Understand this story
The Department of Automation at Tsinghua University and GigaAI propose the State World Models method, with the paper titled "State World Models: Execute the Rules, Generate the Pixels". The method separates rule execution from pixel generation: a stateful runtime first settles events, then hands them to a frozen video model for rendering, requiring no training or fine-tuning. On 50 counterfactual twin sample pairs, the best paired joint compliance rate among 11 off-the-shelf backends is 32%; after the same frozen MiniMax-H3 is connected to this harness, PAIRFOLLOW reaches 92%, PAIRSTRICT reaches 90%, and TWINCONTRAST reaches 96%. The paper also reports that on WBench, setting adherence rises from 71.1 to 87.2, while image quality, physical plausibility, and text-driven categories are roughly on par with the mean.
AI-generated from public reports
Developments
1 confirmed developments
Tsinghua and GigaAI propose the State World Models method of settling rules first and then rendering
Source:English text is being prepared.
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 21
Tencent Cloud Developer Community
Tsinghua and GigaAI propose State World Models: rules settle first, then render, paired compliance rate rises from 32% to 92%
A paper from the Department of Automation at Tsinghua University and GigaAI, "State World Models: Execute the Rules, Generate the Pixels," proposes separating rule execution from pixel generation: a stateful runtime first settles events, then hands them to a frozen video model for rendering, with zero training and zero fine-tuning. On 50 pairs of counterfactual twin samples, the best paired joint compliance rate among 11 off-the-shelf backends is 32%; the same frozen MiniMax-H3, connected to this harness, reaches PAIRFOLLOW 92%, PAIRSTRICT 90%, and TWINCONTRAST 96%. The paper also reports that on WBench, setting adherence rises from 71.1 to 87.2, while image quality, physical plausibility, and text-driven categories remain roughly on par with the average.