecessary hesitation. Appropriate caution in hazardous situations and avoiding unnecessary hesitation are compatible objectives, although the navhard results show that their balance still requires improvement.

Weakness 2: Oracle Best-of-6 Evaluation and the Contribution of nuReasoning-NR

We thank the reviewer for carefully examining the evaluation protocol and comparative claims.

Oracle best-of-N evaluation is a commonly used diagnostic in driving VLA research. Both AutoVLA and NoRD report oracle selection from six generated trajectories. Our best-of-6 follows this existing practice to assess candidate-generation potential under a fixed sampling budget. Specifically, we use benchmark scores to select candidates, rather than an internal learned scorer. We will explicitly label the starred results as Oracle Best-of-6 and distinguish them from single-output performance.

Recent NAVSIM v1 results already occupy a high-scoring, closely grouped range. For context, Qwen-Drive-1.0, built on Qwen3.5-4B, reports 90.7 PDMS and 91.4 with oracle best-of-6 on NAVSIM v1.1. Our paper reports 90.8 and 93.3, respectively, using a weaker Qwen2.5-VL-3B, which further demonstrates our SOTA performance.

Our central contribution concerns learning from negative and recovery behavior. On nuReasoning-NR, adding NR supervision increases SpanVLA’s average PDMS from 60.6 to 69.6, with improvements on both negative scenarios (65.1 → 74.1) and recovery scenarios (56.1 → 65.1). Ego progress also improves from 58.6 to 66.9. These are continuous, single-output results. Moreover, applying NR supervision to AutoVLA improves its average PDMS from 59.1 to 64.1, providing evidence beyond our own architecture.

These results motivate nuReasoning-NR as a complementary benchmark that explicitly measures behavior in negative and recovery scenarios. Our contribution therefore extends beyond small differences in aggregate NAVSIM scores.

Weakness 3: Reward Hyperparameters and the SFT Objective

We thank the reviewer for carefully examining the training formulation. We clarify the reward settings, the intended SFT weighting, and the role of the additional action loss below.

Reward hyperparameters. Both the negative-behavior penalty and recovery reward use a weight of 5 and a proximity threshold of 1.7 m. Figure S4 presents the corresponding sensitivity analysis.

Mixed supervision for reasoning and action generation. Our SFT data contain both reasoning-annotated samples with action targets and samples with action targets only. In the intended formulation, the LM loss has weight 1 for samples with CoT and 0 for samples without CoT, while action supervision remains active in both cases.

Input/context tokens are excluded from both target losses. When enabled, the LM loss covers supervised assistant-output tokens, including reasoning, formatting, and action tokens. The action loss selects only action-codebook tokens. Mixing these target formats is intended to develop both reasoning-assisted planning and direct action generation, providing a foundation for adaptive reasoning in SFT training.

Intentional action-token reweighting. The overlap between the losses is intentional. The standard five-second trajectory contains ten action-codebook tokens, while reasoning length varies across samples. As reasoning becomes longer in the slow thinking samples, action tokens account for a smaller fraction of the full-response average loss. The additional, separately normalized action loss strengthens trajectory supervision and prevents its contribution from being diluted by longer reasoning sequences.

   

Question 1: Citation Error

Thank you for catching this citation error. The unresolved citation in Table S1 should refer to DriveVLA-W0 (Li et al., 2025a), which is already included in the bibliography. We will correct the citation.

   

Question 2: Conservatism and Progress across Benchmarks

We thank the reviewer for highlighting the relationship between safety, progress, and hesitant behavior.

On navhard, we observe more conservative and safe driving after adding NR supervision. The NR configuration improves no-at-fault collision from 86.9 to 89.8, driving-direction compliance from 87.1 to 94.8, and time-to-collision from 82.7 to 88.2. Overall EPDMS increases from 40.1 to 40.8, accompanied by a substantial EP reduction from 85.5 to 63.4.

DrivoR [1] also mentioned that high ego progress leads to worse EPDMS in navhard, so they suggest that prioritizing ego progress too heavily can hurt EPDMS scores, especially in the second stage of navhard, where perturbations to the ego vehicle’s position and heading call for a cautious return to a safe path, and aggressive forward motion increases the risk of collisions. Our qualitative observations also show that the policy becomes more conservative to be consistent with the safety-oriented training objective.

Reducing unnecessary hesitation and exercising appropriate caution are compatible objectives. Our motivation is to correct context-dependent undesirable behavior: making progress when conditions permit, while avoiding unsafe maneuvers in hazardous situations. However, the navhard results show that the current balance favors safety at an efficiency cost.

The same aggregate trade-off is not observed on nuReasoning-NR, and its metric breakdown is already reported in Table 1. Adding NR supervision improves EP from 58.6 to 66.9, average PDMS from 60.6 to 69.6, and comfort from 92.5 to 93.2; the remaining reported metrics improve or remain unchanged. These results demonstrate that NR supervision can support improvements in both safety and progress. Improving the efficiency and comfort of the more conservative navhard behavior while retaining its safety gains remains a focus of our future work.

   

Reference

[1] Kirby, E., Boulch, A., Xu, Y., Yin, Y., Puy, G., Zablocki, É., ... & Cord, M. (2026). Driving on registers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 32058-32069).

Reviewer_rJGqOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a limitation of driving VLAs: they usually learn only from successful demonstrations, with little training on which behaviors to avoid or how to recover when things go wrong. The authors introduce nuReasoning-NR, a dataset with 30K positive samples, 3K negative samples, and 3K recovery samples. They also propose NR-RFT, which penalizes predictions that stay close to undesirable driving trajectories and rewards predictions that follow expert corrections. To speed up trajectory generation, the model starts from the historical trajectory and uses information from selected VLM layers to generate continuous actions. Experiments cover NAVSIM v1/v2, nuReasoning, and nuReasoning-NR. On nuReasoning-NR, adding negative and recovery data raises the average PDMS from 60.6 to 69.6. The action bridge also substantially reduces trajectory generation time.

Critical strengths

Negative and recovery data improve planning in difficult scenarios, and the benefit extends to more than one model.In Table 1 (p. 7), adding NR data raises SPANVLA's average PDMS from 60.6 to 69.6 and AutoVLA's from 59.1 to 64.1. The left panel of Figure 4 further shows a score of 64.2 when only negative samples are added, rising to 69.6 when recovery samples are included as well. These results support the paper's central idea: learning from mistakes and the subsequent corrections is useful alongside learning from successful demonstrations The action bridge generates trajectories much faster while also achieving higher planning scores.In Table 4 (p. 8), the time to generate a 10-waypoint trajectory drops from AutoVLA's 0.40 seconds to 0.08 seconds. Table 5 (p. 9) also isolates the effect of historical-trajectory initialization: with the same cache configuration, PDMS rises from 86.4 to 90.8 while generation time stays at 0.08 seconds. The paper therefore demonstrates a speed improvement and provides a controlled comparison showing the benefit of historical-trajectory initialization.

Critical weaknesses

The paper does not clearly explain how reinforcement fine-tuning improves the continuous trajectories used at inference.** Section 4.4 and Appendix D.1 state that rewards are computed from trajectories decoded from action tokens, and that these rewards are used to update the VLM. Section 4.1, however, says that a separate continuous action module generates trajectories at inference. Appendix C.1 only describes how this module is trained during supervised fine-tuning; it does not explain whether it is updated further before or after reinforcement fine-tuning. The updated VLM could improve continuous trajectories through its internal features, so this does not establish that the method is invalid. Still, the authors need to spell out the training sequence and evaluation setup so readers can verify that the reported planning gains and generation speedup are both achieved by the same final model.

Questions for authors
  1. Please clarify the order of the training stages, which modules are updated at each stage, and how trajectories are generated for Tables 1, 3, and 4 and Figure 4. Is the continuous action module retrained after reinforcement fine-tuning? If the main results all come from the final model using continuous outputs, clarifying the existing procedure would largely address this concern. If the planning gains mainly come from trajectories decoded from action tokens, it is necessary to establish whether those gains carry over to the continuous outputs.

  2. Why does adding NR data improve safety metrics while noticeably reducing progress and comfort? In Table S3 (p. 23), the overall EPDMS rises from 40.1 to 40.8. In the second stage, the collision, driving-direction, and time-to-collision metrics improve, but Ego Progress (EP) drops from 85.5 to 63.4 and Extended Comfort (EC) from 67.4 to 50.8. Please explain what driving behaviors account for these changes. If the model mainly reduces risk by driving less, the loss in driving efficiency needs to be explained. If it avoids dangerous situations while still completing the driving task, that would better support the claim that it handles difficult scenarios more effectively.

5 comments

Weakness and Question 1:

We thank the reviewer for their careful reading of our paper and for highlighting the aspects that require further clarification. We hope the explanation below clarifies the training sequence, the connection between NR-RFT and continuous action generation, and the evaluation setup.

The continuous action module is trained again after NR-RFT. We refer to the training strategy of Alpamayo from NVIDIA [1]. The complete training sequence is:

  1. SFT: Train the VLM language component and continuous action module with supervised objectives, while keeping the vision encoder frozen.
  2. NR-RFT: Update the VLM language policy through LoRA using rewards computed from trajectories decoded from action tokens. The continuous action module is not updated during this stage.
  3. Post-RFT flow matching: Freeze the updated VLM and train the continuous bridge and flow-matching head.

Thus, NR-RFT influences continuous outputs through both the updated VLM conditioning used to train the post-RFT bridge.

The SpanVLA results in Tables 1 and 3 and Figure 4 use continuous outputs from the final model associated with each training configuration. Table 4 uses continuous outputs for the flow-matching and L1-head variants, and action-token decoding for the autoregressive baselines. We will further make it more clear in the revised version.

For the quality–speed comparison, Table 4 reports 90.8 PDMS and 0.08-second action generation for the 10-waypoint flow-matching configuration. The 21.5× action-generation speedup concerns the 50-waypoint configuration.

 

Question 2: Safety Prioritization in Challenging Scenarios

Safety remains a central challenge in navhard. High ego progress alone does not establish successful driving: a trajectory can achieve substantial progress while still resulting in a collision. Our objective therefore emphasizes handling these difficult situations more safely.

Adding NR data yields consistent improvements across three key safety-related metrics: no-at-fault collision increases from 86.9 to 89.8, driving-direction compliance from 87.1 to 94.8, and time-to-collision from 82.7 to 88.2. These gains indicate improvements across multiple aspects of safety.

DrivoR [2] also mentioned that high ego progress leads to worse EPDMS in navhard, so they suggest that prioritizing ego progress too heavily can hurt EPDMS scores, especially in the second stage of navhard, where perturbations to the ego vehicle’s position and heading call for a cautious return to a safe path, and aggressive forward motion increases the risk of collisions. Our qualitative observations also show that the policy becomes more conservative to be consistent with the safety-oriented training objective.

Overall, EPDMS increases from 40.1 to 40.8 despite the reductions in progress and comfort. These results support improved safety-related performance in challenging scenarios, accompanied by more conservative behavior. Improving efficiency and comfort while retaining these safety gains is a focus of our future work.

 

Reference

[1] Wang, Y., Luo, W., Bai, J., Cao, Y., Che, T., Chen, K., ... & Pavone, M. (2025). Alpamayo-r1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail. arXiv preprint arXiv:2511.00088.

[2] Kirby, E., Boulch, A., Xu, Y., Yin, Y., Puy, G., Zablocki, É., ... & Cord, M. (2026). Driving on registers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 32058-32069).

AuthorsOpenReview ↗

Follow-Up to Your Review

Dear Reviewer rJGq,

Thank you again for the time and effort you put into reviewing our work!

We understand that this might be a super busy period for you. We wanted to briefly let you know that we have posted point-by-point responses to your questions and concerns.

Whenever you have a chance, we would be grateful for your feedback on whether the responses address your concerns. We would be happy to clarify anything that remains open.

Thank you!

AuthorsOpenReview ↗

Follow-Up to Your Review (Second Reminder)

Dear Reviewer rJGq,

We are following up on our point-by-point responses, which we posted several days ago. We believe these materials further address your concerns.

As the end of the Author-Reviewer Discussion phase is fast approaching, please let us know whether our responses address your main concerns or if any points still need clarification. We are available to discuss them here.

Thank you again for your time and effort devoted to this review!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.