Pith.

Computer Vision and Pattern Recognition · arXiv:2611.06120 v1 · submitted 4 Nov 2026 · 35 references

ReactWorld: Structuring Interaction Responses for Embodied Video Generation

Hanwei Liu, Huijia Zhang, Hanzhang Wang, Guangyu Liu, Hanbei Zhan, Linjie Hu, Bin Xiao, Xuan Zhou

show 1 more author

Dong Zhang

The paper says

“Embodied video world models predict scene changes caused by robot actions, yet generated object responses may remain inconsistent with task intent and contact dynamics. Critical responses in embodied scenes often concentrate in local regions and specific time intervals.”

more of the abstract

Inspired by chain-of-thought reasoning, we introduce ReactWorld, an embodied video generation model guided by structured interaction anticipation. In the generation backbone's feature space, an interaction response field predictor anticipates how the interaction will unfold, explicitly describing participant roles, future motion response intensity, and local motion changes. A response feedback adapter forms increments to object features from the difference between the object branch's outputs with predicted and zeroed local dynamics inputs. It also converts predicted displacements into positional biases for temporal attention. These two mechanisms guide object feature updates and information retrieval across frames, respectively. Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories. On EZS-Bench, it achieves a Domain score of 86.98 and an average score of 81.87.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 5 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: minor revision.

A genuinely well-built generation method, but the Qwen3-VL labeler–judge overlap on RBench is real and deserves an explicit cross-check before the 3.29-point claim is taken at face value. This line is wrong

3 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 35 references say.

The paper says, at §4.1 and §3.4

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The same model family (Qwen3-VL) is used for both offline role labeling (§3.4) and the RBench evaluation (§4.1, using Qwen3-VL-235B-A22B-Instruct). The LLM Usage Statement transparently acknowledges this, but the paper does not discuss whether this overlap could inflate reported improvements.

more of objection 1

If the evaluation VLM systematically favors videos whose motion or appearance patterns align with the labeling VLM's biases, the 3.29-point gain over PhysisForcing may partly reflect consistency rather than genuine interaction plausibility. The human evaluation on contact plausibility (Table 3) partially mitigates this, but the improvement over PhysisForcing for contact plausibility (4.06 vs 3.78, p=0.140) is not statistically significant. The authors should add a discussion of this potential confound and, if possible, report RBench scores evaluated with a different VLM family as a robustness check.

The paper says, at §3.4

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The automatically generated labels for roles, future response intensity, and motion (using Qwen3-VL, SAM3, and CoWTracker) are not quantitatively evaluated. Section 3.4 states 'Response labels are constructed offline from training videos' but provides no accuracy metrics or human agreement study.

more of objection 2

If these labels contain systematic errors or noise, the interaction response field predictor will learn biased patterns, which could degrade generation guidance even if the overall benchmark scores increase due to other factors. The authors should report validation statistics (e.g., agreement with human annotations on a sample) or discuss the expected impact of label noise.

The paper says, at Eq. (5) and Table 5

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The zero-dynamics reference (Eq. 5) is a key conceptual contribution, but Table 5 shows that removing it ('No zero reference') still yields a Domain score of 85.12 vs 86.98 for the full model. The improvement is modest and the ablation does not test whether the zero-dynamics operation is specifically beneficial compared to a simpler subtraction baseline (e.g., using a learned constant or a separate MLP).

more of objection 3

The authors should clarify why the zero-dynamics design is important beyond the empirical 1.86-point Domain gain, or provide additional analysis (e.g., measuring the magnitude of ∆_k^t under different conditions).

6 minor objections

The paper says, at §3.2, Eq. (2)

Pith says: minor objection 1 of 6

The notation for the role map R uses softmax over four classes, but the text lists only three interaction roles (r, d, a) plus background b. Clarify that the softmax is over four channels.

The paper says, at §3.3, Eq. (7)

Pith says: minor objection 2 of 6

The expression 'o selects the direct target or affected object with the larger gate' is ambiguous. It is not explicitly defined whether this selection is per-spatial-location or per-frame. Please clarify.

show 4 more

The paper says, at §4.3, Fig. 6

Pith says: minor objection 3 of 6

The x-axis label 'Bias strength multiplier' is not linked to the actual β value. State the default β in the caption or text for reproducibility.

The paper says, at §3.4

Pith says: minor objection 4 of 6

The sentence 'Response supervision grounds structured interaction anticipation in labeled targets' is awkward. Consider rephrasing for clarity.

The paper says, at References

Pith says: minor objection 5 of 6

Several references (e.g., [18] RecCAR, [23] VPT) are listed only as 'arXiv preprint' without author names or years in the citation keys. Verify that these are correctly cited as published or preprint and ensure author lists are complete if they exist.

The paper says, at Figure 2

Pith says: minor objection 6 of 6

The figure is informative but the font in panel (b) and (c) is very small. Consider increasing font size for readability in the camera-ready version.

What would settle it

Re-evaluate the 400 RBench examples with an alternative VLM evaluator (e.g., GPT-4V, Gemini 2.0 Flash, or a non-Qwen open-source VLM). Compute the average RBench-style scores for ReactWorld and PhysisForcing separately. If ReactWorld's margin over PhysisForcing remains at least 2.0 points under the alternative evaluator, the concern is substantially mitigated. If the margin drops below 1.0 points or reverses, the reported gains are likely partially confounded.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 5 Nov 2026

  1. Reader holds up confidence moderate
  2. Skeptic partly agrees: says it holds up only with conditions 18 s · 13,590 tokens read, 1,509 written Potential confound between the automatic labeling pipeline (Qwen3-VL) and the RBench evaluation metric (Qwen3-VL-235B) could inflate observed improvements.
  3. Editor wrote the verdict above 1 min 8 s · 14,490 tokens read, 10,586 written
  4. Referee minor revision: 3 major objections, 6 minor 20 s · 13,904 tokens read, 1,923 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

First take: the reader's ACCEPT is right, and the stress-test's labeler-evaluator worry is the most important thing to know about this paper. It is a real unaddressed overlap, but it does not sink the central claim.

What is new: the zero-dynamics reference in Eq. 5 is a genuinely clean idea. Hold features and context fixed, and form the object update as the difference between the object branch output with predicted dynamics and with zeroed dynamics. The update is identically zero without a dynamics signal, which stops the model from learning to ignore its conditioning. The displacement-guided temporal attention bias (Eqs. 7–8) is a real addition too, and the directional intervention in Fig. 5 shows the feedback path carries causal signal rather than correlation. The ablations are unusually careful: component-by-component, an amplitude-matched control for the zero reference, a bias-strength sweep, insertion depth, all on the same data and budget. Video-FT, the same-backbone same-data baseline, is the right control, and ReactWorld beats it by 5.57 points average (b13). Reporting Holm-adjusted p-values and a transparent LLM usage statement help as well.

Now the soft spots, in proportion. The stress-test gets a detail wrong — Table 3 reports three human metrics, not four — but the substance stands. Qwen3-VL assists role labeling in §3.4, and the same model family judges RBench in §4.1. The paper discloses both uses without ever discussing the overlap. That matters most exactly where the wins are largest: spatial relations (+14.5 over Video-FT) and visual reasoning (+18.3) are the VLM-judgment-heavy categories. This does not make the result an artifact — EZS-Bench Domain is judged by Qwen2.5-VL-72B and ReactWorld still leads PhysisForcing by 2.49, and the motion and displacement labels come from CoWTracker, not Qwen3-VL — but the authors should add a cross-judge evaluation on RBench or at least a direct discussion. I also agree with the reader that the offline label accuracy is unquantified: the whole anticipation branch is supervised by Qwen3-VL + SAM3 + CoWTracker labels with no agreement numbers. One correction to the reader: RBench is not fully independent of the training data. RoVid-X appears on both sides, as a training source and as the RBench dataset; the episode-level split helps, but I would ask how much of the 400-example test set overlaps the training distribution. The 2D-representation limit is acknowledged in the conclusion (b199); deferred supplementary details are a minor annoyance.

Who this is for: embodied AI and physics-aware video generation researchers. The zero-dynamics reference mechanism alone is worth citing. Send it to peer review; reviewers should push on the labeler–judge overlap, label accuracy, and the RBench train/test overlap.

The strongest objection

The central claim relies on RBench scores evaluated by Qwen3-VL-235B-Instruct (§4.1). The same model family (Qwen3-VL) is used offline for role labeling (§3.4): 'We use Qwen3-VL to assist task-related role assignment'. If the evaluation VLM systematically favors videos whose appearance aligns with the labeling VLM's biases—for instance, preferring certain motion patterns or object appearances that the labeling pipeline also produces—then the reported 3.29-point improvement over PhysisForcing may partly reflect consistency between labeling and evaluation rather than genuine gains in interaction plausibility. The human evaluation (Table 3) partially mitigates this, but it covers only a small subset (50 evaluators rating 4 metrics) and the advantage over PhysisForcing in contact plausibility (4.06 vs 3.78, p=0.140) is not significant. The paper does not discuss or control for this overlap.

The reader's weakest_assumption focused on label accuracy (noise in automatic annotations degrading guidance). My concern is distinct: the potential circularity between the labeling VLM and the evaluation VLM, which could inflate apparent improvements even if labels are accurate. The reader did not flag this eval-label overlap, so agreement is partial. The paper is otherwise well-executed with thorough ablations, human eval, and code release plan. The concern is concrete and testable; if it survives the test, acceptance is justified. Hence I recommend conditional acceptance pending the re-evaluation described.

The referee's report

If the reported gains hold, ReactWorld provides a principled way to inject task-relevant local dynamics into video generation for embodied scenarios, addressing a known limitation of global motion guidance. The idea of predicting interaction roles, response intensity, and displacement in feature space, and then using those predictions as internal conditions (rather than external control signals), is well-motivated. The method is tested on two established benchmarks with multiple baselines, including a fine-tuned version of the same backbone, and includes ablation studies that isolate the contributions of each component. The human evaluation, while limited in sample size, adds credibility. The paper is clearly written and the figures effectively illustrate the architecture.

Pith's reading of the paper (a machine's summary, not the authors')

ReactWorld introduces a method for embodied video generation that predicts how interactions will unfold in the feature space of the generation backbone. It predicts participant roles (robot, direct target, affected object), future motion response intensity, and interval motion (support and displacement) from intermediate features. These predictions are fed back into generation through two mechanisms: object feature updates computed as the difference between outputs with predicted versus zeroed local dynamics, and displacement-based biases that guide temporal attention across frames. The result is more task-consistent and physically plausible video generation, as measured by improvements across all nine categories on RBench and higher Domain scores on EZS-Bench.

What carries the argument. The central mechanism is the combination of an interaction response field predictor and a response feedback adapter. The predictor, composed of four lightweight heads, reads intermediate features from the diffusion backbone to estimate: participant role masks (robot, direct target, affected object, background), future motion response intensity per role (continuous value in [0,1] per location), interval motion support (binary or continuous motion region per time step), and interval 2D displacement. The feedback adapter uses these predictions in two paths: (1) it constructs role- and response-weighted context features, then computes object-feature increments as the difference between the object branch's output with predicted local dynamics versus zeroed dynamics, thus isolating the effect of predicted motion; (2) it converts predicted displacement into a positional bias added to query–key logits in temporal attention over a fixed local window, steering cross-frame information retrieval toward locations where motion is expected.

Core claim. The paper demonstrates that explicitly anticipating structured interaction dynamics—roles, future response intensity, and local motion—within the denoising feature space and using those predictions to condition object feature updates and cross-frame attention significantly improves the quality of embodied video generation. The interaction response field predictor operates on intermediate backbone features within each denoising forward pass, outputting role maps, continuous response values, motion support, and 2D displacements. The response feedback adapter then updates video features by taking the difference between the object branch's outputs with predicted and zeroed local dynamics inputs, and adds displacement-guided positional biases to temporal attention. On RBench, ReactWorld achieves 77.51 average, beating PhysisForcing by 3.29 points; on EZS-Bench, it achieves a Domain score of 86.98. Both human evaluation and ablations confirm that the structured anticipation and feedback contribute to task progress, contact plausibility, and physical plausibility.

The premise it rests on. The interaction response labels—participant roles, motion intensity, and displacements—are constructed offline using Qwen3-VL, SAM3, and CoWTracker without quantified accuracy; any systematic noise or bias in these labels would be learned by the predictor and could degrade the generation guidance.

If the paper is right (Pith's inferences, not the authors')

Structured interaction anticipation can serve as a general framework for embedding task-relevant dynamics into any video generation backbone, potentially improving consistency for robotic manipulation preview and planning.

The zero-dynamics reference trick—taking the output difference between predicted and null conditions to isolate the effect of dynamics—provides a clean way to inject conditional information into feature updates without disrupting the backbone's learned representations.

Displacement-guided temporal attention biases could be extended to longer-range motion or combined with optical flow predictions to handle complex articulated motion.

Because the predictor operates entirely in feature space and requires only the initial image and task instruction at inference, the method is applicable to settings where explicit trajectory or role annotations are unavailable at test time.

The reliance on automatically generated labels (roles, response intensity, displacements from Qwen3-VL, SAM3, and CoWTracker) without quantified accuracy means that any systematic bias in those labels could be learned by the predictor and limit generalization to scenes or tasks unlike the training distribution.

The 2D interaction representation, as the authors acknowledge, leaves room for explicit 3D reasoning; extending to 3D motion and relations could further improve spatial consistency, especially for tasks involving depth changes or occlusions.

If the labeled training data were augmented with manual verification or active learning to correct the most uncertain labels, the predictor's accuracy—and thus the overall generation quality—might improve further, especially for tasks with subtle interaction boundaries.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Embodied video world models predict scene changes caused by robot actions, yet generated object responses may remain inconsistent with task intent and contact dynamics. Critical responses in embodied scenes often concentrate in local regions and specific time intervals. Inspired by chain-of-thought reasoning, we introduce ReactWorld, an embodied video generation model guided by structured interaction anticipation. In the generation backbone's feature space, an interaction response field predictor anticipates how the interaction will unfold, explicitly describing participant roles, future motion response intensity, and local motion changes. A response feedback adapter forms increments to object features from the difference between the object branch's outputs with predicted and zeroed local dynamics inputs. It also converts predicted displacements into positional biases for temporal attention. These two mechanisms guide object feature updates and information retrieval across frames, respectively. Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories. On EZS-Bench, it achieves a Domain score of 86.98 and an average score of 81.87.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-05

The authors' main claim, as the classifier picked it out 1.00

“Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.13 · human participants 0.04 · controlled comparison 0.84 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Embodied video world models predict scene changes caused by robot actions, yet generated object responses may remain inconsistent with task intent and contact dynamics. background 0.69 0.11 0.04 hedged possibility 0.85
Critical responses in embodied scenes often concentrate in local regions and specific time intervals. background 0.99 0.25 0.02 asserted 0.99
Inspired by chain-of-thought reasoning, we introduce ReactWorld, an embodied video generation model guided by structured interaction anticipation. contribution 0.98 0.88 0.06 asserted 1.00
In the generation backbone's feature space, an interaction response field predictor anticipates how the interaction will unfold, explicitly describing participant roles, future motion response intensity, and local motion changes. method 0.90 0.88 0.05 asserted 1.00
A response feedback adapter forms increments to object features from the difference between the object branch's outputs with predicted and zeroed local dynamics inputs. method 0.99 0.88 0.03 asserted 1.00
It also converts predicted displacements into positional biases for temporal attention. method 0.99 0.87 0.02 asserted 1.00
These two mechanisms guide object feature updates and information retrieval across frames, respectively. method 0.96 0.65 0.04 asserted 1.00
Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories. result 1.00 0.96 0.66 asserted 1.00
On EZS-Bench, it achieves a Domain score of 86.98 and an average score of 81.87. result 1.00 0.94 0.12 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.82

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-05: 1 request, 6,832 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (7)
Figure 1 of the paper
Figure 1. Structured interaction responses and generation performance. (a) Roles, local motion support, interval dynamics, and subsequent states in garment folding. Overlay opacity encodes support strength, and arrows denote 2D displacement. (b) Robot and target displacements across tasks and viewpoints. (c) RBench scores on five task categories, and average scores on RBench and EZS-Bench. Abstract Embodied video world models predict scene changes caused by robot actions, yet generated object responses may remain inconsistent with task intent and contact dynamics. Critical responses in…
Figure 2 of the paper
Figure 2. Overview of ReactWorld. Top: response prediction and generative feedback within each denoising forward pass. Bottom: (a) predicting roles, future responses, and interval dynamics to construct interaction context and local dynamic conditions; (b) extracting object￾response increments relative to a zero-dynamics reference and updating features by role; and (c) converting gated, clipped displacements into positional biases within a fixed candidate window. Biases are added to query–key logits before softmax and feature aggregation, followed by bounded residual feedback. Scene and…
Figure 3 of the paper
Figure 3. Structured response predictions and displacement biases. (a) The future response field referenced to the first frame, together with object roles, motion support, and displacement at frames 0, 56, and 76, overlaid on the corresponding RGB frames. (b) The same query and a shared 7 × 7 candidate window for the interval from frame 56 to 60, showing attention without displacement bias, the displacement bias, and attention with bias. Red and green circles indicate the query and displacement endpoint, respectively. the ball on it is the affected object. These roles distinguish directly…
Figure 4 of the paper
Figure 4. Embodied task generation across methods. Columns show Wan2.2, Cosmos3-Nano, ABot-PhysWorld, PhysisForcing, and ReactWorld. Rows show stacking cups, dispensing jam, bagging apples, and opening an oven
Figure 5 of the paper
Figure 5. Directional interventions on feedback displacement. (a) Normal feedback. (b,c) Leftward and rightward offsets applied to the feedback displacement input shift the robot arm’s motion in the corresponding directions. 0 0.5 1 1.5 2 4.35 4.45 4.55 4.65 (a) PSS 0 0.5 1 1.5 2 4.14 4.16 4.18 4.20 (b) TAC Bias strength multiplier (×)
Figure 6 of the paper
Figure 6. Displacement bias strength. Filled markers denote the default setting. A multiplier of zero disables positional bias while retaining object feature updates and content-based temporal attention
Figure 7 of the paper
Figure 7. Insertion depth. (a) Input image and PCA visualiza￾tions of backbone features at the same forward call and time posi￾tion. (b) EZS-Bench Domain scores for different insertion blocks. Insertion depth

Figures are the authors', reproduced from the paper.

References (35)
  1. [1] AgiBot-World-Contributors, Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, et al. AgiBot World Colosseo: A Large- scale Manipulation Platform for Scalable and Intelligent Em- bodied Systems. arXiv preprint arXiv:2503.06669, 2025. 6 on PithPith's reviewarXiv
  2. [2] Shuai Bai, Y uxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, et al. Qwen3-VL Technical Report. arXiv preprint arXiv:2511.21631, 2025. 6 on PithPith's reviewarXiv
  3. [3] Shuai Bai, Keqin Chen, Xuejing Liu, et al. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923, 2025. 6 on PithPith's reviewarXiv
  4. [4] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2Act: Human Video Generation in Novel Scenarios enables Generalizable Robot Manipulation. In Proceedings of the 9th Conferenc... on Pith
  5. [5] Nicolas Carion, Laura Gustafson, Y uan-Ting Hu, Shoubhik Debnath, Ronghang Hu, et al. SAM 3: Segment Anything with Concepts. arXiv preprint arXiv:2511.16719, 2025. 6 on PithPith's reviewarXiv
  6. [6] Chi-Lam Cheang, Guangzeng Chen, Y a Jing, Tao Kong, Hang Li, Yifeng Li, Y uxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Y ang, Hanbo Zhang, and Minzhao Zhu. GR- 2: A Generative Video-Language-Action Model with Web- Scale Knowledge for Robot Manipulation. arXiv preprint arXiv:2410.0... on PithPith's reviewarXiv
  7. [7] Hila Chefer, Uriel Singer, Amit Zohar, Y uval Kirstain, Adam Polyak, Y aniv Taigman, Lior Wolf, and Shelly Sheynin. VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models. In Proceed- ings of the 42nd International Conference on Machin... on Pith
  8. [8] Tianxiao Chen, Hanmo Chen, Huajin Chen, Bo Li, Qi Y e, and Peng-Tao Jiang. VIPER: Visual In-Context Physics Rea- soning for Physically Plausible Video Generation. arXiv preprint arXiv:2607.23472, 2026. 2 on PithPith's reviewarXiv
  9. [9] Y uzhi Chen, Ronghan Chen, Dongjie Huo, Y andan Y ang, Dekang Qi, Haoyun Liu, Tong Lin, Shuang Zeng, Junjin Xiao, Xinyuan Chang, Feng Xiong, Xing Wei, Zhiheng Ma, and Mu Xu. ABot-PhysWorld: Interactive World Foundation Model for Robotic Manipulation with Physics Alignment, on Pith
  10. [10] Y ufan Deng, Zilin Pan, Hongyu Zhang, Xiaojie Li, Ruoqing Hu, Y ufei Ding, Yiming Zou, Y an Zeng, and Daquan Zhou. Rethinking video generation model for the embodied world, on Pith
  11. [11] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Y usuf Aytar, Michael Rubinstein, Chen Sun, Oliver Wang, Andrew Owens, and Deqing Sun. Motion Prompting: Controlling Video Generation with Motion Tra- je... on Pith
  12. [12] Joel Jang, Seonghyeon Y e, Zongyu Lin, Jiannan Xiang, Jo- han Bjorck, Y u Fang, Fengyuan Hu, Spencer Huang, Kaushil Kundalia, Y en-Chen Lin, Loïc Magne, Ajay Mandlekar, Avnish Narayan, Y ou Liang Tan, Guanzhi Wang, Jing Wang, Qi Wang, Yinzhen Xu, Xiaohui Zeng, Kaiyuan Zheng, R... on Pith
  13. [13] Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Y e, Niloy Mitra, and Duygu Ceylan. Track4Gen: Teaching Video Dif- fusion Models to Track Points Improves Video Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3 on Pith
  14. [14] Zihang Lai, Eldar Insafutdinov, Edgar Sucar, and Andrea V edaldi. CoWTracker: Tracking by Warping instead of Cor- relation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. 6 on Pith
  15. [15] Y ue Liao, Pengfei Zhou, Siyuan Huang, Donglin Y ang, Shengcong Chen, Y uxin Jiang, Y ue Hu, Jingbin Cai, Si Liu, Jianlan Luo, Liliang Chen, Shuicheng Y an, Maoqing Y ao, and Guanghui Ren. Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation. arXiv pre... on PithPith's reviewarXiv
  16. [16] NVIDIA. Cosmos World Foundation Model Platform for Physical AI. arXiv preprint arXiv:2501.03575, 2025. 2 on PithPith's reviewarXiv
  17. [17] NVIDIA. Cosmos 3: Omnimodal World Models for Physical AI. arXiv preprint arXiv:2606.02800, 2026. 6, 7 on PithPith's reviewarXiv
  18. [18] RECCAR. RecCAR: Closing the Cross-Attention Gap in Joint Video Generation. arXiv preprint arXiv:2609.27901, on PithPith's reviewarXiv
  19. [19] Y u Shang, Xin Zhang, Yinzhou Tang, Lei Jin, Chen Gao, Wei Wu, and Y ong Li. RoboScape: Physics-informed Embodied World Model. In Advances in Neural Information Processing Systems, 2025. 2 9 on Pith
  20. [20] Xiaoyu Shi, Zhaoyang Huang, Fu-Y un Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Motion-I2V: Consistent and Controllable Image-to-Video Generation with Explicit Motion Modeling. Research pub- lica... on Pith
show 15 more
  1. [21] Rongze Tang, Jianjie Fang, Zhaolu Wang, Ziyou Wang, Xvyuan Liu, Haisheng Su, Xin Zhang, Wei Wu, Chen Gao, Y ong Li, and Zhibo Chen. IMPACT: Attention Is the Interac- tion Map for Scalable Interaction-Aware World Model Train- ing. arXiv preprint arXiv:2609.00161, 2026. 1, 2 on PithPith's reviewarXiv
  2. [22] Team Wan, Ang Wang, Baole Ai, et al. Wan: Open and Advanced Large-Scale Video Generative Models. arXiv preprint arXiv:2503.20314, 2025. 6, 7 on PithPith's reviewarXiv
  3. [23] VPT. Enhancing Video Physical Consistency via Role-aware Joint Training and Modality-decoupled Denoising. arXiv preprint arXiv:2607.04653, 2026. 2, 3 on PithPith's reviewarXiv
  4. [24] Shuai Wang, Y axin Feng, Xuekun Jiang, Shihan Tian, Ningyu Y an, Xing Shen, Chaoyang Lyu, Hui Wang, Y un- song Zhou, Hanqing Wang, Jiangmiao Pang, Y ang Xiang, Xing Gao, Chunhua Shen, and Weinan Zhang. GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation... on PithPith's reviewarXiv
  5. [25] Zhouxia Wang, Ziyang Y uan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A Uni- fied and Flexible Motion Controller for Video Generation. In ACM SIGGRAPH 2024 Conference Papers, 2024. 3 on Pith
  6. [26] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V . Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Pro- cessing Systems, 2022. 2 on Pith
  7. [27] Weijia Wu, Zhuang Li, Y uchao Gu, Rui Zhao, Y efei He, David Junhao Zhang, Mike Zheng Shou, Y an Li, Tingting Gao, and Di Zhang. DragAnything: Motion Control for Anything Using Entity Representation. In Computer Vision – ECCV 2024, pages 331–348, 2024. 3 on Pith
  8. [28] Zhexiao Xiong, Yizhi Song, Liu He, Wei Xiong, Y u Y uan, Feng Qiao, and Nathan Jacobs. PhysAlign: Physics-Coherent Image-to-Video Generation through Fea- ture and 3D Representation Alignment. arXiv preprint arXiv:2603.13770, 2026. 2 on PitharXiv
  9. [29] Sherry Y ang, Yilun Du, Seyed Ghasemipour, Jonathan Tomp- son, Leslie Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning Interactive Real-World Simulators. In Interna- tional Conference on Learning Representations , 2024. 2 on Pith
  10. [30] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUW A: Fine-grained Control in Video Generation by Integrating Text, Image, and Trajectory. arXiv preprint arXiv:2308.08089, 2023. 3 on PithPith's reviewarXiv
  11. [31] Peiwen Zhang, Y ufan Deng, Shangkun Sun, Juncheng Ma, Duomin Wang, Jonas Du, Zilin Pan, Y e Huang, Hao Liang, Songyan Huang, Ruihua Zhang, Enze Xie, Ming-Y u Liu, and Daquan Zhou. PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation. arXiv preprint arXiv:... on PithPith's reviewarXiv
  12. [32] Xiangdong Zhang, Jiaqi Liao, Shaofeng Zhang, Fanqing Meng, Xiangpeng Wan, Junchi Y an, and Y u Cheng. Vide- oREPA: Learning Physics for Video Generation through Re- lational Alignment with Foundation Models. In Advances in Neural Information Processing Systems, 2025. 2 on Pith
  13. [33] Zhenghao Zhang, Junchao Liao, Menghao Li, Zuozhuo Dai, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented Diffusion Transformer for Video Gener- ation. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2025. 3 on Pith
  14. [34] Siyuan Zhou, Yilun Du, Jiaben Chen, Y andong Li, Dit-Y an Y eung, and Chuang Gan. RoboDreamer: Learning Composi- tional World Models for Robot Imagination. In Proceedings of the 41st International Conference on Machine Learning , pages 61885–61896, 2024. 1, 2 on Pith
  15. [35] Fangqi Zhu, Hongtao Wu, Song Guo, Y uxiao Liu, Chilam Cheang, and Tao Kong. IRASim: A Fine-Grained World Model for Robot Manipulation. arXiv preprint arXiv:2406.14540, 2024. 1, 2 10 on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment