EvoWorld — Out-of-Sight World Evolution

Brain-guided generation

Watch EvoWorld observe, reason, and evolve.

Based on the user prompt, the visual renderer generates a segment, the VLM observes the realized world, updates memory, reasons about the event, and guides the next generation.

Case studyRuined cathedral · Colossal gargoyle lands

ObserveRememberReasonRenderVerify

The challenge

A coherent world is more than a remembered frame.

When the observer looks away, new events should still alter the world. A successful model must execute the switched event, preserve the subject and scene, and render a causally plausible change when the camera returns.

Comparison of out-of-sight evolution failures across video world models

01

Look away

The camera leaves a visible subject and establishes an observation gap.

02

Change the world

A new event is introduced mid-rollout, without revealing the future response.

03

Return and reveal

The revisited subject must remain recognizable and show a plausible consequence.

The framework

A reasoning brain.
A visual world renderer.

EvoWorld separates world-state reasoning from rendering. A VLM continuously records what the renderer actually generated, then reasons about an appropriate subject state after an event is observed.

EvoWorld framework with VLM brain, autoregressive visual renderer, Identity-Anchored Attention, and two-stage training

01

VLM brain reasoning

The brain tracks persistent entities and world states, then reasons about event consequences before rendering each revisit.

02

Event rendering

Events emerge clearly and evolve while the focal subject remains outside the current view.

03

Causal subject reactions

Revisited subjects preserve their identity while responding plausibly to the intervening event.

Data & benchmark

Built for evolution beyond the frame.

EvoWorld-Data provides segment-level supervision for event realization, long-range continuity, and behavioral change. EvoWorld-Bench tests the full event–revisit–reaction chain.

~12kcurated videos

65event types

63scene types

14reaction types

120benchmark cases

Inside EvoWorld-Data

Structured examples.
Visible evolution.

Selected training videos from EvoWorld-Data's photorealistic and game domains illustrate its range of subjects, environments, camera trajectories, and event-driven changes. Click any example to inspect the full clip.

EvoWorld-DataPhotorealistic domain↗

EvoWorld-DataGame domain↗

Benchmark results

Consistent worlds.
Strong evolution.

EvoWorld improves out-of-sight reactive evolution while preserving visual quality and consistency.

Core task success90.8%

Core task success requires Revisit, Event, and Causal scores to equal 1.

Revisit success100.0

Event success97.5

Causal plausibility94.2

Reaction quality85.0

Subject consistency99.2

Background consistency98.3

EvoWorld-Bench evaluation

Scores in percent · higher is better

Method	Core	Revisit	Event	Causal	React.	Visual	Subj.	Bg.	Motion
LingBot-World	15.8	92.5	77.5	29.2	31.3	68.8	87.6	94.6	98.0
LingBot-World-v2	20.0	76.7	68.3	27.1	25.4	76.7	82.6	92.5	98.4
YUME-1.5	15.8	62.5	86.7	22.5	20.0	65.4	72.9	88.7	99.0
SANA-WM	13.3	88.3	67.5	26.7	24.2	53.8	67.7	90.8	98.7
Echo-WM	16.7	60.8	66.7	22.5	21.7	82.9	97.6	89.2	98.7
Evoke	20.0	95.8	80.0	32.9	36.3	78.3	93.5	95.4	99.0
Evoke-Turbo	21.7	98.3	89.2	32.5	29.6	65.4	97.0	96.2	99.1
EvoWorld ours	90.8	100.0	97.5	94.2	85.0	83.8	99.2	98.3	99.5

EvoWorld world evolution demos

Watch the world change.

Event A colossal gargoyle lands in the cathedral.

Reaction The knight raises his shield and braces.

Event Lava flows into the stone passage.

Reaction The explorer climbs onto the steps.

Event A shower pipe bursts and floods the floor.

Reaction The man steps back with a towel in hand.

Event A spectral knight forms from the marsh fog.

Reaction The ranger stands and draws her sword.

Event A dark monster appears by the boulder.

Reaction The archer nocks an arrow and takes aim.

Event Water bursts from the canal and floods the path.

Reaction The backpacker turns around and lowers the newspaper.

Citation

Build worlds that remember, reason, and evolve.

If you find EvoWorld useful, please cite our paper.

@article{bie2026evoworld,
  title   = {EvoWorld: Towards Out-of-Sight World Evolution in Generative Video Models},
  author  = {Bie, Yequan and Jiang, Dengyang and Chen, Zhixuan and Liang, Xiaozhuan and Che, Haoxuan and Liu, Quande and Chen, Xi and Chen, Hao},
  journal = {arXiv preprint arXiv:2611.06748},
  year    = {2026}
}

The world should keep evolving—beyond the visible state.