Physical Evolution Supervision for Embodied World Models
Process-level supervision for physically consistent target evolution.IGR restores target-instance consistency, while TIA promotes cross-frame target consistency.
Matched rollouts under the same instruction compare Standard SFT, IGR, and the full EVEWorld model.
InstructionUse the right hand to pick up red tomato from upper black tray of plastic shelf to inside of brown paper bag.
InstructionUse the right hand to pick up the bread slice from the toaster and place it on the blue plate.
InstructionMove the boxed drink the drink box with brown circular print onto the coaster the coaster made of smooth wood.
InstructionBring large block, medium block, and small block to the center using the left arm, the right arm, and the left arm.
InstructionPut the drink container with tapered shape squarely on top of the smooth round palm-sized coaster.
Arrows indicate the conceptual addition of supervision components; variants use the same base initialization.
Quantitative and qualitative results are available in the paper; we summarize the strongest evidence here.
Our FlowWAM-based EVEWorld submission, listed as Supervision_WM, achieves an EWMScore-P of 70.17 on the official WorldArena 2.0 Track 1 leaderboard, ranking 17th overall and 6th in JEPA Similarity.
Track 1 evaluates simulator video quality under WorldArena 2.0, with EWMScore-P reporting the difficulty- and out-of-distribution-adjusted aggregate score.
Protocol reference rollouts from the public evaluation harness — external baseline material rather than EVEWorld outputs — are collected on the protocol reference page.
Reading these numbers. MLR is a dedicated target-instance consistency diagnostic rather than a composite measure of video quality, and it should be read together with motion, appearance, and instruction-following metrics. Not every component metric improves uniformly: IGR and TIA provide complementary supervision, and their relative contributions can depend on the backbone. Full profiles, the DreamGenBench splits, the component ablation, and the training setup are reported in the paper.
Standard frame-level reconstruction emphasizes local visual fidelity but provides limited process-level supervision over target evolution. This leaves two complementary gaps: target-instance consistency and cross-frame consistency.
Frame-level reconstruction provides weak supervision for maintaining the target-instance state throughout manipulation.
Nt = N0 is necessary but not sufficient.
Correct instance count alone does not guarantee coherent target identity and appearance across time.
The perturbed target region occupies only a small fraction of the full frame, allowing its reconstruction signal to be diluted by the surrounding scene. In controlled corruption experiments, Standard SFT retains 92–99% of the injected duplicate area.
IGR provides direct restoration supervision for the correct target-instance state.
Correct target count alone does not establish correspondence between the target at adjacent frames. Consequently, locally plausible predictions may still drift in appearance, geometry, or identity over time.
TIA introduces explicit cross-frame alignment of target representations.
EVEWorld retains the host model's native representation and introduces two complementary forms of evolution supervision during post-training: IGR promotes target-instance consistency through restoration supervision, while TIA promotes cross-frame consistency through temporal feature alignment.
Restoring the correct target-instance state from count-perturbed demonstrations.
clean demo → + duplicate → count-edited demo → world model → clean target state
Aligning target representations across adjacent frames.
frame t−1 target → local matching → feature transport → ΔH → frame t representation
Training-only localization supervision. At inference, cross-frame matching operates directly on the model's hidden features and requires neither target localization nor an external tracker.
MLR measures persistent deviations from the initial target-instance count while excluding robot-supported occlusion.
Only rollouts whose target is detected in the conditioning observation enter the denominator, so every model is evaluated on the same detector-eligible set.
Computed per eligible video and reported as a percentage; lower is better. The adjusted count Ñt is the raw GroundingDINO count after robot-supported occlusions are discounted.
Eligibility is fixed from the conditioning observation. A rollout is counted as Model Laziness only when the adjusted target count remains inconsistent for at least two consecutive sampled timestamps.
MLR is a dedicated consistency diagnostic rather than a composite measure of overall video quality. It should therefore be interpreted jointly with motion, visual-quality, and instruction-following metrics. A low MLR alone does not imply strong dynamics or task execution.
Additional comparisons highlight instance consistency and cross-frame target consistency across representative manipulation tasks.
Starting from the same pretrained backbone, we compare Standard SFT, IGR, and the full EVEWorld model.
InstructionUse the right hand to pick up green apple from bottom shelf to top shelf.
InstructionUse the left hand to pick up blue clock from bottom wooden shelf to center of teal plate.
Connections illustrate supervision variants rather than sequential checkpoint initialization.
Representative side-by-side rollouts under matched instructions.
InstructionPlace the white boxdrink with sturdy lid on the light gray streaked wooden coaster.
InstructionTake the bottle with green cap with arms, put it in the dustbin with white inner bag, repeat the same for the transparent bottle and the rounded bottle with a white neck.
InstructionGrab large block, medium block, small block, move them, and arrange by size centrally.
InstructionPlace the brown and white boxdrink on the wooden coaster.
EVEWorld improves target-instance consistency across controlled ablations, backbone transfer, and stochastic inference.
01 · Main result
−85.7%MLR
11.11% → 1.59%
DreamGenBench · vs. matched Standard SFT
| Method | MLR (%) ↓ | Gemini-IF (%) ↑ |
|---|---|---|
| GigaWorld-0 | 12.70 | 60.19 |
| Standard SFT | 11.11 | 53.57 |
| EVEWorld | 1.59 | 60.85 |
Relative to the matched Standard-SFT baseline, EVEWorld reduces persistent target-instance violations while preserving instruction fidelity.
02 · Component ablation
IGR + TIAboth components
1.59 MLR · 60.85 Gemini-IF
DreamGenBench · GigaWorld-based setting
| Variant | IGR | TIA | MLR (%) ↓ | Gemini-IF (%) ↑ |
|---|---|---|---|---|
| Standard SFT | – | – | 11.11 | 53.57 |
| IGR only | ✓ | – | 4.76 | 51.85 |
| TIA only | – | ✓ | 7.94 | 55.29 |
| EVEWorld | ✓ | ✓ | 1.59 | 60.85 |
IGR strengthens target-instance consistency, while TIA adds cross-frame supervision; combining both gives the strongest result on the GigaWorld-based setting.
03 · Cross-backbone
47.89% → 35.21%MLR
Standard SFT → EVEWorld
FlowWAM · RoboTwin · held-out actions
| Variant | PSNR (dB) ↑ | Flow-EPE ↓ | MLR (%) ↓ |
|---|---|---|---|
| Standard SFT | 11.687 | 2.828 | 47.89 |
| + IGR | 12.544 | 2.487 | 40.03 |
| + TIA | 13.445 | 1.850 | 22.54 |
| EVEWorld | 12.765 | 2.207 | 35.21 |
TIA is particularly effective on this architecture, while the full EVEWorld model still improves substantially over Standard SFT, showing backbone-dependent component contributions.
04 · Robustness
6 / 6paired trials improve
PBench Robot QA
Gemini-3.8-Flash re-evaluation
Across three matched inference seeds and two rollout horizons, EVEWorld improves all six evaluated pairs.
Together, these results support evolution supervision across controlled ablations, backbone transfer, and stochastic inference.
If you find this work useful, please cite it as follows.
@software{eveworld2026,
title = "{EVEWorld}: Physical Evolution Supervision for Embodied World Models",
author = "{Anonymous Author(s)}",
year = {2026},
url = "{https://behappya.github.io/EVEWorld/}",
note = "Code and project page released for anonymous review."
}