acceptodds
Under review as a conference paper at ICLR 2027

EVEWorld: Physical Evolution Supervision for Embodied World Models

Abstract

Embodied world models enable scalable simulation of embodied interactions for robot learning. However, existing models are prone to Model Laziness, as they focus on visual fidelity at the expense of physical reasoning and lack process-level supervision over the temporal dynamics of manipulated objects. In this work, we propose EVEWorld, a physical evolution-supervision framework for physically consistent target evolution. EVEWorld consists of two components: Instance-Guided Restoration (IGR) and Temporal Instance Alignment (TIA). First, IGR promotes instance consistency through restoration supervision. Second, TIA promotes cross-frame consistency by aligning target instances across adjacent frames. We further introduce the Model Laziness Rate (MLR), a metric that measures persistent violations of instance consistency in generated trajectories. Extensive experiments on DreamGenBench, EWMBench, and PBench demonstrate the effectiveness of EVEWorld, notably achieving an 87.5% reduction in MLR compared with GigaWorld-0. On the WorldArena 2.0 Track 1 leaderboard, our model ranks 6th in JEPA Similarity and 17th overall, which further validates the performance of our evolution supervision strategy. Source code is available at an anonymous link: https://anonymous.4open.science/r/EVEWorld-release-C8B4.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, Embodied AI, Video Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Qyz7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies failures of embodied video world models in robot manipulation, where the manipulated target is spuriously duplicated, disappears, or deforms across frames. The authors refer to violations of target-instance count conservation as “Model Laziness” and propose EVEWorld, which consists of two components. Instance-Guided Restoration (IGR) copy-pastes an additional target instance into demonstration videos and trains the model to restore the clean video with a region-weighted denoising objective. Temporal Instance Alignment (TIA) performs local feature matching, cross-frame feature transport, and correspondence supervision at a selected Transformer layer. The detector and target localizations serve only as offline training supervision; at generation time, TIA operates directly on the model’s hidden features.

The authors further propose the Model Laziness Rate (MLR), which uses the initial target count as its reference, exempts under-counts caused by robot occlusion, and records a violation only when it persists over two consecutive sampled frames. The main experiments fine-tune GigaWorld-0 on 92 GR1 videos and evaluate on DreamGenBench, with additional experiments on WorldArena, EWMBench, PBench, and FlowWAM/RoboTwin. Compared with a Standard SFT baseline trained under the same budget, EVEWorld reduces MLR on DreamGenBench from 7/63 to 1/63 and improves Gemini-IF from 53.57% to 60.85%. On WorldArena 1.0, MLR drops from 43/157 to 21/157.

Critical strengths
  1. A well-motivated problem with a targeted solution. The paper studies a failure mode of clear practical relevance: when world models are used as data engines or policy evaluators, duplicated or vanishing objects are a real obstacle. The motivation is concrete, and each component maps directly onto the targeted failure. IGR turns an extra instance and its clean counterpart into a paired restoration task. TIA addresses the cross-frame inconsistencies that can remain even when the instance count is preserved.

  2. Broad evaluation with well-matched controls. The main paper evaluates on multiple benchmarks that cover in-domain performance, cross-domain and cross-distribution generalization, a second backbone, and robustness across random seeds. Each comparison uses a Standard SFT control with the same initialization, data, and training budget, so the gains can be attributed to the proposed supervision rather than to the training setup.

Critical weaknesses
  1. IGR may delete legitimate instances, and the full method does not seem to transfer well to other backbones. IGR is trained only to remove synthetically added instances, yet the paper reports no retention or false-deletion rates in real multi-instance scenes (N₀ > 1), nor whether the evaluation set contains such cases. Moreover, although Sec. 5.3 cites FlowWAM as evidence of cross-backbone generalization, TIA alone is best on every metric there (MLR 22.54%), and adding IGR degrades all five (MLR 35.21%; Table 5).

  2. The core evidence gives limited support to the main conclusions, and the reliability of the current MLR computation is questionable. MLR on DreamGenBench is computed on only 63 eligible samples (Table 1 caption), so the 85.7% and 87.5% reductions correspond to violation events falling from 7 to 1 and from 8 to 1. MLR is also computed automatically with a single detector, GroundingDINO, and the paper reports no quantitative agreement (e.g., precision and recall) with human labels. The lower MLR therefore cannot be clearly attributed to fewer genuine physical violations.

Questions for authors
  1. (Clarification) Could the authors analyze whether IGR training causes the model to erroneously delete legitimate instances? Reporting how many of the eligible samples have N₀ > 1 would help assess this.

  2. (Suggested experiment) Could the authors report MLR on DreamGenBench with sufficient statistical power, for example by using more random seeds? The same question applies to the WorldArena 1.0 comparison in Table 3.

  3. (Suggested experiment) Could the authors validate MLR with human inspection and annotation? As a cheaper alternative, does the ranking of models stay the same when MLR is computed with a different detector that is not used during training?

  4. (Clarification) In Table 6, IGR alone lowers overall Gemini-IF (51.85 vs. 53.57 for Standard SFT), and TIA alone raises it only slightly (55.29), yet their combination reaches 60.85, with Env jumping from 51.72 to 72.41. In addition, on FlowWAM (RoboTwin, Table 5), TIA alone performs best on every metric (MLR 22.54%), while adding IGR to TIA degrades all five metrics (MLR 35.21%). Why does combining the two components bring a large gain on GigaWorld-0 but perform worse than TIA alone on FlowWAM?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Qyz7 Weaknesses

We're truly grateful for your thoughtful feedback and support. Below, we've done our best to respond to each of your points.

W1. Instance Preservation and Cross-Backbone Generalization

We thank the reviewer for raising this important point. We directly audited legitimate-instance preservation in multi-instance scenes with three independent human annotators (Table 21). Across WorldArena 1.0/2.0, the 55 shared N0 > 1 inputs yield 165 videos across three generators per benchmark. Majority-vote labels exclude valid visibility changes and instruction-required removal.

BenchmarkMulti-instance inputsReleased lossSFT lossEVEWorld loss
WorldArena 1.023/1003/234/233/23
WorldArena 2.032/1006/324/325/32

Released denotes GigaWorld-0/FlowWAM. EVEWorld matches or improves upon these baselines and matches SFT's pooled multi-instance loss count (8/55). Across 326 audited videos/model, disappearance decreases by 60.0% (25 to 10), alongside fewer duplications (45 to 24; Table 20).

We compared FlowWAM LoRA ranks 32/64/128 and full fine-tuning on the same 2,250 training episodes. E7 results (1,974 steps; TIA -> EVEWorld; Table 56):

AdaptationPSNR ↑SSIM ↑LPIPS ↓Flow-EPE ↓MLR (%) ↓
Rank-3213.320 -> 12.6110.773 -> 0.7610.346 -> 0.3801.945 -> 2.32424.13 -> 37.42
Full FT14.059 -> 14.1620.797 -> 0.8010.280 -> 0.2781.369 -> 1.42417.32 -> 16.39

The EVEWorld-minus-TIA MLR gap shrinks from +13.29 to +0.41 and +0.03 percentage points at ranks 32/64/128, then reverses to -0.93 under full FT. Under full FT, EVEWorld improves PSNR, SSIM, LPIPS, and MLR over TIA at E7 (Table 56). EVEWorld has lower MLR than TIA at every full-FT checkpoint from E4 to E8; TIA retains lower Flow-EPE, and IGR-only has slightly lower MLR at E5/E7/E8 (Table 55). At E6, EVEWorld improves all five metrics over matched SFT, including MLR from 39.45% to 15.81% (Table 5).

The joint benefit of IGR and TIA depends on adaptation capacity. Q4 gives the GigaWorld component ablation, FlowWAM checkpoint trajectories, and a mechanism interpretation.

EVEWorld has fewer violations than SFT at every tested GigaWorld step from 150 to 500 (Table 35); Q4 gives the complete IF sweep.

W2. Sample Size and Evaluation Reliability

We expanded DreamGenBench from one to four matched inference seeds on the same 63 eligible tasks (252 videos/model). EVEWorld has fewer violations than both baselines at every seed (Table 14):

SeedOfficial GR1SFTEVEWorld
0045/637/631/63
0616/638/632/63
1275/636/632/63
20265/638/631/63
Pooled21/25229/2526/252
MLR (%)8.3311.512.38
Seed SD (pp)0.791.520.92

SD is the sample SD across seeds. The pooled reduction versus SFT is 9.13 percentage points (79.3% relative). On 157 shared eligible WA1 inputs, violations fall from 43/157 to 21/157: 27.39% to 13.38%, with respective 95% Wilson CIs (%) [21.02, 34.84] and [8.92, 19.58] (Table 16). On 60 WA1 tasks x 3 seeds, human MLR falls from 26.67% to 10.00% (Table 25).

We validated MLR against three-human-rater majority votes on 978 videos (2,934 judgments). Raters independently checked occlusion, valid exits, hidden-instance reveals, and overlap separation. Every rater finds fewer EVEWorld violations on every benchmark (Table 23). GroundingDINO detects 162/165 genuine violations and flags only 15/813 human-negative videos (Table 22):

EvaluatorPrecision (%)Recall (%)Agreement (%)
GroundingDINO91.5398.1898.16
Gemini92.9095.1597.96
Qwen94.0896.3698.36

Human-consensus MLR falls on all three benchmarks: DreamGenBench 9.52% -> 2.38%, WA1 26% -> 10%, WA2 26% -> 17% (SFT -> EVEWorld; Tables 17-19).

EVEWorld has the lowest MLR under all five evaluators on the 60-task comparison (Table 25). See Q3 for the full five-evaluator comparison. With OWLv2 target detection, MLR falls from 25.00% (SFT) to 15.56% (EVEWorld); robot localization and SAM2 tracking stay fixed. This comparison uses one human judgment per video, with some labels reused from the main audit.

EVEWorld improves policy learning and planning in real and simulated environments with both backbones. ID results, SFT → EVEWorld (%; Tables 10–11); Wan/Giga: Wan2.2/GigaWorld-0:

PlatformPolicy: WanPolicy: GigaPlanning: WanPlanning: Giga
Real Robot38.3→50.030.9→43.319.2→30.611.6→23.6
RoboTwin74.2→81.273.0→80.723.0→31.021.0→26.0

Policies trained only on original demonstrations score 27.2% on Piper (Real Robot) and 70.0% on RoboTwin. EVEWorld improves policy learning over these baselines and matched SFT on both platforms.

AuthorsOpenReview ↗

Response to Reviewer Qyz7 Questions

Thank you very much for taking the time to review and for your support. We try our best to address your questions as follows.

Q1. Legitimate-Instance Preservation

IGR reconstructs the clean video, retaining legitimate instances in its training target. Three independent human-raters assess generated videos: 23/100 WA1 and 32/100 WA2 inputs have N0 > 1, yielding 165 videos across three generators per benchmark. W1 gives the loss counts and exclusions for valid visibility changes.

EVEWorld matches SFT's pooled multi-instance loss count (8/55) and reduces disappearance by 60.0% across the full audit (W1; Tables 20-21). IGR-only also lowers total MLR under all five evaluators (Q3); MLR combines duplication and disappearance. As auxiliary scene-quality evidence, WA1 Background Consistency rises from 88.67 to 88.91 and EWMBench Scene from 84.38 to 86.37 (Tables 3-4).

Q2. Additional Seeds and Statistical Support

The improvement holds across all four DreamGenBench inference seeds and in the WorldArena comparison. W2 reports per-seed counts and SD on the same 63 tasks, pooled counts, and WA1's 157-input results with 95% Wilson intervals. DreamGen MLR decreases by 9.13 percentage points (79.3% relative) versus SFT. Seed SD measures variation across repeated generation; Wilson intervals quantify each model's violation rate.

On 60 fixed WA1 tasks with three inference seeds, human judgments give 48/180 violations for SFT and 18/180 for EVEWorld (Table 25). This supplements the original 157-input comparison.

Q3. Human Validation and Alternative Detector

Human consensus confirms fewer EVEWorld violations on all three benchmarks (W2: 978-video audit and precision/recall; Tables 17-23). Raters checked legitimate visibility changes and required violations at two consecutive sampled timestamps.

EVEWorld has the lowest MLR under all five evaluators on 60 WA1 tasks x 3 inference seeds (180 videos/model; Table 25; %):

EvaluatorOfficial GR1SFTIGRTIAEVEWorld
Human21.1126.6717.2218.8910.00
Gemini21.1123.8920.0018.3311.67
Qwen20.0026.1119.4419.4412.22
GroundingDINO23.8925.5619.4421.1111.67
OWLv227.7825.0021.1126.1115.56

OWLv2 replaces GroundingDINO for target detection/counting at evaluation; GroundingDINO robot localization and SAM2 tracking remain fixed. This comparison uses one human judgment/video, with some seed-004 labels reused from the main audit.

Q4. Cross-Backbone Complementarity of IGR and TIA

The combination improves Overall/Env/Object beyond either component alone (GigaWorld, 250 steps; Table 6; Gemini-IF, %):

VariantOverallEnvObjectBehavior
SFT53.5751.7242.0067.02
IGR51.8549.4336.6769.50
TIA55.2955.1742.6768.79
EVEWorld60.8572.4150.3364.89

EVEWorld gains 5.56/17.24/7.66 points over TIA on Overall/Env/Object; Behavior decreases by 3.90. Env/Object measure instruction following with unseen environments/objects.

IGR learns content restoration; TIA supplies temporal correspondence. Their combination must learn both which instances to preserve and how to propagate them. GigaWorld uses full FT; the original FlowWAM used Rank-32 LoRA. The FlowWAM sweep supports capacity-dependent joint benefits (W1): limited low-rank updates may constrain restoration and alignment; full FT offers greater adaptation capacity.

W1 gives the E7 five-metric comparison. EVEWorld MLR trajectories (%, lower is better; --: unavailable; Tables 52-55):

EpochRank-32Rank-64Rank-128Full FT
E433.5029.5024.1620.96
E533.9427.1721.1817.87
E636.2423.8421.6315.81
E737.4224.4422.2816.39
E8--25.5023.0617.62

Under full FT, EVEWorld has lower MLR than TIA at every E4-E8 checkpoint, alongside higher PSNR/SSIM; LPIPS is lower at E4-E7, while TIA retains lower Flow-EPE (Table 55). Table 5 reports E6; W1 uses E7, the last checkpoint shared by all settings.

GigaWorld training dynamics (Table 36; Gemini-IF, %):

StepSFT OverallEVE OverallEVE EnvEVE ObjectEVE Behavior
15057.7860.4371.2050.2664.60
20060.2053.4065.1039.6660.80
25053.5760.8572.4150.3364.89
30044.2045.9558.0038.0046.97
35041.6052.5064.3544.3553.85
40039.8060.3072.0049.1065.00
45038.3068.0078.5059.1870.90
50037.2067.6078.2058.6370.60

EVEWorld's IF recovers after step 300 and peaks at 450, while SFT declines. Main results retain the matched 250-step budget. EVEWorld has fewer violations at every scanned step from 150 to 500 (Table 35).

AuthorsOpenReview ↗

Additional Downstream Utility and Sincere Appreciation

Additional Downstream Utility and Sincere Appreciation

We sincerely thank you for your thoughtful feedback and constructive suggestions, which helped us further strengthen our work. Beyond improving physical consistency in video generation, EVEWorld brings consistent performance gains in downstream policy learning and action planning across Piper real-robot and RoboTwin simulation tasks, demonstrating its practical utility as both a data engine and a world-model-based planner.

1. EVEWorld as a Data Engine

We train downstream π0.5 policies using SFT- or EVEWorld-generated data, with matched initialization and training budgets. For Piper, each task uses 100 real demonstrations and 100 generated trajectories; RoboTwin uses 40 original demonstrations and 25 generated action-supervision episodes. Wan and Giga denote Wan2.2 and GigaWorld-0, respectively. All scores are percentages (%).

Platform / SplitOriginal-onlyWan SFT → EVEGiga SFT → EVE
Piper / ID27.238.3 → 50.030.9 → 43.3
Piper / OOD-appearance21.934.5 → 49.327.2 → 40.0
Piper / OOD-position5.315.9 → 27.310.6 → 20.0
RoboTwin / ID59.061.8 → 71.759.3 → 64.0
RoboTwin / OOD-appearance39.342.2 → 49.041.2 → 47.2
RoboTwin / OOD-position22.724.0 → 28.223.2 → 27.3

EVEWorld outperforms matched SFT across all 12 comparisons, on both real robots and simulation, including ID and OOD settings. It also outperforms the original-data-only baseline on every split.

2. EVEWorld for Action Planning

EVEWorld also improves executable action planning using frozen world models and task-specific action decoders on Piper real robots and RoboTwin. Wan and Giga denote Wan2.2 and GigaWorld-0, respectively. All scores are percentages (%).

Platform / SplitWan SFT → EVEGiga SFT → EVE
Piper / ID19.2 → 30.611.6 → 23.6
Piper / OOD-appearance11.9 → 27.96.6 → 17.2
Piper / OOD-position5.3 → 9.30.0 → 0.0
RoboTwin / ID23.0 → 31.021.0 → 26.0
RoboTwin / OOD-appearance22.0 → 29.016.0 → 17.0
RoboTwin / OOD-position16.0 → 20.011.0 → 12.0

EVEWorld improves 11 out of 12 planning comparisons, with one tie and no regressions, demonstrating that improved physical evolution modeling translates into better downstream executable planning.

Real-robot demonstrations are available on our anonymous project page under "Real Robot."

These results are now included in the revised manuscript (Section 5 and Appendix C.5).

We sincerely hope that EVEWorld can provide useful insights and foundations for future research in embodied world models, while contributing to advances in practical downstream robotic applications.

We greatly appreciate your time and valuable feedback throughout the review process. If you find that these additional results strengthen the contribution of our work, we would be truly grateful if you would consider revisiting your assessment and, if appropriate, raising your score. Thank you again for your consideration!

Reviewer_U6bBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies unsupported target duplication and disappearance in embodied video generation, referring to these target-count inconsistencies as Model Laziness. It proposes EVEWorld, which combines Instance-Guided Restoration (IGR) and Temporal Instance Alignment (TIA). IGR inserts additional target instances into training videos and reconstructs the clean videos using a region-weighted restoration objective. TIA matches and transports features between adjacent frames within the Transformer, with target locations providing alignment supervision. The paper also introduces MLR, the fraction of rollouts whose target count deviates from its initial value at two consecutive sampled timestamps after accounting for robot-induced occlusion. On DreamGenBench, EVEWorld reduces MLR from 11.11% to 1.59% and improves Gemini-IF from 53.57% to 60.85% over Standard SFT with matched initialization, training data, and training steps. Experiments also examine domain transfer, different training distributions, a second backbone, evaluation thresholds, and inference randomness. On FlowWAM, the full method improves over Standard SFT but underperforms TIA alone on every reported metric.

Critical strengths
  1. The supervision targets a concrete failure mode and is supported by ablations. IGR provides a direct restoration target for surplus instances, while TIA introduces adjacent-frame correspondence supervision. Ablations of regional weighting and restoration indicate that the gains cannot be attributed solely to assigning greater weight to target regions. In controlled restoration experiments, IGR reduces the retained fraction of injected instances from approximately 92–99% to 23–42%, providing further evidence for the restoration objective.
  2. The core comparisons are controlled and account for evaluation variability. Standard SFT, EVEWorld, and the main component ablations share initialization, training data, and training steps. The paper reports MLR confidence intervals, sensitivity to occlusion and persistence thresholds, repeated Gemini judgments, and matched inference-seed comparisons. These checks strengthen the evidence for the main gains, although the instruction-following improvement varies across seeds.
  3. The evaluation covers several settings. On WorldArena 1.0, MLR decreases from 43/157 to 21/157 while aggregate video metrics improve. Experiments with another training distribution, a second backbone, and multiple PBench inference seeds also show gains. The evidence therefore extends beyond a single benchmark, although improvements are not uniform across metrics or component configurations.
Critical weaknesses
  1. MLR has limited coverage and lacks systematic validation against human judgments. MLR covers only 63 of 126 DreamGenBench tasks and 157 of the 817 WorldArena tasks retained after instruction filtering. A shared evaluation set supports fair comparisons, but detector-based filtering may exclude small targets, occluded objects, or otherwise difficult scenes, leaving generalization to these tasks unclear. The paper also does not report systematic agreement between the primary MLR labels and human annotations. Non-robot occlusion and targets moving out of frame may be mistaken for disappearance, complicating interpretation of this central metric.
  2. The complementarity of the two components depends strongly on the setting. On FlowWAM, TIA alone outperforms the full method on all five metrics. On DreamGenBench, the combined method’s Behavior Gemini-IF is lower than Standard SFT and either component alone. These differences remain unexplained, making it unclear when IGR and TIA should be combined and when restoration supervision may weaken the benefits of temporal alignment.
  3. Direct evidence for identity preservation and reduced deformation remains largely qualitative. The paper presents distortion examples assessed through VLM-assisted review and partial manual verification, but does not report systematic identity or deformation error rates, evaluation size, or annotation agreement. Feature-matching probes, subject consistency, and optical-flow metrics offer related evidence, but do not directly establish how much these errors are reduced in the generated videos.
Questions for authors
  1. How do detector-excluded tasks differ from eligible tasks in target size, occlusion, and operation type? Is there a human-annotated evaluation quantifying MLR false positives and false negatives, particularly for non-robot occlusion and targets moving out of frame?
  2. Does the full method’s underperformance relative to TIA alone on FlowWAM persist across seeds or checkpoints? Can an analysis of corruption strength or loss weights explain this result?
  3. How many videos were reviewed for deformation and identity consistency, and how were the examples selected? Are aggregate error rates and human annotation agreement available for Standard SFT, IGR-only, and the full method?
  4. What development set was used to choose CFG, regional weights, and other design settings? How was it separated from the final test sets?
3 comments
AuthorsOpenReview ↗

Response to Reviewer U6bB Weaknesses

Thank you so much for your careful review and encouragement. We've done our best to address your concerns below.

W1. Evaluation coverage and metric validity

Three-rater human review confirms lower EVEWorld MLR on all three benchmarks. The 978-video audit extends WorldArena evaluation beyond the all-generator detector intersection, includes occlusions, and uses no relocation-only filter. DreamGenBench retains the same 63 relocation tasks across two seeds; Q1 explains the task-type and detection criteria.

SFT → EVEWorld MLR (%); human majority vote (Tables 17–19):

BenchmarkHumanGroundingDINOGeminiQwen
DreamGenBench9.52→2.3811.90→2.3810.32→2.389.52→2.38
WorldArena 1.026→1027→1228→926→11
WorldArena 2.026→1727→1826→1726→19

The improvement survives human checks of legitimate occlusion and exits. GroundingDINO/Gemini/Qwen agree with consensus on 98.16/97.96/98.36% of videos, with FP/FN counts of 15/3, 12/8, and 10/6 (Table 22; in Q1). OWLv2 replaces the target detector and confirms lower MLR, 25.00→15.56%, on 60 WorldArena 1.0 tasks × 3 seeds; robot localization and SAM2 remain fixed (Table 25).

EVEWorld improves policy learning and planning in real and simulated environments with both backbones. SFT → EVEWorld (%; Tables 10–11); Wan/Giga: Wan2.2/GigaWorld-0:

PlatformPolicy: WanPolicy: GigaPlanning: WanPlanning: Giga
Real Robot38.3→50.030.9→43.319.2→30.611.6→23.6
RoboTwin74.2→81.273.0→80.723.0→31.021.0→26.0

Piper (Real Robot) : stage completion; RoboTwin: task success.

W2. Component complementarity

Adaptation capacity changes the FlowWAM ordering with TIA weight fixed at 0.1. At E7 (1,974 steps), EVEWorld-minus-TIA MLR gaps shrink from +13.29 to +0.41/+0.03 points at ranks 32/64/128, then reverse to -0.93 under full FT (Table 56). Q2 gives all checkpoints and the corruption analysis.

Full FT at E6 (Table 5):

MetricSFTIGRTIAEVEWorld
PSNR ↑12.42214.11014.05014.209
SSIM ↑0.7630.8000.7970.803
LPIPS ↓0.3780.2790.2860.274
Flow-EPE ↓2.3511.5841.3411.430
MLR (%) ↓39.4516.0916.8815.81

EVEWorld leads both components on 4/5 metrics at E6. Under full FT, it also outperforms TIA on PSNR/SSIM and MLR throughout E4–E8, and LPIPS at E4–E7; TIA retains lower Flow-EPE (Table 55). Joint training benefits from greater adaptation capacity; TIA-only remains effective under the original Rank-32 setting.

DreamGenBench ablation at 250 matched steps (Table 6; Gemini-IF and MLR, %):

MetricSFTIGRTIAEVEWorld
Overall IF ↑53.5751.8555.2960.85
Env IF ↑51.7249.4355.1772.41
Object IF ↑42.0036.6742.6750.33
Behavior IF ↑67.0269.5068.7964.89
MLR ↓11.114.767.941.59

Overall, Env, and Object improve beyond either component. Gains over TIA are 5.56/17.24/7.66 points; Behavior is 3.90 points lower. Env/Object test unseen environments/objects. Behavior Qwen-IF improves from 68.09 to 76.60 versus SFT at the same checkpoint (Table 2).

Training trajectory, SFT → EVEWorld, seed 004 (Tables 35–36); main results use 250 steps:

StepOverall IF (%)Behavior IF (%)Violations / 63
15057.78→60.4370.20→64.607→4
20060.20→53.4071.50→60.806→1
25053.57→60.8567.02→64.897→1
30044.20→45.9552.30→46.978→2
35041.60→52.5049.60→53.859→1
40039.80→60.3047.60→65.008→2
45038.30→68.0045.80→70.909→1
50037.20→67.6044.50→70.6010→2

EVEWorld exceeds SFT on Behavior Gemini-IF throughout 350–500 steps while maintaining lower MLR at every tested step. Behavior's ordering thus depends on training stage.

W3. Deformation and identity preservation

All 978 videos were independently reviewed by three raters (2,934 annotations): 326/model, comprising 63 DreamGenBench tasks × seeds 004/061 and 100 random initially target-present videos per WorldArena benchmark (seed 004). Majority-vote results (Table 24), SFT → EVEWorld:

BenchmarkDeformationIdentity driftCombined (%)
DreamGenBench24/126→9/12615/126→6/12630.95→11.90
WorldArena 1.023/100→8/10015/100→5/10038→13
WorldArena 2.025/100→7/10016/100→4/10041→11
Pooled72/326→24/32646/326→15/32636.20→11.96

Released GigaWorld-0/FlowWAM baselines have 83/326 deformation and 53/326 identity-drift cases. EVEWorld reduces deformation by 66.7% and identity drift by 67.4% versus SFT, with both lower on every benchmark. These count-preserving errors are separate from MLR; Q3 explains the annotation scope and IGR-only evidence.

AuthorsOpenReview ↗

Response to Reviewer U6bB Questions

We're truly grateful for your thoughtful feedback and support. Below, we've done our best to respond to each of your points.

Q1. Evaluation coverage and MLR reliability

DreamGenBench first filters by operation type: 54 hold/lift tasks lack a destination; 9 static tasks lack source/destination. The remaining 63 relocation tasks pass an initial-target detection check (Appendix A.1).

The broader WorldArena audit confirms the gains (W1), including occlusions without additional size filtering (Appendix A.6).

GroundingDINO against three-rater consensus on 978 videos (Table 22):

BenchmarkTPFNFPTN
DreamGenBench2726343
WorldArena 1.06215232
WorldArena 2.07304223
Total162315798

Recall is 98.18%, precision 91.53%, and FPR 1.85% (15/813). Raters verify initial counts and visibility changes, requiring violations at two consecutive sampled timestamps. Review covers changes beyond the detector's 15% target–robot overlap exemption:

FP causeVideos
Non-robot occlusion3
Valid out-of-frame exit6
Hidden-object reveal1
Overlap separation3
Other2

Non-robot occlusion and exits explain 9/15 FP; human-confirmed gains remain (W1).

Q2. Checkpoint robustness and component interactions

All settings use 2,250 RoboTwin episodes, 282 steps/epoch, and TIA weight fixed at 0.1. TIA/EVEWorld MLR (%) (Tables 52–55; five metrics in W2):

EpochRank-32Rank-64Rank-128Full FT
E422.90/33.5024.63/29.5023.63/24.1621.52/20.96
E522.59/33.9423.29/27.1721.72/21.1819.15/17.87
E623.15/36.2423.76/23.8421.32/21.6316.88/15.81
E724.13/37.4224.03/24.4422.25/22.2817.32/16.39
E8—24.81/25.5022.79/23.0618.11/17.62

TIA leads all five metrics at Rank-32 across E4–E7; full-FT EVEWorld has lower MLR across E4–E8. E7 is the last shared checkpoint; Rank-32 ends there.

TIA supplies a dedicated matching pathway; IGR learns to suppress extras and reconstruct affected content. Limited LoRA update directions may constrain joint learning. A pasted duplicate may also compete with the true target for correspondence, requiring learning what to preserve as well as what to propagate (Appendix E.2).

On GigaWorld, longer training yields higher Behavior Gemini-IF than SFT at 350–500 steps (W2), supporting sufficient optimization alongside adequate adaptation capacity to realize EVEWorld's benefits.

Separate corruption probe: retained injected-duplicate area (%), before → after 50 IGR updates from SFT (Table 29):

Noise σOpacity 0.5Opacity 1.0
0.399.2→42.198.6→41.9
0.895.5→32.596.8→26.8
1.592.2→27.593.5→22.9

Duplicate retention falls by 57.5–75.5% across all six settings, with reduced residual-error alignment. This supports restoration across corruption strengths.

Q3. Deformation, identity preservation, and human evaluation

W3 reports direct deformation/identity measurements on all 978 videos, with random WorldArena sampling and three-rater majority voting. The two mutually exclusive labels capture unsupported shape/identity changes at preserved count.

Every rater independently finds lower count MLR on each benchmark (Table 23).

Two representative matched tasks (Appendix F.3) compare SFT losing the target, IGR preserving count but distorting it, and EVEWorld preserving both count and appearance. Separately, human MLR is 26.67/17.22/18.89/10.00% for SFT/IGR/TIA/EVEWorld on 60 WorldArena 1.0 tasks × 3 seeds (Table 25; one judgment/video).

Q4. Development and configuration selection

Official training/test partitions do not overlap: 92 GR1 training videos and 126 DreamGenBench evaluation tasks (Appendix C.1). CFG=7.0 was fixed a priori using the official default; the sweep did not select the final configuration. Sensitivity at step 250 (Table 34; %):

CFGIF ↑MLR ↓
1.028.989.52
2.550.914.76
5.058.786.35
7.060.851.59

Regional weights were selected on a development set disjoint from the final test set; DreamGenBench test results were not used (Table 26; %; weights: ordinary:restoration).

Region / WeightIF ↑MLR ↓
Target path / 1:259.823.34
Target path / 1:359.963.08
Interaction / 1:260.422.06
Interaction / 1:360.571.94

Interaction coverage improves both metrics at either weight, covering arm/gripper restoration beyond the target path. Unit-mean normalization fixes the average weight, preserving spatial emphasis.

TIA's layer is selected by offline target-matching EPE (Table 27; token cells): block 23 minimizes EPE for DreamGen/AgiBot (0.55/5.72), block 12 for FlowWAM (3.42).

AuthorsOpenReview ↗

Additional Downstream Utility and Sincere Appreciation

Additional Downstream Utility and Sincere Appreciation

We sincerely thank you for your thoughtful feedback and constructive suggestions, which helped us further strengthen our work. Beyond improving physical consistency in video generation, EVEWorld brings consistent performance gains in downstream policy learning and action planning across Piper real-robot and RoboTwin simulation tasks, demonstrating its practical utility as both a data engine and a world-model-based planner.

1. EVEWorld as a Data Engine

We train downstream π0.5 policies using SFT- or EVEWorld-generated data, with matched initialization and training budgets. For Piper, each task uses 100 real demonstrations and 100 generated trajectories; RoboTwin uses 40 original demonstrations and 25 generated action-supervision episodes. Wan and Giga denote Wan2.2 and GigaWorld-0, respectively. All scores are percentages (%).

Platform / SplitOriginal-onlyWan SFT → EVEGiga SFT → EVE
Piper / ID27.238.3 → 50.030.9 → 43.3
Piper / OOD-appearance21.934.5 → 49.327.2 → 40.0
Piper / OOD-position5.315.9 → 27.310.6 → 20.0
RoboTwin / ID59.061.8 → 71.759.3 → 64.0
RoboTwin / OOD-appearance39.342.2 → 49.041.2 → 47.2
RoboTwin / OOD-position22.724.0 → 28.223.2 → 27.3

EVEWorld outperforms matched SFT across all 12 comparisons, on both real robots and simulation, including ID and OOD settings. It also outperforms the original-data-only baseline on every split.

2. EVEWorld for Action Planning

EVEWorld also improves executable action planning using frozen world models and task-specific action decoders on Piper real robots and RoboTwin. Wan and Giga denote Wan2.2 and GigaWorld-0, respectively. All scores are percentages (%).

Platform / SplitWan SFT → EVEGiga SFT → EVE
Piper / ID19.2 → 30.611.6 → 23.6
Piper / OOD-appearance11.9 → 27.96.6 → 17.2
Piper / OOD-position5.3 → 9.30.0 → 0.0
RoboTwin / ID23.0 → 31.021.0 → 26.0
RoboTwin / OOD-appearance22.0 → 29.016.0 → 17.0
RoboTwin / OOD-position16.0 → 20.011.0 → 12.0

EVEWorld improves 11 out of 12 planning comparisons, with one tie and no regressions, demonstrating that improved physical evolution modeling translates into better downstream executable planning.

Real-robot demonstrations are available on our anonymous project page under "Real Robot."

These results are now included in the revised manuscript (Section 5 and Appendix C.5).

We sincerely hope that EVEWorld can provide useful insights and foundations for future research in embodied world models, while contributing to advances in practical downstream robotic applications.

We greatly appreciate your time and valuable feedback throughout the review process. If you find that these additional results strengthen the contribution of our work, we would be truly grateful if you would consider revisiting your assessment and, if appropriate, raising your score or confidence. Thank you again for your consideration!

Reviewer_tVMnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes restoration learning (IGR) and adjacent-frame alignment (TIA) to reduce failures in which video world models duplicate or erase manipulation targets, and introduces MLR (Model Laziness Rate) to measure the rate of these failures. It post-trains GigaWorld-0 on 92 DreamGen GR1 robot videos and compares it with Standard SFT, i.e., conventional fine-tuning with the same data and number of steps.

Critical strengths
  • The paper defines a common failure in measurable terms. It defines object duplication and disappearance as changes relative to the initial count, allowing failures to be counted per rollout without ground-truth future videos. Using this criterion, it shows that for the evaluable items in DreamGen (63 items where the target was detected in the first frame), all five zero-shot models fail on 11–29% (Table 1). This is a meaningful measurement framework for quantifying failures that directly matter when video world models are used as data generators or simulators.
Critical weaknesses
  1. The importance of the problem and its practical value have not been demonstrated. The problem definition requires the initial target count to remain unchanged throughout (p.3), so tasks where the count legitimately changes, such as cutting or assembly, are outside its scope from the outset. Even within this scope, decreases in count are exempted only when the robot occludes the target, while all increases are counted as violations (p.5). Thus, physically valid scenes, such as an initially occluded object of the same type becoming visible or two identical objects overlapping and appearing as one, can also be counted as failures. IGR is also trained only to restore corrupted videos with an extra copy of the target pasted in (Eq. 2), and does not directly address disappearance. More fundamentally, no experiment shows whether reducing these failures makes generated videos more useful for policy learning or planning. WorldArena, which is used for evaluation, also includes evaluations of policy learning from generated videos (data engine) and their use for planning (action planner), but the paper omits these and reports only video quality metrics.
  2. The evidence that the problem has been solved relies on a weak baseline and an unvalidated metric. Standard SFT has worse instruction following (IF) than the base model, GigaWorld-0 (Gemini-assessed IF 60.19→53.57), while EVEWorld is at the base level (60.85; Table 1), so the IF improvement restores what SFT degraded. In supplementary Tables 3 and 9, the official GR1 fine-tuned model (Official GR1 Fine-tune) has higher IF (63.10), and the gap in failed videos is small (5 vs. 1 out of 63), yet it is omitted from Table 1 in the main text. The remaining evidence, MLR, counts objects using a detector, but neither counting accuracy nor agreement with human judgments is reported.
Questions for authors
  1. Are the generated videos more useful for policy learning or planning? Does EVEWorld outperform SFT in WorldArena's data engine and action planner evaluations?
  2. How closely does MLR agree with human judgments, and do the results hold when counts are obtained with a different detector?
  3. Which checkpoint is Official GR1 Fine-tune, and how do its data and training settings differ from Standard SFT? Does the MLR improvement hold when it is included as a baseline in Table 1 and the comparison is repeated across seeds?

I will raise my score if question 1 establishes utility for policy learning and planning, and the improvements hold under the MLR validation and baseline comparisons in questions 2–3.

7 comments
AuthorsOpenReview ↗

Response to Reviewer tVMn Weaknesses

Thank you very much for taking the time to review and for your support. We try our best to address your weaknesses as follows.

W1. Problem Scope and Practical Utility

We evaluated both data-engine and action-planner utility on Piper real robots and WorldArena's RoboTwin tasks. Cells show SFT→EVEWorld (%) for Piper stage completion and RoboTwin task success; Wan/Giga denote Wan2.2/GigaWorld-0. Gains in all 12 policy-learning and 11/12 planning comparisons (one tie) demonstrate practical utility across backbones and ID/OOD settings (Tables 10–11). Real-robot experiment (Section 5.5) and videos are available on abstract anonymous repository project-page under “Real Robot.”

Platform / splitPolicy: WanPolicy: GigaPlanning: WanPlanning: Giga
Real Robot / ID51.9→70.837.3→53.719.2→30.611.6→23.6
Real Robot / OOD appearance62.1→72.838.6→54.711.9→27.96.6→17.2
Real Robot / OOD position21.9→43.316.6→32.65.3→9.30→0
RoboTwin / ID34→4715→2123→3121→26
RoboTwin / OOD appearance33→4411→1822→2916→17
RoboTwin / OOD position29→389→1216→2011→12

We conducted a human audit of duplication and disappearance across 978 videos with three independent human annotators, yielding 2,934 annotations. Cells show SFT→EVEWorld. Both errors decrease on every benchmark, demonstrating improved target preservation alongside duplicate removal (Tables 20, 23).

BenchmarkVideos/modelDuplicationDisappearanceHuman MLR (%)
DreamGenBench1269→34→19.52→2.38
WorldArena 1.010018→810→326→10
WorldArena 2.010018→1311→626→17

We audited occlusion, valid exits, hidden-object reveals, and overlap separation with the three human raters. Our GroundingDINO+SAM2 protocol does not count an under-count as a disappearance if a missing target has ≥15% overlap with the robot mask,which reduces false detections (Q2). Automated evaluators achieve 97.96–98.36% agreement with human consensus(Table 22). These results support that EVEWorld reduces genuine count violations after accounting for legitimate visibility changes.

False-positive causeGroundingDINOGeminiQwen
Non-robot occlusion331
Legal exit634
Hidden-object revelation111
Overlap separation313
Other241
Total151210

W2. Baseline Strength and Evaluation Reliability

We extended matched training to 500 steps to compare against GigaWorld-0 and the released Official GR1. Official GR1 refers to the released checkpoint trained with the official recipe. As this recipe yielded relatively low IF for our SFT and EVEWorld runs, we increased their training budget from 200 to 250 steps and fixed the budget at 250 steps for all results reported in Table 1.

Train stepSFT-IF (%) ↑SFT-MLEVE-MLEVE-IF (%) ↑
15057.787/634/6360.43
20060.206/631/6353.40
25053.577/631/6360.85
30044.208/632/6345.95
35041.609/631/6352.50
40039.808/632/6360.30
45038.309/631/6368.00
50037.2010/632/6367.60

Our results suggest that the effectiveness of combining IGR and TIA depends on both training budget and adaptation capacity. With the GigaWorld backbone, EVEWorld reaches an IF of 68.00 after 450 training steps, surpassing Official GR1. The same trend is observed with FlowWAM: TIA-only outperforms EVEWorld at rank 32, while EVEWorld's benefits become increasingly evident as model capacity increases.

FlowWAMTIA-onlyEVEWorld
Rank-3224.1337.42
Rank-6424.0324.44
Rank-12822.2522.28
Full FT17.3216.39

EVEWorld achieves a lower MLR than Standard SFT and Official GR1 across all four inference seeds. EVEWorld also improves IF over SFT at every seed and reduces pooled MLR by 79.3% versus SFT and 71.4% versus Official GR1 (Table 14). Across 150–500 training steps, it retains fewer violations (Table 35). These results support consistent gains across sampling and training stages.

ModelPooled violationsMLR (%) ↓
Official GR121/2528.33
Standard SFT29/25211.51
EVEWorld6/2522.38

We measured MLR precision and recall using majority-vote annotations from three independent human annotators. Across 978 videos, GroundingDINO detects 162/165 genuine violations and flags only 15/813 human-negative videos (Table 22).

EvaluatorPrecision (%)Recall (%)Agreement (%)
GroundingDINO91.5398.1898.16
Gemini92.9095.1597.96
Qwen94.0896.3698.36
AuthorsOpenReview ↗

Response to Reviewer tVMn Questions

Thank you very much for taking the time to review and for your support. We try our best to address your questions as follows.

Q1. Downstream Policy Learning and Planning

We trained π0.5 policies with SFT- or EVEWorld-generated videos under matched data and training budgets. Cells show SFT→EVEWorld (%): stage completion on the real robot and task success on RoboTwin. EVEWorld improves performance in most task–split comparisons.

PlatformTaskSplitWanGiga
Real robotFill pen holderID64.0→78.038.0→54.0
OOD appearance64.0→72.044.0→56.0
OOD position24.0→40.020.0→32.0
Stack cover blocksID39.8→63.636.5→53.4
OOD appearance60.2→73.633.2→53.4
OOD position19.8→46.613.2→33.2
RoboTwinAdjust_bottleID18.0→26.06.0→8.0
OOD appearance14.0→24.02.0→4.0
OOD position12.0→20.02.0→2.0
click_beilID50.0→68.024.0→34.0
OOD appearance52.0→64.020.0→32.0
OOD position46.0→56.016.0→22.0

We also executed world-model plans on both platforms. EVEWorld improves most planning results, with ties to SFT occurring mainly in OOD-position settings.

PlatformTaskSplitWanGiga
Real robotFill pen holderID12.0→28.010.0→24.0
OOD appearance4.0→16.00.0→8.0
OOD position4.0→12.00.0→0.0
Stack cover blocksID26.4→33.113.2→23.1
OOD appearance19.8→39.813.2→26.4
OOD position6.6→6.60.0→0.0
RoboTwinAdjust_bottleID14.0→22.010.0→16.0
OOD appearance14.0→20.06.0→6.0
OOD position8.0→10.04.0→4.0
click_beilID32.0→40.032.0→36.0
OOD appearance30.0→38.026.0→28.0
OOD position24.0→30.018.0→20.0

Q2. Human Validation and Evaluator Robustness

We validated MLR against three-rater consensus on 978 videos, obtaining 91.53% precision and 98.18% recall for GroundingDINO. Agreement is 98.16% (W2). Counts below list raters 1/2/3 over 126/100/100 videos per model; every rater finds fewer EVEWorld violations, supporting consistent human-confirmed improvements (Table 23).

BenchmarkSFTEVEWorld
DreamGenBench13/12/123/3/3
WorldArena 1.027/26/2611/10/10
WorldArena 2.026/26/2617/16/17

We compared five evaluators—human annotations, Gemini, Qwen, GroundingDINO, and OWLv2—on 60 WorldArena 1.0 tasks across three inference seeds (180 videos/model). EVEWorld achieves the lowest MLR (%) under all five evaluators (Table 25).

EvaluatorOfficial GR1SFTEVEWorld
Human21.1126.6710.00
Gemini21.1123.8911.67
Qwen20.0026.1112.22
GroundingDINO23.8925.5611.67
OWLv227.7825.0015.56

We tested MLR sensitivity to the occlusion and persistence rules in the GroundingDINO–SAM2 pipeline. The occlusion threshold specifies the minimum fraction of each missing target’s reference area overlapped by the robot mask to exempt an under-count. Cells report violating videos out of 63 for k=1/2/3 consecutive sampled timestamps. EVEWorld maintains lower MLR than SFT across all 12 settings (Table 13).

Mask-overlap thresholdSFT (k=1/2/3)EVEWorld (k=1/2/3)
0.1011/5/42/1/1
0.1514/7/53/1/1
0.2017/9/63/1/1
0.2520/10/74/3/1

Q3. Controlled Baseline Comparisons

We clarify that Official GR1 is the released 200-step checkpoint trained with the official recipe. For our SFT and EVEWorld runs, we increased the training budget from 200 to 250 steps We additionally extended matched training to 500 steps: EVEWorld exhibits fewer ML violations than SFT and reaches a higher IF than Official GR1 at step 450 (W2).

We further examined training budget and adaptation capacity (see W2 for detailed settings and results). EVEWorld maintains fewer ML violations than SFT across evaluated checkpoints and exceeds the reported Official GR1 mean IF at step 450. On FlowWAM, its MLR disadvantage relative to TIA-only narrows with increasing LoRA rank and reverses under full fine-tuning.

We compared all three models on the same 63 tasks across four inference seeds. The table shows mean IF (%) as SFT→EVEWorld and Model Laziness counts. EVEWorld improves IF over SFT at every seed, establishing a consistent count-preservation advantage over baselines across inference seeds (Tables 14, 31).

SeedIF: SFT→EVEWorldOfficial GR1SFTEVEWorld
00453.57→60.855/637/631/63
06157.41→58.736/638/632/63
12754.50→61.265/636/632/63
202652.18→60.815/638/631/63
Pooled MLR (%)—8.3311.512.38

Thank you for the additional experiments. The MLR reduction now looks robust. However, I cannot conclude that EVEWorld helps policy learning:

  1. The world model and IDM appear to be trained on all 200 Piper demonstrations (Sec. 5.5), including R2, whose "generated counterparts" form the synthetic data (App. C.5.1). The mixed setting thus uses R2 indirectly, so its gain over the real-only 100-demo baseline does not show data-engine utility. On RoboTwin, no baseline trained without generated data is reported, and the setup is not described.
  2. App. C.5.1 states that three single-view π0.5 policies are trained independently, which is unusual, yet a single score is reported.

I keep my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary

Dear Area Chair,

We sincerely thank all reviewers for their constructive feedback. We believe that EVEWorld can offer useful insights for future research on embodied world models and provide a practical reference for exploring their downstream applications in robotics.

Paper Summary

We propose EVEWorld, a physical evolution supervision framework combining IGR and TIA to reduce target-instance duplication and disappearance in embodied world models, and introduce the Model Laziness Rate (MLR) to quantify such failures. Experiments across three backbones validate its effectiveness, achieving an 85.7% reduction in MLR on DreamGenBench while improving instruction following. We further demonstrate its practical value as both a data engine and an action planner for downstream tasks on real robots and in simulation, achieving up to a 160.6% relative improvement over matched SFT in real-robot planning stage completion under OOD appearance shifts.

Rebuttal Summary

1. Validation of MLR

A common concern across the three reviews was the reliability of our proposed metric, particularly regarding human verification and ambiguous edge cases. In response, we conducted an extensive human audit across 978 videos, yielding 2,934 annotations. Lastly,our evaluation achieves 98.16% video-level agreement with human consensus , supporting the reliability of the proposed evaluation protocol.

2. Effectiveness of EVEWorld in Reducing MLR

We evaluate MLR using five complementary evaluation approaches. EVEWorld consistently achieves the lowest MLR across these evaluators, demonstrating the robustness of its improvements to different evaluation methods. Moreover, the non-zero MLR observed across different generators and benchmarks suggests that Model Laziness is a recurring and general phenomenon in embodied world models, rather than an isolated issue specific to a particular model or evaluation setting.

3. Downstream applications

Following reviewer tVMn's suggestion, we evaluate Wan2.2 and GigaWorld-0 on real robots and in simulation under ID, OOD-appearance, and OOD-position conditions.

  • Data engine: Generated trajectories improve policy performance in all 12 settings over SFT augmentation and original-only training, with up to 88.7% relative improvement over SFT in real-robot OOD-position stage completion.
  • Planner: Real-robot planning stage completion improves by up to 160.6% relative to SFT under OOD appearance shifts.

4. Robustness of EVEWorld

  • Robustness across training checkpoints. We examined EVEWorld under different training checkpoints. Across checkpoints from 150 to 500 training steps, EVEWorld consistently maintains substantially lower MLR than matched Standard SFT. This demonstrates that the improvement is robust across training stages rather than a particular checkpoint.

  • Complementary contributions of IGR and TIA. Our component ablation further demonstrates that IGR and TIA address distinct but complementary aspects of physical consistency. Controlled tests reduce injected-duplicate retention by 66.4%. DreamGenBench ablations show that IGR and TIA jointly improve instance consistency and instruction following.

We sincerely appreciate the Area Chair's time and consideration of our work. We hope that EVEWorld can provide a useful reference for future research on embodied world models and their potential applications in robotics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.