acceptodds
Under review as a conference paper at ICLR 2027

Zebra3D: Efficient Hybrid Foundation Models for Scalable 3D Reconstruction

Abstract

Feed-forward 3D foundation models reconstruct scenes from unposed images, but the quadratic cost of global attention limits their scalability to large view sets. We present Zebra3D, a layerwise sparse–linear hybrid attention framework that improves the efficiency of pretrained 3D foundation models without training from scratch. Across VGGT, HunyuanWorld-Mirror, and , we observe a broadly consistent depth-wise pattern: early stages are relatively tolerant to linear replacement, while sensitive middle-to-late stages tend to exhibit more concentrated attention. Zebra3D exploits this structure by assigning linear-centered attention to tolerant stages and tile-sparse attention to sensitive stages, using landmark summaries to preserve unselected context. Closed-form affine fitting enables training-free conversion, while optional post-conversion training combines task supervision with output and relational feature distillation. Across all three backbones, Zebra3D achieves strong reconstruction and camera-pose accuracy at high attention sparsity, with training-free variants remaining close to their dense counterparts. End-to-end speedups increase with input view count, reaching up to at 1,024 views. After post-conversion training, Zebra3D-VGGT reaches 80.5% RealEstate10K AUC@30 at 94% mean attention sparsity, within 1.1 percentage points of dense VGGT fine-tuned on the same data and schedule.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3d reconstruction, 3d models, sparse attention, linear attention, efficient inference, knowledge distillation, hybrid models

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

This paper presents Zebra3D, a practical method that converts pretrained 3D reconstruction transformers into hybrids of linear and tile-sparse attention for effiicient reconstruction. Zebra3D assigns linear attention to stages that tolerate replacement, while sensitive stages retain selected token-level interactions and landmark summaries of omitted tiles. Then a centered affine correction fitted on calibration data enables conversion without gradient updates. The paper also presents optional refinement that combines task supervision with distillation. The authors provide extensive experiments in the paper across several benchmarks and compare the method with wide baselines. The paper largely retain the accuracy of the dense-model, while an aggressive configurations provide larger acceleration.

Critical strengths
  1. The core contribution of Zebra3D is a framework that demonstrates useful and efficient conversion across multiple pretrained backbones and measurable end-to-end acceleration. he conservative configurations retain near-dense pose accuracy across three backbones, while a separate refined VGGT configuration achieves 5.60× speedup with comparable reconstruction results, supporting a practical quality–efficiency trade-off.

  2. The paper provide detailed experiments and ablations that validate the proposed architecture. The sensitivity-guided layer allocation is supported by controlled comparisons, providing direct evidence that where the operators are placed affects reconstruction performance. The explanations and discussions also provide thorough analysis of the gains and is informative for evaluating where the gains come from.

  3. The initialization and context-preservation mechanisms are well-motivated and supported by strong evidence in the experiments sections. The paper breakdowns the effectiveness of adding landmarks to tile-sparse attention and shows that the components contribute more than implementation convenience. This further proves that the strategy is practical and highly reuseable for feed-forward dense backbones deployments.

Critical weaknesses
  1. The specific contribution of relational distillation remains insufficiently isolated in the ablation. In Tables 3(c), the authors do not include the complete objective with only the relational term removed, while the dense GT-FT control omits all distillation. This makes it difficult to distinguish relational preservation from the other feature and output matching losses. I suggest adding this experiments in the experiments section.

  2. The current efficiency presentation leaves the exact practical operating costs unclear. The results of efficiency in the paper lack absolute latency and peak-memory measurements. I suggest adding paired measurements directly to make the reported accuracy–runtime trade-offs easier to assess and reproduce.

  3. Long-sequence accuracy is not established as broadly as long-sequence runtime. Figure 5 reports speedups at 1,024 views for all three benchmarks, whereas Table 2 provides quantitative long-sequence geometry only for refined VGGT variants. I suggest adding the corresponding results since short-sequence pose results cannot establish equivalent quality retention at these larger scales. This would help readers to better understand the quality-efficiency tradeoff of the proposed methods.

  4. The paper mainly provides metrics in the tables to prove the effectiveness, I would suggest adding more qualitative comparisons with its inference efficiency for better paper presentation. In addition, I would also suggest the authors include more in-the-wild examples either in the main paper or the appendix for a more comprehensive comparisons.

  5. There are some misalignments in the papers: a) A small numerical discrepancy should be corrected. For Pi3 at 0.5L/0.8S, RE10K AUC-30 is 0.866 in Table 1 and 7 but 0.867 in Table 12 b) The calibration-set reporting needs clarification. Figure 8 states that results are averaged over 11 held-out scenes, whereas Appendix B.6 describes 13 scenes for reference profiling and fitting. c) A minor typo exists in section 4.3, “Therefor” should be “Therefore.”

Questions for authors

The paper already validates the architecture effectiveness on 3 different backbones, I wonder how does the proposed method perform if using VGGT-Omega as a backbone. In addition, as mentioned in the weakness, how does the complete refinement objective perform when only the relational loss is removed, keeping all other losses, data, and optimization settings unchanged?

5 comments

We sincerely thank the reviewer for the careful assessment and for recognizing Zebra3D’s practical conversion framework, sensitivity-guided allocation, and context-preservation mechanisms. In response, we ran the requested experiments, and they support our original claims. ** [Answer 1]** presents a matched ablation of the relational loss. [Answer 2] reports peak GPU memory for all backbones and methods up to 1,000 views. [Answer 3] extends the ScanNet-50 evaluation to HYWM and π³. ** [Answer 4]** clarifies the visual comparisons, and [Answer 5] corrects the reporting inconsistencies and typo. [Answer 6] evaluate Zebra3D on two additional backbones, VGGT-Omega and Fast3R. We hope these additional experiments and clarifications address the reviewer’s concerns.

[Answer 1] Relational distillation (W1/Q2)

We trained the requested control by removing only (Lrel\mathcal{L}_{\mathrm{rel}}), keeping the remaining losses, their weights, the training data, and the optimization settings unchanged. The loss weights are not renormalized.

The full feature objective is Lfeat=0.25L1+0.50Lcos⁡+0.25Lrel\mathcal{L}_{\mathrm{feat}} = 0.25\mathcal{L}_{1} + 0.50\mathcal{L}_{\cos} + 0.25\mathcal{L}_{\mathrm{rel}} while the control uses Lfeatw/o rel=0.25L1+0.50Lcos⁡\mathcal{L}_{\mathrm{feat}}^{\text{w/o rel}} = 0.25\mathcal{L}_{1} + 0.50\mathcal{L}_{\cos}. Both retain the same task and output-distillation losses.

(a) Multi-view reconstruction and pose (Table R17). Without the relational loss, reconstruction degrades on 7Scenes and on ScanNet-50 at every tested sequence length (100 to 1000 frames). On NRGBD and RE10K the variant without it is slightly better, so the gain is not uniform across benchmarks, but it is consistent across all lengths of the long-sequence benchmark. The relational loss preserves the teacher's similarity structure between tokens of different views, and the resulting features are used for every view pair in a long sequence; its effect is visible on ScanNet-50 at all four lengths. A scene-level reconstruction score, however, averages over many view relations, which compresses per-pair improvements into a small aggregate gain. We therefore measure the per-pair effect directly with two-view pose estimation (Table R18), where the model must relate a single pair of views.

Table R17: Full objective versus removing only the relational term. VGGT, 0.5L/0.6S.

Objective7S Acc. ↓NRGBD Acc. ↓RE10K AUC@30 ↑ScanNet CD, 1,000 frames ↓500 frames ↓300 frames ↓100 frames ↓
Full0.01520.02350.80490.4510.4440.4350.430
Without (Lrel\mathcal{L}_{\mathrm{rel}})0.01690.02280.80610.4540.4450.4440.433

(b) Two-view pose estimation. To measure the per-pair effect behind the long-sequence gain, we evaluated MegaDepth-1500 [1] and ScanNet-1500 [2], where the model must relate a single pair of views directly. Here the full objective outperforms the variant without the relational loss on all eight metrics, with significant gains on six (Table R18). On MegaDepth, it even slightly exceeds the dense teacher at the three larger angular thresholds, while the control remains below the teacher.

Table R18: Two-view relative pose estimation. Zebra3D-VGGT(0.5L/0.6S). An asterisk indicates a statistically significant improvement over training without the relational loss (95% paired-bootstrap confidence interval excludes zero). Bold compares the student variants.

ModelMegaDepth AUC@5 ↑AUC@10 ↑AUC@20 ↑AUC@30 ↑ScanNet AUC@5 ↑AUC@10 ↑AUC@20 ↑AUC@30 ↑
Dense teacher0.40540.61010.76440.84430.36130.58510.75240.8374
Full0.3960*0.6108*0.7674*0.8484*0.3503*0.5727*0.74290.8312
Without (Lrel\mathcal{L}_{\mathrm{rel}})0.38780.60010.76000.84240.34310.56860.74160.8302

Summary. The matched ablation shows that the relational loss consistently improves long-sequences reconstruction (Table R17). Two-view pose estimation, which removes the compensation from intermediate views, isolates the underlying effect and shows it clearly (Table R18). This confirms that the long-sequence gain stems from better-preserved cross-view token relations, as the loss is designed to do. We will add these experiments to the revised manuscript.

[1] Li, Zhengqi, and Noah Snavely. "Megadepth: Learning single-view depth prediction from internet photos." 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2018.

[2] Dai, Angela, et al. "Scannet: Richly-annotated 3d reconstructions of indoor scenes." 2017 IEEE conference on computer vision and pattern recognition (CVPR). IEEE, 2017.

AuthorsOpenReview ↗

Official Comment by Authors

[Answer 2] Inference cost and peak GPU memory (W2)

We added peak GPU memory measurements for all three backbones across 100–1,000 views, comparing Zebra3D with dense inference, A3D, and BlockSparse3D.

Table R19: Peak GPU memory in GiB.

BackboneMethod100 views300 views500 views700 views1,000 views
VGGTDense13.429.645.862.084.2
A3D13.429.745.962.184.3
BlockSparse3D13.629.847.675.3128.0
Zebra3D (0.5L/0.6S)13.529.745.962.188.0
Zebra3D (0.7L/0.8S)13.529.846.062.284.4
HYWMDense16.337.959.480.9110.4
A3D16.437.959.581.0110.5
BlockSparse3D16.538.159.681.2136.1
Zebra3D (0.5L/0.6S)16.438.059.581.0110.5
Zebra3D (0.7L/0.8S)16.538.059.681.1110.5
π³Dense8.217.126.034.848.1
A3D7.817.126.034.948.2
BlockSparse3D8.619.938.162.3109.7
Zebra3D (0.5L/0.6S)7.314.622.831.946.7
Zebra3D (0.7L/0.8S)7.214.422.130.343.6

Zebra3D uses approximately the same memory as dense inference on VGGT and HYWM, and less memory on π³ (Table R19). Zebra3D also uses substantially less memory than BlockSparse3D at larger view counts.

We will add the reported speedups, using the same configurations and input lengths.

[Answer 3] Long-sequence reconstruction on HYWM and π³ (W3)

We have added ScanNet-50 reconstruction results for HYWM and π³ across 100–1,000 frames, reporting quality and speed for the same configurations.

Table R20: ScanNet-50 long-sequence reconstruction with HYWM and π³.

BackboneMethod1,000 frames CD ↓Speedup ↑500 frames CD ↓Speedup ↑300 frames CD ↓Speedup ↑100 frames CD ↓Speedup ↑
HYWMDense0.4561.00×0.4491.00×0.4321.00×0.4311.00×
Zebra3D (0.5L/0.8S)0.4533.26×0.4412.86×0.4302.54×0.4281.81×
Zebra3D-KD (0.7L/0.8S)0.4595.21×0.4594.11×0.4393.17×0.4312.01×
π³Dense0.6081.00×0.6041.00×0.6081.00×0.6111.00×
Zebra3D (0.5L/0.8S)0.6094.11×0.6103.72×0.6211.83×0.6181.65×
Zebra3D-KD (0.7L/0.8S)0.6047.04×0.6036.05×0.6082.89×0.6042.25×

The new results show substantial acceleration without refinement. Training-free Zebra3D improves HYWM’s reconstruction quality at every evaluated length. On π³, its error is nearly identical to dense inference at 1,000 frames. These configurations run 3.26× and 4.11× faster, respectively (Table R20).

The more aggressive refined configuration provides further acceleration. On π³, it matches or improves dense reconstruction quality at every evaluated length and reaches 7.04× speedup. On HYWM, it reaches 5.21× speedup with a small increase in reconstruction error.

We will add this table so that the long-sequence runtime comparisons are accompanied by geometry measurements for all three backbones.

[Answer 4] Qualitative comparisons and in-the-wild examples (W4)

The manuscript includes long-sequence reconstruction comparisons in Figure 6 and Appendix Figure 11, along with KITTI and NRGBD examples in Appendix Figure 12.

We will add more qualitative comparisons in the appendix, including in-the-wild captures.

[Answer 5] Reporting corrections and typos (W5)

Thank you for catching these points. We will address each in the revision:

  • π³ RE10K result: We will check the underlying evaluation output and use the same correctly rounded value throughout manuscript Tables 1, 7, and 12.
  • Calibration scene counts: We will clarify that 13 scenes were used for the Figure 8 averages and for reference profiling and fitting in Appendix B.6, and correct any inconsistent reporting.
  • Typo: We will change “Therefor” to “Therefore” in Section 4.3.
AuthorsOpenReview ↗

Official Comment by Authors

[Answer 6] Results on VGGT-Omega and Fast3R (Q1)

(a) VGGT-Omega. We have evaluated Zebra3D on VGGT-Omega and compared it with the dense model, A3D, and BlockSparse3D.

Table R21: VGGT-Omega. Sparsity is reported at 300 and 10 views.

MethodSparsity, 300 / 10 views7S Acc. ↓7S Comp. ↓NRGBD Acc. ↓NRGBD Comp. ↓RE10K AUC-30 ↑RE10K RTA-5 ↑CO3Dv2 AUC-3 ↑CO3Dv2 AUC-30 ↑
Dense0% / 0%0.0130.0130.0250.0100.8780.8010.4810.881
A3D73% / 66%0.0130.0130.0250.0110.8690.7830.3810.855
BlockSparse3D70% / 70%0.0130.0130.0340.0130.8630.7710.3670.855
Zebra3D (0.4L/0.7S)78% / 72%0.0130.0130.0240.0100.8760.7970.4500.874

Zebra3D matches or improves every reported quality metric relative to both efficient baselines while using greater sparsity (Table R21). It also stays close to dense VGGT-Omega, matching its 7Scenes reconstruction errors and remaining within 0.2 percentage points on RE10K AUC@30.

(b) Fast3R. We also test Fast3R, whose global-attention architecture provides a different test from VGGT-Omega. We select its layer allocation using its own sensitivity measurements.

Table R22: Fast3R reconstruction and pose estimation.

MethodMean sparsity at 300 views7Scenes Acc. ↓NRGBD Acc. ↓RE10K AUC-30 ↑CO3Dv2 AUC-30 ↑
Dense Fast3R0%0.0360.0720.6050.767
A3D80.6%0.1150.2450.1320.006
Zebra3D (0.5L/0.6S)79.5%0.0520.1090.4500.549

At similar sparsity, Zebra3D outperforms our A3D port on all four metrics, reducing reconstruction accuracy errors by approximately 55% on both datasets (Table R22). A gap to dense Fast3R remains, but Zebra3D preserves much more of its reconstruction and pose accuracy.

These additional experiments bring the evaluation to five backbones and show that the method also applies beyond architectures that alternate frame-wise and global attention. We will add both comparisons to the revised manuscript.

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

This paper introduces Zebra3D, a framework for scaling up pretrained feed-forward 3D reconstruction models through training-free attention conversion. It combines linear and sparse attention based on layer-wise sensitivity analysis. The framework also supports optional post-conversion refinement through knowledge distillation. Experiments demonstrate competitive reconstruction accuracy and substantial inference speedups compared with existing baselines.

Critical strengths
  1. Clear motivation and well-supported methodological design. The problem motivation is clear, and the empirical analysis reveals consistent trends in attention concentration and sensitivity across different layers. These observations provide a reasonable justification for the proposed hybrid attention design.

  2. Extensive experimental evaluation. The method is evaluated on three different pretrained backbones, multiple linear/sparse attention configurations, and common 3D reconstruction benchmarks. The results demonstrate promising accuracy–efficiency trade-offs, with substantial inference speedups while maintaining competitive reconstruction accuracy. The ablation studies further support the effectiveness of the proposed components.

Critical weaknesses
  1. Robustness to the choice of calibration set. The sensitivity analysis is conducted using a relatively small calibration set of 13 scenes. Although Appendix C.3 evaluates the stability of layer-wise sensitivity rankings across different calibration datasets, it does not provide an end-to-end quantitative analysis of how the choice of calibration set affects downstream reconstruction performance. Such an analysis would help establish the robustness of the proposed method. Additionally, it would be interesting to investigate how layer-wise sensitivity changes with the number of input views and whether the current layer allocation remains optimal for larger image collections.

  2. Missing GPU memory evaluation. The paper does not report GPU memory consumption for different numbers of input views. Given that scalability is a central motivation, quantitative memory comparisons against the dense backbones and existing efficient methods would provide a more complete assessment of the practical benefits of Zebra3D.

Minor presentation concern: The main results section is somewhat difficult to read due to the large number of quantitative values presented in the text. Since the tables already contain the detailed results, a more high-level, qualitative discussion would improve readability.

Questions for authors
  1. Could the authors provide further analysis of the choice of calibration set, particularly how different calibration sets affect downstream reconstruction accuracy? Additionally, how does the layer-wise sensitivity ranking change with the number of input views? (See W1.)

  2. Could the authors provide additional long-sequence reconstruction results for HYWM and π3\pi^3, as well as GPU memory consumption comparisons across different input view counts? These evaluations would help further demonstrate the scalability and generalizability of the proposed method.

5 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer for the thoughtful assessment and for highlighting the motivation for our sensitivity-guided design and the breadth of our evaluation. In response, we conducted additional experiments that quantify how the choice of calibration scenes affects downstream reconstruction accuracy, characterize how layer sensitivity and allocation vary with the number of calibration samples and extend our scalability evaluation with peak GPU memory comparisons and long-sequence reconstruction results for HYWM and π³. Below, [Answer 1] reports the GPU memory and [Answer 2] presents the calibration-set and view-count analyses, and [Answer 3] additional long-sequence reconstruction results on HYWM and Pi3 backbones. We will incorporate these findings into the revised manuscript and streamline the main results section to highlight the key accuracy–efficiency trade-offs, with detailed numbers in the tables. We hope the new evidence addresses the reviewer’s concerns.

[Answer 1] Peak GPU memory across input-view counts (W2/Q2)

We have added peak GPU memory measurements for all three backbones, comparing Zebra3D with dense inference, A3D, and BlockSparse3D over 100–1,000 input views.

Table R11: Peak GPU memory in GiB.

BackboneMethod100 views300 views500 views700 views1,000 views
VGGTDense13.429.645.862.084.2
A3D13.429.745.962.184.3
BlockSparse3D13.629.847.675.3128.0
Zebra3D (0.5L/0.6S)13.529.745.962.188.0
Zebra3D (0.7L/0.8S)13.529.846.062.284.4
HYWMDense16.337.959.480.9110.4
A3D16.437.959.581.0110.5
BlockSparse3D16.538.159.681.2136.1
Zebra3D (0.5L/0.6S)16.438.059.581.0110.5
Zebra3D (0.7L/0.8S)16.538.059.681.1110.5
π³Dense8.217.126.034.848.1
A3D7.817.126.034.948.2
BlockSparse3D8.619.938.162.3109.7
Zebra3D (0.5L/0.6S)7.314.622.831.946.7
Zebra3D (0.7L/0.8S)7.214.422.130.343.6

The principal practical benefit is substantial acceleration with a memory footprint generally comparable to dense inference. On π³, Zebra3D additionally reduces peak memory across every evaluated input length. On VGGT and HYWM, memory remains close to dense and A3D.

These measurements complement the runtime results by showing that Zebra3D’s acceleration does not generally require a substantial increase in peak GPU memory. We will add these measurements and clarify that the main benefit of Zebra3D is much faster inference at a similar memory footprint.

AuthorsOpenReview ↗

Official Comment by Authors

[Answer 2] Robustness to calibration dataset and sample count (W1/Q1)

We extend manuscript Appendix C.3 with downstream evaluations, varying the calibration dataset and scene count.

(a) Choice of calibration dataset. Table R12 compares Zebra3D (0.5L/0.8S) allocations selected using different datasets.

Table R12: Downstream performance with different calibration datasets at 0.5L/0.8S. “Reference” denotes the released configuration.

BackboneCalibration7S Acc. ↓7S Comp. ↓NRGBD Acc. ↓NRGBD Comp. ↓RE10K AUC-30 ↑RE10K RTA-5 ↑CO3Dv2 AUC-3 ↑CO3Dv2 AUC-30 ↑
VGGTReference0.01750.02700.02080.01380.72950.50720.53470.8920
7Scenes0.01750.02700.01550.01080.72710.50430.52780.8906
DTU0.01700.02660.01550.01080.72680.50250.52860.8907
RE10K0.01680.02540.02440.01640.72750.49880.51420.8897
NRGBD0.01740.02700.01540.01080.72730.50500.53640.8914
CO3Dv20.01700.02660.01550.01080.72680.50250.52860.8907
HYWMReference0.01300.02440.01880.01330.85000.75830.37640.8543
7Scenes0.01320.02460.01900.01340.85190.76120.36050.8524
DTU0.01300.02440.01880.01330.84990.75970.37670.8542
RE10K0.01320.02470.01930.01350.84420.74830.36210.8520
NRGBD0.01290.02450.01910.01350.85090.75970.36870.8519
CO3Dv20.01290.02450.01890.01330.85000.75800.36570.8524
π³Reference0.01210.01820.01470.00960.86640.79230.54180.8954
7Scenes0.01250.01770.01520.00970.86400.78640.54020.8950
DTU0.01320.01820.01820.01040.86180.78170.52120.8927
RE10K0.01250.01770.01520.00970.86400.78640.54020.8950
NRGBD0.01210.01820.01470.00960.86660.79240.54240.8953
CO3Dv20.01240.01840.01490.00960.86410.78720.54060.8946

Performance is broadly stable across calibration domains (Table R12). Pose AUC@30 changes little; reconstruction varies more, with several datasets improving VGGT’s NRGBD results. We test 0.5L because its assignments vary more than at 0.7L (manuscript Appendix C.3).

(b) Number of calibration scenes. We also repeat the analysis with smaller calibration sets, measuring both agreement with the reference allocation and downstream accuracy.

Table R13: Fraction of stages assigned the same linear/sparse operator as the reference configuration.

Calibration scenesVGGTHYWMπ³Mean
175.00%75.00%88.89%79.63%
275.00%83.33%100.00%86.11%
483.33%91.67%88.89%87.96%
883.33%91.67%100.00%91.67%

Table R14: Downstream performance with different calibration scene counts at 0.5L/0.8S.

BackboneScenes7S Acc. ↓7S Comp. ↓NRGBD Acc. ↓NRGBD Comp. ↓RE10K AUC-30 ↑RE10K RTA-5 ↑CO3Dv2 AUC-3 ↑CO3Dv2 AUC-30 ↑
VGGT10.01380.02540.02770.01370.73000.50850.51660.8869
20.01540.02640.01560.01050.72840.50550.52880.8908
40.01570.02640.01570.01040.73030.50980.53350.8912
80.01700.02550.01620.01080.72700.49880.52100.8902
Reference (13)0.01750.02700.02080.01380.72950.50720.53470.8920
HYWM10.01410.02260.02140.01370.83840.73640.33520.8473
20.01280.02360.01920.01340.84560.74980.34720.8504
40.01290.02360.01940.01350.84690.75230.34530.8495
80.01320.02460.01910.01340.84590.75030.37250.8534
Reference (13)0.01300.02440.01880.01330.85000.75830.37640.8543
π³10.01250.01820.01770.01010.86510.78940.52000.8922
20.01210.01820.01470.00960.86650.79240.54200.8953
40.01560.01650.03470.01260.84620.74640.29540.8560
80.01210.01820.01470.00960.86650.79240.54200.8953
Reference (13)0.01210.01820.01470.00960.86640.79230.54180.8954

Even small calibration sets can produce good allocations, and average agreement with the reference increases as more scenes are included (Tables R13–R14). Performance does not improve monotonically with scene count. The results show that the reference configuration does not require a large calibration corpus.

AuthorsOpenReview ↗

Official Comment by Authors

[Answer 3] Long-sequence reconstruction on HYWM and π³ (Q2)

We have added ScanNet-50 results for HYWM and π³ across 100–1,000 frames, reporting reconstruction quality and speed for the same configurations. These experiments demonstrate that Zebra3D’s long-sequence benefits extend beyond VGGT and are achievable both with and without refinement.

Table R15: ScanNet-50 long-sequence reconstruction with HYWM and π³.

BackboneMethod1,000 frames CD ↓Speedup ↑500 frames CD ↓Speedup ↑300 frames CD ↓Speedup ↑100 frames CD ↓Speedup ↑
HYWMDense0.4561.00×0.4491.00×0.4321.00×0.4311.00×
Zebra3D (0.5L/0.8S)0.4533.26×0.4412.86×0.4302.54×0.4281.81×
Zebra3D-KD (0.7L/0.8S)0.4595.21×0.4594.11×0.4393.17×0.4312.01×
π³Dense0.6081.00×0.6041.00×0.6081.00×0.6111.00×
Zebra3D (0.5L/0.8S)0.6094.11×0.6103.72×0.6211.83×0.6181.65×
Zebra3D-KD (0.7L/0.8S)0.6047.04×0.6036.05×0.6082.89×0.6042.25×

The new results show substantial acceleration without refinement. Training-free Zebra3D improves HYWM’s reconstruction quality at every evaluated length. On π³, its error is nearly identical to dense inference at 1,000 frames. These configurations run 3.26× and 4.11× faster, respectively (Table R15).

The more aggressive refined configuration provides further acceleration. On π³, it matches or improves dense reconstruction quality at every evaluated length and reaches 7.04× speedup. On HYWM, it reaches 5.21× speedup with a small increase in reconstruction error. Thus, both backbones offer useful choices between preserving more accuracy and running faster.

Effect of refinement. We also compare each model before and after refinement at the same configuration, so the effect of refinement is separate from the change in conversion budget.

Table R16: Effect of refinement on ScanNet-50 at 0.7L/0.8S. Bold indicates the better result within each backbone and sequence length.

BackboneMethod1,000 frames CD ↓500 frames CD ↓300 frames CD ↓100 frames CD ↓
VGGTZebra3D0.4800.4670.4590.449
Zebra3D-KD0.4600.4490.4470.439
HYWMZebra3D0.4650.4500.4290.434
Zebra3D-KD0.4590.4590.4390.431
π³Zebra3D0.6240.6080.6110.614
Zebra3D-KD0.6040.6030.6080.604

Refinement improves reconstruction in ten of twelve settings, including every evaluated length for VGGT and π³ (Table R16). HYWM improves at the shortest and longest lengths, with regressions at the intermediate lengths. Refinement is therefore useful for recovering quality after aggressive conversion.

We will incorporate both tables into the revised manuscript.

[Answer 4] Improving the presentation of the main results (W3)

Thank you for this suggestion. We will reduce the number of values repeated in the text and leave the detailed comparisons in the tables. The discussion will focus on what the results show: how well training-free conversion preserves accuracy, how speedups grow with sequence length, and when refinement is worth the additional training.

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

The paper targets the quadratic cost of global attention in feed-forward 3D foundation models. Profiling each global-attention stage, it observes that middle stages are both sensitive to linear replacement and highly concentrated in attention, while early and late stages are tolerant and diffuse. Building on this, ZEBRA3D converts a pretrained model into a sparse–linear hybrid: tolerant stages use linear-centered attention initialized by a closed-form ridge fit, and sensitive stages use tile-sparse softmax attention whose unselected tiles are summarized by mean-key/value landmarks. The paper reports accuracy close to the dense models at high attention sparsity, with end-to-end speedups that grow with the number of input views.

Critical strengths
  • An insightful and consistent diagnosis. The paper shows that a stage's sensitivity to linear replacement tracks how concentrated its attention is, and this depth-wise pattern holds across all three backbones. The resulting stage ranking is also fairly stable across calibration datasets, which makes the layer-wise choice of operator measurable rather than ad hoc.
  • A practical, training-free conversion. The closed-form centered fit and the landmark-augmented tile-sparse attention require no gradient updates, so a pretrained model can be converted and used directly. The main design choices are each examined in ablations.
Critical weaknesses
  • The advantage over existing efficient methods is not consistent across backbones and settings. ZEBRA3D is better than the available baselines in some settings but only comparable in others, and its fastest configuration is not its most accurate one.

    • Short sequences (Table 1, 10 views). On HYWM and π³, ZEBRA3D matches or outperforms the available baselines, although these baselines are the authors' ports of VGGT methods, applied without changes (Appendix B.2). On VGGT, no method dominates: among training-free methods, FastVGGT and SAF3R are best on 7Scenes, A3D on RE10K, and ZEBRA3D on NRGBD and CO3Dv2. At this length no method is meaningfully faster than the dense model (Fig. 5), so this is an accuracy-only comparison.
    • Long sequences (Table 2, VGGT on ScanNet-50). ZEBRA3D improves on FastVGGT at a similar speed, but at a speed similar to A3D it reaches a similar Chamfer distance, and at 100–300 frames A3D is faster than every ZEBRA3D configuration. Only the most aggressive configuration (0.7L/0.8S) is clearly faster at 500–1,000 frames. These ZEBRA3D results are also obtained after refinement, whereas the baselines are training-free.
    • Headline claims. Accuracy claims mainly use 0.5L/0.6S, while speed claims use 0.7L/0.8S (abstract, Fig. 1b, Fig. 5). At 1,000 frames, for instance, 0.5L/0.6S runs at 2.30×, compared with 3.83× for A3D and 3.97× for FastVGGT. An accuracy–speed plot for each backbone at long sequences, covering all configurations and baselines and separating training-free from refined models, would make the comparison much clearer.
  • The benefit of the refinement stage (ZEBRA3D-KD) is unclear. At 0.5L/0.6S, the refined model does not consistently improve over the training-free model and in several cases underperforms it, for example on CO3Dv2 for all three backbones and on NRGBD for VGGT (Table 1). Its largest gain, on VGGT RE10K, even exceeds the dense VGGT, and the dense VGGT fine-tuned on the same data shows a similar gain. This suggests that the gain comes from the fine-tuning data rather than from recovering accuracy lost in conversion. Refinement may be more useful in the aggressive 0.7L setting, as the π³ results suggest, but training-free 0.7L results are not reported for VGGT and HYWM, nor in Table 2. It would help to report these results, to present refined models separately from the training-free baselines (Table 1, Fig. 1b), and to discuss in which settings refinement is worth its additional training cost.

  • The sparse-attention ablation does not fully support the claimed advantage over A3D's operator. Table 3(b) reports this ablation only on π³, where tile-sparse attention with landmarks outperforms A3D's subsampled global attention (SGA). On VGGT, however, SGA achieves higher RE10K AUC@30 at both 0.5L and 0.7L and is competitive on 7Scenes (Appendix Table 17), although the proposed operator is generally better on NRGBD. Since each part of Table 3 reports a single backbone or budget, summarizing the ablations across all three backbones in the main text would give a more reliable picture.

Questions for authors
  • Can you add training-free 0.7L rows for VGGT and HYWM to Table 1 and training-free rows to Table 2? In which settings does refinement improve over the training-free model?
  • Does the depth-wise sensitivity pattern, and the method, carry over to geometry models outside the VGGT family, for example one built entirely on global attention such as Fast3R?
6 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer for the careful assessment and for recognizing our sensitivity analysis and practical training-free conversion. We have added matched-sparsity comparisons, long-sequence results for HYWM and π³, and evaluations on Fast3R and VGGT-Omega. Below, [Answer 1] compares Zebra3D with existing methods, [Answer 2] examines when refinement helps, [Answer 3] summarizes the sparse-attention ablations, and [Answer 4] presents the additional backbones. We will incorporate these findings into the revised manuscript. We hope this additional evidence addresses the reviewer’s concerns and would be grateful if the reviewer could reconsider their assessment in light of the new results.

[Answer 1] Advantage over existing efficient methods (W1)

(a) Short sequences: comparisons at comparable sparsity. To complement Table 1, we added two controlled comparisons at ten views to test whether Zebra3D's accuracy gains over A3D persist at comparable sparsity (Tables R1-R2). First, we keep A3D's published recipe, including its uncompressed pose-reference frame, and adjust Zebra3D to nearly match its sparsity (74% vs. 75%). Second, we subsample A3D's reference frame to match Zebra3D at 79% sparsity. Both comparisons lead to the same conclusion on VGGT: Zebra3D improves CO3Dv2 AUC-3 by 9.0-12.0 percentage points over A3D across the two comparisons, with comparable RE10K RTA-5 (Table R1).

Table R1: Training-free VGGT at ten views.

SparsityMethodRE10K RTA-5 ↑CO3Dv2 AUC-3 ↑
75%A3D, published recipe0.5350.468
74%Zebra3D (0.5L/0.5S)0.5300.558
79%A3D, reference frame subsampled0.5260.434
79%Zebra3D (0.5L/0.6S)0.5250.554

On HYWM, Zebra3D leads on both metrics at matched sparsity and retains both gains at 89% sparsity against published A3D at 75% (Table R2). Manuscript Table 1 uses A3D’s published reference-frame setting, which achieves higher accuracy than the reference-subsampled variant in Tables R1–R2. We will clarify this distinction and include the comparable-sparsity results in the revision.

Table R2: Training-free HYWM at ten views. Differences are Zebra3D − A3D; brackets report uncertainty intervals.

ComparisonSparsity: Zebra3D / A3DΔ RE10K RTA-5 ↑Δ CO3Dv2 AUC-3 ↑
Zebra3D (0.5L/0.6S) vs. published A3D79% / 75%+0.029 [+0.024, +0.034]+0.078 [+0.072, +0.083]
Zebra3D (0.5L/0.6S) vs. A3D with reference frame subsampled79% / 79%+0.042 [+0.036, +0.048]+0.109 [+0.103, +0.115]
Zebra3D (0.5L/0.8S) vs. published A3D89% / 75%+0.018 [+0.013, +0.023]+0.054 [+0.049, +0.059]

(b) Long sequences: reconstruction quality and speed. We agree that the trade-off depends on sequence length and configuration. On VGGT, A3D is faster at 100–300 frames. At 1,000 frames, Zebra3D’s fastest configuration reaches 5.60× but has higher CD than A3D (Table 2). To test whether Zebra3D can accelerate long-sequence reconstruction without refinement beyond VGGT, we added ScanNet-50 results for HYWM and π³ at 100–1,000 frames (Table R3).

In Table R3, at 1,000 frames, training-free Zebra3D (0.5L/0.8S) achieves 3.26× and 4.11× speedups over dense HYWM and π³, respectively, with slightly lower CD on HYWM and nearly identical CD on π³. The more aggressive refined configuration (0.7L/0.8S) delivers larger speedups on both backbones and matches or improves on dense π³’s CD at every tested length. Answer 2 isolates the effect of refinement at a fixed configuration.

Table R3: ScanNet-50 long-sequence reconstruction with HYWM and π³.

BackboneMethod1,000 frames CD ↓Speedup ↑500 frames CD ↓Speedup ↑300 frames CD ↓Speedup ↑100 frames CD ↓Speedup ↑
HYWMDense0.4561.00×0.4491.00×0.4321.00×0.4311.00×
Zebra3D (0.5L/0.8S)0.4533.26×0.4412.86×0.4302.54×0.4281.81×
Zebra3D-KD (0.7L/0.8S)0.4595.21×0.4594.11×0.4393.17×0.4312.01×
π³Dense0.6081.00×0.6041.00×0.6081.00×0.6111.00×
Zebra3D (0.5L/0.8S)0.6094.11×0.6103.72×0.6211.83×0.6181.65×
Zebra3D-KD (0.7L/0.8S)0.6047.04×0.6036.05×0.6082.89×0.6042.25×

(c) Presentation. We will label headline results by backbone, configuration, and setting, report peak speedup separately from RE10K accuracy. We will add accuracy–speed plots with all evaluated methods by backbone, marking training-free and refined variants.

AuthorsOpenReview ↗

Official Comment by Authors

[Answer 2] Benefit of refinement under aggressive conversion (W2/Q1)

We added the requested training-free VGGT and HYWM results at 0.7L/0.8S, together with training-free long-sequence results for all three backbones. Comparing each model before and after refinement shows that refinement is most useful when aggressive conversion has reduced accuracy.

(a) Short sequences. Refinement improves both RE10K pose metrics and NRGBD accuracy error on all three backbones (Table R4). The clearest recovery is on π³, where every reported metric improves and NRGBD accuracy error falls by approximately 60%. VGGT also benefits substantially on RE10K, gaining 8.7 percentage points in AUC@30.

Table R4: Effect of refinement at the same 0.7L/0.8S configuration. Bold marks the better result within each backbone; ties are bolded in both rows.

BackboneMethod7S Acc. ↓7S Comp. ↓NRGBD Acc. ↓NRGBD Comp. ↓RE10K AUC-30 ↑RE10K RTA-5 ↑CO3Dv2 AUC-3 ↑CO3Dv2 AUC-30 ↑
VGGTZebra3D0.0170.0250.0290.0180.7180.4590.4620.881
Zebra3D-KD0.0160.0250.0250.0190.8050.6540.4030.872
HYWMZebra3D0.0130.0240.0230.0140.8390.7390.3300.841
Zebra3D-KD0.0140.0220.0190.0140.8470.7520.2950.838
π³Zebra3D0.0210.0200.0430.0150.8360.7260.2560.841
Zebra3D-KD0.0130.0180.0170.0100.8440.7510.4400.876

The benefit depends on the task: CO3Dv2 performance decreases for VGGT and HYWM, and a few reconstruction metrics show smaller regressions. At the more conservative 0.5L/0.6S setting, refinement also gives less consistent gains (manuscript Table 1). These results help identify when the additional training is worthwhile.

(b) Long sequences. Refinement improves VGGT and π³ at every evaluated length, and all three backbones at 1,000 frames (Table R5). HYWM shows a smaller, mixed benefit, improving at 100 and 1,000 frames but regressing at the intermediate lengths.

Table R5: ScanNet-50 Chamfer distance before and after refinement at 0.7L/0.8S. Bold indicates the better result within each backbone and sequence length.

BackboneMethod1,000 frames ↓500 frames ↓300 frames ↓100 frames ↓
VGGTZebra3D0.4800.4670.4590.449
Zebra3D-KD0.4600.4490.4470.439
HYWMZebra3D0.4650.4500.4290.434
Zebra3D-KD0.4590.4590.4390.431
π³Zebra3D0.6240.6080.6110.614
Zebra3D-KD0.6040.6030.6080.604

(c) When is refinement worthwhile? For conservative conversion, the training-free model already preserves much of the dense model’s accuracy. Refinement becomes more useful at aggressive budgets, where it can recover a substantial part of the lost quality. It requires 3,000 optimizer updates, including a frozen-teacher forward pass during training, and adds no inference-time computation.

The reviewer also raises an important distinction between recovering conversion losses and benefiting from the training data. On VGGT RE10K, both the refined hybrid and the fine-tuned dense model improve, so we do not attribute the full gain to recovery after conversion (manuscript Table 1). ScanNet shows a different pattern: dense ground-truth fine-tuning worsens reconstruction (manuscript Table 2), while refinement improves the aggressive Zebra3D model (Table R5). This supports the value of the refinement recipe beyond simply training on additional data.

AuthorsOpenReview ↗

Official Comment by Authors

[Answer 3] Sparse-attention ablation across backbones (W3)

We summarize SGA and TileSparse + Landmark across all three backbones at both linear-attention budgets, using 0.75S throughout. Our operator improves HYWM and π³ pose accuracy at both budgets and NRGBD reconstruction across all backbones at 0.5L.

Table R6: RE10K pose estimation, AUC@30 ↑.

BackboneLinear fractionSGATileSparse + Landmark
VGGT0.5L0.7390.733
0.7L0.7270.721
HYWM0.5L0.8430.852
0.7L0.8300.841
π³0.5L0.8470.868
0.7L0.7590.837

Table R7: NRGBD reconstruction, mean accuracy error ↓.

BackboneLinear fractionSGATileSparse + Landmark
VGGT0.5L0.0290.018
0.7L0.0400.029
HYWM0.5L0.0200.019
0.7L0.0240.023
π³0.5L0.0170.014
0.7L0.0430.043

The clearest advantage is on NRGBD: TileSparse + Landmark gives lower reconstruction error in five of six settings and ties SGA in the remaining setting at the reported precision (Table R7). The gains are largest on VGGT, where it reduces error by roughly one-third. On RE10K, it leads on HYWM and π³, while SGA is slightly better on VGGT (Table R6). SGA also remains competitive on 7Scenes (manuscript Appendix Table 15).

Separately, manuscript Table 3(b) shows that adding landmarks substantially improves sparse attention. We will summarize the cross-backbone comparisons in the main text to distinguish this benefit from the task-dependent comparison with SGA.

[Answer 4] Generalization to Fast3R and VGGT-Omega (Q2)

We have added Fast3R experiments to test a different architecture. Selecting layers using Fast3R’s own sensitivity measurements works substantially better than transferring VGGT’s layout.

(a) Sensitivity-guided allocation on Fast3R. On Fast3R, the middle stages tolerate linear replacement best, while the early and late stages are more sensitive. We therefore select the allocation from Fast3R’s sensitivity measurements and compare it with front-sparse placement and a direct transfer of VGGT’s placement.

Table R8: Fast3R layer-placement ablation at 0.5L/0.8S, approximately 89% sparsity.

Placement7Scenes Acc. ↓NRGBD Acc. ↓RE10K AUC-30 ↑CO3Dv2 AUC-30 ↑
Sensitivity-guided0.0540.1120.4440.499
Front-sparse0.0930.1790.1350.218
VGGT placement transferred0.0810.1740.0040.001

Sensitivity-guided placement performs best on every metric (Table R8). Transferring VGGT’s layout severely reduces pose accuracy, showing why the allocation should follow each backbone’s measured sensitivity.

(b) End-to-end Fast3R conversion. We also compare Zebra3D with dense Fast3R and A3D at similar sparsity.

Table R9: Fast3R reconstruction and pose estimation.

MethodMean sparsity at 300 views7Scenes Acc. ↓NRGBD Acc. ↓RE10K AUC-30 ↑CO3Dv2 AUC-30 ↑
Dense Fast3R0%0.0360.0720.6050.767
A3D80.6%0.1150.2450.1320.006
Zebra3D (0.5L/0.6S)79.5%0.0520.1090.4500.549

Zebra3D outperforms A3D on all four metrics, reducing reconstruction errors by approximately 55% on both datasets at similar sparsity (Table R9). A3D’s use of frame-wise attention in early layers may explain its larger accuracy loss, given Fast3R’s sensitivity there.

(c) Additional backbone: VGGT-Omega. We also evaluate Zebra3D on VGGT-Omega, another recent VGGT-based model.

Table R10: VGGT-Omega. Sparsity is reported at 300 and 10 views.

MethodSparsity, 300 / 10 views7S Acc. ↓7S Comp. ↓NRGBD Acc. ↓NRGBD Comp. ↓RE10K AUC-30 ↑RE10K RTA-5 ↑CO3Dv2 AUC-3 ↑CO3Dv2 AUC-30 ↑
Dense0% / 0%0.0130.0130.0250.0100.8780.8010.4810.881
A3D73% / 66%0.0130.0130.0250.0110.8690.7830.3810.855
BlockSparse3D70% / 70%0.0130.0130.0340.0130.8630.7710.3670.855
Zebra3D (0.4L/0.7S)78% / 72%0.0130.0130.0240.0100.8760.7970.4500.874

Zebra3D matches or improves every reported quality metric relative to both efficient baselines while using greater sparsity (Table R10). It also stays close to dense VGGT-Omega, matching its 7Scenes reconstruction errors and remaining within 0.2 percentage points on RE10K AUC@30.

These additions extend the evaluation to five backbones. Fast3R has a different sensitivity pattern from VGGT, HYWM, and π³, but the same measurement-based allocation approach remains effective. We will clarify this distinction in the revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers and the AC for their constructive feedback. We have uploaded a revised manuscript in which all changes are highlighted in blue. The main revisions are listed below.

  1. Comparison with efficient baselines (AKhk W1)
    • New Appendix A.5 (Tables 7-8) compares Zebra3D with A3D at comparable sparsity on VGGT and HYWM, including an A3D variant whose reference frame is also subsampled. Appendix B.2 clarifies that the main tables use A3D's published recipe.
  2. Long-sequence reconstruction on all backbones (AKhk W1; mygd Q2; gdLY W3)
    • New Appendix Table 6 reports ScanNet-50 Chamfer distance and end-to-end speedup for HYWM and π³ at 100-1,000 frames, for training-free and refined variants of every configuration, compared with the dense backbones. Figure 7 shows the 1,000-frame trade-off for all three backbones. The "Long-sequence reconstruction" paragraph (Section 4.2) summarizes the results on all three backbones.
  3. When refinement helps (AKhk W2/Q1)
    • Table 1 adds training-free Zebra3D (0.7L/0.8S) for VGGT and HYWM, and Table 2 adds training-free VGGT (0.7L/0.8S), so that each refined model can be compared with its training-free counterpart.
    • The "Refinement and operating points" paragraph (Section 4.2) compares each model before and after refinement at a fixed configuration, and discusses when the additional training is worthwhile.
  4. Sparse-attention ablation across backbones (AKhk W3)
    • The "Attention operators" paragraph (Section 4.3) now summarizes the comparison with SGA across all three backbones and both linear fractions. New Appendix Table 22 collects these results.
  5. Additional backbones (AKhk Q2; gdLY Q1)
    • New Appendix E.4 (Tables 33-35) adds Fast3R, including a layer-placement ablation, and VGGT-Omega. This extends the evaluation to five backbones. A summary is added to Section 4.2.
  6. Peak GPU memory (mygd W2/Q2; gdLY W2)
    • New Appendix Table 10 reports peak GPU memory for all three backbones from 100 to 1,000 views, summarized in the "Efficiency across backbones" paragraph (Section 4.2).
  7. Robustness to the calibration set (mygd W1/Q1)
    • Appendix C.3 adds downstream results for allocations selected with different calibration datasets and with different numbers of calibration scenes (Tables 14-16), summarized in Section 4.3.
  8. Relational distillation (gdLY W1/Q2)
    • Table 3(c) adds a matched ablation that removes only the relational loss. Appendix D.4 adds long-sequence ScanNet-50 and two-view pose results on MegaDepth-1500 and ScanNet-1500 (Tables 28-29), summarized in the "Distillation" paragraph (Section 4.3).
  9. Qualitative comparisons and in-the-wild examples (gdLY W4)
    • New Appendix Figure 13 (Appendix F.1) compares dense VGGT and Zebra3D-KD (0.5L/0.6S) on five additional video sequences, including two in-the-wild aerial videos.
  10. Presentation and corrections
    • Section 4.2 now repeats fewer values from the tables and focuses on what the results show, including how speedups grow with sequence length and when refinement is worth the additional training (mygd W3).
    • Headline results in the abstract, introduction, and figures now state the configuration (0.5L/0.6S or 0.7L/0.8S) and the setting (training-free or refined) (AKhk W1). Figure 1(b) now shows training-free and refined Zebra3D separately (AKhk W2).
    • We unified the π³ RE10K AUC-30 and VGGT RE10K RTA-5 values across tables, corrected the calibration scene count in Figure 8 (13 scenes), and fixed typos (gdLY W5).

Index by reviewer.

  • AKhk: W1 → items 1, 2, 10; W2/Q1 → items 3, 10; W3 → item 4; Q2 → item 5.
  • mygd: W1/Q1 → item 7; W2/Q2 → items 2, 6; W3 → item 10.
  • gdLY: W1/Q2 → item 8; W2 → item 6; W3 → item 2; W4 → item 9; Q1 → item 5; W5 → item 10.

Table and figure numbering.

The new appendix tables shift the numbering of later tables. Tables cited in our responses map as follows: Appendix Table 7 → Table 11, Table 12 → Table 19, and Table 15 → Table 23. Tables 1-5, Figures 1-12, and all existing section numbers are unchanged; Figure 13 and Appendices A.5 and E.4 are new.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.