acceptodds
Under review as a conference paper at ICLR 2027

ActionRoPE: Action as Coordinates for Spatially Consistent Control in Planar-Camera Game World Models

11 views

Abstract

Game world models have made rapid progress toward real-time interactive generation. Yet action control is still predominantly implemented as feature conditioning, leaving the model to learn from training data how each action should move the visual world. This formulation is poorly matched to a broad class of planar-camera worlds, such as top-down and isometric games. In these worlds, player motion induces a geometrically well-defined planar camera displacement. We observe that the key to spatially consistent control in such worlds is to represent the effect of each action directly in world coordinates. To this end, we introduce ActionRoPE, a parameter-free action interface that converts action-driven camera displacement into world-coordinate rotary positions. As a result, tokens corresponding to the same world location retain the same positional address across time. These world coordinates further expose a spatial asymmetry in generation: regions covered by the initial observation are geometrically known, while newly revealed regions remain to be synthesized. We therefore propose Coordinate-Conditioned Noise Scheduling (CNS), which assigns token-wise denoising schedules according to this known–new structure. It allows observed regions to stabilize earlier, while new content is generated with reference to them. Compared with four representative action-conditioning mechanisms, our approach follows the commanded motion more faithfully and reduces the revisit drift by 66% over the strongest baseline. The same interface transfers across video backbones from 1.3B to 33B parameters and to out-of-distribution scenes. Project Page: https://anonymous.4open.science/w/ActionRoPE-7722/

Primary area:
generative models
Keywords:
RoPE, Game World Model

Reviews

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

For planar-camera game world models (top-down/isometric), the paper argues that action control should be geometric rather than learned feature conditioning. ActionRoPE is a parameter-free interface that converts action-driven camera displacement into world-coordinate rotary positions, so that a static world location keeps the same positional address across frames. Building on this, Coordinate-Conditioned Noise Scheduling (CNS) exploits the resulting known/new partition to denoise first-frame-observed regions earlier than newly revealed ones. Using clips collected from 50.9 hours of gameplay, the method is trained on Wan2.2-TI2V-5B and compared against four action-injection mechanisms on motion fidelity, direction compliance, and closed-loop revisit, plus cross-backbone (1.3B–33B) and OOD-scene transfer.

Critical strengths
  1. A simple, parameter-free control representation. Writing displacement into RoPE avoids a separate learned action-conditioning pathway. Results across backbones support the usefulness of the interface beyond a single generator.
  2. A useful closed-loop evaluation. The revisit-drift protocol directly tests spatial consistency. The training-versus-sampling control for CNS (Appendix B.4) further supports the conclusion that its benefit stems from training rather than from a sampling-time change alone.
Critical weaknesses
  1. Action information is unequal for two baselines. Prompt and Cross-Attn discard motion magnitude, confounding the speed-following comparison. AdaLN receives continuous displacement and provides the strongest drift baseline, so its comparison remains informative.
  2. The full-model gain combines positional encoding with a different training recipe. The CNS ablation is useful, but a continuous-displacement baseline with the same known/new conditioning and loss treatment is missing. Such a comparison would show how much of the improvement requires ActionRoPE, rather than the accompanying region-dependent training.
Questions for authors
  1. Full-command baselines. Do the motion-fidelity gains persist when Prompt and Cross-Attn also receive displacement magnitude under the same backbone and training data?
  2. ActionRoPE vs. recipe. Since AdaLN already gets continuous displacement, can it use the same known/new mask, noise schedule, and loss weighting? Comparing it with ActionRoPE+CNS would isolate the coordinate-based positional encoding.
  3. Comparable revisit distances. For the out-and-back trajectories in Table 1, please report displacement at the turning point alongside final drift, to confirm the revisit advantage does not merely reflect greater generated motion.
4 comments
AuthorsOpenReview ↗

Response to Reviewer FNPM

We thank the reviewer for recognizing the parameter-free design and closed-loop revisit evaluation of ActionRoPE. We address the three concerns below.

W1 & Q1. Full-command baselines

All methods receive the same source camera displacement. To preserve their original action interfaces and implementations, Prompt (Incantation) and Cross-Attn (Matrix-Game) convert it into discrete directions (Appendix C.3). We thank the reviewer for highlighting this distinction and will clarify it in the revision.

We further evaluate full-command variants preserving displacement magnitude under identical training settings:

MethodSlope →1R² ↑Angle ↓PSNR ↑SSIM ↑NCC ↑DINO ↑Drift ↓
Prompt (Full)0.903−0.621.14°22.030.6370.7430.86318.2
Cross-Attn (Full)0.878−1.172.09°19.090.4820.4550.77751.7
ActionRoPE0.9810.9520.68°23.940.6910.8310.8927.8

These information-matched comparisons further demonstrate the effectiveness of world-coordinate positional encoding for accurate motion control and spatial consistency.

W2 & Q2. ActionRoPE vs. training recipe

We further apply the identical training recipe to AdaLN, including the known/new mask, CNS noise schedule, and 2× region-weighted loss, while preserving its continuous-displacement conditioning. We compare the results under the same backbone and training protocol:

MethodPSNR ↑SSIM ↑NCC ↑DINO ↑Drift ↓
AdaLN (Original)21.110.6040.6850.86222.6
AdaLN + Mask + CNS + Loss×218.390.4830.3890.731110.0
ActionRoPE + Mask, w/o CNS22.810.6520.7880.88511.7
ActionRoPE + Mask + CNS23.940.6910.8310.8927.8

The matched training recipe degrades AdaLN's revisit performance, highlighting the importance of world-coordinate positional alignment. Meanwhile, the ActionRoPE ablation keeps Mask and Loss×2 unchanged, demonstrating the complementary benefit of CNS for revisit consistency.

Q3. Comparable revisit distances

We additionally report the commanded and generated displacement at the turning point (frame 40), alongside the final residual drift, across the same evaluation trajectories in main paper.

MethodCommanded distance (px)Generated distance (px)Final drift (px) ↓
Prompt170.9151.926.6
Linear170.9126.035.3
AdaLN170.9157.722.6
Cross-Attn170.9141.740.2
Ours170.9160.57.8

Under identical commanded distances, ActionRoPE achieves both the closest outbound displacement to the target and the lowest final revisit drift. These results jointly demonstrate accurate motion execution and spatial consistency.

AuthorsOpenReview ↗

Follow-up Response to Reviewer FNPM

Dear Reviewer FNPM,

We sincerely appreciate the time and effort you have dedicated to reviewing our paper. As the rebuttal discussion session draws to a close, we kindly request you to review our submitted rebuttal and welcome any further questions that you might have.

Thank you once again for your valuable feedback and consideration.

Thank you for the detailed rebuttal and additional experiments. The experiments addresses my concerns and reinforces my original positive assessment of the paper, so I will keep my positive score.

AuthorsOpenReview ↗

Thank you for your review

We’re glad our response helped clarify the concerns. Thank you for taking the time to discuss our work and for your constructive feedback.

Reviewer_fCA9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ActionRoPE, a parameter-free action-control interface for planar-camera game world models. Instead of injecting actions as learned features, ActionRoPE converts action-induced camera displacement into world-aligned rotary positional embeddings, improving spatial consistency across frames. The authors further introduce Coordinate-Conditioned Noise Scheduling (CNS) to distinguish previously observed and newly revealed regions. Experiments demonstrate improved motion fidelity and revisit consistency over four action-conditioning baselines, with additional results across different video backbones and out-of-distribution scenes.

Critical strengths
  • Simple and well-motivated geometric formulation: ActionRoPE directly incorporates action-induced displacement into positional embeddings without additional control parameters. The approach is elegant and exploits the geometric structure of planar-camera environments.

  • Strong empirical results and portability: The method achieves substantially better motion fidelity and revisit consistency than the evaluated baselines. The CNS ablation further supports its effectiveness, and experiments across four backbones demonstrate promising architectural portability.

Critical weaknesses
  • Insufficient comparisons and isolation of contributions: ActionRoPE is only compared against feature-based action-conditioning methods, without evaluating geometry-aware camera-control approaches such as ReRoPE or SCoPE. Moreover, the Prompt and Cross-Attn baselines receive only discrete directions without speed information, potentially disadvantaging them. The combined use of ActionRoPE, CNS, the known/new mask, and region-weighted losses also makes it difficult to isolate the contribution of the proposed coordinate representation.

  • Limited evaluation of generalizability: Experiments focus on a single game with approximately rigid planar camera motion, primarily evaluating straight-line movements and simple round trips. Despite the claimed spatial consistency, performance under longer, more complex trajectories and imperfect action-to-camera calibration remains unclear.

Questions for authors
  • ActionRoPE explicitly models 2D translation, while Figure 1(c) suggests support for camera rotation. How does the method handle rotation, and could the coordinate formulation be extended to zooming?

  • Can the authors compare against geometry-aware camera-control methods such as ReRoPE or SCoPE under matched planar-camera settings?

  • Can the authors isolate the individual contributions of ActionRoPE, CNS, the known/new mask, and region-weighted training losses?

  • How does the method perform under longer, more complex trajectories with multiple turns and revisits, or when the action-to-camera calibration is inaccurate?

5 comments
AuthorsOpenReview ↗

Response to Reviewer fCA9

We thank the reviewer for recognizing ActionRoPE's geometric formulation, strong empirical results, and portability.

Q1. Camera rotation and zooming

ActionRoPE naturally supports planar rotation and zoom by mapping screen-space tokens to world coordinates before RoPE. Extending Eq. (3) gives

pkijw=1zkR(−θk)(pijs−bk),\mathbf p^w_{kij}=\frac{1}{z_k}R(-\theta_k)(\mathbf p^s_{ij}-\mathbf b_k),

Here θk\theta_k, zkz_k, bk\mathbf b_k denote rotation, zoom, and translation; θk=0\theta_k=0, zk=1z_k=1, bk=−Δk/s\mathbf b_k=-\Delta_k/s recover Eq. (3). No additional data supervision is required. Zero-shot videos on unseen scenes are on our updated anonymous project page. Mean errors are 3.5° for commanded ±20° rotations and 9.0% for 1.25×/0.8× zoom.

W1 & Q2. Camera-aware RoPE comparison

Our setting differs: ReRoPE and SCoPE address general 3D camera control, while ActionRoPE targets 2.5D planar-camera game world models. A 2D displacement admits multiple 3D camera configurations, complicating direct transfer. For fairness, we adapt both to the game engine's planar-camera geometry under the same evaluation protocol:

MethodSlope →1R² ↑Angle ↓PSNR ↑SSIM ↑NCC ↑DINO ↑Drift ↓
ReRoPE-planar0.9160.250.90°21.130.6120.7010.87024.7
SCoPE-planar0.8180.721.31°22.000.6410.7390.87721.6
ActionRoPE + CNS0.9810.9520.68°23.940.6910.8310.8927.8

These results support our setting-specific design: simple 2D world-coordinate alignment achieves superior control and revisit consistency.

W1 & Q3. Baseline fairness and component contributions

Baseline fairness. All methods receive the same source camera displacement. Prompt (Incantation) and Cross-Attn (Matrix-Game) convert it to discrete directions to preserve their original interfaces (Appendix C.3); We will clarify this. We also test full-command variants preserving magnitude under matched training:

MethodSlope →1R² ↑Angle ↓PSNR ↑SSIM ↑NCC ↑DINO ↑Drift ↓
Prompt (Full)0.903−0.621.14°22.030.6370.7430.86318.2
Cross-Attn (Full)0.878−1.172.09°19.090.4820.4550.77751.7
ActionRoPE0.9810.9520.68°23.940.6910.8310.8927.8

Component contributions. ActionRoPE establishes world-coordinate correspondence; Mask and CNS improve region-aware generation. All ablations below retain the same 2× region-weighted loss:

ConfigurationSlope →1R² ↑Angle ↓PSNR ↑SSIM ↑NCC ↑DINO ↑Drift ↓
ActionRoPE + Mask, w/o CNS1.0090.9560.67°22.810.6520.7880.88511.7
ActionRoPE + CNS, w/o Mask1.0280.9260.63°23.120.6650.7940.88111.5
ActionRoPE + Mask + CNS0.9810.9520.68°23.940.6910.8310.8927.8

We also apply the same Mask, CNS schedule and Loss×2 to AdaLN:

ConfigurationSlope →1R² ↑Angle ↓PSNR ↑SSIM ↑NCC ↑DINO ↑Drift ↓
AdaLN (Original)0.784−0.521.92°21.110.6040.6850.86222.6
AdaLN + Mask + CNS + Loss×20.5000.0321.0°18.390.4830.3890.731110.0

Mask and CNS contribute complementary revisit gains, while the matched-recipe AdaLN comparison highlights the importance of world-coordinate alignment.

Region-weighted loss. Loss×2 emphasizes newly revealed regions, without directly supervising return displacement. We isolate its contribution to new-region appearance:

ConfigurationColor consistency ↑Style distance ↓
ActionRoPE + CNS + Loss×163.6%0.445
ActionRoPE + CNS + Loss×264.9%0.423

This improves new-region color and style consistency.

W2 & Q4. Complex trajectories and calibration

We further evaluate longer, more complex trajectories with our distilled few-step autoregressive model, including multiple turns and repeated revisits:

TrajectoryDuration (s)RevisitsPSNR (dB) ↑Drift (px) ↓
Single revisit30119.69.4
Multiple revisits60319.213.1

Results demonstrate consistent geometry beyond short round trips.

Regarding calibration, ActionRoPE explicitly follows the supplied camera displacement, independently of the upstream action-to-camera mapping. Motion-fidelity and speed-scaling experiments (Table 1, Appendix B.2) demonstrate accurate command following across a wide range of displacements. Calibration errors affect the supplied displacement, while ActionRoPE preserves an explicit, adjustable control interface.

AuthorsOpenReview ↗

Follow-up Response to Reviewer fCA9

Dear Reviewer fCA9,

We sincerely appreciate the time and effort you have dedicated to reviewing our paper. As the rebuttal discussion session draws to a close, we kindly request you to review our submitted rebuttal and welcome any further questions that you might have.

Thank you once again for your valuable feedback and consideration.

AuthorsOpenReview ↗

Follow-up Response to Reviewer fCA9

Dear Reviewer fCA9,

As the discussion period closes within 24 hours, we would appreciate it if you could let us know whether our additional experiments and clarifications address your concerns. If any questions remain, we would be happy to respond promptly before the discussion closes.

Thank you again for your time and feedback.

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

This paper proposes ActionRoPE, which incorporates action-induced camera motion into the RoPE embedding for planar-camera worlds, so that the spatial positional encoding of each static background location stays aligned across time. It also introduces Coordinate-Conditioned Noise Scheduling (CNS), which separates regions observed in the first frame from newly revealed regions and applies a per-location noise schedule. Compared against four action-conditioning methods implemented on the same Wan backbone, the paper reports better control of motion magnitude and better revisit consistency, and it shows results on several backbones.

Critical strengths
  1. Low parameter overhead and use across multiple backbones. ActionRoPE adds no learnable control branch, and the parameters needed for CNS’s mask input are small. Results on four backbones of different scales and architectures suggest the approach does not depend on a single model.
  2. Consistent gains under the reported evaluation settings. In Table 1, the full method performs best on all motion-magnitude, direction, and revisit metrics.
Critical weaknesses
  1. It is unclear whether the outcome of an action should be predetermined when generating from action conditions. The effect of an action depends on the surrounding environment, for example when a wall blocks the path or when the character is swimming underwater. The method does not model these interactions with the environment. It computes the outcome with a simple rule, which reduces the diversity of the generated results. More broadly, the method addresses only character-centric 2D camera movement rather than general actions.
  2. CNS may work in static scenes, but it does not account for dynamic objects or for interactions between the character and its surroundings. In particular, it does not allow the environment to change. Tasks such as pick-and-place change the environment, and the method does not handle this.
Questions for authors
  1. At inference, the user’s action is converted into camera motion by a fixed rule, regardless of the scene. What does the model generate when the environment should change the outcome of the action, for example when the character walks into a wall or moves through water? Can the same action produce different outcomes in different scenes? Can the approach extend to actions other than character-centric 2D camera movement?
  2. The generated videos are only a few seconds long, and the revisit test covers only a short round trip. Why is this length enough to evaluate action control? Do you have results on longer videos?
  3. ActionRoPE also takes camera motion as input, not the action itself. Under pure 2D translation, what is the practical difference from existing camera-RoPE methods? Why are these methods not included in the comparison?
  4. The known region is defined only from the first frame. In long video generation across multiple chunks, what counts as the known region? Do you have results showing that CNS still helps in this setting?
  5. CNS assumes that the known region does not change and stabilizes it first. What happens when it should change, for example with moving objects, or with character–environment interactions such as pick-and-place? Does CNS suppress these changes? Do you have experiments that test this?
10 comments
AuthorsOpenReview ↗

Response to Reviewer QVEh

We thank the reviewer for recognizing the effectiveness and portability of ActionRoPE. We address the concerns below.

W1 & Q1. Environment-dependent action outcomes

ActionRoPE enables explicit modeling and control of environment-dependent motion. For observed environments, gameplay data provides both actions and their resulting displacements. For example, a movement command producing zero displacement at a wall reveals a motion constraint that can be recovered from the data and applied at inference. The same principle extends to speed changes in water and other environment-dependent motion.

For unseen environments, implicit action-conditioning methods need to infer such constraints through learned dynamics, with no guarantee of correct behavior. In contrast, ActionRoPE allows motion outcomes to be explicitly specified and verified: the same action can stop at a wall or deliberately pass through it by controlling the spatial RoPE displacement. This is shown by our wall-interaction and variable-speed experiments on the updated anonymous project page. Automatic discovery of constraints in unseen scenes remains a separate dynamics-modeling problem; An upstream controller can provide scene-dependent displacements for ActionRoPE to execute.

Q2. Longer video generation

Our 81-frame evaluation follows fixed-length video generation practice in Wan-based works, including SCoPE, Matrix-Game 4.0, and LingBot-World. We train and evaluate all bidirectional methods under the 81-frame protocol, ensuring a fair comparison without confounding factors from autoregressive distillation.

To evaluate long-horizon consistency, we extend ActionRoPE to our autoregressive model and conduct longer evaluations with single and repeated revisits:

TrajectoryDuration (s)RevisitsPSNR (dB) ↑Drift (px) ↓
Single revisit30119.69.4
Multiple revisits60319.213.1

These results, together with videos on the updated anonymous project page, demonstrate that ActionRoPE maintains spatial consistency across extended trajectories and repeated revisits, beyond fixed-length bidirectional evaluation.

Q3. Comparison with camera-aware RoPE

Our work targets a different setting: ReRoPE and SCoPE are designed for general 3D camera control, while ActionRoPE focuses on 2.5D planar-camera game world models. Direct transfer is nontrivial: the same 2D displacement may correspond to multiple 3D camera configurations, introducing ambiguity and potentially unfair assumptions.

For a fair comparison, we adapt these methods to the planar-camera configuration closest to the original game engine's geometry and evaluate them under the same protocol:

MethodSlope →1R² ↑Angle ↓PSNR ↑SSIM ↑NCC ↑DINO ↑Drift (px) ↓
ReRoPE-planar0.9160.250.90°21.130.6120.7010.87024.7
SCoPE-planar0.8180.721.31°22.000.6410.7390.87721.6
Ours0.9810.9520.68°23.940.6910.8310.8927.8

These results highlight the advantage of our setting-specific design. By exploiting planar-camera geometry, ActionRoPE achieves accurate control through direct 2D coordinate alignment with a simpler formulation. It provides a simple yet effective solution for the largely underexplored setting of 2.5D planar-camera game world models.

Q4. Known regions across multiple chunks

We consistently define "known" as regions observed in the original first frame, not those generated in previous chunks. This avoids treating model-generated content as equally reliable as the initial observation, while ActionRoPE maintains world-coordinate alignment across chunks independently of the known/new classification.

We further evaluate CNS over 20–40 s trajectories, including revisits to the initial region:

MethodRevisit PSNR (dB) ↑Drift (px) ↓
ActionRoPE w/o CNS19.821.2
ActionRoPE + CNS21.216.9

These results confirm that CNS remains beneficial for long-horizon consistency, as returning to the initial region reactivates its known-region denoising schedule.

W2 & Q5. Dynamic objects and changing environments

ActionRoPE aligns spatial addresses, not visual content. The same world location retains its positional address, while objects and scene content remain free to evolve.

Similarly, CNS does not freeze known regions. As explained in Sec. 3.3, it only adjusts token-wise denoising schedules and does not copy or overwrite observed content. Dynamic-scene experiments on the updated anonymous project page demonstrate moving objects and changing visual content alongside consistent static backgrounds. General object-manipulation dynamics, such as pick-and-place, remain beyond the current evaluation scope.

Thank you for the detailed responses. I still have some unresolved concerns and would appreciate further clarification on the following points.

Regarding Q1. Action outcomes and interactions with the scene

Could the authors clarify how the upstream controller works in practice? How does it determine action outcomes arising from interactions with the scene? If a rule-based model or game engine predicts these outcomes in advance, and ActionRoPE then generates video from the resulting displacement only to preserve background consistency, the scene interactions are handled externally. This could substantially weaken the claimed novelty and contribution of ActionRoPE as an action-conditioned world model. Please clarify what ActionRoPE itself learns about action outcomes and how these outcomes are determined at inference.

Regarding Q4. Consistency across the explored world

Evaluating consistency only against the first frame seems tailored to the evaluation setup. In a world model, consistency should also be evaluated when the user revisits regions generated in previous chunks. This consistency is important to the user’s actual experience of exploring the world. The architecture and experiments are not designed to preserve this form of world consistency. The results therefore demonstrate first-frame consistency, but do not establish consistency across the evolving world.

AuthorsOpenReview ↗

Follow-up Response to Reviewer QVEh

We thank the reviewer for the thoughtful follow-up and constructive feedback. We appreciate the opportunity to further clarify our design and evaluation.

Q1. Action outcomes and scene interactions

We respectfully clarify that ActionRoPE does not independently determine scene-dependent action outcomes. Allowing the generative model to override the prescribed displacement would undermine our claim of precise and spatially consistent control in a well-defined planar-camera world. ActionRoPE instead learns to generate the corresponding visual evolution under the specified motion.

Compared with end-to-end implicit Wan baselines, which couple action outcomes, action conditioning, and visual synthesis, ActionRoPE intentionally decouples action outcomes from conditioning and visual synthesis. This allows outcomes to be determined by explicit, deterministic rules, making them more reliable and verifiable.

This design 1) advances over end-to-end implicit baselines that produce action outcomes implicitly and without guaranteed precision; 2) aligns with the logic of real game engines and recent agentic world models (WorldMind¹, GameDirector², AgentGarten³); and 3) shares the motivation of camera-aware RoPE methods (ReRoPE⁴, SCoPE⁵), which take prescribed camera trajectories as input regardless of the visuals.

Compared with recent agentic world models, exposing a control interface alone also does not guarantee faithful execution. ActionRoPE takes a further step by enabling players to influence generation more precisely and spatially consistently through this interface, instead of letting the generative model implicitly determine action outcomes. This is demonstrated by our matched comparisons and wall-interaction evaluations.

Q4. Consistency across the explored world

Our main evaluation follows the common practice of closed-loop revisit protocol in SCoPE⁵ (Sec. 5.3, "Closed-Loop Revisit Consistency") and other methods. Since different methods may generate different content in newly explored regions, evaluating revisits against the shared initial observation provides a consistent reference for fair cross-method comparison.

We appreciate the reviewer's suggestion to further evaluate consistency across previously generated regions. We additionally evaluate revisits to regions synthesized in earlier chunks, using previously generated frames in a 30s rollout as references:

ReferenceRevisit gapGenerated PercentagePSNR ↑SSIM ↑NCC ↑
Earlier generated frame20–30 s23%22.630.6100.584
Earlier generated frame10–20 s87%21.640.5230.573
Earlier generated frame0–10 s100%25.980.6840.613

These results demonstrate that ActionRoPE maintains revisit consistency beyond the initial observation, including previously generated regions revisited after 20–30 seconds and views containing substantial newly generated content. This further supports spatial consistency across the evolving generated world, beyond first-frame revisit evaluation.

We thank the reviewer again for the valuable feedback and welcome any further questions or discussion.

References

[1] Deng et al., WorldMind: Decoupled Game World Model for State-Aware NPC Behavior, 2026. https://arxiv.org/abs/2608.21439

[2] Lin et al., GameDirector: Decoupling Gameplay Logic from Rendering for Player-Configurable Game World Models, 2026. https://arxiv.org/abs/2609.25652

[3] Chi et al., AgentGarten: Code Worlds for Evolving Agents, 2026. https://arxiv.org/abs/2610.12374

[4] Li et al., ReRoPE: Repurposing RoPE for Relative Camera Control, 2026. https://arxiv.org/abs/2602.08068

[5] Yin et al., SCoPE: Sightline-Coordinate Positional Encoding for Video Diffusion Transformers, 2026. https://arxiv.org/abs/2606.27345

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Rebuttal for All Reviewers

We thank all three reviewers for their thoughtful feedback and constructive suggestions.

Recognized strengths. Reviewers highlighted ActionRoPE's parameter-free geometric formulation, strong empirical results, cross-backbone portability, and effective closed-loop revisit evaluation. We appreciate their recognition of our approach to explicit action control through world-coordinate positional encoding.

We have conducted additional experiments and provided clarifications addressing the major concerns:

  1. Information-matched baseline comparisons [fCA9, FNPM]. We clarify that all methods receive the same source camera displacement, while Prompt and Cross-Attn originally discretize it to preserve their published action interfaces. We additionally evaluate full-command variants retaining displacement magnitude. ActionRoPE outperforms both variants across all eight metrics, confirming the effectiveness of coordinate-based control under information-matched comparisons.

  2. Comparison with camera-aware RoPE [QVEh, fCA9]. We adapt ReRoPE and SCoPE to planar-camera configurations matching the game engine. ActionRoPE achieves superior motion fidelity and revisit consistency across all eight metrics, demonstrating the effectiveness of direct 2D world-coordinate alignment for planar-camera game world models.

  3. Component contributions and training fairness [fCA9, FNPM]. Additional ablations isolate the contributions of Mask, CNS, and region-weighted loss. We also apply the identical training recipe to AdaLN. The matched-recipe comparison highlights the importance of world-coordinate alignment, while Mask and CNS provide complementary revisit improvements. Region-weighted loss further improves color and style consistency in newly generated regions.

  4. Long-horizon consistency and CNS [QVEh, fCA9]. We extend ActionRoPE to autoregressive generation and evaluate 30-second single-revisit and 60-second multiple-revisit trajectories, achieving revisit drift of 9.4 px and 13.1 px, respectively. A separate 20–40 s evaluation confirms that CNS remains beneficial across multiple chunks, improving revisit PSNR from 19.8 to 21.2 dB and reducing drift from 21.2 to 16.9 px.

  5. Extended control and evaluation fairness [QVEh, fCA9, FNPM]. We demonstrate environment-dependent motion, dynamic scene content, and zero-shot rotation and zoom control without additional data supervision. We also report actual turning-point displacements across all baselines. ActionRoPE achieves the closest displacement to the command and the lowest final revisit drift, confirming that its consistency advantage is not due to shorter generated trajectories.

These additional experiments and visualizations are available on our updated anonymous project page. We will incorporate the results and clarifications into the revised paper. We hope they address the reviewers' concerns and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.