In plain words
The paper introduces Stateful Harness, a framework for making generated videos follow game rules when the right outcome depends on information hidden from view. It checks the game’s state and rules first, then asks a frozen video model to show the resulting event.
A vision-language model suggests what changed in the clip, and the rules check those suggestions before the next turn.
In 50 paired cases, the strongest unwrapped video systems showed both required endings in only 32% of pairs. With the same MiniMax-H3 model, the full framework reached 92% on that measure.
The research
When a generated interaction depends on information the viewer cannot see, how can its video show the right consequence rather than merely a plausible response?
The question matters for rule-governed scenes where the same visible action should lead to different outcomes depending on the world’s state.
Stateful Harness resolves each action under fixed game rules, then uses a frozen video model to render the event and inform the next turn.
The authors tested 50 paired cases where the same opening image and action required different endings because of a hidden value. A multimodal judge checked the clips, and a pair counted only when both showed their required endings. Under Rule-only instructions, the strongest unwrapped systems reached 32% on this measure.
For MiniMax-H3, stating the required ending raised that score to 64%, but only 4% of pairs met the stricter rule that also excludes forbidden endings. With the same frozen model, the full harness reached 92% and 90% on the two measures, supporting the combined approach on these cases.
The evidence comes from 50 designed hidden-state cases, plus three illustrative multi-turn episodes. It supports the approach for the tested interactions, but does not establish performance across arbitrary worlds, rules, or longer episodes.
Analysis
Deep
- Original paper authors
- Chan Yan Ki, Yanran Zhang, Wenzhao Zheng, Xiaofeng Wang, Zheng Zhu
- Submitted to arXiv
- License
- CC BY 4.0