acceptodds
Desk rejected at ICLR 2027

Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Video world models are expected to predict how the physical world evolves, yet they often produce visually plausible videos that violate basic physical principles. Existing approaches commonly assume that natural language is insufficient to represent the physical knowledge required for reliable generation, and therefore introduce additional visual, latent, numerical, or planning-based signals. We revisit this assumption and introduce Physis-Lang, a self-evolving framework that treats physical language as a shared and optimizable representation across data curation, model training, and video generation. Physis-Lang represents physical processes through language that describes their relevant entities, causes, interactions, governing principles, temporal evolution, and effects. To improve this representation, we construct PhysCapBench, which decomposes physical processes into atomic assertions and evaluates captions using recall and precision. An agentic loop iteratively analyzes assertion-level errors and refines the instruction used to produce physical captions. Physis-Lang further converts model deficiencies into textual descriptions and uses language-guided retrieval to identify visually diverse videos that cover missing physical processes. Experiments on four physical video benchmarks with Wan and Cosmos backbones demonstrate consistent improvements in physical plausibility. Notably, starting from open-source Cosmos backbones, our Physis-Lang-enhanced models surpass the leading proprietary Veo 3.1 model.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video generation, world model, agents, physics-aware video generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_aPGGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Physis-Lang, a self-evolving framework that leverages physical language as a shared representation for data curation, model training, and video generation. The framework aims to improve the physical understanding of video world models through language-guided data retrieval and physical-aware learning.

Critical strengths

1、The proposed Physis-Lang framework provides a unified perspective by introducing physical language as a shared representation across data curation, model training, and video generation, which is a meaningful attempt to bridge physical reasoning and generative modeling.

2、The proposed language-guided data curation strategy, physical language representation, and self-evolving training paradigm are evaluated through multiple experiments, demonstrating the effectiveness of the framework in enhancing physical consistency and video generation capabilities.

Critical weaknesses

1、It is unclear whether the 71K videos selected from WISA-80K are obtained through the proposed language-guided retrieval strategy or through other filtering procedures. The authors should clarify the relationship between the proposed retrieval mechanism and the final training data construction pipeline.

2、To better demonstrate the effectiveness of the proposed language-guided retrieval mechanism, the authors could conduct additional experiments on a public physics-related video dataset. By retrieving different amounts of data using the proposed strategy and evaluating their impact on model training, the authors could more convincingly verify that the proposed method can select informative physical samples rather than relying on manually curated data.

3、The source and composition of the retrieval corpus used in the deficiency-guided retrieval pipeline are unclear. Although the authors state that 112K videos are collected through the proposed retrieval strategy, it is not specified from which video database these samples are retrieved, how large the candidate pool is, and how the retrieval corpus is constructed. More details are needed to better understand and reproduce the data construction process.

4、The proposed self-evolving retrieval pipeline relies on accurately identifying model deficiencies. However, the paper does not provide sufficient analysis of the deficiency diagnosis process. Since the identified deficiencies are directly used to generate retrieval queries, inaccurate diagnosis may lead to suboptimal data augmentation. The authors are encouraged to evaluate the reliability of deficiency identification and analyze how detection errors affect the retrieval results and model performance.

Questions for authors

see Weaknesses

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer aPGG (1/2)

W1: The authors should clarify the relationship between the proposed retrieval mechanism and the final training data construction pipeline.

Language-guided retrieval mechanism and general-purpose data filtering are two independent procedures in our training data construction pipeline. The 71K videos selected from WISA-80K are obtained through general quality filtering, not through the proposed language-guided retrieval mechanism.

Specifically, before training, we apply a general-purpose filtering stage to process training data. We compute motion scores using UniMatch optical flow [1] and aesthetic scores using the LAION aesthetic predictor [2], retaining only clips satisfying both criteria: (1 ≤ motion_score ≤ 70) and (aesthetic_score ≥ 0.7) (normalized to [0, 1]). We further use PySceneDetect [3] to identify and remove multi-shot videos, and exclude samples with incomplete metadata. This process yields 71K high-quality WISA videos from original WISA-80K. This filtering procedure was briefly described in Appendix B.5.1 of the original manuscript, and we have expanded this section in the revised manuscript to provide more detailed implementation details.

The two components play complementary roles: language-guided retrieval identifies videos targeting physical deficiencies in the pre-trained generator, while general-purpose filtering ensures the quality of training videos from both sources.

W2: Retrieving different amounts of data using the proposed strategy and evaluating their impact on model training.

To further validate the effectiveness and generalizability of our proposed language-guided retrieval mechanism, we conduct additional experiments using PE-Video[4][5], which is a public video dataset consisting of 998k samples, independent of the original retrieval gallery. We re-run our language-guided retrieval pipeline on this new gallery to identify physics-informative training videos.

Specifically, we apply our retrieval pipeline to 14 physical-deficiency categories (we added Appendix I in the revised paper to visualize them), retrieving at most 10K videos per category. This yields a total of 117K newly retrieved videos. To further investigate the effect of the amount of retrieved training data, we uniformly sample 20K videos from this collection to construct a smaller training subset.

We keep the WISA-only training set (71K videos) fixed and vary the amount of newly retrieved data, comparing three settings: (1) WISA-only, without additional retrieved data; (2) WISA-only + Retrieved (20K), incorporating 20K newly retrieved videos; and (3) WISA-only + Full Retrieved (117K), incorporating the full set of 117K newly retrieved videos.

The results are summarized below:

Training DataPhyGenBenchPhysics-IQ Verified
WISA-only68.1240.68
WISA-only + Retrieved (20K)69.7941.56
WISA-only + Full Retrieved (117K)71.6742.11

The results demonstrate two key findings:

  1. Effectiveness on a new public gallery. Incorporating only 20K newly retrieved videos improves performance over the WISA-only baseline by +1.67 points on PhyGenBench and +0.88 points on Physics-IQ Verified. Importantly, these videos are retrieved automatically from a new gallery, providing additional evidence that our method can identify useful physics-related training samples.

  2. Consistent improvements with increasing retrieval scale. Increasing the retrieved data from 20K to 117K further improves performance, yielding overall gains of +3.55 points on PhyGenBench and +1.43 points on Physics-IQ Verified compared with the WISA-only baseline. The consistent improvements across both benchmarks demonstrate the benefits of incorporating additional physics-informative samples identified through our retrieval mechanism.

Overall, these experiments on a new, independently constructed public-data gallery provide further evidence of the effectiveness and generalizability of our language-guided retrieval mechanism, demonstrating that it can identify useful physics-related training data beyond the original retrieval gallery.

[1] H. Xu et al. Unifying Flow, Stereo and Depth Estimation, IEEE TPAMI, 2023.

[2] C. Schuhmann et al. LAION Aesthetic Predictor, 2022.

[3] B. Castellano et al. PySceneDetect, 2014.

[4] Daniel Bolya et al. Perception Encoder: The best visual embeddings are not at the output of the network, arXiv, 2025

[5] Jang Hyun Cho et al. PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding, arXiv, 2025

AuthorsOpenReview ↗

Author Response to Reviewer aPGG (2/2)

W3: Which video database these samples are retrieved, how large the candidate pool is, and how the retrieval corpus is constructed.

The candidate pool consists of 4M fetched YouTube video clips. They cover diverse categories, including conversations (65.4%), travel and outdoor scenes (11.8%), lifestyle and daily activities (10.7%), entertainment and media (5.5%), animals (3.4%), and educational and instructional content (3.3%).

The retrieval training dataset contains 111,792 unique videos across 14 physical-deficiency categories. We construct the dataset by processing categories in ascending order of their available candidate pool sizes, starting from the least represented category (low gravity), and collecting up to 10K unique videos per category. Its composition is as follows: human body motion, light, contact collision, rigid body motion, fluid dynamics, cloth deformation, physical material properties, and buoyancy each contribute 10K videos (8.95% per category, 71.56% combined). The remaining categories include chemical processes (9.4K, 8.41%), thermal processes (7.7K, 6.90%), fracture mechanics (6.2K, 5.56%), soft-body deformation (5.8K, 5.17%), elasticity (2.7K, 2.37%), and low gravity (26 videos, 0.02%). More details of how we constructed the 112K dataset are in Section 3.2.

W4: The authors are encouraged to evaluate the reliability of deficiency identification.

In our pipeline, the diagnosis results were manually reviewed before being used to construct retrieval queries. This human verification step ensured that the reported deficiency in each category was genuinely present in the corresponding video and prevented clearly inaccurate diagnoses from propagating into the retrieval and data-augmentation stages.

To demonstrate examples for the verification process, we have added Appendix I to the revised manuscript. The appendix presents a representative example for each deficiency category using six temporally ordered key frames, together with a brief explanation of the observed physical inconsistency. These examples provide direct evidence for the validity of the diagnosed deficiencies and clarify how each category is identified.

Paper revision

We have revised the manuscript, with all changes highlighted in blue. Specifically, we clarified the training data construction and filtering pipeline in Appendix B.5.1, and expanded Appendix I with examples of manually verified physical deficiencies.

Dear Reviewer aPGG,

With about one day left in the discussion period, we wanted to follow up on our rebuttal. Following your suggestions, we added retrieval experiments on the independent public PE-Video dataset at two data scales (20K and 117K), with improvements on both evaluated benchmarks. We also clarified the WISA filtering procedure, the retrieval corpus and its composition, and the manual verification of diagnosed physical deficiencies.

We would appreciate hearing whether these experiments and clarifications address your concerns. If any questions remain, we would be happy to discuss them further before the deadline.

Thank you again for your constructive feedback.

Best regards,

The Authors

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

This paper proposes Physis-Lang, a framework that leverages self-evolving physical language to improve the physical plausibility of video generation. The proposed framework integrates physics-aware caption refinement, language-guided data retrieval, and physical reasoning. Extensive experiments demonstrate its effectiveness across multiple benchmarks and model architectures.

Critical strengths

(1)This paper introduces a novel and well-motivated perspective on using language as a unified representation of physical knowledge. (2)The proposed framework is well-designed, and the construction of PhysCapBench provides additional value. (3)Extensive evaluations and ablation studies demonstrate consistent improvements across different benchmarks and model backbones.

Critical weaknesses

(1)The reliance on GPT-based evaluation could benefit from additional validation against human judgments. (2)Further comparisons with random data selection would better demonstrate the advantages of the proposed retrieval strategy.

Questions for authors

See Weekness.

2 comments

W1: The reliance on GPT-based evaluation could benefit from additional validation against human judgments.

Setup. We ran a blind pairwise study on 40 randomly sampled video pairs (20 from VideoPhy-2 and 20 from PhyGenBench), each generated by pretrained Cosmos3 and our Physis-Lang. Five CS PhD students independently chose the more physically plausible video in each pair or a tie, with the left/right order randomized. For each pair, the video that receives more votes is taken as the human preference; if both receive the same number of votes, the pair is considered as a tie.

Human scoring and preference for Physis-Lang. Each benchmark contributed 100 votes (20 video pairs × 5 annotators). On VideoPhy-2, 41% of votes favored Physis-Lang, 28% favored pretrained Cosmos3, and 31% indicated a tie, yielding a net preference of +13 percentage points for Physis-Lang. On PhyGenBench, 51% favored Physis-Lang, 33% favored pretrained Cosmos3, and 16% indicated a tie, yielding a net preference of +18 percentage points. In the blind study, Physis-Lang received more preference votes than pretrained Cosmos3 on both benchmarks.

Agreement among human annotators. A majority of five evaluators prefers one video in 35 of the 40 pairs (17 of 20 on VideoPhy-2 and 18 of 20 on PhyGenBench; the other 5 pairs are tie). In all of these 35 pairs, at most one annotator chose the opposite video that the majority did not prefer. Therefore, it shows our human judgments are almost aligned.

The alignment between human raters and GPT-5.5. We utilize the sign test to quantify the alignment. The test assumes that for each pair, the evaluations of humans and GPT-5.5 are equally likely to agree or disagree. Under this assumption, the probability of getting the actual result is p=3.4×10−4p = 3.4 \times 10^{-4}, well below the common threshold of extremely strong alignment (0.001)(0.001).

We have added and highlighted these details for this study in App. G. Further, in the examples in Fig. 13, GPT-5.5 not only gives accurate scores, but also explains the proper reason for scoring, which matches human judgment.

W2: Further comparisons with random data selection

We trained Cosmos3-Nano on WISA augmented with the same amount of randomly sampled videos from the same retrieval pool. We compared this model with the WISA-only and language-guided retrieval results reported in the paper.

Benchmark ↑WISA+ RandomΔ vs. WISA+ Language-GuidedΔ vs. WISALanguage-Guided − Random
PhyGenBench68.1267.50−0.6271.04+2.92+3.54
Physics-IQ Verified40.6840.57−0.1143.41+2.73+2.84

Δ denotes the absolute score difference.

  • Language-guided retrieval outperforms random selection on both benchmarks. PhyGenBench increases by 3.54 points and Physics-IQ Verified increases by 2.84 points.
  • Random augmentation does not improve upon WISA alone. PhyGenBench decreases from 68.12 to 67.50, while Physics-IQ Verified remains nearly unchanged, decreasing from 40.68 to 40.57.

Paper revision

All corresponding changes are highlighted in blue in the revised PDF. We have added the human study setup, results, and detailed analysis in Appendix G.

Dear Reviewer vVMr,

With about one day left in the discussion period, we wanted to follow up on our rebuttal. Following your suggestions, we added a blind human study to assess alignment with human judgments and a controlled comparison with random data selection, using the same number of videos from the same retrieval pool. Language-guided retrieval outperformed random selection on both evaluated benchmarks.

We would appreciate hearing whether these additions address your concerns. If anything remains unclear, we would be happy to clarify further before the deadline.

Thank you again for your helpful feedback.

Best regards,

The Authors

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

This paper proposes Physis-Lang, a framework that uses structured physical language as a shared representation across data curation, model training, and inference for video generation. An agentic loop refines the instruction for writing physics captions, guided by a new benchmark, PhysCapBench, and a physics-aware critic. The resulting captions are used to build a 183K-video training set. Experiments on four physical video benchmarks show consistent gains for Wan2.1 and the Cosmos3 family, and the fine-tuned Cosmos3-Nano outperforms Veo 3.1 on three of the four benchmarks.

Critical strengths

The idea is simple and practical: the method improves physical plausibility without changing model architectures or training objectives, which makes it easy to apply to different video generators. The experiments are thorough. The evaluation covers four widely used benchmarks for both text-to-video and image-to-video generation, two model families, and three model scales from 4B to 64B. It also compares against strong open-source models, a leading proprietary model, and several recent physics-focused methods. The ablations are well organized and look at prompting, fine-tuning, negative prompts, retrieved data, and the evolution iterations separately.

Critical weaknesses

W1. The physics reasoning at test time comes from GPT-5.5 rather than from the video model. All contributions of the paper center on physics captions, while the four benchmarks only provide short official prompts. According to Sec. 3.1 and App. B.2, the captioning instruction is optimized once. At test time, a GPT-5.5 upsampler then expands each benchmark prompt using this fixed instruction. This design is reasonable, but it creates a gap between training and testing. During training, the captioner sees the full video and describes what actually happened. At test time, the upsampler only sees the prompt and the first frame, so the physics reasoning is GPT-5.5’s prediction of the outcome. In Fig. 8, the upsampler often writes down the very answers that the benchmarks are designed to test, such as the violet flame of burning potassium, the stone sinking into the water, and the cup collapsing under the kettlebell. The negative prompts also rule out the wrong outcomes. As a result, the reported scores reflect the combined system of GPT-5.5 and the video model. Tab. 7 shows that on PhyGenBench, prompting alone already brings +5.62, while fine-tuning adds +3.75 on top of that. Since the baselines, including Veo 3.1, do not receive the same text, the comparison may not be fully fair.

W2. The replacement of the official evaluators lacks human validation. For VideoPhy-2 and PhyGenBench, the official evaluators are replaced with GPT-5.5. App. G supports this choice with two qualitative examples and with the observation in Tab. 10 that GPT-5.5 spreads the scores more widely. However, being more discriminative does not necessarily mean being more accurate. The claim that GPT-5.5 agrees with human judgment is also not supported by quantitative evidence. Since the main conclusions rely on these scores, this is an important gap.

W3. GPT-5.5 plays several roles in the pipeline, which may bias the evaluation. GPT-5.5 serves as the captioner, the upsampler, the diagnosis agent, and the evaluator. Both the training captions and the inference prompts are written by GPT-5.5, and the generated videos are then scored by GPT-5.5. The evaluator may therefore prefer videos that match its own descriptions and style. This could enlarge the gap between the proposed method and the baselines.

W4. The largest gain in caption quality may come from a higher frame rate rather than from instruction evolution. On PhysCapBench, the F1 score rises from 78.64 to 87.82, a total gain of 9.18 points. Of this, 6.19 points come from the step between iteration 8 and iteration 9. Fig. 6 shows that iteration 9 also raised the frame sampling rate from 2 fps to 4 fps. Over iterations 1 to 8, the score mostly stays between 78 and 82. It is therefore unclear how much of the improvement comes from the evolved instruction itself and how much comes from denser input frames.

Questions for authors

Q1. On the source of test-time physics reasoning. I would appreciate it if the authors could clarify two points: whether each benchmark prompt is upsampled once or sampled several times and then selected, and what inputs the upsampler receives for each benchmark. It would also be very helpful to report additional results:

The fine-tuned model with an upsampler that only describes the scene and does not predict the outcome. The baseline, like Veo 3.1 given the same physics reasoning. The fine-tuned models using only the official prompts. On Physics-IQ Verified, how often the upsampler predicts the correct outcome, and how the generation scores differ when the prediction is right or wrong.

Together, these results would show how much of the gain the video model has learned itself and how much is provided by GPT-5.5 at test time.

Q2. On human validation of the evaluator. It would greatly strengthen the paper to run a small blind human study. Reporting the human scores, the agreement among annotators, and the agreement between GPT-5.5 and the human ratings would make the evaluation much more convincing.

Q3. On evaluator bias. I suggest re-scoring a subset of the generated videos with an evaluator that is not used anywhere else in the pipeline. Since Gemini already serves as the caption critic, a strong open-source VLM may be a better choice. Reporting its agreement with GPT-5.5, and whether the ranking of the methods stays the same, would help rule out this concern. The human study in Q2 could also serve as a reference here.

Q4. On the frame-rate change at iteration 9. Can you make an ablation that separates the two factors. This would show whether the self-evolution loop itself brings a clear gain.

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer uTxe (1/2)

Q1: Ablation on Test-time Upsample

Whether each benchmark prompt is upsampled once or several times and then selected, and what inputs the upsampler receives for each benchmark.

Each prompt is upsampled exactly once, with no candidate sampling or selection. The upsampler receives the official prompt of each benchmark. For the I2V benchmarks, it also receives the switch frame (Physics-IQ Verified) or the first frame (PhyGround); PhyGenBench and VideoPhy-2 are text-only. The upsampler never sees any ground-truth future frame.

The fine-tuned model with an upsampler that only describes the scene and does not predict the outcome. The fine-tuned models using only the official prompts.

We examine the contributions of generator fine-tuning and physics reasoning. Under the same official base prompt, our fine-tuned model improves the score from 40.23 to 42.27 (+2.04), demonstrating that the generator itself benefits from the physical knowledge acquired during training. Further upsampling this prompt with our physics_reasoning, without the additional outcome-prediction, increases the score to 42.85 (+0.58).

The baseline, like Veo 3.1 given the same physics reasoning.

We evaluate Veo 3.1 using Physis-Lang prompts. Its score improves from 34.99 with the official prompt to 36.71 with Physis-Lang (+1.72), demonstrating that our physics-aware prompt upsampling is beneficial to different video generation models. Under the same full Physis-Lang prompts, our fine-tuned Cosmos3-Nano achieves 43.41, outperforming Veo 3.1 by 6.70 points.

On Physics-IQ Verified, how often the upsampler predicts the correct outcome, and how the generation scores differ when the prediction is right or wrong.

We manually examine all 66 scenes in Physics-IQ Verified: 62 of the 65 determinable predictions are completely correct, 3 are partially correct, and none is incorrect; one case is undeterminable. The generated videos in all four partially determined or ambiguous cases follow the predicted outcomes. We further randomly inspect 10 correctly predicted cases and observe that the generated videos also closely follow the upsampled captions, confirming the influence of the upsampler's physical reasoning on video generation.

Q2: Reporting the human scores, the agreement among annotators, and the agreement between GPT-5.5 and the human ratings

Setup. We ran a blind pairwise study on 40 randomly sampled video pairs (20 from VideoPhy-2 and 20 from PhyGenBench), each generated by pretrained Cosmos3 and our Physis-Lang. Five CS PhD students independently chose the more physically plausible video in each pair or a tie, with the left/right order randomized. For each pair, the video that receives more votes is taken as the human preference; if both receive the same number of votes, the pair is considered as a tie.

Human scoring and preference for Physis-Lang. Each benchmark contributed 100 votes (20 video pairs × 5 annotators). On VideoPhy-2, 41% of votes favored Physis-Lang, 28% favored pretrained Cosmos3, and 31% indicated a tie, yielding a net preference of +13 percentage points for Physis-Lang. On PhyGenBench, 51% favored Physis-Lang, 33% favored pretrained Cosmos3, and 16% indicated a tie, yielding a net preference of +18 percentage points. In the blind study, Physis-Lang received more preference votes than pretrained Cosmos3 on both benchmarks.

Agreement among human annotators. A majority of five evaluators prefers one video in 35 of the 40 pairs (17 of 20 on VideoPhy-2 and 18 of 20 on PhyGenBench; the other 5 pairs are tie). In all of these 35 pairs, at most one annotator chose the opposite video that the majority did not prefer. Therefore, it shows our human judgments are almost aligned.

The alignment between human raters and GPT-5.5. We utilize the sign test to quantify the alignment. The test assumes that for each pair, the evaluations of humans and GPT-5.5 are equally likely to agree or disagree. Under this assumption, the probability of getting the actual result is p=3.4×10−4p = 3.4 \times 10^{-4}, well below the common threshold of extremely strong alignment (0.001)(0.001).

We have added and highlighted these details for this study in App. G. Further, in the examples in Fig. 13, GPT-5.5 not only gives accurate scores, but also explains the proper reason for scoring, which matches human judgment.

AuthorsOpenReview ↗

Author Response to Reviewer uTxe (2/2)

Q3: Re-scoring a subset of the generated videos with an evaluator that is not used anywhere else in the pipeline

To further examine potential evaluator bias, we introduce Qwen3.8-27B, an independent open-source VLM that is not involved in any other component of our pipeline, as an additional evaluator. We evaluate the same video generators using both GPT-5.5 and Qwen3.8-27B, with results summarized below.

Video GeneratorPhyGenBench (GPT-5.5)PhyGenBench (Qwen)VideoPhy-2 All (GPT-5.5)VideoPhy-2 All (Qwen)VideoPhy-2 Hard (GPT-5.5)VideoPhy-2 Hard (Qwen)
Veo 3.165.6358.3368.8772.4258.4363.48
Cosmos3-Nano61.6756.2560.4165.3148.3152.25
Physis-Lang71.0462.7168.0271.2462.3664.61

The two evaluators produce exactly the same ranking of the three video generators across all three evaluation metrics:

  • PhyGenBench: Physis-Lang > Veo 3.1 > Cosmos3-Nano.
  • VideoPhy2 All: Veo 3.1 > Physis-Lang > Cosmos3-Nano.
  • VideoPhy2 Hard: Physis-Lang > Veo 3.1 > Cosmos3-Nano.

In particular, both evaluators consistently rank Physis-Lang above the pretrained Cosmos3-Nano baseline across all metrics. These results suggest that our main conclusions are robust to the choice of evaluator, mitigating the concern that the observed improvements are specific to GPT-5.5.

Furthermore, our human study in Q2 provides an independent reference for evaluator reliability. This provides additional evidence that the evaluation results are broadly aligned with human judgments.

Q4: Can you make an ablation that separates the two factors.

To isolate the effect of instruction evolution, we re-evaluate Iteration 9 at the original 2 FPS. Its F1 score reaches 85.20, compared with 78.64 for Iteration 1 at the same frame rate, yielding a 6.56-point improvement. Increasing the frame rate of Iteration 9 from 2 FPS to 4 FPS further improves the F1 score to 87.82 (+2.62 points). Notably, the frame-rate adjustment was itself proposed by the evolution agent based on its diagnosis, demonstrating that the self-evolution loop can discover improvements not only to the instructions but also to the input configuration. We have included this ablation in Appendix A.

Paper revision

We have revised the manuscript based on the reviewer's suggestions, with all corresponding changes highlighted in blue in the revised PDF. Specifically, we have added the human study details and results in Appendix G. We have also included the frame-rate ablation in Appendix A.

Dear Reviewer uTxe,

With about one day left in the discussion period, we wanted to follow up on our rebuttal. Following your suggestions, we added prompt and frame-rate ablations, a comparison with Veo 3.1 using the same physics-aware prompts, a blind human study, and an independent evaluation with Qwen to examine potential evaluator bias.

We would appreciate hearing whether these additional experiments and clarifications address your concerns. If any questions remain, we would be happy to discuss them further before the deadline.

Thank you again for your thoughtful and constructive feedback.

Best regards,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.