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

cs.CVOct 2, 2026

EVEWorld: Physical Evolution Supervision for Embodied World Models

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

Organizations: China Merchants Group · Nanyang Technological University

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.

Figures & tables

Figure 1: Embodied world models face the challenge of Model Laziness , where instance consistency is violated by spurious duplication or disappearance during interaction. Red denotes Standard SFT, which exhibits such failures, while Green denotes EVEWorld , which better maintains physical consistency throughout the interaction. Dashed arrows trace target evolution over time.

Figure 2: Comparison of Standard SFT, IGR, and EVEWorld . (a) Existing video world models ( GigaWorld Team et al., 2025 ; Yang et al., 2025 ; Team Wan et al., 2025 ; NVIDIA et al., 2025 ) , including our Standard SFT baseline, may exhibit instance-consistency failures, such as target duplication at the goal. (b) Our IGR variant improves instance consistency, but cross-frame consistency failures such as target distortion may remain. (c) Our full EVEWorld model combines IGR and TIA to maintain physically consistent target evolution throughout the interaction. Relative to Standard SFT, EVEWorld reduces MLR by 85.7% and improves overall Gemini-IF by 13.6%. Additionally, the inputs consist of an initial image and an instruction $c$ .

Figure 3: Overall architecture of EVEWorld . Given a clean demonstration $x$ and instruction $c$ during training, IGR constructs a count-edited video $\tilde{x}$ , which is processed by the video diffusion transformer to restore $x$ . Within the transformer, TIA operates at the selected layer $\ell^{\star}$ to perform cross-frame instance matching and residual refinement, after which the refined features pass through the remaining layers to produce the final output.

Figure 4: CFG sensitivity across training stages on DreamGenBench. Gemini-IF and MLR are evaluated under different classifier-free guidance weights

w

, with each curve denoting a training checkpoint.

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Example of occlusion-aware MLR evaluation. The rollout shows target detections and robot occlusion across sampled timestamps. Under-counts supported by sufficient target–robot overlap are excluded from violation evidence, whereas under-counts that persist without sufficient overlap are retained as target-disappearance evidence.

Figure 6: Examples of MLR counting. Red marks a persistent count violation retained by MLR. Green marks an under-count excluded by the occlusion-aware evaluation. A disappearance is counted only when under-count evidence persists outside occlusion-supported timestamps.

Figure 7: Count-preserving cross-frame distortion. Red marks Standard SFT, where the target deforms despite preserving its instance count. Green marks EVEWorld , which preserves target shape and temporal consistency.

Model / config.	Detector v1 (%)	MLR (%) $\downarrow$	Events / elig.	Peak dev.
GigaWorld-0 (Pretrained)	62.1	12.70	8 / 63	0.23
Official GR1 Fine-tune	68.8	7.94	5 / 63	0.14
Standard SFT	63.2	11.11	7 / 63	0.27
EVEWorld (step 100)	70.4	4.76	3 / 63	0.17
EVEWorld (seed 6666, step 250)	66.7	3.17	2 / 63	0.15
EVEWorld (u3, step 300)	68.0	3.17	2 / 63	0.03

Appendix

Table 10: Detector calibration for MLR. Detector v1 denotes the raw count-deviation rate before spatial merging, occlusion-aware under-count handling, and temporal persistence. Peak dev. denotes the mean maximum raw count deviation,

\max_{t}|N_{t}-N_{0}|

, over eligible videos. MLR and Events / elig. are computed using the final occlusion-aware protocol.

Figure 8: Persistence of instance consistency violations across models. (a) Fraction of detector-eligible DreamGenBench videos with occlusion-aware count violations lasting at least

k

consecutive sampled timestamps; (b) cumulative fraction of videos that have exhibited such a violation by each sampled timestamp.

Model	Events / eligible	MLR (%) $\downarrow$	95% Wilson CI	Coverage (%)
GigaWorld-0 (Pretrained)	8 / 63	12.70	[6.58, 23.11]	50.00
Official GR1 Fine-tune	5 / 63	7.94	[3.44, 17.27]	50.00
Standard SFT	7 / 63	11.11	[5.49, 21.20]	50.00
Wan2.2-TI2V-5B	11 / 63	17.46	[10.04, 28.62]	50.00
Wan2.2-I2V-A14B	7 / 63	11.11	[5.49, 21.20]	50.00
CogVideoX1.5-5B-I2V	18 / 63	28.57	[18.90, 40.70]	50.00

Appendix

Table 12: Cross-model MLR on DreamGenBench. Events / eligible denotes the number of Model Laziness cases over the shared 63-item set. 95% Wilson CI denotes the confidence interval of MLR, and Coverage denotes the fraction of detector-eligible prompts.

Model	Events / eligible	MLR (%) $\downarrow$	95% Wilson CI	Coverage (%)
GigaWorld-0 (Pretrained)	43 / 157	27.39	[21.02, 34.84]	19.22
Standard SFT	43 / 157	27.39	[21.02, 34.84]	19.22
EVEWorld	21 / 157	13.38	[8.92, 19.58]	19.22

Appendix

Table 13: Cross-domain MLR on WorldArena 1.0. Results use the same occlusion-aware protocol and the same shared set of 157 detector-eligible videos as main-text Table 3 , with 95% Wilson confidence intervals.

Figure 9: Regional weighting in IGR. The final v4 restoration region and the actual pasted-instance region receive three times the reconstruction weight of ordinary regions before unit-mean normalization.

Variant	MLR (%) $\downarrow$	Env $\uparrow$	Object $\uparrow$	Behavior $\uparrow$	Overall $\uparrow$
IGR w/o spatial emphasis	6.35	56.32	44.00	82.27	61.11
Spatial emphasis only	14.29	34.48	34.67	62.41	44.97
IGR only	4.76	49.43	36.67	69.50	51.85
EVEWorld	1.59	72.41	50.33	64.89	60.85

Appendix

Table 16: IGR restoration controls on DreamGenBench. The first two variants isolate spatial weighting and duplicate-restoration supervision. IGR w/o spatial emphasis removes the spatial weighting, and Spatial emphasis only applies the weighting without the duplicate-restoration target. Env, Object, Behavior, and Overall report Gemini-IF instruction-following accuracy in percentages.

Figure 10: Controlled restoration analysis of IGR. Injected duplicates probe restoration behavior under different corruption-noise levels $\sigma$ and paste opacities $\alpha$ . (a) Standard SFT largely retains the injected duplicate, whereas IGR substantially improves duplicate removal across corruption strengths. (b) Residual errors under Standard SFT remain strongly aligned with the injected duplicate, while IGR reduces this alignment after restoration training.

Opacity $\alpha$	$\sigma$	Ret. before $\downarrow$	Ret. after $\downarrow$	Rel. reduction (%) $\uparrow$	Dir. cos. before / after
0.5	0.3	0.992	0.421	57.5	0.983 / 0.627
0.5	0.8	0.955	0.325	66.0	0.950 / 0.378
0.5	1.5	0.922	0.275	70.2	0.907 / 0.263
1.0	0.3	0.986	0.419	57.5	0.993 / 0.695
1.0	0.8	0.968	0.268	72.3	0.980 / 0.406
1.0	1.5	0.935	0.229	75.5	0.961 / 0.295

Appendix

Table 17: Robustness of duplicate removal across corruption strengths. Retention is the fraction of injected duplicate area preserved by the reconstruction. Dir. cos. measures the alignment between the residual error and the injected duplicate. Before and after denote the Standard SFT checkpoint and the model after 50 IGR fine-tuning steps, respectively.

		Gemini-IF (%) $\uparrow$
Model	Repeats	Overall	Env	Object	Behavior
CogVideoX1.5-5B-I2V ( Yang et al., 2025 )	2	5.56 $\pm$ 1.12	15.52	1.00	4.26
Wan2.2-TI2V-5B ( Team Wan et al., 2025 )	2	10.32 $\pm$ 1.12	17.24	5.00	11.70
Wan2.2-I2V-A14B ( Team Wan et al., 2025 )	2	15.87 $\pm$ 3.37	20.69	13.00	15.96
Cosmos-Predict2-2B ( NVIDIA et al., 2025 )	2	24.60 $\pm$ 1.12	31.03	28.00	17.02
GigaWorld-0 ( GigaWorld Team et al., 2025 )	6	60.19 $\pm$ 2.90	54.60	54.33	69.86

Appendix

Table 18: Gemini-IF judge variance across models on DreamGenBench. Overall reports mean

\pm

SD over repeated Gemini-IF evaluations of the same generated videos. Env, Object, and Behavior report the corresponding mean scores on each split. Repeats denotes the number of independent Gemini-IF evaluations.

Seed	Judge repeats	Standard SFT (%)	EVEWorld (%)	$\Delta$ (pp)
004	6	53.57 $\pm$ 2.92	60.85 $\pm$ 1.79	$+$ 7.28
061	3	57.41 $\pm$ 2.79	58.73 $\pm$ 1.37	$+$ 1.32

Appendix

Table 19: Matched inference-seed comparison of Standard SFT and EVEWorld on DreamGenBench. Gemini-IF reports mean

\pm

SD over repeated judge evaluations at each inference seed.

\Delta

denotes the improvement of EVEWorld over Standard SFT in percentage points.

Seed	6	12	20	42	70	Mean $\pm$ SD
Local-11	57.12	56.89	57.57	57.37	57.46	57.28 $\pm$ 0.28

Appendix

Table 20: Inference-seed variance under the full-1000 transfer protocol. Local-11 is evaluated from the same FlowWAM checkpoint across five inference seeds. Mean

\pm

SD reports variation across seeds.

Component	Seed 0	Seed 17	Seed 35	Seed 42	Seed 70	Mean $\pm$ SD
Image Quality	58.95	59.41	59.35	59.37	57.46	58.91 $\pm$ 0.83
Aesthetic	44.30	44.18	44.56	44.38	44.35	44.35 $\pm$ 0.14
Dynamic Degree	29.59	31.44	29.79	27.88	33.17	30.37 $\pm$ 2.01
Flow Score	18.65	20.57	17.97	15.67	20.11	18.59 $\pm$ 1.95
Motion Smoothness	68.64	68.79	68.14	68.15	71.28	69.00 $\pm$ 1.31
Subject Consistency	73.84	76.67	73.22	80.08	75.56	75.87 $\pm$ 2.72

Appendix

Table 21: Per-component inference-seed variance under the transfer protocol. Component scores are evaluated across five inference seeds on a 50-prompt probe set. Mean

\pm

SD reports variation across seeds.

CFG weight $w$	Train step
	50	100	150	200	250	300
1.0	26.08 / 9.52	39.32 / 6.35	26.91 / 4.76	26.08 / 9.52	28.98 / 9.52	31.87 / 6.35
2.5	45.95 / 4.76	54.23 / 11.11	61.68 / 4.76	47.19 / 3.17	50.91 / 4.76	55.06 / 4.76
5.0	50.91 / 7.94	57.53 / 9.52	57.53 / 3.17	51.32 / 1.59	58.78 / 6.35	48.85 / 4.76
7.0	52.15 / 4.76	66.23 / 9.52	60.43 / 6.35	53.40 / 1.59	60.85 / 1.59	45.95 / 3.17

Appendix

Table 22: CFG sensitivity across training stages on DreamGenBench. Each cell reports Gemini-IF

\uparrow

/ MLR

\downarrow

in percentages under different CFG weights

w

and training steps.

w{=}1.0

denotes inference without classifier-free guidance. The shaded column marks the step-250 checkpoint used in the main experiments. The shaded entry at step 250 and

w{=}7.0

corresponds to the main EVEWorld configuration reported in main-text Table 1 .

Model / horizon	Severity $\downarrow$	Any shortcut $\downarrow$	Missing stage $\downarrow$	$n$
GigaWorld-0 SFT, 3.8s	0.94	0.32	0.30	92
GigaWorld-0 SFT, 7.8s	1.79	0.63	0.60	92
GigaWorld-0 Pretrained, 3.8s	1.17	0.39	0.38	92
GigaWorld-0 Pretrained, 7.8s	1.97	0.67	0.66	92
Wan2.2-TI2V-5B, 3.8s	2.30	0.82	0.72	92

Appendix

Table 24: Long-horizon process evaluation across models and output durations. Qwen-based evaluation is performed on 92 matched prompts. Severity measures shortcut severity. Any shortcut and Missing stage report the fractions of generations containing a shortcut or a missing stage.

Route	Matched control	MLR (%) $\downarrow$		Rel. $\Delta$	95% CI
		Control	Route
PhysicsLatent	Pretrain 5.8s	17.4	16.2	$-$ 7.0	[ $-$ 51.5, $+$ 69.1]
EAG	weight 0	15.4	38.5	$+$ 150.0	[0.0, $+$ 500.0]
LAD-LoRA	LoRA control	3.8	3.8	0.0	[0.0, 0.0]
Causal Frontier	control50	30.8	41.7	$+$ 35.4	[ $-$ 42.9, $+$ 300.0]
ICH-D	Pretrain (4 seeds)	11.2	13.1	$+$ 16.6	[ $-$ 19.4, $+$ 70.8]

Appendix

Table 25: Comparison with alternative designs. Each route is evaluated against its matched control using the same prompts, conditioning frames, and inference seeds. Panel (a) reports MLR for each route and its matched control. Panel (b) reports Gemini-IF as mean

\pm

SD over three judge repeats. Rel.

\Delta

denotes the relative change from the matched control, and 95% CIs are computed by paired bootstrap. MLR values are reported to one decimal place; Rel.

\Delta

and the CIs are computed from unrounded values.

Method	BLEU $\uparrow$	CLIP $\uparrow$	Diversity $\uparrow$	Logic $\uparrow$	Scene $\uparrow$
GigaWorld-0	33.12	89.09	2.77	100.00	89.58
Standard SFT	34.44	86.86	3.47	100.00	84.38
EVEWorld	34.23	87.74	3.26	100.00	86.37

Appendix

Table 26: Complete EWMBench results. Panel (a) reports semantic and scene components. Panel (b) reports motion components and the official Motion, Semantics, and Overall scores. Scores are rescaled for readability: component and Motion scores are multiplied by 100, and Semantics and Overall by 10.

Epoch	PSNR (dB) $\uparrow$	SSIM $\uparrow$	Core-8 $\uparrow$	PSNR (dB) $\uparrow$	SSIM $\uparrow$	Core-8 $\uparrow$	$\Delta$ PSNR (dB)
	Control			EVEWorld
0	16.79	0.807	47.8	18.99	0.857	43.3	+2.20
1	15.38	0.785	47.0	20.23	0.871	48.0	+4.85
2	14.92	0.757	50.5	21.14	0.885	45.9	+6.22
3	15.76	0.785	50.1	18.44	0.839	48.7	+2.68
4	15.02	0.770	48.2	18.02	0.834	47.9	+3.00

Appendix

Table 27: FlowWAM fidelity across checkpoints on a fixed RoboTwin slice. Each checkpoint is evaluated on episode 45 from each of 50 tasks. Core-8 denotes the mean of eight no-reference metrics from the same evaluator.

\Delta

PSNR reports the gain of EVEWorld over the control.

Figure 11: Target-instance consistency under Model Laziness. Red marks Standard SFT, where Model Laziness violates target-instance consistency during manipulation. Green marks EVEWorld , which maintains target-instance consistency throughout the interaction.

Panel of Figure 12 — Figure 12: Additional Model Laziness cases. Red marks Standard SFT, where target-instance consistency is violated during manipulation. Green marks EVEWorld , which maintains target-instance consistency throughout the interaction.

Figure 13: Cross-frame target consistency. Standard SFT and IGR-only show target duplication, appearance drift, or deformation ( red ), while the full EVEWorld preserves target identity and appearance across matched interaction stages ( green ).

Figure 14: Cross-frame target consistency. Standard SFT and IGR-only show target duplication, appearance drift, or deformation ( red ), while the full EVEWorld preserves target identity and appearance across matched interaction stages ( green ).

Figure 15: Instance inconsistency. The 24 fixed MLR timestamps are shown in temporal order. Check marks denote count-consistent states, while crosses denote count deviations. The first cross marks the onset; an MLR event is counted only after two consecutive violating timestamps to suppress isolated detection errors.

Figure 16: Cross-frame inconsistency. The 24 fixed timestamps are shown in temporal order. Check marks denote consistent target states, while crosses mark persistent appearance, geometry, or identity changes. The target count is preserved, so these failures are recorded separately from MLR.

Figure 17: Physical consistency. The 24 fixed timestamps are shown in temporal order. Check marks indicate consistent target states throughout the rollout. The target preserves both its instance count and cross-frame appearance, with no persistent count violation or distortion.

Figure 18: Instance inconsistency. The 24 fixed MLR timestamps are shown in temporal order. Check marks denote count-consistent states, while crosses denote count deviations. The first cross marks the onset; an MLR event is counted only after two consecutive violating timestamps to suppress isolated detection errors.

Figure 19: Cross-frame inconsistency. The 24 fixed timestamps are shown in temporal order. Check marks denote consistent target states, while crosses mark persistent appearance, geometry, or identity changes. The target count is preserved, so these failures are recorded separately from MLR.

Figure 20: Physical consistency. The 24 fixed timestamps are shown in temporal order. Check marks indicate consistent target states throughout the rollout. The target preserves both its instance count and cross-frame appearance, with no persistent count violation or distortion.

Embodied World Models

Embodied world models predict how environments change under an agent’s actions to support planning, learning, simulation, and control.

See all 165 papers

ABot-PhysWorld: Interactive World Foundation Model for Robotic Manipulation with Physics Alignment
Mar 24, 2026Yuzhi Chen, Ronghan Chen, Dongjie Huo +11Embodied World Models Action-Conditioned Video Generation
AMAP CV Lab Alibaba Group
Time-JEPA: Grounding Video World Models with Physical Time Series
Oct 14, 2026Shenyi Li, Ziyang Zhang, Wen Wu +2Video World Models Physical Reasoning in Video
Harbin Institute of Technology · Shanghai Artificial Intelligence Laboratory · Tsinghua University
$ω$ -EVA: Envision, Verify, and Act with Latent Interactive World Models
Jun 8, 2026Zhenguo Sun, Yu Sun, Hande Huang +1Robot Policy Learning World Models for Robotics

Mar 24, 2026J/K move · Enter open · S save

ABot-PhysWorld: Interactive World Foundation Model for Robotic Manipulation with Physics Alignment

Yuzhi Chen, Ronghan Chen, Dongjie Huo, Yandan Yang+10

AMAP CV Lab Alibaba Group

Embodied World Models Action-Conditioned Video Generation Physical Consistency in Video Generation Video Diffusion Models Robot Manipulation Benchmarks

Open paper AskPDF arXiv

Video-based world models offer a powerful paradigm for embodied simulation and planning, yet state-of-the-art models often generate physically implausible manipulations - such as object penetration and anti-gravity motion - due to training on generic visual data and likelihood-based objectives that ignore physical laws. We present ABot-PhysWorld, a 14B Diffusion Transformer model that generates visually realistic, physically plausible, and action-controllable videos. Built on a curated dataset of three million manipulation clips with physics-aware annotation, it uses a novel DPO-based post-training framework with decoupled discriminators to suppress unphysical behaviors while preserving visual quality. A parallel context block enables precise spatial action injection for cross-embodiment control. To better evaluate generalization, we introduce EZSbench, the first training-independent embodied zero-shot benchmark combining real and synthetic unseen robot-task-scene combinations. It employs a decoupled protocol to separately assess physical realism and action alignment. ABot-PhysWorld achieves new state-of-the-art performance on PBench and EZSbench, surpassing Veo 3.1 and Sora v2 Pro in physical plausibility and trajectory consistency. We will release EZSbench to promote standardized evaluation in embodied video generation.

Abstract

Embodied World Models

Explore similar work

ABot-PhysWorld: Interactive World Foundation Model for Robotic Manipulation with Physics Alignment

Time-JEPA: Grounding Video World Models with Physical Time Series

ωωω-EVA: Envision, Verify, and Act with Latent Interactive World Models

$ω$ -EVA: Envision, Verify, and Act with Latent Interactive World Models