Video World Models · 2026-10-05
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
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
EVEWorld mitigates “Model Laziness” in embodied world models by adding physical evolution supervision. It combines Instance-Guided Restoration (IGR) for instance consistency via count-edited restoration training and Temporal Instance Alignment (TIA) for cross-frame consistency by aligning target instances across adjacent frames. The authors report large MLR reductions vs GigaWorld-0 and strong WorldArena 2.0 Track 1 leaderboard placement.
Why it matters
Embodied world models can succeed visually while failing physically through target duplication/disappearance (instance consistency) and identity drift/deformation (cross-frame consistency). EVEWorld targets these specific rollout-level failures with explicit training losses and a metric (MLR) that measures persistent instance-consistency violations, aiming to improve physical target evolution without sacrificing instruction following.
Method
- Instance-Guided Restoration (IGR): insert an extra target instance into demonstrations (count-edited samples) and train the model to restore the corresponding clean latent, with higher spatial weight on restoration-related regions.
- Temporal Instance Alignment (TIA): at a selected transformer layer, align target features across adjacent frames via local instance locating and feature transport with a residual refinement; uses offline target locations only for training supervision.
- Model Laziness Rate (MLR): a metric measuring persistent violations of target-instance consistency over a rollout using the observed initial count, with two-frame persistence criterion and an occlusion-aware exclusion for under-counts.
Limitation
MLR intentionally measures changes in target-instance count and therefore does not capture failures that preserve N0 while corrupting the target itself (distortion is assessed separately).
Abstract (from arXiv)
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.
Related papers
- EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models
- Causal world models under physical parameter shifts
- EmbodiedSmith: Scaling Embodied Data through Recursive Self-Improvement Flywheel in Simulation
- Video2World: Benchmarking Coding Agents for Interactive World Modeling from Embodied Videos
- Time-JEPA: Grounding Video World Models with Physical Time Series
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-02 · Indexed by Pub-AI: 2026-10-05 UTC
EVEWorld: Physical Evolution Supervision for Embodied World Models
Preprint posted: 2026-10-02 · Original record ↗
- arxiv
- 2610.03374
Explicit source versions (1)
- v1 · 2026-10-02
- Preprint available
Source date: 2026-10-02 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.
