acceptodds
Under review as a conference paper at ICLR 2027

Restoring Spatial Distance Feedback in Token-Native Driving VLAs

Abstract

Discretizing continuous trajectory waypoints directly into a Vision Language Action (VLA) model's vocabulary allows it to treat motor control as standard autoregressive generation. This unified formulation ensures that policy execution directly inherits the pre-training capabilities and scaling law benefits of the underlying Vision Language Model (VLM), bypassing the need for specialized continuous regression heads. However, optimizing these tokenized action spaces with standard categorical cross-entropy imposes an isometric loss landscape that penalizes all incorrect tokens uniformly, regardless of their physical error in metric space. We demonstrate that this loss geometry causes closed-loop policy failure in models trained with discrete K-means tokens (K-DISK) despite high action reconstruction accuracy. We address this bottleneck through two core contributions: 1) adapting Ordered Action Tokenization (OAT) from robotics to autonomous driving, where continuous trajectory proximity is mapped to shared sequence prefixes, and 2) Coarse Trajectory Anchor Regularization (CTAR), a differentiable metric loss that backpropagates smooth- trajectory errors directly through the first coarse token via a frozen decoder. This provides the explicit metric feedback missing from pure cross-entropy. Evaluated on CARLA Bench2Drive and on NAVSIM v2, our method yields significant gains over standard K-DISK (+33.57 DS, +0.0979 EPDMS) and matches the performance continuous action heads (+1.55 DS, +0.0425 EPDMS), demonstrating that VLAs can achieve closed-loop fidelity without departing from native token architectures.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action, autonomous driving, ordered action tokenization, discrete action tokens

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nBkZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the optimization challenges of token-native VLAs for autonomous driving. The authors argue that cross-entropy loss does not account for the physical distance between discrete action tokens. As a result, predictions with different trajectory errors may receive the same loss, leading to poor closed-loop driving performance.

To address this issue, the authors adapt Ordered Action Tokenization (OAT) to driving trajectories, using coarse-to-fine token sequences to establish a relationship between shared prefixes and spatial distance. They further propose Coarse Trajectory Anchor Regularization (CTAR), which introduces continuous spatial distance supervision for the first coarse token through a frozen trajectory decoder.

Experiments on CARLA Bench2Drive and NAVSIM v2 show substantial improvements over K-DISK. The proposed method also matches or outperforms the evaluated continuous control baselines on the overall metrics.

Critical strengths
  1. The paper addresses a meaningful problem. Rather than simply modifying the model architecture, the authors focus on the mismatch between classification loss and physical distance in discrete action spaces. In particular, K-DISK achieves lower trajectory reconstruction error but higher actual prediction error, which provides strong motivation for the proposed approach.

  2. The method is well motivated and shows strong experimental results. OAT and CTAR address discrete action prediction from two complementary perspectives: ordered action representations and continuous distance supervision. The method achieves clear improvements on two autonomous driving benchmarks, and the CTAR ablation demonstrates the importance of this component.

Critical weaknesses
  1. Inference efficiency is not quantitatively evaluated. The proposed method requires autoregressive generation of 32 action tokens, which may introduce substantial inference overhead compared with a continuous regression head. Although the paper discusses the impact of inference latency on evaluation, it does not report actual inference time or throughput. Given the real-time requirements of autonomous driving, a latency comparison with the continuous baseline is needed to assess the trade-off between performance gains and inference cost.

  2. Some experimental results are not sufficiently analyzed. On NAVSIM v2, although the proposed method achieves a higher EPDMS, its Extended Comfort score drops substantially from 0.747 for the continuous baseline to 0.373. The paper does not adequately explain this degradation.

Questions for authors
  1. What are the actual inference latency and throughput when generating 32 action tokens autoregressively? Compared with the continuous regression baseline, can the proposed method meet the same real-time control requirements?

  2. Why does OAT + CTAR achieve a substantially lower Extended Comfort score than the continuous baseline? Is this degradation related to trajectory smoothness, acceleration, or jerk? It would be helpful to provide further analysis of the underlying causes.

  3. CTAR applies distance supervision only to the first coarse token, while cross-entropy for the subsequent tokens still lacks explicit spatial distance information. Could the authors further explain the rationale for supervising only the first token? Could applying distance supervision to multiple prefixes further improve trajectory quality?

3 comments

We thank the reviewer for the thoughtful questions and constructive feedback. We address each point below and have updated the revised PDF accordingly:

1. Inference Latency, Throughput, and Real-Time Control (Weakness 1 & Question 1): We acknowledge that our study does not benchmark wall-clock latency or throughput: neither OAT + CTAR nor the compared VLM baselines are real-time systems in their current configurations, as all share the same InternVL2-1B backbone processing 1,024 visual tokens plus autoregressive language tokens before action emission.

  • Generating 32 action tokens sequentially adds 32 short KV-cached decoding steps over a single-pass continuous MLP head, on top of the shared InternVL2-1B visual prefill and text generation.
  • Unlike flat tokenizers that must emit all tokens before decoding, OAT's nested prefix structure supports anytime prefix decoding , letting a single trained model truncate generation at any k without retraining to trade off steps for latency, e.g., zero-shot truncation to 10 tokens (69%69\% fewer action steps) still beats 10-token K-DISK on CARLA scenario44 (35.91 vs. 29.65 DS; Section 4.5).
  • Keeping control inside the native discrete vocabulary ensures deployment directly inherits standard LLM accelerations (speculative decoding, KV-caching, and quantization). Added an explicit real-time latency limitation/future-work note in Section 5.

2. Root Cause of Lower Extended Comfort (0.373 vs. 0.747) on NAVSIM v2 (Weakness 2 & Question 2): We explicitly analyze this degradation as a limitation in Sections 4.3 and 5 (alongside Appendix Section A.9):

  • Single-frame Comfort remains high (C = 0.962 vs. 0.976; Table 2), with zero lateral acceleration inflation (1.0x baseline) and superior spatial compliance (DAC = 0.853 vs. 0.753, LC = 0.964 vs. 0.940, TTC = 0.942 vs. 0.931). Individual trajectories decoded at each tick are physically smooth.
  • Unlike C, EC measures temporal consistency across consecutive replanning frames. Because our policy decodes memorylessly at each tick without conditioning on the prior frame's predicted prefix, minor visual shifts near a codebook boundary can step argmax selection between adjacent discrete codes. As shown in Appendix Section A.9, our deployed NAVSIM codec (1.182 m ADE) has residual quantization error that is along-track dominated (2.1–3.1x along-to-cross track ratio). Discrete longitudinal position/velocity jumps between frames t and t+1 induce 6.4–13.6x longitudinal jerk inflation across ticks, penalizing EC; whereas continuous heads vary smoothly with small visual perturbations.
  • In closed-loop CARLA execution, vehicle inertia and tracking controllers filter this jitter (73.75 DS, 56.4% SR; Table 1). Prior-tick action prefix conditioning and cross-frame smoothness losses are highlighted as future work.

3. Rationale for Supervising Only t=1 in CTAR & Multi-Prefix Extension (Question 3):

  • Under teacher forcing (Eq. 1), steps t>1 already condition on ground-truth prefixes, localizing candidate trajectories within a bounded spatial neighborhood where CE focuses on local refinement. By contrast, t=1 has no conditioning prefix. It is an unconstrained global decision setting macro trajectory scale and direction, where pure CE assigns identical loss to all non-target codes regardless of physical error.
  • Supervising t=1 requires only one forward/backward pass through frozen D per step and avoids chaining straight-through estimator approximations across multiple autoregressive steps.
  • Extending CTAR to a small set of coarse prefixes (k1,2,4k_{1,2,4}, evaluating on soft distributions at 1…k1 \dots k with zero-padded suffixes) is an avenue for future work to provide finer multi-scale curvature gradients at the cost of k decoder passes.
Area_Chair_hzodOpenReview ↗

Please help to join the discussion

Dear Reviewer nBkZ,

The discussion deadline is approaching. Please help to join the discussion and offer the reviewer's opportunity to further respond to your concerns.

Thanks!

AC

Thank you for the detailed response. The analysis of cross-frame inconsistency and longitudinal jerk provides a reasonable explanation for the drop in Extended Comfort. The rationale for applying CTAR only to the first token is also convincing. My main concern remains inference efficiency. Although the authors provide additional results on prefix truncation, actual latency and throughput measurements are still missing, making it difficult to assess the overhead relative to continuous regression heads and the trade-off between performance and efficiency. Overall, I appreciate the value of the problem and the experimental results. However, as inference efficiency remains insufficiently evaluated, I will maintain my original rating.

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

This paper investigates the mismatch between categorical cross-entropy supervision and physical trajectory distance in token-native vision-language-action models for autonomous driving. It adapts Ordered Action Tokenization (OAT) to represent trajectories as coarse-to-fine token sequences and introduces Coarse Trajectory Anchor Regularization (CTAR), which provides explicit trajectory-distance supervision for the first coarse token through a frozen decoder. Evaluations on CARLA Bench2Drive and NAVSIM v2 report substantial improvements over a K-means tokenization baseline and competitive performance against the evaluated continuous-output models. Reconstruction analyses and ablation studies examine the gap between tokenization fidelity and policy performance, along with the contributions of CTAR and decoder freezing.

Critical strengths
  • The gap between action representation and policy learning is a nice motivation.
  • The paper borrows OAT from robotics and introduces CTAR to address the lack of spatial distance feedback in standard categorical cross-entropy. The two components fit together well.
  • Large performance improvements over K-DISK on both CARLA Bench2Drive and NAVSIM v2.
  • Nice figures and captions. They make the main idea and method easy to follow.
Critical weaknesses
  • The “zero-gradient plateau” description of cross-entropy needs clarification. Cross-entropy still provides gradients with respect to logits; what it lacks is explicit information about physical trajectory distance.
  • The comparison against K-DISK changes several factors at the same time, including the tokenizer, token sequence length, and training loss. This makes it hard to tell how much each change contributes to the gains.
  • Table 2: Higher aggregate scores do not mean improvements across all aspects. In particular, the extended comfort score is much lower than that of the continuous baseline (0.373 vs. 0.747).
Questions for authors

Typo: CATR in Tab2. It should be CTAR.

2 comments

We thank Reviewer 2 for the careful reading and constructive feedback. We address each point below and have incorporated the corresponding updates in the revised PDF:

1. Clarifying "Zero-Gradient Plateau" in Cross-Entropy: You are completely right: categorical cross-entropy always produces non-zero logit gradients. What has a flat (zero) gradient is the metric sensitivity of the loss with respect to physical trajectory error: because all incorrect indices incur identical loss regardless of their physical distance, shifting probability mass from a catastrophic far-miss (5 m away) to a near-miss (5 cm away) produces zero reduction in cross-entropy loss, leaving logit gradients uninformative about geometric proximity. We have revised the wording in the Introduction to state this precisely.

2. Isolating Tokenizer, Sequence Length, and Training Loss vs. K-DISK: We isolate all three factors across controlled ablations in Table 3 and Section 4.5 on CARLA scenario44:

  • Tokenizer vs. Loss at Fixed 32-Token Length (Table 3): Holding the 32-token vocabulary and learned decoder fixed at matched training steps (105k), switching from unstructured K-DISK to ordered tokens under pure CE (OAT w/o CTAR) improves Driving Score from 35.50 → 48.50 DS (+13.00 DS), and adding the CTAR metric loss yields a further +26.69 DS (48.50 → 75.19 DS).
  • Metric Loss Alone on K-DISK (Section 4.5): Training 10-token K-DISK with distance-based Gaussian soft-label CE improves scenario44 performance at step 35k from 29.65 → 33.99 DS (+4.34 DS; Route Completion 43.94% → 49.84%), confirming that metric-aware supervision helps flat K-means tokens, yet still trails OAT + CTAR (83.49 DS at 35k; 75.19 DS at 105k) because K-DISK lacks full-horizon coarse-to-fine prefix compression.
  • Equal 10-Token Budget Comparison (Section 4.5): Increasing K-DISK from 10 to 32 tokens would not fix its bottleneck because 10-token K-DISK already achieves near-lossless reconstruction (0.0146 m ADE vs. 0.315 m for 32-token OAT; Fig. 4b), whereas its failure stems from 1.840 m policy prediction error. Conversely, using OAT's anytime prefix decodability to truncate our 32-token OAT + CTAR policy zero-shot at inference to 10 tokens (without retraining) still outperforms 10-token K-DISK (35.91 vs. 29.65 DS, +6.25 DS; RC 59.01% vs. 43.94%) and 10-token soft-label K-DISK (33.99 DS).

3. Table 2 Sub-Metrics & Lower Extended Comfort (0.373 vs. 0.747): We agree that higher aggregate EPDMS (0.2861 vs. 0.2403) does not imply uniform gains across every sub-metric, and we explicitly acknowledge lower two-frame Extended Comfort (EC = 0.373 vs. 0.747) as a limitation in Section 4.3 and Section 5:

  • Single-frame physical Comfort remains high (C = 0.962 vs. 0.976, with 1.0x lateral acceleration relative to baseline) alongside large gains in spatial safety/compliance (DAC = 0.853 vs. 0.753, LC = 0.964 vs. 0.940, TTC = 0.942 vs. 0.931).
  • The EC drop occurs specifically across consecutive replanning frames (t→t+1t \rightarrow t+1) because our policy decodes memorylessly at each tick over an along-track-dominated discrete codec (1.182 m ADE; Appendix Sec. A.9), causing discrete longitudinal step jumps across ticks when visual inputs cross quantization boundaries.
  • In closed-loop CARLA execution, vehicle inertia and low-level tracking controllers filter this high-frequency jitter (73.75 DS, 56.4% SR; Table 1). Conditioning generation on prior-tick action prefixes and adding cross-frame smoothness losses is highlighted as future work.

4. Typo (CATR → CTAR) in Table 2: Thank you for catching this typo. We have corrected the spelling and verified consistent naming throughout the revised manuscript.

Thank you for the response and the additional analyses. I will maintain my positive score.

Reviewer_Gr8POpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses issues with cross-entropy loss for driving VLAs, where high action reconstruction accuracy may not correlate with errors during closed-loop performance. The authors first adapt ordered (i.e., coarse-to-fine) action tokenization from prior work, as well as introduce CTAR – a coarse trajectory anchor regularization mechanism which optimizes a continuous Smooth-L1 loss when decoding the first predicted token into a coarse trajectory.

Results demonstrate improvements over the continuous head baseline on Bench2Drive (CARLA) and NavSimv2.

Critical strengths

[S1] Formulation: Many driving VLAs are studied in open-loop. The paper analyzes limitations for closed-loop driving, specifically the error penalization in CE may not correlate with important physical errors in closed-loop.

[S2] Simple method: CTAR leverages a frozen trajectory decoder while retaining discrete action prediction, such that the approach is straightforward and does not require modifying the underlying VLM architecture.

Critical weaknesses

There are several significant concerns around motivation, clarity, and analysis.

[W1] Unclear justification for OAT

[W1.1] Unclear distance-aware supervision: Section 3.1 argues that incorrect early tokens alter subsequent autoregressive contexts, leading to larger accumulated cross-entropy losses (Eq. 4). However, Eq. (1) conditions predictions on ground-truth prefixes, suggesting teacher-forced training, where incorrect predictions do not alter subsequent training contexts. Therefore, it remains unclear how the proposed training objective produces the claimed distance-aware loss behavior.

[W1.2] Unclear spatial distance guarantee: Section 3.1 claims that trajectories sharing longer token prefixes have a bounded physical distance. However, the proposed bound is derived from average reconstruction error, not the worst-case error, which does not guarantee the same bound for individual trajectory pairs. The theoretical justification for this claim therefore appears incomplete.

[W2] Insufficient isolation of the proposed contributions

[W2.1] Missing metric-aware K-DISK baseline: The paper demonstrates substantial improvements with OAT+CTAR, but does not evaluate whether similar metric-aware supervision could improve K-DISK. Do the gains require ordered tokenization or is the primarily result from introducing continuous trajectory supervision?

[W2.2] Confounded action representations: OAT uses 32 tokens and a learned trajectory decoder, whereas K-DISK uses 8 or 10 tokens with a different decoding mechanism. These differences make it difficult to isolate the contribution of ordered tokenization, despite K-DISK's better reconstruction accuracy.

[W3] Limited performance improvements and evaluation discrepancies

[W3.1] Lack of competitiveness: While there is a clear improvement over the continuous-head baseline on Bench2Drive and NAVSIMv2, both results are generally underwhelming compared to current SOTA (on both leaderboards, e.g., far behind most methods today for NAVSIMv2). This raises concerns around complementarity with current SOTA VLAs.

[W3.2] Inconsistent performance reporting: The abstract reports a +33.57 DS improvement over K-DISK, whereas Table 1 implies +35.77 DS and Section 4.2 reports +39.69 DS. The conclusion also cites a different K-DISK score. These need to be explained more clearly.

[W3.3] Evaluation protocol discrepancies: The official SimLingo checkpoint achieves 72.20 DS under the authors' evaluation setup, compared to its published 85.07 DS. This substantial discrepancy, which is also inconsistent with independently reported SimLingo results, e.g., in ``Latent-Centroid Steering’’ by Hu et al. This raises concerns about evaluation protocol consistency and reproducibility. The authors should explain the difference and clarify whether their evaluation settings match those used in the published benchmarks.

[W3.4] Incomplete Bench2Drive reporting: Bench2Drive evaluates driving performance through DS, SR, efficiency, comfort, and several ability-specific metrics. However, the authors report only DS for their model. Reporting additional metrics, particularly SR and comfort, would provide a more complete understanding of its driving performance and enable more meaningful comparisons with other baselines.

[W3.5] Inconsistent CARLA benchmark settings: The authors use different Bench2Drive evaluation subsets across their experiments. Table 1 reports results on the full Bench2Drive benchmark, while Figure 4a uses only the 44-route version. Furthermore, it is unclear which benchmark version is used in Table 3, since the reported OAT+CTAR score of 75.19 does not match the other reported results. These differences should be clearly explained to ensure that the results are comparable.

[W3.6] Incomplete hyperparameter reporting and analysis: The paper does not provide complete details of the training hyperparameters, despite stating in the main text that they are reported in the appendix. Furthermore, no hyperparameter sensitivity analysis is provided, making it difficult to assess the robustness of the proposed method to different training configurations.

[W4] Practical efficiency and driving quality

[W4.1] Missing inference latency: Despite emphasizing compatibility with standard LLM-serving infrastructure, the paper does not report inference latency. Since OAT requires autoregressive generation of 32 tokens per trajectory, the sequential decoding process may introduce substantial latency compared with continuous regression heads. Without inference-time measurements, its practicality for real-time autonomous driving remains unclear.

[W4.2] Degraded driving comfort: On NAVSIM v2, OAT+CTAR achieves substantially lower Extended Comfort than the continuous baseline (0.373 vs. 0.747, Table 2). Appendix A.9 also reports 6.4-13.6× jerk inflation, raising concerns about trajectory smoothness despite improvements in the composite EPDMS.

Minor Weaknesses:

[MiW1] Inconsistent method naming: CTAR' is repeatedly written as CATR' including Tables 1-2 and parts of Section 4.

[MiW2] Inconsistent NAVSIM terminology: Table 2 describes the evaluation as open-loop,' while Sections 4.1 and 4.3 describe it as pseudo-closed-loop.'

[MiW3] Unclear training-data accounting: Section 4.1 reports approximately 59,600 NAVSIM scenes, while Appendix A.1 reports 152,495 unmixed anchors. Would help to clarify the relationship between these.

[MiW4] Inconsistent training-curve reporting: Figure 3a reports results at different training-step intervals across methods. OAT+CTAR has a data point every 10K steps, while Continuous Path-Speed has a data point every 20K steps, and K-DISK follows an irregular evaluation schedule. Furthermore, the final OAT+CTAR data point is missing. The authors should clarify whether all methods received comparable training budgets and explain the missing result.

Questions for authors
  • Clarify methodological reasoning and justification
  • Isolate the contributions (e.g., K-DISK with CTAR)
  • Missing comparison with SOTA
  • Clarify evaluation discrepancies
2 comments

We thank the reviewers for their feedback. Updates are incorporated in the revised PDF:

[W1.1] Teacher-Forced Supervision: Under teacher forcing (Eq. 1), the model conditions on ground-truth prefixes (Eq. 4 describes sequence-level NLL / inference divergence), and token position maps to spatial resolution: (1) for near-misses, coarse tokens (t<k) incur ℓt≈0\ell_t \approx 0 and loss concentrates on refinement tokens (t>k), pinpointing updates; (2) for far-misses, large losses fire immediately at coarse tokens (t<k), driving macro updates; (3) at t=1 (no prefix), CTAR supplies directional metric feedback (+26.69 DS, Table 3). Revised in Sec. 3.1.

[W1.2] Expected vs. Worst-Case Bound: Clarified in Sec. 3.1 that E[∥a(A)−a(B)∥∣z(A)(1:k)=z(B)(1:k)]≤2ϵk\mathbb{E}[\lVert \mathbf{a}^{(A)} - \mathbf{a}^{(B)} \rVert \mid z^{(A)}(1{:}k) = z^{(B)}(1{:}k)] \le 2\epsilon_k is an expected-case property, not a worst-case guarantee, empirically backed on NAVSIM (P(<10 m∣k≥1)=0.983P(<10\,\text{m} \mid k \ge 1) = 0.983; App. Sec. 6.3, Fig. 6).

[W2.1] Metric-Aware K-DISK Ablation: We trained K-DISK with Gaussian soft-label CE and evaluated at step 35k on scenario44 (Sec. 4.5). Soft-label CE improves K-DISK from 29.65->33.99 DS (+4.34 DS; RC 43.94%->49.84%), confirming isometric CE is a bottleneck, yet still trails full OAT+CTAR (83.49 DS at 35k; 75.19 DS at 105k) and even zero-shot 10-token OAT+CTAR (35.91 DS, 59.01% RC). OAT+CTAR succeeds because: (1) OAT compresses full-horizon macro geometry into t=1 for CTAR vs. local step tokens; (2) hierarchical refinement (t>1) organizes kinematic hidden states (App. Fig. 8); (3) gradients are density-invariant without kernel tuning.

[W2.2] Isolating Tokenization & Token Budget: (1) Table 3 holds the 32-token vocab/decoder fixed: OAT alone beats K-DISK (35.50 48.50 DS), and CTAR adds +26.69 DS (75.19 DS). (2) 10-token K-DISK already has 21.6 better codec ADE (0.0146 m vs. 0.315 m), so token count is not its bottleneck. (3) Zero-shot truncating 32-token OAT+CTAR to 10 tokens at inference (lower bound vs. native 10-token training) beats 10-token K-DISK by +6.25 DS (35.91 vs. 29.65) and +15.07% RC (Sec. 4.5).

[W3.1] Leaderboard Positioning & Complementarity: Our goal is a controlled comparison of action interfaces on a lightweight 1B single-camera VLM (InternVL2-1B, no multi-view/LiDAR/BEV/RL). Here, OAT+CTAR matches vision-only SimLingo on Bench2Drive-220 (73.75 vs. 72.20 DS; 66% vs. 59% completion) and beats TransFuser-LTF (0.2861 vs. 0.2310 EPDMS) and continuous regression (0.2403) on reactive navhard_two_stage, and stays orthogonal to scaling backbones, sensors, CoT, and RL with full LLM serving compatibility.

[W3.2] Consistent CARLA Reporting: Standardized Abstract, Intro, and Conclusion to Bench2Drive-220 (73.75 vs. 37.98 DS, +35.77 DS; Table 1) and explicitly labeled scenario44 comparisons in Sec. 4.2 (75.19 vs. 35.50 DS matched at 105k [+39.69]; 75.19 vs. 41.62 DS vs. K-DISK's 65k peak [+33.57]).

[W3.3] SimLingo Calibration & Fair Protocol: For a strictly fair comparison, all policies share the exact same InternVL2-1B architecture, training configs, and unified ~10 Hz wrapper rather than comparing across mismatched setups. Official simlingo repo docs also note public epoch=013 weights are a retrained reproduction rather than original paper weights; evaluated in our harness, epoch=013 scores 72.20 DS uniformly (72.60 scenario44, 72.10 remaining 173), matching OAT+CTAR (73.75 DS; +2.14 5.20 paired, n=211) (Sec. 4.2).

[W3.4] Full Bench2Drive Metrics: Expanded Table 1 with SR (56.4% vs. 50.9% SimLingo, 39.1% Cont., 9.5% K-DISK), Efficiency (247.0%), and Comfortness (27.4%).

[W3.5] CARLA Subsets: Bench2Drive-220 (n=220) is for final headline results (Table 1); stratified scenario44 (n=44, 1 route/scenario) enables tractable sweeps (Figs. 3a, 4a) and ablations (Table 3, at 105k steps). Clarified in Sec. 4.1, 4.5, and Table 3.

[W3.6] Hyperparameters: Added full hyperparameter/data table in Appendix Table 11.

[W4.1] Latency: Neither our model nor compared VLM baselines (InternVL2-1B) target real-time edge latency, which remains future work; OAT supports anytime prefix decoding (k<32) to trade latency for accuracy without retraining and preserves LLM serving accelerations (Sec. 5).

[W4.2] Extended Comfort (EC): Single-frame Comfort is high (C=0.962 vs. 0.976, 1.0 lateral accel.), while lower two-frame EC (0.373 vs. 0.747) reflects memoryless tick-by-tick decoding over an along-track-dominated coarse codec (1.182 m ADE; App. Sec. A.9), causing longitudinal step jumps across ticks; closed-loop CARLA controllers filter this jitter (73.75 DS), and temporal prefix conditioning is noted as future work (Sec. 4.3, 5).

[MiW1–MiW4]: Addressed in revision, clarifying ~59,600 14-chunk anchor samples vs. 152,495 full pool (Sec. 4.1, App. 6.1) and identical 15-epoch (~95k-step) budgets with failed 95k OAT eval omitted (Fig. 3a).

Area_Chair_hzodOpenReview ↗

Please help to join the discussion

Dear Reviewer Gr8P,

The discussion deadline is approaching. Please help to join the discussion and offer the reviewer's opportunity to further respond to your concerns.

Thanks!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.