acceptodds
Under review as a conference paper at ICLR 2027

Dream-Lift-VLA: Generative 2D-to-3D World Dreaming for Vision-Language-Action Models

Abstract

Vision-language-action (VLA) models offer a promising path toward generalist robot policies. A key challenge is that sparse action supervision provides limited guidance for capturing fine-grained future scene changes. To address this limitation, recent methods incorporate world modeling into VLAs by jointly predicting future 2D videos and 3D geometry. However, these two forms of supervision are inherently ordered, with future 2D prediction first capturing current-to-future scene evolution within the same modality and future 3D prediction subsequently grounding this evolution across modalities in 3D space. Parallel training ignores this order by imposing 3D grounding before the 2D scene evolution has been sufficiently learned. To address this issue, we propose **Dream-Lift-VLA**, a progressive VLA learning framework that includes two phases: **Dream** and **Lift**. Specifically, the Dream phase uses world embeddings to condition a pretrained video diffusion model for future-video generation, enabling the VLA to learn fine-grained scene dynamics and manipulation semantics. Building upon these learned dynamics, the Lift phase uses the world embeddings to condition a pretrained depth diffusion model for future-depth generation, further shaping the VLA representations with 3D spatial awareness. Empowered by this progressive 2D-to-3D representation learning, our policy develops stable representations of both manipulation semantics and 3D geometry. Extensive experiments on diverse simulation and real-world tasks demonstrate that Dream-Lift-VLA consistently outperforms state-of-the-art approaches. The code of this work will be released soon.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Manipulation, Vision-Language-Action Models, Generative World Modeling

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_2DwbOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dream-Lift-VLA, a progressive framework that first uses future-video supervision to learn scene dynamics and subsequently introduces future-depth supervision to strengthen spatial representations. The learned world embeddings support action generation, while both generative branches are removed during inference. Experiments evaluate the framework across simulation benchmarks and real-world manipulation tasks.

Critical strengths

1.The evaluation covers multiple simulation benchmarks and real-world tasks under both in-distribution and out-of-distribution conditions, demonstrating competitive performance and consistent overall improvements over the compared baselines. 2.The ablation studies are extensive, examining training order, training duration, supervision targets, and generative backbones.

Critical weaknesses
  1. Although the curriculum ablations support the effectiveness of the proposed training schedule, the rationale for learning future 2D dynamics before introducing depth supervision remains insufficiently explained. In particular, the paper provides limited analysis of how this progression improves representations relevant to action prediction, beyond the observed performance gains.
  2. The contribution relative to DreamVLA also warrants clearer characterization, as the comparison does not fully control for differences in the backbone and pretrained generative components. Consequently, the extent to which the improvements over DreamVLA stem from the proposed curriculum remains difficult to assess.
Questions for authors

1.Why should future video prediction precede depth prediction for VLA learning? Can the authors provide further analysis showing how this ordering improves representations relevant to action prediction? 2How sensitive is the advantage over joint training to the relative weights of the video and depth objectives? Does progressive training remain advantageous when these weights are carefully tuned? 3.Do the findings from the training-order ablations on MetaWorld also hold on LIBERO and real-world tasks? Such results would help establish the broader applicability of the proposed curriculum. 4. Despite the empirical gains, I find the methodological contribution somewhat incremental, with the novelty appearing to lie primarily in the integration and training schedule of existing components. I currently lean toward rejection, but remain open to revising my assessment if the rebuttal convincingly addresses the concerns above and better substantiates the methodological novelty.

4 comments
AuthorsOpenReview ↗

Response to the Reviewer 2Dwb (1/3)

We thank the reviewer for the constructive feedback. Below, we analyze the learned representations, present controlled experiments with DreamVLA, evaluate sensitivity to loss weights and the generalizability of the curriculum, and clarify our methodological contributions.

[W1, Q1] Further Representation Analysis

A: We thank the reviewer for highlighting the need to explain why future-video learning should precede future depth supervision and how this order affects action-relevant representations. To assess the benefits of the proposed curriculum, we follow prior works [1, 2] and use linear probes to examine the future-dynamics and action-relevant information encoded in the learned representations. Furthermore, we visualize attention maps to show which image regions the action tokens attend to. Finally, we discuss possible mechanisms behind the observed differences.

Both probe experiments compare Joint, Dream, Dream→Lift, and Lift→Dream, using the same 600 observations for fitting and all 96 observations from 12 unseen expert trajectories across six tasks for testing. The VLM remains frozen, and neither the original action head nor proprioceptive states are used. We evaluate prediction errors using standard root mean square error (RMSE) and mean absolute error (MAE), with lower values indicating better performance [3].

(1) Future-dynamics probe. We mean-pool the final-layer features of the six world tokens and train a linear probe to predict the 3D displacement of the primary manipulated object or tool from t to t+4. RMSE is computed in centimeters against simulator-recorded ground-truth displacements from the expert trajectories.

(2) Action probe. We separately mean-pool the final-layer world-token and action-token features, concatenate them, and train a linear probe to predict the current and next expert actions. We report translational-action RMSE and gripper-action MAE.

Compared with joint training and reverse-order training, the Dream stage yields lower future-displacement and translational-action prediction errors, although gripper prediction does not improve. The subsequent Lift stage further reduces all three errors, with gripper MAE decreasing from 0.0660 to 0.0403. These results suggest that our curriculum learns representations that better capture future dynamics, improve geometric understanding, and support action prediction.

ModelFuture Object Displacement RMSE (cm) ↓Translational Action RMSE ↓Gripper Action MAE ↓
Joint1.00510.35620.0565
Lift→Dream0.93130.35540.0652
Dream0.88790.33400.0660
Dream→Lift0.87080.31270.0403

(3) Attention Visualization. We visualize attention from action tokens to visual tokens using the same observations and instructions. In the presented examples (https://anonymous.4open.science/r/dream-lift-vla-20374/figures/attention-visualization.png), Joint training yields relatively diffuse attention, whereas the Dream stage encourages clearer attention to manipulated objects and a stronger response to the target placement location in Shelf Place (red box). The subsequent Lift stage further strengthens the focus on manipulated objects.

AuthorsOpenReview ↗

Response to the Reviewer 2Dwb (2/3)

[W1, Q1] Further Representation Analysis (continued)

Analysis:

(1) Under the same 60k-step training budget, joint SVD+PPD training with loss weights of (0.25, 0.25) achieves a success rate of 67.3% (see our response to Q2), close to the 68.0% achieved by PPD-only training. This suggests that the model may not fully benefit from video supervision when both objectives are introduced from the outset. One possible explanation is that introducing depth supervision from the outset in joint training may encourage the model to prioritize overall scene geometry for action prediction without sufficiently focusing on manipulated objects, targets, and their changes, making it difficult to develop a robust understanding of manipulation semantics. In contrast, the Dream stage first develops action-relevant semantic understanding, providing semantic and dynamic context for future-depth prediction. The Lift stage uses the dynamics learned during Dream to understand the resulting spatial relationships, further refining its understanding of manipulation semantics from a spatial perspective. Moreover, reversing the order introduces geometric supervision before Dream’s semantic and dynamic context is established, potentially weakening the grounding of learned manipulation semantics in task-relevant spatial relationships.

The attention maps support this interpretation: attention under joint training is relatively diffuse, whereas the Dream stage focuses more clearly on manipulated objects and target locations, and the Lift stage maintains and further strengthens the focus on manipulated objects. Reverse-order training shows some signs of semantic understanding, with attention to objects in certain cases, but remains limited in precise localization.

(2) Moreover, joint training can fit actions in scenes seen during training (as discussed in our response to Reviewer g5bK’s comment W2), but it does not fully develop stable representations of manipulation semantics, which limits generalization. Our curriculum learning approach builds on a more stable understanding of manipulation semantics and incorporates spatial relations relevant to manipulation, reducing reliance on scene-specific cues. In reverse-order training, the semantics learned during Dream may remain insufficiently grounded in spatial relationships, leading to less precise localization and higher probe errors.

The probe results further support this interpretation: compared with joint training and reverse-order training, our curriculum achieves lower errors in both the future-dynamics and action probes on unseen trajectories.

[1] AdaPolicy: Action-Driven Anticipatory Robot Policies via Next-Chunk Prediction, arXiv 2026.
[2] Vision-Language Models Unlock Task-Centric Latent Actions, arXiv 2026.
[3] Root mean square error (RMSE) or mean absolute error (MAE)? – Arguments against avoiding RMSE in the literature, Geosci. Model Dev. 2014.

[W2] Additional Controlled Experiments

A: Thank you for the comment. To distinguish the contributions of the training curriculum from those of the generative components, we conduct two groups of controlled experiments based on DreamVLA on MetaWorld.

Settings. The first retains DreamVLA’s original components but replaces joint supervision with a two-stage curriculum: learning future dynamics and semantics first, followed by depth geometry, with action supervision retained in both stages. The second keeps the same policy backbone and curriculum learning but replaces the auxiliary prediction components with SVD and PPD. The total training budget remains 40 epochs, divided into 20 epochs per stage. Each stage uses an initial learning rate of 2×10−42\times10^{-4} and cosine decay, following DreamVLA’s official configuration.

Result. Without changing the model components, the curriculum learning yields absolute success-rate gains of 4.7% on MetaWorld. Introducing SVD and PPD provides further gains of 3.3%. These results further support the effectiveness of the proposed curriculum and highlight the benefits of leveraging pretrained generative components.

MethodTotal EpochsMetaWorld SR (%)
DreamVLA4066.0
DreamVLA + Two-Stage Curriculum20+2070.7
DreamVLA + Two-Stage SVD–PPD20+2074.0
AuthorsOpenReview ↗

Response to the Reviewer 2Dwb (3/3)

[Q2] Sensitivity to Loss Weights

A: We conduct additional loss-weight sensitivity experiments for joint training, varying both the total auxiliary-loss weight and the relative weights of video and depth prediction. With a total weight of 1.0, (0.5, 0.5), (0.2, 0.8), and (0.8, 0.2) achieve success rates of 65.3%, 65.3%, and 66.0%, respectively. With a total weight of 0.5, (0.25, 0.25), (0.1, 0.4), and (0.4, 0.1) achieve 67.3%, 66.0%, and 66.7%. Although reweighting improves joint training from 65.3% to 67.3%, it remains below progressive training’s 82.0%. Thus, joint training exhibits some sensitivity to loss weighting, but progressive training retains a substantial advantage across the tested configurations.

(λ2D,λ3D)(\lambda_{\mathrm{2D}},\lambda_{\mathrm{3D}})Total WeightAvg. SR (%)
Action-only baseline068.0
(0.5,0.5)(0.5,0.5) (original)1.065.3
(0.2,0.8)(0.2,0.8)1.065.3
(0.8,0.2)(0.8,0.2)1.066.0
(0.25,0.25)(0.25,0.25)0.567.3
(0.1,0.4)(0.1,0.4)0.566.0
(0.4,0.1)(0.4,0.1)0.566.7

[Q3] Curriculum Generalizability

A: Thank you for the suggestion. We extend the training-order ablations to LIBERO and real-world demonstrations, evaluating on LIBERO, LIBERO-Plus, and four real-world tasks. The table reports average success rates (%). Dream→Lift achieves absolute gains of 2.7%, 9.7%, and 13.7% over joint training across these three settings, respectively, and consistently outperforms Lift→Dream. These results reproduce the trend observed on MetaWorld and support the broader applicability of our training curriculum.

Training StrategyLIBERO SR (%)LIBERO-Plus SR (%)Real-World ID SR (%)
Dream→Lift (ours)99.083.785.0
Lift→Dream97.077.773.8
Joint Dream+Lift96.374.071.3

[Q4] Method Contributions

A: Thanks for your valuable suggestions, which have helped us further clarify our contributions. Our method introduces a novel concept and architecture, as reviewer u75x also noted: “The idea to ‘2D-to-3D imagine’ is interesting and novel to the best of my knowledge.” We further clarify our contributions from three perspectives, namely conceptual novelty, architectural design, and extensibility.

(1) Concept. In world modeling for VLAs, no prior work has explored our progressive world-modeling paradigm. Learning future semantics first and then developing geometry foresight has been shown to be important for action generation, yet all previous methods have overlooked this point.
(2) Architecture. We introduce a novel two-stage generative VLA architecture that deliberately leverages pretrained generative priors through tailored conditioning and adaptation, connecting world modeling to action learning rather than simply assembling existing components.

  • Native-pathway conditioning: We carefully design an interface in which PPD’s native VFM features attend to world embeddings, integrating Dream’s future dynamics while preserving its native geometric context (Table 9).
  • Generative priors for policy learning: We deliberately adapt pretrained video and depth generators jointly with the VLA, harnessing their native generative capabilities to shape action-relevant dynamic and spatial representations (Table 10).
  • From depth estimation to spatial foresight: We explicitly repurpose PPD from current-depth estimation to future-depth prediction, grounding VLA representations in future geometry and encouraging the policy to anticipate action-relevant spatial changes rather than only encode the current scene (Table 7).

(3) Extensibility. Our design supports extensions across generative backbones (Table 11), supervision paradigms (Appendix B.3, Table 16), and VLA baselines (our reply to W2). These results support progressive world modeling as a reusable learning principle, offering new methodological insight and a strong, extensible baseline for the VLA community.
We will incorporate these clarifications into the revised manuscript to help readers better understand Dream-Lift-VLA’s design rationale and novelty.

We thank the reviewer again for the constructive suggestions, which help us clarify Dream-Lift-VLA’s contributions and strengthen the experimental evidence, and we will include these results and our discussion in the latest version.

AuthorsOpenReview ↗

With Appreciation for Your Feedback and Consideration #20374.

Dear Reviewer 2Dwb,

Thank you again for your thoughtful and constructive feedback. In our rebuttal, we have analyzed the learned representations through linear probes and attention visualizations, added controlled DreamVLA experiments to distinguish the benefits of the curriculum from those of the generative components, evaluated loss-weight sensitivity, extended training-order ablations to LIBERO, LIBERO-Plus, and real-world tasks, and clarified our conceptual and architectural contributions.

We hope that our response helps address your concerns, and we would sincerely appreciate any further feedback.

Best regards,

The Authors

Reviewer_g5bKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper trains a VLA policy in two future-prediction stages: world queries drive a frozen, LoRA-adapted video diffusion model first, then a depth diffusion model takes over to predict future depth; neither runs at inference. The claim is that 2D dynamics should be learned before 3D geometry. MetaWorld 82.0%, LIBERO 99.0%, LIBERO-Plus 83.7%, real-robot OOD 78.1% vs about 42% for the strongest baselines.

Critical strengths
  1. What convinces me most is Table 5, not the headline success rates. Going 2D→3D gives 82.0%, direct 3D 69.3%, joint training 64.7%, 3D→2D 68.7%, and 2D followed by joint 72.0%. So the ordering itself matters; the gain is not just "add another supervision signal."
  2. The evaluation is unusually broad for this kind of paper, real-robot OOD included (78.1% vs about 42%). Since both generative branches are dropped at test time, latency stays low at 76.8 ms, the best among the compared methods, and swapping in Wan2.2 and DepthFM still gives 81.3%. The recipe is not welded to SVD and PPD.
Critical weaknesses
  1. The depth supervision is partly self-referential: PPD generates the Stage-II targets offline, and the adapted PPD is then trained to predict them. Whether this injects new geometric information or mostly distills the teacher into the policy is unclear, and whatever depth errors the teacher has become the ceiling.
  2. Why does joint training (64.7%) do worse than even the action-only baseline (68.0%)? Loss weighting, gradient conflict, optimization difficulty; the paper doesn't really say. As long as that is unexplained, I can't tell whether the lesson is "2D must come first" or just "this particular joint objective interferes with itself," and the main claim rests on that distinction.
Questions for authors
  1. Were the LIBERO/LIBERO-Plus baselines re-run under your protocol? If not, how do checkpoint selection, evaluation counts and action chunking differ?
  2. Why does joint 2D+3D training fall below the action-only baseline? Any gradient-conflict measurements or other evidence?
  3. For the robot experiments: same 100 demonstrations per task for π0.5 and DreamVLA, fine-tuned with their official recipes? Do the OOD gaps survive a change of demonstration set?
7 comments
AuthorsOpenReview ↗

Response to the Reviewer g5bK (1/4)

We thank the reviewer for the constructive feedback. Below, we compare depth supervision targets, analyze joint and progressive training, clarify the LIBERO and LIBERO-Plus evaluation protocols, and report additional real-world OOD results.

[W1] Depth Target

A: We thank the reviewer for the question about PPD-generated depth supervision. We address these concerns by citing the existing decoder results in Appendix B.1, Table 14, and adding comparisons using simulator ground-truth and RealSense depth to examine the role of the pretrained predictor and the impact of the depth-supervision source.

First, using PPD-generated depth labels, a decoder trained from scratch improves success from 72.0% to 79.3% (Appendix B.1, Table 14). Using PPD as the predictor further raises success to 82.0%. This comparison suggests that future geometric supervision drives much of the improvement, while PPD’s pretrained priors provide further benefits.

Second, replacing pseudo-depth with GT depth on MetaWorld preserves 82.0% success, while using RealSense depth in real-world experiments reduces average success from 85.0% to 83.8%. We observe substantial noise around object boundaries in RealSense depth, which may compromise supervision quality. These results suggest that PPD-generated depth maps provide sufficient geometric detail for effective policy supervision.

Together, the above results show that using PPD’s own predictions as labels does not prevent the policy from benefiting from future geometric supervision, whose effectiveness is further supported by the independent GT-depth comparison.

SettingDepth SupervisionAvg. SR (%)
MetaWorldPPD pseudo-depth82.0
MetaWorldGT depth82.0
Real-worldPPD pseudo-depth85.0
Real-worldRealSense depth83.8

[W2, Q2] Joint Training

A: We thank the reviewer for highlighting the need to distinguish the benefits of our curriculum from optimization issues in joint training. We address this concern through loss-weight sensitivity experiments, gradient and convergence analyses, and representation probes and visualizations, followed by a discussion of possible underlying mechanisms.

Weights and Optimization. As shown below, we evaluate six joint-training configurations, varying both the total auxiliary-loss weight and the relative weights of the 2D and 3D objectives. With a total weight of 1.0, (0.5, 0.5), (0.2, 0.8), and (0.8, 0.2) achieve 65.3%, 65.3%, and 66.0%, respectively. Reducing the total weight to 0.5 yields 67.3%, 66.0%, and 66.7% for (0.25, 0.25), (0.1, 0.4), and (0.4, 0.1). Although reweighting improves joint training, the best tested configuration achieves only 67.3%, remaining below the action-only baseline of 68.0% and our progressive curriculum’s 82.0%.

We also monitor gradient cosine similarities throughout the full 60k-step training run. The mean similarities for Action–Video, Action–Depth, and Video–Depth are +0.000153, +0.001143, and +0.000458, respectively, all close to zero. Following the method for quantifying gradient conflict strength in [1], we further measure the proportion of sampled cosine similarities below −0.01, which is 0% for all three pairs. These measurements do not indicate pronounced gradient conflict under this criterion. Furthermore, the action losses (https://anonymous.4open.science/r/dream-lift-vla-20374/figures/action-loss-curve.png) of both training paradigms decrease substantially and stabilize, showing no clear difficulty in optimizing the action objective.

(λ2D,λ3D)(\lambda_{\mathrm{2D}},\lambda_{\mathrm{3D}})Total WeightAvg. SR (%)
Action-only baseline068.0
(0.5, 0.5), original setting1.065.3
(0.2, 0.8)1.065.3
(0.8, 0.2)1.066.0
(0.25, 0.25)0.567.3
(0.1, 0.4)0.566.0
(0.4, 0.1)0.566.7
AuthorsOpenReview ↗

Response to the Reviewer g5bK (2/4)

[W2, Q2] Joint Training (continued)

Representation. To assess the benefits of the proposed curriculum, we follow prior works [2, 3] and use linear probes to examine the future-dynamics and action-relevant information encoded in the learned representations. Furthermore, we visualize attention maps to show which image regions the action tokens attend to.

Both probe experiments compare Joint, Dream, Dream → Lift, and Lift → Dream, using the same 600 observations for fitting and all 96 observations from 12 unseen expert trajectories across six tasks for testing. The VLM remains frozen, and neither the original action head nor proprioceptive states are used. We evaluate prediction errors using standard root mean square error (RMSE) and mean absolute error (MAE), with lower values indicating better performance [4].

(1) Future-dynamics probe. We mean-pool the final-layer features of the six world tokens and train a linear probe to predict the 3D displacement of the primary manipulated object or tool from t to t+4. RMSE is computed in centimeters against simulator-recorded ground-truth displacements from the expert trajectories.

(2) Action probe. We separately mean-pool the final-layer world-token and action-token features, concatenate them, and train a linear probe to predict the current and next expert actions. We report translational-action RMSE and gripper-action MAE.

Compared with Joint, Dream achieves lower future-displacement and translational-action prediction errors, although gripper prediction does not improve. Adding Lift further reduces all three errors, with gripper MAE decreasing from 0.0660 to 0.0403. These results suggest that our curriculum learns representations that better capture future dynamics, improve geometric understanding, and support action prediction.

ModelFuture Object Displacement RMSE (cm) ↓Translational Action RMSE ↓Gripper Action MAE ↓
Joint1.00510.35620.0565
Lift → Dream0.93130.35540.0652
Dream0.88790.33400.0660
Dream → Lift0.87080.31270.0403

(3) Attention Visualization. We visualize attention from action tokens to visual tokens using the same observations and instructions. In the presented examples (https://anonymous.4open.science/r/dream-lift-vla-20374/figures/attention-visualization.png), Joint training yields relatively diffuse attention, whereas the Dream stage encourages clearer attention to manipulated objects and a stronger response to the target placement location in Shelf Place (red box). The subsequent Lift stage further strengthens the focus on manipulated objects.

AuthorsOpenReview ↗

Response to the Reviewer g5bK (3/4)

[W2, Q2] Joint Training (continued)

Analysis:

(1) Under the same 60k-step training budget, joint SVD+PPD training with loss weights of (0.25, 0.25) achieves a success rate of 67.3%, close to the 68.0% achieved by PPD-only training. This suggests that the model may not fully benefit from video supervision when both objectives are introduced from the outset. One possible explanation is that introducing depth supervision from the outset in joint training may encourage the model to prioritize overall scene geometry for action prediction without sufficiently focusing on manipulated objects, targets, and their changes, making it difficult to develop a robust understanding of manipulation semantics. In contrast, the Dream stage first develops action-relevant semantic understanding, providing semantic and dynamic context for future-depth prediction. The Lift stage uses the dynamics learned during Dream to understand the resulting spatial relationships, further refining its understanding of manipulation semantics from a spatial perspective.

The attention maps support this interpretation: attention under joint training is relatively diffuse, whereas the Dream stage focuses more clearly on manipulated objects and target locations, and the Lift stage maintains and further strengthens the focus on manipulated objects.

(2) Moreover, joint training can fit actions in scenes seen during training, but it does not fully develop stable representations of manipulation semantics, which limits generalization. Our curriculum learning approach builds on a more stable understanding of manipulation semantics and incorporates spatial relations relevant to manipulation, reducing reliance on scene-specific cues.

The probe results further support this interpretation: compared with joint training, our curriculum achieves lower errors in both the future-dynamics and action probes on unseen trajectories.

[1] Recon: Reducing Conflicting Gradients from the Root for Multi-Task Learning, ICLR 2023.
[2] AdaPolicy: Action-Driven Anticipatory Robot Policies via Next-Chunk Prediction, arXiv 2026.
[3] Vision-Language Models Unlock Task-Centric Latent Actions, arXiv 2026.
[4] Root mean square error (RMSE) or mean absolute error (MAE)? – Arguments against avoiding RMSE in the literature, Geosci. Model Dev 2014.

[Q1] Evaluation on LIBERO/LIBERO-Plus

A: Thank you for the question. As stated in Appendix A.4, following prior work [1–3], most external baseline results on LIBERO and LIBERO-Plus are taken from published papers or official reports. For WoG, which did not report LIBERO results, we perform downstream fine-tuning following its official repository (https://github.com/Selen-Suyue/WoG).

Checkpoint selection. The cited baseline reports do not explicitly specify their checkpoint-selection criteria. For WoG and our method, we consistently evaluate the final training checkpoint on both LIBERO and LIBERO-Plus.

Evaluation counts. On LIBERO, all reported results are based on 50 rollouts per task except the original DreamVLA result, which is based on 20. To eliminate this discrepancy, we re-evaluate its public checkpoint (https://huggingface.co/WenyaoZhang/DreamVLA) with 50 rollouts per task, obtaining 91.7%, compared with the reported 92.6%. On LIBERO-Plus [4], all methods directly evaluate their LIBERO-trained checkpoints on all 10,030 perturbed instances, without additional fine-tuning.

Action chunking. Our method uses 8 steps, while our WoG reproduction uses 16, following its official configuration. The other methods use their published or official settings: π₀: 50; π₀.₅: 10; VLA-JEPA: 7; DreamVLA: 3; Fast-WAM: 32; and Motus: 16. These differences reflect model-specific training and execution configurations, as the appropriate action chunk length varies across models.

MethodSpatialObjectGoalLongAvg. SR (%)
DreamVLA (Originally Reported)97.594.089.589.592.6
DreamVLA (Re-evaluated)95.095.890.285.891.7

[1] VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model, ECCV 2026.
[2] DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge, NeurIPS 2025.
[3] Fast-WAM: Do World Action Models Need Test-time Future Imagination? NeurIPS 2026.
[4] LIBERO-Plus: A Progressive Robustness Benchmark for Visual-Language-Action Models, CVPR 2026.

Reviewer_u75xOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a two-stage training framework to make VLA models 3D grounded. Built on an off-the-shelf InternVL3-2B with a DiT flow-matching head, the model uses learnable "world queries" for learning future-prediction. In the first Dream stage, the world embeddings condition a LoRA-finetuned video diffusion model to predict the future RGB frames. In the next, Lift, stage, a LoRA-adapted monodepth diffusion model predicts the future endpoint depth map. At inference time, both of these models are removed, leading effectively to zero extra inference cost. The method has strong results on multiple datasets: MetaWorld, LIBERO, LIBERO-Plus, and 4 real-world Franka tasks.

Critical strengths
  • The idea to "2D-to-3D imagine" (i.e., decoupling 2D dynamics from 3D depth lifting) how a task would be solved in the environment at hand is interesting and novel to the best of my knowledge.
  • The proposed 3D grounding strategy has zero inference-time overhead while still providing the gains through the learned representations.
  • The experiments are very strong, showing clear and substantial improvements over the state of the art on multiple datasets. I will get back to this in the weaknesses.
  • The out-of-distribution experiments are also strong, demonstrating that the proposed method will likely work out of the box.
Critical weaknesses

In general, I like the paper but there are a couple things to be clarified before making the final decision.

  • It seems like a little of cherry picking is going on (the authors can defend themselves easily by showing actual numbers) in Table 12. The appendix figures (9, 10, ...) are showing significantly more scenarios than what Table 12 has numbers for. It may simply be that the authors chose some so that the table fits into the 2-column template, but I would still like to see all numbers.
  • The proposed method shares the exact VLM + DiT architecture of WoG (if I haven't misunderstood). Still, its Action-only baseline outperforms WoG by 15.3% on MetaWorld and 25.3% on LIBERO-Plus before any world modeling is added. This needs some explanation.
  • Some papers, e.g. [a], show that pi0.5 achieves 85.7% score at LIBERO-Plus, while in the paper's Table 3 it only gets 80.3%. The 85.7% of the same method on the same dataset is actually better than the authors' method with its 83.7%.
  • In the end of the day, "3D Lifting" is simply a 2.5D pseudo-depth self-distillation without actually leveraging 3D geometry, multi-view information, etc. The target predicted depth is scale-normalized and is generated offline by the exact same frozen PPD model. This creates a self-distillation bias favoring PPD over standalone decoders and lacks metric scale. However, this probably is fine given that the inference do not rely on this future depth. However, I would soften the wording around "3D lifting"

[a] Zhang, Z., Li, Z., Rahmati, B., Yang, R.H., Ma, Y., Rasouli, A., Pakdamansavoji, S., Wu, Y., Zhang, L., Cao, T. and Wen, F., 2026. Do world action models generalize better than vlas? a robustness study. arXiv preprint arXiv:2603.22078.

Questions for authors
  • What is the reason for showing different results in Table 12 and Fig.9,10,...?
  • What is the reason for WoG lagging so much behind?
  • What is the reason for pi0.5 having different scores than reported in the literature?
4 comments
AuthorsOpenReview ↗

Response to the Reviewer u75x (1/2)

We thank the reviewer for the constructive feedback on the real-world evaluation, the WoG comparison, the reported π0.5\pi_{0.5} results, and the scope of 3D lifting. Below, we provide the complete real-world OOD results, analyze WoG’s performance on LIBERO and LIBERO-Plus, report updated LIBERO-Plus results for π0.5\pi_{0.5} using the official checkpoint, and clarify the wording on 3D lifting.

[W1, Q1] Complete Evaluation

A: Thank you for the comment. Due to space constraints in the two-column layout, we reported only two tasks per OOD setting, although we evaluated additional tasks. The complete results are provided below.

Each entry reports success rates (%) under the in-distribution (ID) setting → the corresponding out-of-distribution (OOD) setting.

Put Bear in Box

MethodBackgroundLightHeight
π0.5\pi_{0.5}80.0 → 50.080.0 → 45.080.0 → 30.0
DreamVLA75.0 → 55.075.0 → 50.075.0 → 45.0
Dream-Lift-VLA95.0 → 80.095.0 → 90.095.0 → 85.0

Serve Food

MethodBackgroundLightNovel Object
π0.5\pi_{0.5}65.0 → 55.065.0 → 60.065.0 → 25.0
DreamVLA60.0 → 40.060.0 → 50.060.0 → 30.0
Dream-Lift-VLA85.0 → 80.085.0 → 80.085.0 → 65.0

Stack Bowls

MethodBackgroundLightNovel ObjectHeight
π0.5\pi_{0.5}70.0 → 65.070.0 → 60.070.0 → 55.070.0 → 35.0
DreamVLA65.0 → 55.065.0 → 55.065.0 → 50.065.0 → 35.0
Dream-Lift-VLA90.0 → 90.090.0 → 85.090.0 → 85.090.0 → 80.0

Put Hotdog in Drawer

MethodLightHeight
π0.5\pi_{0.5}25.0 → 15.025.0 → 15.0
DreamVLA40.0 → 20.040.0 → 30.0
Dream-Lift-VLA70.0 → 60.070.0 → 65.0

Real-world robot videos are available at https://anonymous.4open.science/r/dream-lift-vla-20374/real-robot-video.mp4.

[W2, Q2] WoG Results

A: Thank you for raising this point. We fine-tuned WoG following the downstream adaptation instructions in its GitHub repository (https://github.com/Selen-Suyue/WoG). Although both methods use a VLM–DiT framework, WoG adopts Prismatic/OpenVLA [1], whereas ours uses InternVL3-2B [2]. Differences in backbone pretraining may contribute to the MetaWorld gap. WoG also underperforms π0.5\pi_{0.5} on MetaWorld (52.7% vs. 71.3%), although π0.5\pi_{0.5} combines a VLM with a flow-matching Transformer action expert without explicit future-world modeling.

For LIBERO-Plus, WoG keeps its OXE-pretrained future-conditioning encoder frozen during LIBERO fine-tuning while continuing to align VLM features with the encoder's outputs. We hypothesize that this alignment may encourage reliance on scene-specific appearance cues. Such reliance may be less problematic in in-distribution LIBERO evaluation but could undermine action robustness under unseen background, lighting, and viewpoint changes in zero-shot LIBERO-Plus evaluation.

To investigate this hypothesis, we fine-tune the official Stage I checkpoint on LIBERO demonstrations without the Stage II future-feature querying and alignment components. LIBERO performance decreases only slightly, whereas LIBERO-Plus success improves from 49.0% to 57.5%, with absolute gains of 20.6%, 13.5%, and 14.9% under camera, lighting, and background variations, respectively. These results are consistent with the hypothesis that alignment may increase sensitivity to appearance changes.

LIBERO

MethodSpatialObjectGoalLongAvg. SR (%)
WoG98.296.896.289.295.1
WoG without Query & Align97.697.095.687.294.4

LIBERO-Plus

MethodCameraRobotLanguageLightBackgroundNoiseLayoutAvg. SR (%)
WoG11.952.676.258.266.317.770.849.0
WoG without Query & Align32.558.976.571.781.220.774.657.5

[1] OpenVLA: An Open-Source Vision-Language-Action Model, CoRL 2024.
[2] https://huggingface.co/OpenGVLab/InternVL3-2B, 2025.

AuthorsOpenReview ↗

Response to the Reviewer u75x (2/2)

[W3, Q3] Reported π0.5\pi_{0.5}

A: Thank you for pointing out this discrepancy. Following common practice in prior work, most of our baseline results on LIBERO and LIBERO-Plus are taken from published papers or official reports. Specifically, our reported π0.5\pi_{0.5} result on LIBERO-Plus was taken from an ACL 2026 paper [1]. Reported scores vary across studies: Being-H0.7 [2] reports 77.4%, while the paper you cited reports 85.7%. To ensure a fair comparison with our method, we re-evaluate π0.5\pi_{0.5} using the official LIBERO checkpoint (https://github.com/Physical-Intelligence/openpi/blob/main/examples/libero/README.md), following the official LIBERO [3] and LIBERO-Plus [4] evaluation protocols also used for our method.

LIBERO

MethodSpatialObjectGoalLongAvg. SR (%)
π0.5\pi_{0.5}, official checkpoint97.698.698.892.897.0
Dream-Lift-VLA99.2100.099.297.699.0

LIBERO-Plus

MethodCameraRobotLanguageLightBackgroundNoiseLayoutAvg. SR (%)
π0.5\pi_{0.5}, official checkpoint68.674.885.195.393.283.985.782.8
Dream-Lift-VLA77.960.887.596.094.592.283.383.7

[1] Stable Language Guidance for Vision–Language–Action Models, ACL 2026.
[2] Being-H0.7: A Latent World-Action Model from Egocentric Videos, NeurIPS 2026.
[3] LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning, NeurIPS 2023.
[4] LIBERO-Plus: A Progressive Robustness Benchmark for Visual-Language-Action Models, CVPR 2026.

[W4] Softened wording

A: We agree with your suggestion. Consistent with your description in Strength 1, we have qualified “3D lifting” as “3D depth lifting” at key locations to clarify its depth-based scope. For example, the Introduction now states “first models future 2D scene evolution and then performs 3D depth lifting of the learned dynamics,” and Section 3.3 is retitled “Stage II: 3D Depth Lifting of Future Dynamics.” We have also softened the corresponding wording in the experimental and conclusion sections, as well as in the Appendix.

We thank the reviewer again for helping us strengthen the experimental evidence and state our claims more precisely, and we will include these results and our discussion in the latest version.

I thank the authors for their thorough rebuttal. It resolved all my concerns. Thus, I increase my rating to Clear Accept.

Thank you for your feedback and for supporting the acceptance of our paper. We sincerely appreciate the constructive comments you provided during the review process, which have helped improve the quality and completeness of our work. Following your suggestion, we will ensure that all details discussed in the rebuttal are properly incorporated into the final version. Thank you again for your time and valuable insights.

Reviewer_Usy9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dream-Lift-VLA, a two-stage training framework for vision-language-action policies. The Dream stage conditions a pretrained video diffusion model (SVD) on VLA world embeddings to generate future videos. The Lift stage starts from the Dream checkpoint and replaces video generation with future-depth generation using a pretrained depth model (PPD). Both generative branches are removed at inference. The method reports 82.0% success on MetaWorld, 99.0% on LIBERO, 83.7% on LIBERO-Plus, and 78.1% in real-world OOD settings.

Critical strengths
  1. The curriculum ablations directly test the main idea. On MetaWorld, video-to-depth training (82.0%) clearly beats joint training (64.7%) and the reverse order (68.7%). Training Stage I longer does not help (72.0% vs. 71.3%), and the gains hold when SVD and PPD are replaced with Wan2.2 and DepthFM.

  2. The gains carry over to LIBERO-Plus (74.3% → 80.0% → 83.7% across stages, on 10,030 test instances) and to real-world OOD settings, where Stage II raises success under height changes from 60.0% to 82.5%. Removing the generative branches at inference keeps the policy efficient.

Critical weaknesses
  1. Alternative explanations for the curriculum effect are not ruled out. There is no action-only → PPD control, so it is unclear whether Stage II needs learned 2D dynamics or simply a trained starting point. The joint variants may also suffer from loss imbalance (if both λ = 0.5, the auxiliary weight is doubled), and it is not stated whether Stage II restarts the learning-rate schedule. Finally, Stage II helps after SVD or Wan2.2, but not after SANA (71.3% → 69.3%) or V-JEPA (71.3% → 67.3%), which suggests compatibility between the two stages matters as much as their order.

  2. The "3D" supervision is pseudo-depth from future RGB frames. The depth targets are produced by the frozen PPD from the future frame and normalized, so they are relative, single-view depth derived from an image Stage I already supervises. The gain may come from the PPD / Depth Anything V2 prior rather than from geometry, which weakens the "2D-to-3D lifting" framing.

  3. The paper offers little analysis of where the gains come from. Tables 4–5 report only averages, with no per-task results or failure-mode analysis. The paper also gives no quantitative quality metrics for the generated videos or depth maps.

Questions for authors
  1. What loss weights, step counts, and learning-rate schedules are used for each variant in Tables 4–5? Does the sequential advantage remain with a rebalanced joint baseline?
  2. What happens if Stage II with PPD starts from the action-only checkpoint?
  3. How does performance change with ground-truth simulator depth or RealSense sensor depth instead of PPD pseudo-labels?
  4. Could the authors give per-task results for Tables 4–5 and describe which failure modes the curriculum fixes?
4 comments
AuthorsOpenReview ↗

Response to the Reviewer Usy9 (1/3)

We thank the reviewer for the constructive questions on our training strategy, depth supervision, and experimental analysis. We address these points by clarifying training settings and adding curriculum ablations, presenting existing decoder results and new depth-source comparisons, and providing task-level results, failure-mode analyses, and future-prediction metrics.

[W1, Q1, Q2] Additional Details and Results

A: We thank the reviewer for highlighting possible alternative explanations for the curriculum’s benefits. We address these concerns by testing action-only initialization, clarifying training settings and evaluating loss reweighting, and examining cross-stage supervision compatibility.

(1) Action-only initialization. We additionally evaluate an action-only→PPD baseline on MetaWorld under the same two-stage training budget (40k + 20k steps). It achieves 71.3% success, compared with 82.0% for Dream→Lift. This suggests that the benefit of Stage II stems not merely from a pretrained policy initialization but also from the future 2D dynamics learned in the Dream Stage.

Training StrategyAvg. SR (%)
Direct PPD (60k)68.0
Action-only→PPD (40k + 20k)71.3
Dream→Lift (40k + 20k)82.0

(2) Weights, steps, and schedules. In Tables 4–5, the action loss has a weight of 1, and each active auxiliary loss has a weight of 0.5. Two-stage variants are trained for 40k steps in Stage I and 20k steps in Stage II, while single-stage variants are trained for 60k steps. We use AdamW with a constant learning rate of 1×10−51\times10^{-5} on MetaWorld and a cosine schedule on LIBERO. On both benchmarks, Stage II inherits only the Stage I VLA weights and starts with a fresh AdamW optimizer and learning-rate schedule using the same configuration as Stage I. Furthermore, we conduct three additional joint-training experiments with (λ2D,λ3D)=(0.25,0.25)(\lambda_{\mathrm{2D}},\lambda_{\mathrm{3D}})=(0.25,0.25), (0.1,0.4)(0.1,0.4), and (0.4,0.1)(0.4,0.1), keeping the total auxiliary-loss weight at 0.5. Under the same 60k-step budget, these variants achieve success rates of 67.3%, 66.0%, and 66.7%, respectively, all below the action-only baseline of 68.0%. Thus, rebalancing the loss weights does not close the gap to our progressive training curriculum.

(λ2D,λ3D)(\lambda_{\mathrm{2D}},\lambda_{\mathrm{3D}})Total Auxiliary WeightAvg. SR (%)
Action-only baseline068.0
(0.5,0.5)(0.5,0.5), original setting1.065.3
(0.25,0.25)(0.25,0.25)0.567.3
(0.1,0.4)(0.1,0.4)0.566.0
(0.4,0.1)(0.4,0.1)0.566.7

(3) Compatibility. We agree that cross-stage supervision compatibility matters. We do not claim that arbitrary sequential combinations are beneficial. As discussed in the original Appendix B.3, Table 16, for the V-JEPA-based Stage I variant, using compatible depth-feature distillation from Depth Anything V2 instead of PPD generation in Stage II improves success from 71.3% to 76.7% (consecutive frames) or 77.3% (endpoint). For SANA, the original Appendix B.2, Table 15 suggests that endpoint-only supervision provides limited temporal context for subsequent future-depth learning. We therefore extend SANA to predict four consecutive future frames. As shown below, Stage II then improves performance from 72.0% to 74.7%, supporting the role of continuous-dynamics supervision in making subsequent 3D lifting effective.

Stage I ObjectiveStage II ObjectiveAvg. SR (%)
SANA (It+HI_{t+H})–71.3
SANA (It+HI_{t+H})PPD (Dt+HD_{t+H})69.3
SANA (It+1:t+H\mathbf{I}_{t+1:t+H})–72.0
SANA (It+1:t+H\mathbf{I}_{t+1:t+H})PPD (Dt+HD_{t+H})74.7
AuthorsOpenReview ↗

Response to the Reviewer Usy9 (2/3)

[W2, Q3] Depth Target

A: Thank you for the question. Our results suggest that PPD-generated depth maps provide sufficient geometric detail for effective policy supervision.

First, using these labels, a decoder trained from scratch improves success from 72.0% to 77.3% with single-frame supervision and 79.3% with multi-frame supervision (Appendix B.1, Table 14). Using PPD as the predictor further raises success to 82.0%. These results suggest that future geometric supervision provides substantial benefits even without a pretrained depth predictor, while PPD offers additional gains.

Second, replacing pseudo-depth with GT depth on MetaWorld preserves 82.0% success, while using RealSense depth in real-world experiments reduces average success from 85.0% to 83.8%. We observe substantial noise around object boundaries in RealSense depth, which may compromise supervision quality.

SettingDepth SupervisionAvg. SR (%)
MetaWorldPPD pseudo-depth82.0
MetaWorldGT depth82.0
Real-worldPPD pseudo-depth85.0
Real-worldRealSense depth83.8

[W3, Q4] Detailed Results and Analysis

A: We thank the reviewer for suggesting more detailed results and analysis. We address this through task-level results, failure-mode analysis, and quantitative evaluation of future-video and depth predictions.

(1) Task-level results. The following tables report success rates for each of the six MetaWorld tasks, following the original order of Tables 4–5.

ModelPick PlaceDisassembleStick PullAssemblyShelf PlaceHand InsertAvg. SR (%)
Action-only (60k)68.070.052.092.056.070.068.0
Stage I (40k)68.092.052.096.068.056.072.0
Stage I (60k)68.088.056.092.068.056.071.3
Stage I + Stage II (60k)72.092.088.0100.068.072.082.0
Training CurriculumPick PlaceDisassembleStick PullAssemblyShelf PlaceHand InsertAvg. SR (%)
Direct 3D dreaming72.072.080.092.036.064.069.3
Direct 3D dreaming (60k)68.072.084.088.036.060.068.0
2D and 3D dreaming72.060.088.072.016.080.064.7
2D and 3D dreaming (60k)76.056.088.072.020.080.065.3
3D to 2D dreaming72.068.084.096.028.064.068.7
3D to 2D and 3D dreaming64.072.084.096.032.068.069.3
2D to 2D and 3D dreaming64.084.084.092.036.072.072.0
2D to 3D dreaming72.092.088.0100.068.072.082.0
AuthorsOpenReview ↗

Response to the Reviewer Usy9 (3/3)

[W3, Q4] Detailed Results and Analysis (continued)

(2) Failure-mode analysis.
We select the Disassemble task to analyze the action-only baseline’s failure modes and how the Dream stage addresses them. On this task, the Dream stage improves the success rate from the baseline’s 70.0% to 92.0%. This task requires grasping the object before lifting it off the peg. In the illustrative comparison in Fig. A (https://anonymous.4open.science/r/dream-lift-vla-20374/figures/failure-mode-analysis-1.png), the baseline-labeled trajectory lifts without establishing a secure grasp, leaving the object on the peg, whereas the Stage I-labeled trajectory successfully removes it. This illustrates Dream’s motivation: future-video prediction encourages the VLA to learn how actions change scene states, rather than merely reproduce gripper movements.

We select the Stick Pull task to analyze the Dream stage’s failure modes and how the subsequent Lift stage addresses them. On this task, the Lift stage improves the success rate from the Dream stage’s 52.0% to 88.0%. In the illustrative comparison in Fig. B (https://anonymous.4open.science/r/dream-lift-vla-20374/figures/failure-mode-analysis-2.png), both trajectories grasp and engage the tool, but the Stage I-labeled trajectory pulls insufficiently: the minimum handle-to-target distance remains 0.1276 m, above the 0.12 m success threshold. The Stage II-labeled trajectory maintains engagement and reaches 0.1174 m, completing the task. This contrast highlights post-grasp spatial precision, consistent with Lift’s use of future-depth supervision to ground learned dynamics in spatial geometry.

(3) Future-prediction quality.
To quantify the quality of future video and depth predictions, we randomly sample 100 observations from six MetaWorld tasks, predicting four consecutive future RGB frames and the endpoint depth for each observation. PSNR and SSIM measure pixel-level fidelity and structural similarity, respectively, with higher values indicating better quality [1]. LPIPS measures perceptual differences, with lower values being better [2]. Depth predictions are aligned to simulator-provided metric ground truth using per-image scale and shift before evaluation [3]. AbsRel and RMSE measure mean absolute relative error and root-mean-square error, respectively, with lower values being better. δ₁ measures the percentage of pixels with accurate depth predictions, with higher values being better [4].

All metrics are averaged over the 100 observations, with video metrics first averaged across the four predicted frames for each observation. These checkpoint-wise results indicate improving future-video fidelity and aligned depth accuracy during training.

Stage I: Video Generation (40k)

StepPSNR ↑ (dB)SSIM ↑LPIPS ↓
1k22.40.7200.120
2k24.00.7700.095
10k25.50.8200.073
40k26.6220.84290.0612

Stage II: Depth Generation (20k)

StepAbsRel ↓RMSE ↓ (m)δ₁ ↑ (%)
1k0.1050.12888.5
2k0.0830.10291.9
10k0.0620.07794.8
20k0.05720.071495.68

[1] Image Quality Assessment: From Error Visibility to Structural Similarity, TIP 2004.
[2] The Unreasonable Effectiveness of Deep Features as a Perceptual Metric, CVPR 2018.
[3] Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation, CVPR 2024.
[4] Depth Map Prediction from a Single Image using a Multi-Scale Deep Network, NeurIPS 2014.

We thank the reviewer again for the constructive suggestions, which help us strengthen the experimental evidence, and we will include these results and our discussion in the latest version.

AuthorsOpenReview ↗

With Appreciation for Your Feedback and Consideration #20374.

Dear Reviewer Usy9,

Thank you again for your thoughtful and constructive feedback. In our rebuttal, we have clarified the training settings, added ablations on policy initialization and loss weighting, examined cross-stage supervision compatibility, compared depth-supervision sources, and provided task-level results, failure-mode analyses, and quantitative evaluations of future video and depth predictions.

We hope that our response helps address your concerns, and we would sincerely appreciate any further feedback.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Appreciation for Your Efforts on Submission #20374

Dear Reviewers,

As the author discussion period draws to a close today, we would be grateful if you could review our rebuttal at your convenience.

Regardless, we sincerely appreciate the time and effort you have devoted to Submission #20374, as well as your service to the robotics community.

The Authors

AuthorsOpenReview ↗

Summary of Discussions

Dear Area Chair,

We sincerely appreciate the valuable time and effort of the AC and the reviewers. We have provided detailed, point-by-point responses to every weakness and question.

Per-reviewer Outcomes:

We appreciate that Reviewer #u75x described our core idea as “interesting and novel” and the experiments as “very strong.” Reviewer #Usy9 noted that the curriculum ablations “directly test the main idea.” Reviewer #g5bK emphasized that “the ordering itself matters” and described the evaluation as “unusually broad.” Reviewer #2Dwb recognized the “competitive performance and consistent overall improvements” and “extensive” ablation studies.

The reviewers also requested further clarification and additional experiments, which we have addressed as summarized below.

Reviewer #u75x:

The reviewer focused on evaluation completeness, baseline discrepancies, and depth-lifting terminology. We provided complete OOD results, baseline analyses and re-evaluations, and more precise wording. The reviewer confirmed that the rebuttal “resolved all my concerns” and raised their rating to 4 (Clear Accept).

Reviewer #g5bK:

The reviewer focused on depth supervision, joint training, and evaluation protocols. We provided depth-source comparisons, optimization and representation analyses, clarified evaluation settings, and repeated real-world training and evaluation. The reviewer confirmed that these experiments “resolve the issues I raised” and raised their rating to 3 (Weak Accept).

Reviewer #Usy9:

The reviewer focused on curriculum controls, depth supervision, and detailed results. We provided targeted ablations, training clarifications, cross-stage compatibility experiments, depth-source comparisons, and task-level analyses.

Reviewer #2Dwb:

The reviewer focused on the curriculum’s rationale, methodological contributions, and generalizability. We provided representation analyses, controlled DreamVLA experiments, loss-weight sensitivity tests, and extended training-order ablations, alongside clarifications of our conceptual and architectural contributions.

Current status: We are grateful that Reviewers #u75x and #g5bK took the time to consider our rebuttal and raised their ratings to 4 (Clear Accept) and 3 (Weak Accept), respectively. We sincerely appreciate their support for the acceptance of our paper. We sincerely look forward to hearing from Reviewers #Usy9 and #2Dwb when they have an opportunity to review our rebuttal.

We hope this summary demonstrates our efforts to thoroughly address all reviewer feedback. We sincerely thank the AC and all reviewers for helping improve our work and will incorporate these results and clarifications into the final version.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.