- Who it helps
- Video-model researchers and simulation evaluators.
- Availability
- The November 4 paper, revised November 5, is public. Linked weights were inaccessible.
Track a changing scene
A vision-language model maintains scene state and proposes consequences; a video renderer depicts them when the scene is revisited.
Read the evaluation boundary
The implementation combines Qwen3.8-27B and LTX-2.5. Reported benchmark gains are author measurements; plausible reactions do not establish a physically accurate simulator.
What you can do
- Read the original method and its limitations before planning your own evaluation. Open reference ↗
Where the evidence stops
No weights downloaded or videos generated. Artifact access and independent replication remain unresolved.
Sources read. Sources read November 9, 2026 UTC. No third-party code, model inference or complete workflow was executed.
Sources & further reading
Paper · paper
EvoWorld research paper ↗
November 4 submission, November 5 revision; world-state reasoning framework.
Yequan Bie and colleagues · page checked 2026-11-09 · Published 2026-11-04
Paper · paper
EvoWorld methods and evaluation ↗
Implementation and evaluation sections read; no reproduction.
Yequan Bie and colleagues · page checked 2026-11-09
More sources reviewed (2)
Hugging Face · analysis
Daily papers ↗
Nine opening paper cards inspected, including EvoWorld.
Hugging Face · page checked 2026-11-09
Hugging Face · analysis
EvoWorld feed signal ↗
Feed identity prompted the out-of-view question. Direct follow-up failed; original arXiv paper was readable.
Hugging Face · search result checked 2026-11-09
Coverage note: Independent execution and qualifying cross-platform attention remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
EvoWorld studies whether generated scenes change plausibly while objects are out of view.