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

Embodied world models · arXiv preprint

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
2 October 2026 · arXiv:2610.03374

Read paper ↗Full project ↗Code & reproduction ↗

A realistic-looking robot video can still duplicate or lose the object being moved. EVEWorld adds restoration and temporal alignment supervision to improve the consistency of the manipulated target throughout the interaction.

11.11% → 1.59%Model Laziness Rate ↓DreamGenBench · 63 shared eligible items
Compared with matched Standard SFT

−85.7%Relative MLR reductionSame initialization, training data and optimization budget as Standard SFT

53.57% → 60.85%Gemini-IF ↑DreamGenBench instruction following
+7.28 percentage points

Three complementary parts

IGR · Restore instances

Insert an extra target instance into a clean training demonstration, then train the model to recover the clean target state.

TIA · Align over time

Locally match and transport target features between adjacent frames to improve cross-frame target consistency.

MLR · Diagnose failures

Measure persistent target-count violations with occlusion-aware handling. This metric does not cover all physical errors.

Matched qualitative comparison

Instruction: use the right hand to pick up the red tomato from the upper black tray of the plastic shelf and place it inside the brown paper bag.

Standard SFT

IGR only

EVEWorld · IGR + TIA

Public qualitative examples from the existing project website. MLR diagnoses persistent target-count violations; it is not a real-robot success rate or a complete physical correctness score.

让目标物体在交互过程中保持一致

画面看起来真实，物体仍可能被无故复制、消失或跨帧变形。EVEWorld通过实例引导修复（IGR）和时序实例对齐（TIA），把目标的一致性纳入过程监督。在DreamGenBench的63个共同有效样本上，相比匹配Standard SFT，MLR从11.11%降至1.59%，相对下降85.7%。