PaperlayerView exact revision on arXiv

Source-linked AI summary

EVEWorld: Physical Evolution Supervision for Embodied World Models

Kaiqi Wang, Songxin Zhang, Zejian Xie, Xiao Xiong, Zhuoyang Song, Ziwei Wu, Jun Yu Lu, Yitan Teng, Ziying Song, Jiaxing Zhang

arXiv:2610.03374v1cs.CV

TL;DR

Embodied world models can prioritize visual reconstruction without process-level supervision for physically consistent target evolution. EVEWORLD combines restoration and temporal alignment supervision; it reports an 85.7% MLR reduction on DreamGenBench and ranks 17th overall and 6th in JEPA Similarity on WorldArena 2.0 Track 1.

  • Problem

    Standard training emphasizes visual reconstruction but provides limited process-level supervision for physically consistent target evolution, leaving gaps in instance and cross-frame consistency.

    Source: §1 · p. 2, §3.2 · p. 3

  • Method

    EVEWORLD combines count-perturbed video restoration through IGR with adjacent-frame target alignment through TIA to promote instance and cross-frame consistency.

    Source: §1 · p. 2, §4.1 · p. 4

  • Results

    85.7% MLR reduction on DreamGenBench; EVEWORLD ranks 17th overall and 6th in JEPA Similarity on the WorldArena 2.0 Track 1 leaderboard.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    Experiments report improved instruction following and generalization gains across training distributions and model backbones.

    Source: §6 · p. 9

    The local, pairwise supervision does not explicitly capture long-horizon state dependencies or complex multi-object interactions.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

Embodied world models have advanced visual and structural modeling, but this does not ensure manipulated targets remain consistent across a rollout. EVEWORLD adds evolution supervision and reports improved target consistency alongside strong leaderboard rankings.

1 INTRODUCTION

  • Despite stronger visual and structural modeling, embodied world models can still duplicate or lose targets, deform them, or drift in identity across frames.The paper calls such instance-consistency failures Model Laziness: frame-level reconstruction can satisfy appearance or endpoint cues without maintaining consistency throughout an interaction.Source: §1 · p. 2
  • EVEWORLD ranks 17th overall and 6th in JEPA Similarity on the WorldArena 2.0 Track 1 leaderboard, with an EWMScore-P of 70.17.Source: §1 · p. 2
  • EVEWORLD combines IGR restoration of count-perturbed videos with TIA alignment of target instances across adjacent frames, and introduces MLR to measure persistent rollout violations.Source: §1 · p. 2

2 RELATED WORK

Prior work strengthens embodied world models through richer representations, conditioning, tracking, and memory, yet consistent evolution of manipulated targets remains challenging. EVEWORLD adds evolution supervision while retaining the host model's native representation.

2 RELATED WORK

  • Prior work improves structural modeling with richer representations and conditioning, while EVEWORLD retains the host model’s native representation and adds post-training evolution supervision.Examples include joint RGB, depth, and normal prediction; geometric action conditioning; and disentangled dynamics and visual synthesis (DreamX Team et al., 2026; Shan et al., 2026).Source: §2 · p. 3
  • Despite advances in correspondence, tracking, object permanence, and persistent memory, manipulated objects can still fail in instance and cross-frame consistency (Dou et al., 2026).EVEWORLD addresses this challenge with evolution supervision.Source: §2 · p. 3

3 WHY EMBODIED WORLD MODELS BECOME LAZY

The formulation treats target-count changes during manipulation as consistency failures relative to the initial count, with explicit exceptions for intentional count changes and robot occlusion. The paper attributes Model Laziness to gaps in instance and cross-frame supervision.

3.1 PROBLEM STATEMENT

  • Model Laziness is a rollout-level mismatch between a target’s count at any time and its initial count; intentional count changes are out of scope, while robot-induced occlusion receives a separate check.Source: §3.1 · p. 3

3.2 CAUSES OF MODEL LAZINESS

  • Standard SFT retains 92%–99% of injected target area, leaving weak restoration supervision because the perturbed region is small relative to the frame.IGR addresses this gap with a direct restoration target and threefold spatial weighting of perturbed regions, as depicted in the architecture overview.Source: §3.2 · p. 3, Figure 3 · p. 4
  • Correcting target count alone does not prevent distortion, fading, or abrupt shifts; the missing link is explicit supervision connecting the same target across frames.TIA addresses this gap by aligning target instances across adjacent frames.Source: §3.2 · pp. 3–4

4 METHOD

EVEWORLD uses training-time restoration and temporal alignment supervision to promote instance and cross-frame consistency. At inference, it takes only the initial image and instruction.

4 METHOD

  • During training, EVEWORLD makes a count-edited video from a clean demonstration and instruction, then trains restoration toward the clean video; inference uses only the initial image and instruction.Source: §4 · p. 4

4.1 INSTANCE-GUIDED RESTORATION

  • IGR inserts an extra target instance and trains the model to recover the clean latent representation, weighting restoration regions three times more than ordinary regions.The count-edited latent is mapped back to the clean latent, providing explicit instance-consistency supervision.Source: §4.1 · pp. 4–5

4.2 TEMPORAL INSTANCE ALIGNMENT

  • TIA aligns target features across adjacent frames through local instance locating and feature transport, then supervises transported features to favor the true next-frame target location.Alignment supervision uses GroundingDINO and target locations during training; at inference, matching and transport use hidden features without target localization or an external tracker.Source: §4.2 · p. 5

5 EXPERIMENTS

Experiments evaluate EVEWORLD on DreamGenBench and complementary generalization and robustness benchmarks, measuring instruction following, target-instance consistency, and other performance dimensions. Results report reduced Model Laziness alongside instruction-following gains, broader benchmark improvements, and component and inference analyses.

5.1 EXPERIMENTAL SETUP AND METRICS

  • Experiments use 92 DreamGen GR1 videos and 126 DreamGenBench tasks, with MLR tracking persistent target-count deviations from the observed initial count and excluding robot-occluded undercounts.Evaluation also spans WorldArena, PBench, EWMBench, and RoboTwin for complementary performance dimensions.Source: §5.1 · p. 5

5.2 MAIN RESULTS

  • MLR falls from 11.11% to 1.59% (85.7% relative reduction), while Qwen-IF rises from 73.81% to 80.16% and Gemini-IF from 53.57% to 60.85% against Standard SFT.The remaining EVEWORLD violations are reported as transient.Source: Table 1 · p. 6, §5.2 · p. 6
  • Qwen-IF improves across all splits, with Behavior rising from 68.09% to 76.60%; Gemini-IF rises on Environment (51.72% to 72.41%) and Object (42.00% to 50.33%).Source: Table 2 · p. 6, §5.2 · p. 6

5.3 GENERALIZATION EVALUATION

  • On FlowWAM, TIA yields consistently larger gains than IGR, and the full model outperforms Standard SFT; SupervisionWM scores 70.17 EWMScore-P on WorldArena 2.0, ranking 17th overall and 6th in JEPA Similarity.The optimal component configuration may depend on architecture.Source: Table 5 · p. 7, §5.3 · pp. 7–8
  • On WorldArena 1.0, EWMScore-local-8 rises from 53.95 to 56.76 while MLR drops from 27.39% to 13.38%; EWMBench Motion and Overall improve by 3.48% and 1.24%.The WorldArena comparison covers unseen scenes and robot embodiments; the EWMBench comparison uses matched AgiBot data and optimization budget.Source: §5.3 · p. 7, Table 3 · p. 7, Table 4 · p. 7

5.4 ABLATION STUDIES AND ANALYSIS

  • IGR reduces MLR from 11.11% to 4.76%, while combined IGR/TIA reaches 1.59% MLR and 60.85 Gemini-IF; EVEWORLD improves all six paired PBench trials, with mean gains of +1.54 and +1.56.Source: §5.4 · p. 8, Table 6 · p. 8, Table 8 · p. 8
  • Output durations from 3.8 to 19.8 seconds raise p50/p95 latency from 8.17/8.40 to 41.74/45.01 seconds and GPU-hour cost per 1,000 videos from 18.40 to 95.34.The paper reports relatively stable throughput and peak memory, with longer rollouts mainly increasing computation.Source: Table 9 · p. 9, §5.4 · p. 9
  • None of five alternative approaches jointly achieves lower Model Laziness and better instruction fidelity; higher CFG generally raises Gemini-IF, but MLR varies non-monotonically across weights and training stages.Source: §5.4 · p. 9, Figure 4 · p. 9

6 CONCLUSION

EVEWORLD combines instance restoration and temporal alignment, with reported reductions in Model Laziness and gains across distributions and backbones. The appendices document MLR, method details, experiments, generalization, and qualitative results.

6 CONCLUSION

  • 85.7% MLR reduction on DreamGenBench accompanies improved instruction following; EVEWORLD also ranks 17th overall and 6th in JEPA Similarity on WorldArena 2.0 Track 1.The conclusion also reports generalization gains across training distributions and model backbones.Source: §6 · p. 9
  • EVEWORLD pairs IGR for instance consistency with TIA for cross-frame consistency, while MLR quantifies target-instance consistency failures called Model Laziness.Source: §6 · p. 9
  • The current formulation uses local, pairwise supervision and does not explicitly capture long-horizon state dependencies or complex multi-object interactions.Future work will expand evolution supervision for longer-horizon and relational dynamics.Source: §6 · p. 9

APPENDIX OVERVIEW

  • Appendices cover MLR and its evaluation, method details, additional experiments, alternative approaches, generalization, qualitative results, and related work.Source: APPENDIX OVERVIEW · p. 17

A.1 MLR: COMPARISON WITH EXISTING EVALUATION METRICS

MLR measures persistent target-count deviations from the observed initial state, exempting robot-supported under-counts. It complements evaluations anchored to other signals and is scoped to tasks where target-instance counts should remain unchanged.

A.1 MLR: COMPARISON WITH EXISTING EVALUATION METRICS

  • MLR isolates target-instance consistency as a dedicated rollout-level measure rather than folding appearance and disappearance into broader physical-plausibility and task-correctness assessments.Source: §A.1 · p. 18
  • MLR needs no ground-truth future, unlike WFM-Eval’s reference-grounded diagnostics, and derives expected counts from the conditioning observation rather than prompt-specified quantities.Source: §A.1 · p. 18
  • Unlike embodied-execution evaluation, MLR directly identifies persistent target-count violations in generated videos without action decoding or simulator execution.Source: §A.1 · p. 18
  • MLR uses the observed initial target count as its reference, exempts robot-occlusion-supported under-counts, retains over-counts, and registers deviations that persist across consecutive sampled timestamps.The evaluation protocol uses fixed detector, query, occlusion, and persistence settings across models.Source: §A.1 · p. 19

A.2 SCOPE AND LIMITATIONS OF MLR

  • MLR applies to manipulation tasks where the initial target count should remain unchanged, including scenes with multiple target instances; intentional count-changing tasks fall outside its formulation.Such tasks would require an instruction-conditioned expected count rather than a fixed initial-state reference.Source: §A.2 · pp. 20–21
  • Because MLR does not capture count-preserving identity drift or deformation, TIA addresses this complementary cross-frame inconsistency by aligning target instances across adjacent frames.Figure 6 contrasts count-preserving distortion under Standard SFT with EVEWORLD’s preserved target shape and temporal consistency.Source: §A.2 · p. 21, Figure 7 · p. 21

A.3 ROBUSTNESS AND SENSITIVITY ANALYSIS OF MLR

Across sensitivity settings, EVEWORLD retains lower MLR than Standard SFT, and cross-model and cross-domain evaluations also report lower MLR for EVEWORLD. The appendix additionally describes optional VLM review as auxiliary to the fixed detector-based protocol.

A.3 ROBUSTNESS AND SENSITIVITY ANALYSIS OF MLR

  • The raw count-deviation rate differs substantially from the final event rate, illustrating the protocol’s use of spatial merging, occlusion-aware filtering, and temporal persistence.Source: §A.3 · p. 22, Table 10 · p. 22
  • EVEWORLD’s sole violation is confined to short persistence lengths, while general I2V baseline violations extend to larger k.Source: §A.3 · p. 22, Figure 8 · p. 22
  • 85.7% relative MLR reduction under the main setting, from 7/63 (11.11%) to 1/63 (1.59%), with EVEWORLD lower than Standard SFT across tested threshold configurations.Across configurations, relative reductions range from 70.0% to 88.9%.Source: §A.3 · p. 23, Table 11 · p. 23

A.4 CROSS-MODEL EVALUATION OF MLR

  • On the shared 63-item DreamGenBench set, EVEWORLD attains 1.59% MLR, while evaluated model families range from 7.94–28.57%.The results report persistent target-instance consistency failures across multiple video-generation and embodied-world-model families.Source: §A.4 · p. 23, Table 12 · p. 23

A.5 CROSS-DOMAIN EVALUATION OF MLR

  • On 157 detector-eligible WorldArena 1.0 videos, MLR decreases from 27.39% under Standard SFT to 13.38% under EVEWORLD.Source: §A.5 · p. 23, Table 13 · p. 24

A.6 OPTIONAL VLM EVALUATION

  • Optional VLM judgments support review of detector-based MLR events but do not replace or modify the frozen protocol or its reported labels and values.The passage treats VLM judgments as supporting evidence, not ground truth, and allows manual review of ambiguous cases.Source: §A.6 · p. 24

B.1 IGR CONSTRUCTION

EVEWORLD combines instance restoration with adjacent-frame feature alignment to supervise physically consistent target evolution. The appendix details how each component is constructed, jointly trained, and evaluated.

B.1 IGR CONSTRUCTION

  • IGR corrupts a clean trajectory by pasting an extra target instance, then trains reconstruction toward the original clean video.The final v4 binary mask gives restoration-related and pasted-instance regions relative weight 3 versus 1 elsewhere, then normalizes the map to unit mean.Source: §B.1 · pp. 24–26, Figure 9 · p. 25

B.2 TIA ALIGNMENT

  • TIA aligns adjacent-frame target features through local matching and feature transport, with its attachment layer selected by an offline consistency probe.The probe selects block 23 for GigaWorld-0/DreamGen and AgiBot, and block 12 for FlowWAM; the first frame remains unchanged during transport.Source: §B.2 · pp. 26–27, Table 15 · p. 28

B.3 JOINT TRAINING AND INFERENCE

  • Joint training combines IGR restoration and TIA consistency losses; at inference, only the first frame, instruction, and benchmark-specific control are provided.The parser, grounder, and corruption generator are used only during training.Source: §B.3 · p. 28

C.1 IMPLEMENTATION DETAILS

  • The appendix evaluates matched post-training controls across DreamGenBench, EWMBench, and RoboTwin using benchmark-specific training budgets and checkpoints.DreamGenBench uses 92 videos for 250 steps, EWMBench 777 clips for 50 steps, and RoboTwin 2,250 episodes for 1,128 steps.Source: §C.1 · pp. 28–29

C.2 IGR ABLATION AND DUPLICATE RESTORATION

  • IGR alone reduces MLR to 4.76%, while full EVEWORLD reaches 1.59%; removing spatial emphasis raises MLR to 6.35%.Without spatial emphasis, Overall instruction-following score is slightly higher (61.11% vs. 60.85%); spatial emphasis alone does not improve instance consistency. In the controlled restoration evaluation, duplicate-area retention falls from 92–99% under Standard SFT to 23–42% after 50 IGR steps across six settings.Source: §C.2 · p. 29, Table 16 · p. 29, Table 17 · p. 29

C.3 EVALUATION STABILITY

  • EVEWORLD’s advantage persists under matched inference seeds and repeated instruction-following judgments; FlowWAM’s Local-11 score varies by at most 0.68 points across five seeds.Motion metrics vary more than appearance metrics, while aggregate Local-11 remains relatively stable.Source: §C.3 · pp. 30–31, Table 19 · p. 30, Table 20 · p. 31

C.4 TRAINING AND LONG-HORIZON BEHAVIOR

Post-training generally reduces instance-consistency violations while improving instruction following, but stronger classifier-free guidance does not reliably improve both. Longer outputs remain more challenging for physical and process-level performance.

Training dynamics

  • As training approaches step 250, MLR generally decreases while Qwen-IF improves, so the measured consistency gain is not accompanied by worse task fidelity.Source: §C.4 · p. 31

Guidance sensitivity

  • Stronger CFG generally raises Gemini-IF, but MLR varies non-monotonically and does not reliably improve with guidance weight.The lowest MLR occurs only at particular combinations of training stage and guidance weight.Source: §C.4 · p. 31, Table 22 · p. 32

Long-horizon behavior

  • PBench physical-QA performance decreases as duration grows from 3.8 s to 19.8 s, with temporal reasoning degrading most strongly.Across durations of at least 9.8 s, Standard SFT improves Domain, Phys, and Time over the pretrained backbone by 1.65, 1.95, and 4.24 points on average, respectively.Source: §C.4 · p. 31, Table 23 · p. 32
  • At 7.8 s, both GigaWorld-0 variants show higher shortcut severity, shortcut frequency, and missing-stage frequency than at 3.8 s.At 3.8 s, Wan2.2-TI2V-5B has larger process errors than either GigaWorld-0 variant; these diagnostics measure failures beyond target-instance count.Source: §C.4 · p. 32, Table 24 · p. 32

D EXPLORATION OF ALTERNATIVE APPROACHES

The explored alternatives do not jointly reduce Model Laziness and improve instruction following. Their individual results show trade-offs across training-time and inference-time approaches, alongside concentrated motion-related gains on EWMBench.

D EXPLORATION OF ALTERNATIVE APPROACHES

D.1 PHYSICSLATENT

  • PhysicsLatent changes MLR from 17.4% to 16.2% but lowers Gemini-IF from 45.65 to 40.22 against its pretrained control.It adds learned process tokens and weak-label heads for manipulation cues.Source: §D.1 · p. 33

D.2 EAG

  • EAG raises Gemini-IF from 41.67 to 45.83 while MLR rises from 15.4% to 38.5% against the weight-0 control.This is a development comparison with 16 prompt–seed pairs, 13 detector-eligible.Source: §D.2 · p. 34

D.3 LAD-LORA

  • LAD-LoRA matches its plain-LoRA control at 3.8% MLR but lowers Gemini-IF from 51.04 to 47.92.The method adds transition energy as a regularizer while keeping the backbone frozen.Source: §D.3 · p. 34

D.4 CAUSAL FRONTIER

  • Causal Frontier raises MLR from 30.8% to 41.7% and lowers Gemini-IF from 27.08 to 12.50 against control50.Its four-frame blocks are committed irreversibly, and prediction errors compound over the rollout at the reported budget.Source: §D.4 · pp. 34–35

D.5 ICH-D

  • ICH-D raises MLR from 11.2% to 13.1% and lowers Gemini-IF from 45.83 to 31.79 against the pretrained model.The authors report that the deletion path introduces its own errors.Source: §D.5 · p. 35

E.1 GENERALIZATION ON EWMBENCH

  • On EWMBench, EVEWORLD improves Motion by 3.48%, Overall by 1.24%, and DYN by 17.94% relative to Standard SFT.Other components remain close to the matched baseline, and gains concentrate in motion-related and aggregate scores.Source: §E.1 · p. 35

E.2 FLOWWAM CHECKPOINT ROBUSTNESS

Across six FlowWAM checkpoints on a fixed RoboTwin slice, EVEWORLD improves reference-fidelity metrics. Qualitative examples also contrast its instance consistency with Standard SFT and IGR-only outputs.

E.2 FLOWWAM CHECKPOINT ROBUSTNESS

  • Across six checkpoints, EVEWORLD improves PSNR by 2.20–6.22 dB and increases SSIM at every checkpoint.The evaluation uses episode 45 from each of 50 tasks.Source: §E.2 · p. 35, Table 27 · p. 36

F.1 INSTANCE CONSISTENCY

  • Qualitative examples show Standard SFT introducing surplus or missing target instances, while EVEWORLD better maintains instance consistency through the rollout.Source: §F.1 · p. 36

F.2 CROSS-FRAME CONSISTENCY

  • In matched manipulation stages, IGR-only can retain cross-frame inconsistencies, whereas adding TIA yields a coherent target across stages.The examples illustrate the complementary roles of restoration and temporal alignment.Source: §F.2 · p. 36

F.3 TARGET EVOLUTION COMPARISON

The comparison shows that preserving target count alone does not ensure cross-frame consistency: IGR-only can still distort the target, while EVEWORLD preserves both. Related work distinguishes this evolution supervision from approaches focused on structural modeling, editing, correspondence, or memory.

F.3 TARGET EVOLUTION COMPARISON

  • IGR-only preserves target count but progressively distorts the target, whereas EVEWORLD preserves both target count and cross-frame consistency.Standard SFT instead loses the target, producing an MLR count violation.Source: §F.3 · pp. 36–37

G RELATED WORK

  • TesserAct, DVG-WM, DreamX-Phi, and ABot-PhysWorld address structural modeling, target selection, object-aware training, or video-level preferences; IGR and TIA directly supervise target evolution.IGR restores count-perturbed demonstrations, while TIA matches and transports manipulated-target features across adjacent frames.Source: G RELATED WORK · p. 43
  • LoVoRA edits objects in supplied videos and ROSE removes specified objects and their scene effects, while IGR uses object insertion as training perturbation for restoration supervision.EVEWORLD predicts from an initial image and instruction without an edit request identifying a spurious target.Source: G RELATED WORK · p. 43
  • Track4Gen, GEM-4D, and memory-based methods address tracking, geometric correspondence, or persistence, while TIA and IGR supervise manipulated-target consistency during prediction.TIA focuses matching and transport on the target; IGR restores count-perturbed videos, including errors not explicitly corrected by geometric correspondence alone.Source: G RELATED WORK · pp. 43–44

Related papers

Loading 2610.03374v1…