acceptodds
Under review as a conference paper at ICLR 2027

VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

Abstract

Models for physical AI tasks, such as robotics and autonomous driving, need not reason solely in language. They can use visual plans—points, bounding boxes, and trajectories—to represent spatial goals and anticipate motion. However, autoregressive vision-language models generate these plans one coordinate token at a time, making richer predictions increasingly costly. We introduce VFlash, a method for fast, accurate visual planning through Flash Primitive Decoding. VFlash exploits the known structure of visual primitives to predict their coordinates jointly, using a lightweight drafter that generates an entire primitive in one forward pass. The layout adapts to the primitive's type, dimensionality, and length, supporting compact boxes and long, variable-length trajectories. Crucially, spatial accuracy need not require reproducing an autoregressive model's exact token sequence. VFlash therefore trains on annotated primitives and directly accepts structurally valid predictions without target-token agreement. Across the evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduces prediction errors relative to autoregressive decoding. Preliminary single-rollout measurements indicate approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Physical AI, Visual Planning, Vision-Language Models, Parallel Decoding, Speculative Decoding

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 4 reviews, above 85% of the others.
Reviewer_hHrwOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces VFlash, a method for fast, accurate visual planning through Flash Primitive Decoding. VFlash exploits the known structure of visual primitives to predict their coordinates jointly, using a lightweight drafter that generates an entire primitive in one forward pass. The layout adapts to the primitive's type, dimensionality, and length, supporting compact boxes and long, variable-length trajectories. Crucially, spatial accuracy need not require reproducing an autoregressive model's exact token sequence. VFlash therefore trains on annotated primitives and directly accepts structurally valid predictions without target-token agreement. Across the evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduces prediction errors relative to autoregressive decoding. Preliminary single-rollout measurements indicate approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline.

Critical strengths

None

Critical weaknesses
  1. As shown in Tables 2, 3, and 5, the performance improvements achieved by the proposed method appear limited and are not sufficiently significant to establish its superiority over competing methods. The authors should provide a more comprehensive analysis of the results to clarify the effectiveness of the proposed approach and identify the scenarios in which it offers substantial advantages.

  2. Although quantitative results are reported, the authors do not provide sufficient visual comparisons to demonstrate the quality and accuracy of the predictions. In particular, visualizations of 2D trajectory prediction, 2D bounding box prediction, 3D trajectory prediction, and dense 3D motion prediction are missing. Such qualitative results are essential for assessing the accuracy, spatial consistency, and temporal coherence of the predicted outputs.

  3. Given the practical relevance of the proposed method to robotics and autonomous driving, the authors should provide qualitative demonstrations illustrating its performance in these application scenarios. For example, visualized demos on real-world robotic platforms or autonomous driving datasets would help demonstrate the method's practical applicability, robustness, and effectiveness in complex environments.

  4. The paper lacks a direct comparison with representative state-of-the-art decoding frameworks, making it difficult to determine whether the observed performance is attributable to the proposed method itself or to other factors in the overall system. The authors should conduct controlled experiments under consistent settings, comparing predictive performance, computational efficiency, and inference latency against existing decoding frameworks to better establish the advantages and contributions of the proposed approach.

Questions for authors
  1. As shown in Tables 2, 3, and 5, the performance improvements achieved by the proposed method appear limited and are not sufficiently significant to establish its superiority over competing methods. The authors should provide a more comprehensive analysis of the results to clarify the effectiveness of the proposed approach and identify the scenarios in which it offers substantial advantages.

  2. Although quantitative results are reported, the authors do not provide sufficient visual comparisons to demonstrate the quality and accuracy of the predictions. In particular, visualizations of 2D trajectory prediction, 2D bounding box prediction, 3D trajectory prediction, and dense 3D motion prediction are missing. Such qualitative results are essential for assessing the accuracy, spatial consistency, and temporal coherence of the predicted outputs.

  3. Given the practical relevance of the proposed method to robotics and autonomous driving, the authors should provide qualitative demonstrations illustrating its performance in these application scenarios. For example, visualized demos on real-world robotic platforms or autonomous driving datasets would help demonstrate the method's practical applicability, robustness, and effectiveness in complex environments.

  4. The paper lacks a direct comparison with representative state-of-the-art decoding frameworks, making it difficult to determine whether the observed performance is attributable to the proposed method itself or to other factors in the overall system. The authors should conduct controlled experiments under consistent settings, comparing predictive performance, computational efficiency, and inference latency against existing decoding frameworks to better establish the advantages and contributions of the proposed approach.

3 comments
AuthorsOpenReview ↗

Response to Reviewer hHrw

We thank the reviewer for the feedback. Below, we identify where VFlash offers the largest benefits, provide the requested visual comparisons and demonstrations, and add controlled comparisons with representative decoding frameworks.

We understand that there are common questions on the Decoder Design and Comparison to Others among all reviewers. As such, we address the common questions and concerns in the General Comment. Please kindly refer to the comment at the top of the page.

W1 / Q1: The improvements in Tables 2, 3 and 5 appear limited. Which scenarios show substantial advantages?

We respectfully note that the improvements are substantial in several settings, which we summarize below by primitive:

  • 3D robot trajectories show large and consistent gains. VFlash improves every ADE, FDE, SR@5 and SR@10 over its AR target across three datasets and three training variants (Table 4). These gains persist when the AR target receives matched training compute (General Comment, Table R1).
  • These gains carry over to downstream control. With the same conditioning interface and action model, MolmoSpaces success is 70.2% with VFlash plans versus 61.0% with AR plans (Table 7), while each planning step on Thor drops from 3,553 to 703 ms.
  • 2D trajectories show consistent gains. VFlash improves all six metrics over its AR target and achieves the best results among all compared methods on VABench-V (Table 2). These gains largely persist against AR with matched compute (General Comment, Table R2).
  • Boxes and dense motion gain speed with comparable accuracy. With the confidence guard, VFlash reaches 61.78 F1 on Dense200 versus 60.10 for AR at 2.59×2.59\times the throughput (Table 3), and dense-motion generation is about 25×25\times faster in a preliminary single-rollout measurement, with mixed accuracy changes (Table 5).

We will add this per-primitive analysis to the revision.

W2, W3 / Q2, Q3: Visual comparisons of the predicted primitives, and qualitative demonstrations in robotics and autonomous-driving scenarios.

All qualitative results are available at this anonymous link. The videos compare AR and VFlash side by side on the same inputs. They cover 3D robot trajectories on real-robot scenes with ground truth overlaid, dense 3D motion, object localization and sub-task planning. The link also includes closed-loop policy rollouts on MolmoSpaces, LIBERO-Plus and LIBERO-PRO, with and without VFlash plans.

W4 / Q4: Comparison with representative decoding frameworks under consistent settings, covering predictive performance, computational efficiency and inference latency.

We address this question with more details in the General Comment. The paper has compared VFlash with DFlash in the same pipeline and hardware platforms (Figure 4, Appendix E), where VFlash is 2.6×2.6\times faster per MolmoSpaces step on Thor. Tables R1–R4 in General Comment add the additional controlled comparisons requested here. In the response, we add three further controls under consistent settings and report predictive performance, efficiency and latency together (Tables R1 to R4): a multi-token-prediction (MTP) head, the speculative-decoding design that current open models such as Qwen3.5 ship; a conventional flow-matching head; and AR with matched training compute.

Under these settings, VFlash is both the fastest and the most accurate decoder. MTP speculative decoding is lossless but speeds up end-to-end decoding by only 2.1×2.1\times over AR, compared with 18.2×18.2\times for VFlash, and with direct acceptance its trajectories are clearly worse than VFlash's. The flow-matching head is less accurate on every 3D metric and 6.2×6.2\times slower than VFlash end to end, and AR with matched compute remains far behind VFlash in 3D. This indicates that VFlash's gains come from its decoding design rather than from other parts of the system. We will add these comparisons to the revision.

Dear Reviewer,

This is a gentle reminder that the discussion phase will close soon. We would appreciate it if you could go over the authors’ rebuttal and share with us whether your original concerns have been resolved, or if you have additional points to raise with the authors.

Thank you for your effort.

Dear reviewer hHrw. Thank you for the thoughtful feedback and constructive comments. As the rebuttal deadline is approaching, we would be grateful if you could take a look at our rebuttal and the updated text to ensure we have addressed your concerns accurately. If that is the case, we would appreciate this to be reflected on the scores. If you have additional remarks, questions or requests, we are more than happy to address them! We appreciate your time and consideration.

Reviewer_r6sUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VFlash, a structured parallel decoding framework built on DFlash to accelerate visual-plan generation. A lightweight drafter, conditioned on frozen VLM features and trained on ground-truth annotations, predicts complete boxes, trajectories, or dense point motion in one pass using predefined primitive layouts. Valid predictions bypass target-token verification, with confidence-based autoregressive fallback for boxes and verified speculative decoding for text. Experiments evaluate spatial prediction, downstream manipulation and driving, and hardware latency.

Critical strengths
  • VFlash reduces per-step latency approximately fivefold on Jetson Thor and DGX Spark for MolmoSpaces, and 3.6-fold on H100. The breakdown includes the target-model pass needed to update cached features, strengthening the evidence for deployment-relevant efficiency.
  • VFlash lowers robot-trajectory errors relative to its autoregressive target across all reported robot-trajectory variants. On MolmoSpaces, it raises success from 62.6% with the original WAM to 70.2%, compared with 61.0% using AR-generated plans from the same VLM. These results suggest that its benefits extend beyond speed to more effective plan-conditioned control.
Critical weaknesses
  • Insufficient isolation of the proposed decoding design. VFlash adds a separately trained five-layer drafter, supervised directly on ground truth, to a frozen AR target. The experiments lack matched parallel-prediction baselines using the same features and comparable training budgets. It therefore remains unclear how much of the accuracy improvement comes from primitive-aligned decoding rather than additional capacity and training.
  • Direct acceptance has task-dependent limitations. Structural validity does not ensure spatial accuracy. On Dense200, direct acceptance reduces mean F1 from 60.10 to 39.09, with AR fallback needed to recover performance. Dense-motion results are also mixed, with lower PWT on all three datasets and worse WorldTrack final displacement error. These findings suggest that the quality benefits of direct acceptance depend on the prediction task.
Questions for authors
  • Do you have comparisons with a conventional parallel coordinate decoder using the same frozen features and comparable training budgets, or with an AR target given additional training computing? These would help isolate the contribution of primitive-aligned decoding.
  • How often do boxes trigger AR fallback, and how sensitive are accuracy and latency to the confidence thresholds? For trajectories, how are structurally valid but spatially inaccurate predictions handled?
5 comments
AuthorsOpenReview ↗

Response to Reviewer r6sU

We thank the reviewer for the careful review and for recognizing both VFlash's deployment-relevant speedup and its benefits for downstream plan-conditioned control. We address the concerns below.

We understand that there are common questions on the Decoder Design and Comparison to Others among all reviewers. As such, we address the common questions and concerns in the General Comment. Please kindly refer to the comment at the top of the page.

W1 / Q1: Comparisons with a conventional parallel coordinate decoder using the same frozen features and comparable training budgets, or with an AR target given additional training compute.

We address this question with more details in the General Comment. We agree that the original experiments did not separate our decoding design from the drafter's extra capacity and training. Please see our General Comment, Tables R1–R4, for the additional controlled comparisons requested here. We ran both requested controls, AR with matched training compute and a conventional parallel head on the same frozen features and training budget, together with a multi-token-prediction head. In 3D, VFlash remains far more accurate than all of them, e.g., InternA1 ADE 0.060 m versus 0.109 m for AR with matched compute and 0.086 m for the flow-matching head. None of these alternatives reproduces VFlash's accuracy or speed.

W2: Structural validity does not ensure spatial accuracy; direct acceptance degrades boxes and some dense-motion metrics.

We agree that structural validity does not ensure spatial accuracy. Direct acceptance only removes the requirement to match the AR target's tokens, so its benefit must be judged by task metrics, and, as the reviewer notes, it depends on the primitive:

  • Boxes. In dense scenes several objects are plausible "next boxes," and parallel slots can mix coordinates of different objects, whereas AR conditions each coordinate on the object already committed. Unconditional acceptance therefore fails (Dense200 F1 60.10 →\rightarrow 39.09), and this is precisely what the confidence guard detects. With the guard, VFlash reaches 61.78 F1 versus 60.10 for AR at 2.59×2.59\times the boxes per second (Table 3).
  • Dense motion. VFlash reduces DAVIS ADE/FDE by about 27%/19% but has lower PWT on all three datasets and a higher WorldTrack FDE. It removes large deviations but places fewer points within the tightest thresholds (Table 5).
  • Trajectories have one target per arm, a declared length and a smooth geometry. Predicting the whole trajectory jointly avoids AR's token-by-token error accumulation, and VFlash improves every 2D and 3D metric over AR (Tables 2 and 4).

We will reframe the claim accordingly, stating consistent accuracy gains for trajectories, and large speedups with comparable accuracy for boxes and dense motion.

Q2: How often do boxes trigger AR fallback, and how sensitive are accuracy and latency to the confidence thresholds?

We re-ran the box experiments to measure both (Table R6). With the default guard, fallback is triggered for 64% of drafted boxes on COCO and 53% on Dense200. Fallback keeps the confident coordinate prefix and lets AR complete only the remaining coordinates, so VFlash still generates 2.8×2.8\times (COCO) and 2.2×2.2\times (Dense200) more boxes per second than AR.

Table R6: Sensitivity to the confidence threshold. Fallback is the fraction of drafted boxes whose remaining coordinates are completed by AR; BPS is boxes per second. Higher F1 and BPS are better.

SettingCOCO F1COCO fallbackCOCO BPSDense200 F1Dense200 fallbackDense200 BPS
AR48.46%n/a3.0260.78%n/a6.70
VFlash, confidence 0.8047.70%62.6%8.6161.64%51.1%16.80
VFlash, confidence 0.90 (default)47.89%63.9%8.5061.50%52.8%14.70
VFlash, confidence 0.9547.87%64.5%8.4761.56%51.3%14.79

Both accuracy and speed are insensitive to the confidence threshold. From 0.80 to 0.95, F1 changes by at most 0.2 points on both datasets, and VFlash stays more than twice as fast as AR.

Q2: How are structurally valid but spatially inaccurate trajectories handled?

They are accepted directly. Such cases are not unique to VFlash, since AR decoding and other parallel heads can also produce them, and how often they occur is what spatial accuracy metrics measure. On these metrics, VFlash achieves the best overall accuracy among these decoders (General Comment, Tables R1 and R2).

Dear Reviewer,

This is a gentle reminder that the discussion phase will close soon. We would appreciate it if you could go over the authors’ rebuttal and share with us whether your original concerns have been resolved, or if you have additional points to raise with the authors.

Thank you for your effort.

Dear reviewer r6sU. Thank you for the thoughtful feedback and constructive comments. As the rebuttal deadline is approaching, we would be grateful if you could take a look at our rebuttal and the updated text to ensure we have addressed your concerns accurately. If that is the case, we would appreciate this to be reflected on the scores. If you have additional remarks, questions or requests, we are more than happy to address them! We appreciate your time and consideration.

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

VFlash addresses the cost of generating visual plans, such as points, bounding boxes, and trajectories in autoregressive vision-language models. Its Flash Primitive Decoding exploits the known structure of these primitives to predict coordinates jointly. Method use a lightweight drafter that generates an entire primitive in one forward pass. Single rollout measurements indicate approximately 25× faster generation for 3D point motion forecasting.

Critical strengths
  • Addresses an important inference bottleneck regarding computational cost. Method reduces the sequential cost of coordinate generation by predicting complete visual primitives in parallel.
  • Comprehensive evaluation. The evaluation goes beyond open-loop coordinate error to include downstream robot manipulation and autonomous driving. Paper include latency measurements across multiple hardware platforms.
Critical weaknesses
  • Limited methodological novelty. The approach combines many components from DFlash and fallback mechanisms closely related to LocateAnything.

  • Missing comparison against a simpler parallel prediction head.

Questions for authors

Q1. Have you compared VFlash against a simple parallel coordinate-prediction head with comparable capacity and training budget? If so, how do accuracy and latency compare?

Q2. When training the drafter, are the target VLM’s context features computed from ground-truth prefixes or from the drafter’s own accepted predictions? Please clarify.

4 comments
AuthorsOpenReview ↗

Response to Reviewer tzZQ

We thank the reviewer for the positive assessment and for recognizing the importance of the inference bottleneck and the comprehensiveness of our evaluation. We address the concerns below.

W1: Limited novelty; components come from DFlash and the fallback is related to LocateAnything.

While VFlash starts from DFlash's parallel drafter, our contribution is a generalist decoding framework built on the known structure of visual primitives, so that a complete primitive can be drafted in one parallel pass and accepted without reproducing the AR target's exact tokens. Neither method provides this, and VFlash differs from them in three ways:

  1. Boundary alignment. Each draft block covers one complete primitive, whose type and length the VLM declares on the fly through a tool call, instead of a fixed-size token window. One decoder thus handles boxes, variable-length 2D/3D trajectories and dense motion within a single response.
  2. Position alignment. Dedicated coordinate tokens or fixed-width numeric fields give every slot a known role (waypoint, axis or delimiter), so the whole primitive can be predicted in parallel with per-slot constraints.
  3. Learning from geometry. The drafter is trained on annotated primitives, and its primitives are accepted without matching the target's tokens, while text keeps verified speculative decoding. DFlash can only accept tokens that agree with the target.

This difference is measurable. In Appendix E, DFlash accepts only 2.2 trajectory tokens per target pass, giving 1,828 ms per MolmoSpaces step on Thor versus 703 ms for VFlash.

W2 / Q1: Have you compared VFlash against a simple parallel coordinate-prediction head with comparable capacity and training budget? How do accuracy and latency compare?

Yes. We compared VFlash with two parallel heads of comparable capacity, trained on the same frozen features with the same training budget, and report the full results in the General Comment (Tables R1 to R4). The first is a flow-matching head built on the drafter's own backbone (598M vs. 537M trainable parameters), and the second is the multi-token-prediction (MTP) head that Qwen3.5 uses for speculative decoding.

Accuracy. Neither head matches VFlash. In 3D, the flow-matching head is behind VFlash on every metric of all three datasets (e.g., InternA1 ADE 0.086 vs. 0.060 m), and with direct acceptance the MTP head is clearly worse than VFlash on every 2D and 3D metric. More Euler steps do not make the flow-matching head more accurate (Table R4).

Latency. With a single sampling step, the flow-matching head is about as fast as VFlash on decoding primitives, but its text remains autoregressive, so the full response is 6.2×6.2\times slower than VFlash's. The MTP head accelerates text but not coordinates and speeds up end-to-end decoding by at most 2.7×2.7\times over AR, compared with 18.2×18.2\times for VFlash. VFlash is the fastest because one drafter accelerates both the text and the primitive.

Q2: Are the target VLM's context features computed from ground-truth prefixes or from the drafter's own accepted predictions during training?

From ground-truth prefixes (teacher forcing). The frozen target encodes the annotated sequence once, and each draft block sees only the target features before its start (Appendix B). At inference, the prefix instead contains the previously accepted outputs, and all reported results are obtained under this setting. We will state this distinction in the revision.

Dear Reviewer,

This is a gentle reminder that the discussion phase will close soon. We would appreciate it if you could go over the authors’ rebuttal and share with us whether your original concerns have been resolved, or if you have additional points to raise with the authors.

Thank you for your effort.

Thank you for addressing my comments and providing further context. I maintain my original score.

Thank you for reviewing our responses and for your continued support. Your feedback helped us strengthen the parallel-head comparisons and clarify both the methodological contributions and the training setup. We will incorporate these results and clarifications into the revision. We appreciate your time and thoughtful feedback.

Reviewer_oHbCOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VFlash, a framework for fast visual planning that predicts complete visual primitives in parallel instead of generating coordinates autoregressively. It supports bounding boxes, trajectories, and dense motion predictions through structure-aware decoding and direct acceptance. Experiments demonstrate substantial speed improvements, competitive spatial accuracy, and promising downstream results in robotics and autonomous driving.

Critical strengths

Significant and practical efficiency improvements. VFlash substantially reduces decoding latency by generating complete visual primitives in parallel. The speedups are measured on Jetson Thor, DGX Spark, and H100, with detailed latency breakdowns that make the results convincing.

Broad experimental evaluation. The method is evaluated across multiple visual prediction tasks and downstream applications. The results demonstrate strong trajectory prediction accuracy and promising performance in robotic manipulation and autonomous driving.

Critical weaknesses

The source of accuracy improvements is unclear. The drafter introduces 537M additional trainable parameters and receives extra training on ground-truth annotations. Therefore, it is difficult to determine whether the improvements come from parallel decoding or additional model capacity and training. Comparisons with training-budget-matched AR models and simpler parallel prediction heads are needed.

The downstream benefits and generality are not fully established. The MolmoSpaces baseline also differs in its conditioning encoder, while LIBERO and driving experiments lack direct AR-plan comparisons. Moreover, direct acceptance significantly reduces bounding box accuracy and produces mixed results on dense motion prediction, suggesting that its effectiveness depends on the primitive type.

Questions for authors

How does VFlash compare with training-budget-matched AR models and simpler non-autoregressive prediction heads?

Could the authors provide more controlled downstream comparisons using AR-generated plans and explain why direct acceptance performs poorly on certain visual primitives?

4 comments
AuthorsOpenReview ↗

Response to Reviewer oHbC

We thank the reviewer for the constructive feedback and for recognizing the practical efficiency gains and the breadth of the evaluation. We address the concerns below.

W1 / Q1: The drafter adds 537M parameters and extra training. How does VFlash compare with training-budget-matched AR models and simpler parallel prediction heads?

We agree that these controls are needed to attribute the accuracy gain. We ran AR with matched training compute, a conventional flow-matching head and a multi-token-prediction (MTP) head, all with the same frozen features and training budget as the VFlash drafter, and report them in the General Comment (Tables R1 to R4). None of them reproduces VFlash's accuracy. For example, on InternA1, VFlash reaches an ADE of 0.060 m, versus 0.109 m for AR with matched compute and 0.086 m for the flow-matching head. These alternatives are also much slower, speeding up end-to-end decoding by at most 2.9×2.9\times over AR, compared with 18.2×18.2\times for VFlash.

W2 / Q2: Could the authors provide more controlled downstream comparisons using AR-generated plans?

Yes. We added an AR-plan baseline on the LIBERO family, in which the same policy is trained and evaluated with plans generated by AR decoding instead of VFlash.

Table R5: LIBERO family with AR-generated plans. Success rate in % (higher is better).

MethodLIBEROLIBERO-PlusLIBERO-PRO
Baseline93.460.740.2
w/o visual plan93.462.645.3
AR plans94.562.345.7
VFlash94.262.647.0

VFlash achieves a success rate similar to AR plans on LIBERO and LIBERO-Plus, while outperforming them under LIBERO-PRO perturbations. It also raises MolmoSpaces success from 61.0% to 70.2%. Beyond faster plan decoding, VFlash thus delivers equal or better downstream control across all evaluated settings, with substantial gains in the most challenging ones. We will add these results to the revision.

W2 / Q2: Why does direct acceptance perform poorly on certain primitives?

Direct acceptance removes the requirement to reproduce the AR target's exact tokens; it does not certify geometry. Whether that helps depends on how ambiguous the primitive is.

  • Trajectories have one target per arm, a declared length and a smooth geometry. Predicting the whole trajectory jointly avoids AR's token-by-token error accumulation, and VFlash improves every 2D and 3D metric over AR (Tables 2 and 4).
  • Boxes in dense scenes are different. When many objects are valid "next boxes," parallel slots can combine coordinates of different objects, whereas AR conditions each coordinate on the object it has already committed to. This is exactly the case the confidence guard detects (Appendix F.4). With the guard, VFlash reaches 61.78 F1 (vs. AR 60.10) at 2.59×2.59\times the boxes/s (Table 3).
  • Dense motion is mixed (Table 5). VFlash reduces the average errors on DAVIS (ADE 1.227 →\rightarrow 0.893 m) but has lower PWT and a higher WorldTrack FDE, i.e., it removes large deviations but places fewer points within the tightest thresholds.

We will state this scope precisely in the revision. The accuracy gains are consistent for trajectories, and for boxes and dense motion VFlash delivers large speedups while keeping accuracy comparable to AR.

Dear Reviewer,

This is a gentle reminder that the discussion phase will close soon. We would appreciate it if you could go over the authors’ rebuttal and share with us whether your original concerns have been resolved, or if you have additional points to raise with the authors.

Thank you for your effort.

I think the authors have addressed my concerns well. I consider this to be a high-quality paper, and I will increase my rating, although my current score is already leaning toward acceptance.

Thank you for revisiting our paper and for your positive assessment. We are glad that our responses addressed your concerns. Your suggestions helped us strengthen the controlled comparisons and clarify the scope of our claims, and we will incorporate these additions into the revision. We appreciate your time and support.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment (Part 1/2)

We thank all reviewers for their constructive feedback. Several reviewers asked whether VFlash's gains come from its decoding design or from the drafter's extra capacity and training, and how VFlash compares with simpler parallel heads and established decoding frameworks (oHbC W1, tzZQ W2, r6sU W1, hHrw W4). We address these questions together here and refer to these results in the individual responses.

We ran three controls, all on the same frozen VLM target and trained with the same training setting as the VFlash drafter.

  • AR + matched compute. To match the drafter's training FLOPs, we continue training the AR VLM target for an additional 32,307 steps on 3D robot trajectories and 10,669 steps on 2D trajectories.
  • Flow-matching head. A parallel coordinate head trained with flow matching, the same objective used by action experts such as GR00T [a]. To isolate this objective, it reuses the VFlash drafter's own backbone, conditions on the same frozen VLM features as the drafter, and trains with the same training setting (598M trainable parameters vs. the drafter's 537M; the extra 61M come from timestep conditioning and a small velocity head). Each coordinate axis is normalized to [-1, 1] by its range in the training data (1st to 99th percentile), the standard per-axis normalization of action heads. It samples coordinates with KK Euler steps; Tables R1 and R2 use K=1K=1, its fastest setting.
  • Multi-token-prediction (MTP) head. The lightweight prediction head that Qwen3.5 [b] attaches to its backbone and that vLLM [c] serves for speculative decoding. It was trained on the same frozen VLM target with the same training setting and drafts two tokens per round, following Qwen3.5's vLLM recipe. As a standard speculative decoder it is lossless (its output equals the AR target's), so we evaluate its accuracy by applying VFlash's direct acceptance at this native draft length, accepting each group of two drafted coordinates without verification.

Table R1: 3D robot trajectories. ADE and FDE in m (lower is better), SR@5 and SR@10 in % (higher is better).

MethodInternA1 ADEInternA1 FDEInternA1 SR@5InternA1 SR@10AgiBot ADEAgiBot FDEAgiBot SR@5AgiBot SR@10MolmoBot ADEMolmoBot FDEMolmoBot SR@5MolmoBot SR@10
AR (paper)0.1300.19636.055.40.0840.12945.674.90.0660.10254.781.3
AR + matched compute0.1090.16544.563.80.0680.11153.179.60.0570.09062.085.1
Flow-matching head0.0860.11838.171.10.0670.09349.984.30.0620.09155.084.0
MTP head, direct acceptance0.1230.18833.856.90.0720.11947.376.70.0700.11050.679.3
VFlash (paper)0.0600.09165.183.50.0360.05779.493.70.0520.08067.187.7

Table R2: 2D trajectories. Errors in normalized image coordinates (lower is better).

MethodShareRobot DFDShareRobot HDShareRobot RMSEVABench-V DFDVABench-V HDVABench-V RMSE
AR0.2020.1750.1420.1300.1250.093
AR + matched compute0.2020.1700.1410.1250.1200.091
Flow-matching head0.2080.1860.1540.1260.1190.088
MTP head, direct acceptance0.2250.1930.1570.1460.1370.108
VFlash0.1940.1620.1330.1260.1160.089

The results lead to three conclusions.

  1. Extra AR training does not close the gap. Matched compute improves the AR target, but VFlash remains far more accurate on every 3D metric. In 2D, VFlash is also better on every metric except VABench-V DFD, where the two are close (0.126 vs. 0.125).
  2. A conventional parallel head with the same features and budget does not match VFlash. In 3D robot trajectories, it is behind VFlash on every metric of all three datasets. In 2D trajectories, VFlash is better on all three ShareRobot metrics, and the two are on par on VABench-V.
  3. An MTP head does not explain the gain either. With direct acceptance, as in VFlash, its trajectories are clearly worse than VFlash's on every 2D and 3D metric.

[a] Bjorck et al., GR00T N1: An Open Foundation Model for Generalist Humanoid Robots, arXiv 2025.

[b] Qwen Team, Qwen3.5: Towards Native Multimodal Agents, https://qwen.ai/blog?id=qwen3.5, 2026.

[c] Kwon et al., Efficient Memory Management for Large Language Model Serving with PagedAttention, SOSP 2023.

AuthorsOpenReview ↗

General Comment (Part 2/2)

Table R3: Latency on 3D robot trajectories. ms per request, batch size 1, one H100.

DecoderVision + prefillTextPrimitiveEnd-to-endSpeedup
AR56.81,220.12,612.73,889.61.0×1.0\times
Flow-matching head, K=1K=157.31,234.137.61,329.02.9×2.9\times
Flow-matching head, K=4K=457.21,228.250.11,335.52.9×2.9\times
Flow-matching head, K=16K=1656.31,228.7100.61,385.62.8×2.8\times
MTP head, speculative decoding56.6441.61,320.81,819.02.1×2.1\times
MTP head, direct acceptance55.4435.3924.21,414.92.7×2.7\times
VFlash51.9126.136.3214.318.2×18.2\times

Latency. At K=1K=1, the flow-matching head is about as fast as VFlash on decoding primitives, but its text remains autoregressive, so the full response is 6.2×6.2\times slower than VFlash's. The MTP head accelerates text but not coordinates, giving only a 2.1×2.1\times end-to-end speedup as a speculative decoder and 2.7×2.7\times with direct acceptance, versus 18.2×18.2\times for VFlash. VFlash is the fastest because one drafter accelerates both the text and the primitive.

Table R4: Accuracy of the flow-matching head versus the number of Euler steps KK. 2D: DFD in normalized image coordinates; 3D: ADE in m (lower is better).

MethodShareRobot DFDVABench-V DFDInternA1 ADEAgiBot ADEMolmoBot ADE
Flow-matching head (K=1K=1)0.2080.1260.0860.0670.062
Flow-matching head (K=2K=2)0.1980.1220.0880.0660.063
Flow-matching head (K=4K=4)0.1970.1420.0940.0710.065
Flow-matching head (K=8K=8)0.2010.1630.0990.0760.068
Flow-matching head (K=16K=16)0.2060.1790.1030.0800.070
VFlash0.1940.1260.0600.0360.052

Sampling steps. More Euler steps make the flow-matching head slower but not more accurate. Its 3D error does not improve with more steps, and in 2D its best step count differs between benchmarks.

In summary, under matched features, data and training budget, neither more AR training, nor a conventional parallel head, nor an MTP head reproduces VFlash's accuracy or speed, indicating that VFlash's gains come from its decoding design rather than from extra capacity or training. We will add these comparisons to the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.