acceptodds
Under review as a conference paper at ICLR 2027

PhysForce: Physics-aware Force Control for Interactive Video World Model

Abstract

Interactive video world models aim to predict how visual environments evolve in response to user interventions, including physically consistent state changes induced by applied forces. Existing methods often represent forces as motion trajectories, which may introduce inappropriate motion priors, causing the model to over-rely on encoded displacement patterns. More fundamentally, the model's internal representations do not reliably localize force-induced state changes, making it difficult to associate applied forces with their resulting physical responses. Also, limited force-annotated data make it difficult to produce physically consistent and temporally stable predictions in unseen scenarios. To address these limitations, we propose PhysForce, a two-stage framework that integrates supervised fine-tuning and reinforcement learning through three complementary components. During supervised fine-tuning, (1) the Latent Force Intervention Encoder maps force location, direction, and magnitude into the video latent space without exposing future motion; and (2) the Physics Attribution Network associates force interventions with localized changes in the latent state, strengthening the learning of physical responses to applied forces. During reinforcement learning, (3) Physics-Aware Reinforcement Learning uses self-supervised rewards on out-of-distribution videos to improve physical consistency and generalization. Extensive experiments show that PhysForce improves force responsiveness, physical consistency, and generalization, demonstrating its effectiveness for interactive video world models.

Primary area:
generative models
Keywords:
Interactive Video World Model, Force-Conditioned Video Generation, Physical Consistency

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_SRP8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces PhysForce, a two-stage, physics-aware force control framework designed to address key challenges in interactive video world models, specifically, the leakage of future motion information caused by trajectory-based force representations, the difficulty in localizing internal representations, and the scarcity of force-annotated data. Stage 1 (SFT) employs a Latent Force Intervention Encoder (LFIE) to encode forces while preventing trajectory leakage, alongside a Physics Attribution Network (PAN) that localizes physical responses through temporal differences. Stage 2 (RL) utilizes Physics-Aware Reinforcement Learning (PARL), which incorporates self-supervised rewards derived from V-JEPA and a "Dynamic Gate" mechanism applied to videos lacking explicit force annotations, thereby enhancing physical consistency and out-of-distribution (OOD) generalization.

Critical strengths

**Comprehensive Evaluation: ** The work demonstrates exceptional empirical rigor through extensive evaluations across 14 metrics on both ID and OOD benchmarks, thorough component-wise ablations, a large-scale user study with 150 participants, and rich qualitative visualizations.

Critical weaknesses
  1. The framework lacks fine-grained temporal control over forces. By simply broadcasting force conditions across time, localized interactions are restricted to initial impulses, and the offline generation pipeline precludes real-time or dynamic adjustments.

  2. The V-JEPA prediction error used in PARL is a representation-based empirical prior rather than an analytical physics solver. While effective as a statistical proxy for consistency, it remains a heuristic that cannot guarantee strict physical plausibility under complex dynamics.

Questions for authors
  1. Since LFIE and PAN operate entirely within 2D image coordinates and video latent spaces without explicit 3D geometry or depth modeling, how does the model handle complex 3D interactions involving severe occlusions, depth-dependent collisions, or out-of-plane forces?

  2. Since the V-JEPA prediction error functions as a representation-based empirical proxy rather than an analytical solver, did the authors observe failure modes where the reward favored visually smooth but physically incorrect behaviors (such as violations of momentum conservation in multi-body collisions)?

  3. While the paper achieves strong empirical performance through a complex multi-component architecture and two-stage training pipeline, the underlying conceptual takeaways could be highlighted more clearly. Could the authors provide a more high-level abstraction or conceptual insight that would help the community generalize these lessons beyond this specific, complex setup?

I am open to raising my rating once my concerns are addressed.

6 comments
AuthorsOpenReview ↗

Feedback to Reviewer SRP8 (1/2)

Thank you for your constructive feedback and valuable suggestions. We have conducted the requested experiments and incorporated the corresponding results and discussions into the revised manuscript. All revisions are highlighted in blue\color{blue}{blue} for ease of reference.

W1. Temporal Control and Streaming Generation

Thank you for suggesting that we extend PhysForce to support interactive force updates, we conduct an experiment to expand our PhysForce on Casual Forcing in Fig.11 in the revised paper. We trained a causal autoregressive variant with casual forcing and applied distribution matching distillation (DMD) to reduce sampling cost. Fig.11(a) shows an example of streaming generation. As shown in Fig. 11(b), the resulting model achieves 16.37 FPS after the first frame, compared with 0.34 FPS for the original model (48.15× speedup), while reducing end-to-end first-frame latency from 236.45 s to 2.17 s (108.96× reduction).

MethodFPS ↑Latency ↓
PhysForce0.34236.45
PhysForce + Casual Forcing16.372.17

Q1. Extending 2D Force Conditioning to 3D Interactions

The current model uses 2D image-space force localization and video latent response modeling. Nevertheless, both LFIE and PAN can be extended to incorporate 3D information. During the rebuutal period, we include a preliminary 3D extension in Fig.10:

  • LFIE: To represent force locations, we extend the original 2D force mask into a 3D volumetric cube, whose center encodes the 3D force application point. To capture force directions, we introduce an additional depth channel and employ an adapter to project the resulting force representation into the same latent dimensionality as the video. This design enables the unified modeling of both 3D point forces and wind forces.
  • PAN: In addition to temporal response supervision, we introduce a depth map alignment loss to encourage the model to capture depth-dependent geometric changes.

With these extensions, PhysForce can generate force-aligned responses to both 3D point forces and 3D wind forces in Fig.10, providing preliminary evidence of its potential for 3D force control.

AuthorsOpenReview ↗

Feedback to Reviewer SRP8 (2/2)

W2 & Q2. Scope and Limitations of the V-JEPA Reward

We thank the reviewer for raising this important concern. We agree that V-JEPA prediction error serves as a representation-level prior rather than an analytical physics solver, and cannot guarantee strict physical plausibility in all complex scenarios.

In PhysForce, Stage 1 first learns force-induced motion and object interactions through supervised fine-tuning on diverse physics-engine-based scenarios from GoalForce and Force Prompting Dataset. In Stage 2, we intentionally introduce a frozen V-JEPA model without force conditioning. This allows RL to leverage V-JEPA's ability to predict future states across diverse real-world scenarios, encouraging the generator to learn more general world dynamics beyond specific force-controlled settings and thus improving generalization. The reward of V-JEPA is also supported by prior work [1], which demonstrates the effectiveness of latent world-model prediction errors in improving physical plausibility, which we have discussed in Appendix A.3.

As shown in Appendix Fig. 17, we observe occasional physical inconsistencies, such as premature energy transfer between objects before contact, which we discuss in the Limitations. Nevertheless, as shown in Fig. 9, V-JEPA guided PARL leverages general future-state prediction priors to improve the learning of world dynamics beyond force-labeled scenarios, reducing unnatural object deformations and jerky motions.

Our qualitative results show reduced abrupt direction changes, interrupted motion, and structural collapse. Quantitatively, PARL improves OOD DCS from 0.625 to 0.782 and ID ST-IoU from 0.275 to 0.332.

[1] Yuan J, Zhang X, Friedrich F, et al. Inference-time physics alignment of video generative models with latent world models[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2026.

Q3. Design Principles and Broader Applicability of PhysForce

  1. High-level conceptual takeaway. The central contribution of PhysForce is a unified intervention–response generalization framework that progressively connects physical intervention, response modeling, and dynamics generalization. Specifically, LFIE defines outcome-independent force conditions without leaking future trajectories; PAN learns localized state transitions, providing response supervision for SFT and spatially grounded rewards for RL; and PARL leverages these rewards to explore beyond force-labeled training data while filtering physically inconsistent outcomes. Thus, the three components form a tightly coupled framework rather than an independent combination of modules, enabling the model to learn not only how forces are applied, but also how objects respond and how these dynamics generalize to unseen scenarios. As a result, PhysForce achieves 36.0% and 61.0% improvements in DCS over existing SOTAs on OOD and ID benchmarks, respectively.

  2. Broader implications for other tasks. (1) This intervention–response paradigm can extend to robotic manipulation and action-conditioned video generation, where task-specific actions serve as intervention inputs and resulting object-state changes provide response supervision. (2) Furthermore, RL post-training can leverage response-aware self-supervised rewards on unlabeled interaction videos to refine learned dynamics beyond action-labeled training data, improving generalization without requiring additional annotated demonstrations.

Dear Reviewer SRP8,

Thank you again for your thoughtful feedback and constructive suggestions. Your comments motivated our additional explorations of streaming generation and 3D interactions, as well as a deeper discussion of the design of our physics-aware reward. We truly appreciate your engagement with our work and would welcome any further thoughts on our response.

Best regards,

The Authors

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

The paper presents PhysForce, a two-stage framework for physics-aware force control in interactive video world models. In the supervised stage, a Latent Force Intervention Encoder (LFIE) injects outcome-agnostic force signals without trajectory leakage, while a Physics Attribution Network (PAN) grounds interventions into localized latent state changes via temporal difference alignment. In the second stage, Physics-Aware Reinforcement Learning (PARL) optimizes the model on force-free videos using a PAN-weighted V-JEPA predictive reward coupled with a heuristic Dynamic Gate to mitigate reward hacking. Experiments on synthetic benchmarks and real-world videos demonstrate improved force controllability and physical consistency.

Critical strengths
  1. Formulating force control as an outcome-agnostic latent intervention effectively eliminates future trajectory leakage inherent in Gaussian-trajectory representations.
  2. The carrier-attribution design linking latent temporal differences to force conditions provides an intuitive mechanism for unsupervised response localization.
  3. Extensive ablations on both Stage 1 and Stage 2 components systematically validate the contribution of individual modules.
Critical weaknesses
  1. The contribution appears primarily incremental. LFIE adapts force-field conditioning, PAN adds learned temporal-response supervision, and PARL combines existing diffusion RL and V-JEPA rewards with regional weighting and filtering. These are reasonable design choices, but the paper does not clearly establish a substantial methodological advance beyond their integration.
  2. PAN is supervised to reconstruct temporal latent differences, which does not necessarily isolate force-induced changes. Quantitative evidence is needed to distinguish its attribution capability from ordinary motion localization.
  3. PhysForce uses an extra RL stage, while the baselines receive matched supervised training. Compute-matched post-training controls and comparisons with relevant recent approaches would strengthen the claimed advantage.
Questions for authors

Please see Weaknesses.

7 comments
AuthorsOpenReview ↗

Feedback to Reviewer XGFS (1/2)

Thank you for recognizing the rationale behind LFIE, PAN, and PARL, and for highlighting the importance of clarifying their complementary roles. We also appreciate your suggestions on strengthening PAN's attribution evidence and ensuring fair post-training comparisons. We have revised the manuscript accordingly, adding further experiments and clarifying the contributions of each component. All changes are highlighted in blue\color{blue}{blue} in the revised manuscript.

W1. Core Novelty of PhysForce

Our central contribution is a unified intervention–response generalization framework for force-driven video generation, in which the three components are tightly connected rather than independently designed.

The key challenge in force-driven video generation is enabling models to accurately understand and respond to forces while generalizing to diverse real-world scenarios. Existing methods often represent forces as moving Gaussian blobs, potentially leaking future-state information, while failing to explicitly localize force-induced state changes, resulting in inaccurate physical responses. Moreover, relying solely on force-labeled simulated data limits generalization to diverse real-world scenarios.

To address these challenges, PhysForce establishes a unified framework that progressively connects physical intervention, response modeling, and dynamics generalization. Specifically, LFIE defines force interventions, providing reliable intervention conditions without prescribing future states. Building upon these conditions, PAN explicitly learns local state transitions, providing targeted supervision for force-induced responses during SFT. Crucially, PAN bridges SFT and RL by using its learned response attribution to construct localized rewards. Based on this connection, PARL leverages PAN-guided rewards to explore additional rollouts beyond the force-labeled training distribution, while filtering physically inconsistent outcomes.

The effectiveness of PhysForce therefore comes not simply from combining three individual modules, but from their progressive and interdependent design. Compared with existing SOTAs, PhysForce achieves substantial DCS improvements of 36.0% on OOD and 61.0% on ID.

W2. Force-Induced Response Attribution in PAN

To clarify how PAN distinguishes force-induced attribution from ordinary motion localization, we highlight both qualitative and quantitative evidence as follows:

  1. In Appendix Fig. 12, we ablate force conditioning in PAN during SFT by removing the force embedding while keeping other inputs fixed. With force conditioning, PAN produces more accurate and localized attribution maps of force-induced state changes, whereas removing it leads to broader and less focused responses. This demonstrates that force information is essential for PAN to accurately localize intervention-induced responses.
  2. In Table 3 of the revised manuscript, we evaluate PAN's ability to localize force-affected regions by comparing it with conventional motion-localization methods, including Optical Flow and SAM. Using IoU, Dice, and normalized centroid error to measure alignment with ground-truth force regions, PAN achieves the best results across all three metrics. These results demonstrate that PAN more accurately localizes force-induced responses than generic motion or segmentation cues.
MethodIoU ↑Dice ↑Centroid Error ↓
Optical Flow0.03890.06580.1154
SAM0.24880.34980.0587
PAN (Ours)0.39560.47130.0295
AuthorsOpenReview ↗

Feedback to Reviewer XGFS (2/2)

W3. Additional Comparisons with RL Baselines

  1. For a fair comparison, we report PhysForce's Stage 1 results (before RL training) on both OOD and ID benchmarks in Tables 1 and 2, respectively. PhysForce achieves consistently better overall performance across mechanical metrics, video quality, and physical consistency. Notably, all methods share the same Wan2.2-5B backbone and training data, ensuring that the performance gains stem from our method design rather than differences in the backbone or data.

Table 1. OOD Evaluation

MethodDCS ↑DA@30° ↑DA@45° ↑MRR ↑RMM ↑CLIP-T ↑PickScore ↑Motion Smoothness ↑Aesthetic Quality ↑Imaging Quality ↑
Wan-Move0.3330.3920.4990.2100.2060.29519.9420.99747.41362.237
Force Prompting0.5750.7140.7480.8101.4530.30620.7970.99647.56862.238
GoalForce0.3920.4900.5600.3890.4690.30120.0460.99647.49862.059
PhysForce Stage1 (Ours)0.6250.7170.7870.9082.0630.30920.8570.99547.50362.406

Table 2. ID Evaluation

MethodDCS ↑DA@30° ↑DA@45° ↑MRR ↑RMM ↑Spatial IoU ↑ST-IoU ↑Weighted IoU ↑MSE ↓
Wan-Move0.4640.4880.5500.4500.3920.3730.1200.2520.005
Force Prompting0.4330.5630.6000.6630.8560.6570.2550.5220.004
GoalForce0.4330.5500.5750.7250.8730.5250.1860.4040.006
PhysForce Stage1 (Ours)0.6520.6380.7380.7581.3080.6540.2750.5380.004
  1. To further evaluate the generalizability of our Stage 2 RL framework, we apply it to the Force Prompting baseline using the same training data as PhysForce. The results on OOD and ID benchmarks are reported in Appendix Table 5. RL consistently improves Force Prompting across most metrics, demonstrating that our RL framework can be effectively generalized to other force-conditioned generation methods.

Table 3. OOD Evaluation

MethodDCS ↑DA@30° ↑DA@45° ↑MRR ↑RMM ↑CLIP-T ↑PickScore ↑Motion Smoothness ↑Aesthetic Quality ↑Imaging Quality ↑
Force Prompting0.5750.7140.7480.8101.4530.30620.7970.99647.56862.238
Force Prompting w/ RL0.7030.8050.8310.8871.6920.31020.9620.99748.24163.312

Table 4. ID Evaluation

MethodDCS ↑DA@30° ↑DA@45° ↑MRR ↑RMM ↑Spatial IoU ↑ST-IoU ↑Weighted IoU ↑MSE ↓
Force Prompting0.4330.5630.6000.6630.8560.6570.2550.5220.004
Force Prompting w/ RL0.5290.6270.6680.7011.0920.7240.2760.5860.004

Dear Reviewer XGFS,

Thank you again for your insightful feedback and constructive suggestions, which prompted us to further strengthen the PAN attribution analysis, extend our RL baseline comparisons, and better articulate the core contributions of our work. At your convenience, we would greatly appreciate any feedback on our response or remaining concerns. Your careful review and constructive guidance have been genuinely valuable to us.

Best regards,

The Authors

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

PhysForce integrates latent-space force encoding, a network designed to model temporal responses, and a physics-aware reinforcement learning mechanism. Built upon the Wan2.2-5B model, this approach enhances responsiveness to directional force and improves alignment with reference motions. The method has been validated in diverse video generation scenarios, where interventions are applied within the image plane.

Critical strengths

The paper shows that adding LFIE helps, but does not sufficiently examine its internal design or compare alternative ways of injecting force conditions. It remains unclear which choices—the four-channel representation, two branches, zero initialization, or channel concatenation—drive the improvement, and whether a simpler force-conditioning interface would work equally well.

Critical weaknesses
  1. The paper shows that LFIE helps, but does not examine its internal design or compare alternative ways of injecting force conditions.
  2. The paper does not show how the generated responses change, especially when constraints mean that a stronger force need not produce proportionally larger motion.
  3. It lacks a quantitative evaluation that fixes the scene, application point, and direction while varying only force magnitude.
Questions for authors
  1. Why is Directional Force Embedding encoded by VAE? How does LFIE compare with a simpler force-conditioning interface?
  2. Please include constrained cases where stronger forces do not necessarily lead to proportionally larger motion.
  3. Can you report response curves across several force magnitudes, with the scene, application point, and direction fixed?
7 comments
AuthorsOpenReview ↗

Feedback to Reviewer hgjG (1/2)

Thank you for recognizing the motivation and presentation of our method, as well as the benefits of LFIE. Your questions help clarify its design and capabilities. We address each point below with additional implementation details and experiments in the revised manuscript. All changes are highlighted in blue\color{blue}{blue} in the revised manuscript.

W1 & Q1. LFIE Design and Alternative Force-Conditioning Interfaces

We have added further quantitative ablation results for LFIE in Appendix Table 4 of the revised manuscript.

  1. The effect of VAE. To avoid trajectory leakage, LFIE uses directional force maps rather than target trajectories. Since force maps are defined on the video frame grid and broadcast along the temporal dimension, we encode them using the pretrained video VAE to obtain latent representations compatible with video latents. This allows the pretrained generator to process force and visual conditions within a shared latent space. The w/o VAE ablation validates this design: LFIE improves all five force-alignment metrics on both OOD and ID benchmarks, with DCS increasing from 0.612 to 0.625 and from 0.549 to 0.652, respectively.

  2. Comparison with other force-conditioning interfaces. We compare LFIE with direct vector addition (Vector), where an MLP-encoded force vector is added to video latents, and cross-attention-based injection (Cross-attn). As shown in Tables 1 and 2, LFIE consistently outperforms both alternatives across all five force-alignment metrics on OOD and ID benchmarks. For example, DCS improves from 0.608 and 0.584 to 0.625 on OOD, and from 0.481 and 0.549 to 0.652 on ID, demonstrating the effectiveness of our conditioning design.

Table 1. OOD Evaluation

MethodDCS ↑DA@30° ↑DA@45° ↑MRR ↑RMM ↑CLIP-T ↑PickScore ↑Motion Smoothness ↑Aesthetic Quality ↑Imaging Quality ↑
Vector0.6080.7050.7840.8571.9360.30720.8050.99647.62562.108
Cross-attn0.5840.5460.6810.8771.9340.30220.8480.99747.40662.227
w/o. VAE0.6120.7100.7230.8941.9960.30520.8510.99547.51762.391
LFIE(Ours)0.6250.7170.7870.9082.0630.30920.8570.99547.50362.406

Table 2. ID Evaluation

MethodDCS ↑DA@30° ↑DA@45° ↑MRR ↑RMM ↑Spatial IoU ↑ST-IoU ↑Weighted IoU ↑MSE ↓
Vector0.4810.5370.6250.7381.1300.6420.2810.5260.005
Cross-attn0.5490.5370.6250.7261.2960.5750.2400.4600.005
w/o VAE0.5490.5500.6330.7321.3010.6080.2530.4940.005
LFIE (Ours)0.6520.6380.7380.7581.3080.6540.2750.5380.004
AuthorsOpenReview ↗

Feedback to Reviewer hgjG (2/2)

W2 & Q2. Force Responses under Scene Constraints

We add Fig. 7 to the revised manuscript, showing the generated videos in Fig. 7(a) and their corresponding PAN responses in Fig. 7(b). With the scene, application point, and force direction fixed, weaker forces (S = 0.1, 0.5) result in limited displacement without reaching the barrier, whereas stronger forces (S = 0.8, 1.0) drive the ball toward the barrier, where it decelerates and eventually settles. Thus, increasing force magnitude affects both motion intensity and contact timing: S = 1.0 reaches the barrier earlier than S = 0.8, but the scene constraint limits further displacement, resulting in similar final positions rather than proportionally larger motion.

PAN captures this temporal difference. At the final timestep, the S = 1.0 sequence exhibits a weaker response due to its earlier settling, whereas the S = 0.8 sequence retains a stronger response because the ball arrives later. Despite their similar final RGB appearances, PAN features distinguish ongoing motion from a settled state, demonstrating the interplay between force magnitude and scene constraints.

W3 & Q3. Quantitative Evaluation of Force Magnitude Sensitivity

We add Fig. 8 to the revised manuscript to address this issue. Fig. 8(a) qualitatively compares the responses under different force magnitudes, while Fig. 8(b) provides the requested quantitative evaluation. Specifically, we vary only S from 0.1 to 1.0, keeping the scene, prompt, application point, direction, and random seeds fixed. Averaged over two scenes and five seeds per scene, RMM increases from 2.068 ± 0.961 to 5.868 ± 1.628 px/frame, with a gradual saturation at higher force magnitudes. These results demonstrate continuous force-magnitude sensitivity while reflecting the limiting effect of physical constraints on displacement under stronger forces.

Force Magnitude S0.10.20.30.40.50.60.70.80.91.0
RMM (Mean ± SD, px/frame)2.068 ± 0.9612.981 ± 0.9383.723 ± 0.9984.236 ± 0.7304.734 ± 0.9055.086 ± 1.1655.613 ± 1.5595.744 ± 1.5935.833 ± 1.5985.868 ± 1.628

Dear Reviewer hgjG,

Thank you again for your encouraging assessment and thoughtful suggestions, which helped us strengthen the LFIE ablations and force response analyses. At your convenience, we would be grateful to hear whether these additions address your questions. We sincerely appreciate your support and valuable guidance.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

To all Reviewers

First and foremost, we sincerely thank all Reviewers, Area Chairs, and Program Chairs for their time and thoughtful feedback. We have provided detailed point-by-point responses to all comments from (Reviewer hgjG, Reviewer XGFS, and Reviewer SRP8).

We are encouraged by the reviewers' positive recognition of our work:

  • Methodology: A well-motivated framework with technically sound designs: LFIE eliminates future-trajectory leakage in force conditioning, PAN explicitly localizes force-induced state changes, and PARL improves physical consistency and generalization beyond force-annotated data.
  • Evaluation: Comprehensive experiments and ablations demonstrate strong force controllability, physical consistency, and OOD generalization, supported by extensive quantitative evaluations.
  • Presentation: Clear motivation, coherent methodological design, and well-organized presentation.

In response to the reviewers' valuable suggestions, we have conducted additional experiments and clarified our methodology and contributions. All corresponding revisions are highlighted in blue\color{blue}{blue} in the revised PDF. The major additions are summarized below:

  • Appendix Table 4: Additional LFIE ablations comparing direct vector addition, cross-attention, and VAE-free encoding, addressing Reviewer hgjG's questions about force-conditioning design.
  • Figures 7–8: Constrained scene visualizations and quantitative response curves across ten force magnitudes, addressing Reviewer hgjG's concerns about scene constraints and magnitude sensitivity.
  • Table 3 and Appendix Figure 12: Quantitative comparisons of PAN with Optical Flow and SAM, together with force-conditioning visualizations, addressing Reviewer XGFS's concerns about response attribution.
  • Appendix Table 5: Additional RL experiments on Force Prompting, evaluating the generalizability of our RL framework on both OOD and ID benchmarks, addressing Reviewer XGFS's concerns about RL comparisons.
  • Figures 10–11: Preliminary 3D force-control experiments and causal streaming generation, addressing Reviewer SRP8's questions about 3D interactions and temporal control.
  • Figure 9, Appendix A.3 and Figure 17: To address Reviewer SRP8's concerns about the V-JEPA reward, we clarify its role as a learned dynamics prior rather than an analytical physics solver, explain how RL improves the generalization of force-driven video generation, and provide a concrete example of collision physical inconsistency.

We sincerely appreciate the reviewers' constructive feedback, which has helped us strengthen our work. We hope the additional experiments and clarifications adequately address their concerns and welcome any further discussion.

Best regards,

The Authors

Dear reviewers, the authors have submitted a rebuttal. Could you please engage in the discussion? Thanks, your AC.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.