Tsinghua's Stateful Harness Pushes Video Rule-Following to 92%

Rule-following in generated video, it turns out, is less a model-capacity problem than a plumbing problem. In a new paper posted to Hugging Face, researchers at Tsinghua University's Department of Automation and GigaAI argue that when a video model is asked to render an action whose outcome depends on state the current frame cannot show, the fix is to keep that state outside the model entirely.

Their framework, Stateful Harness, "maintains an evolving world state under specified rules, renders each resulting event with a frozen video model, and uses the generated clip to update the world state for the next turn." A preview clip from an editable 3D scene feeds spatial and camera guidance alongside the compiled instruction.

To stress-test it, the authors built WorldState-Bench, a benchmark of 50 counterfactual twin pairs that share an initial frame and action but differ in hidden state and required outcome. Under rule-only prompting, "the best of the 11 evaluated video backends produces the required endings for both cases in 32% of pairs." Writing the required endings into the prompt lifts MiniMax-H3 from 32% to 64%; swapping in the full harness around the same frozen model pushes PairFollow to 92% and PairStrict to 90%.

The paper's own framing is blunt: "These results support separating rule execution from video rendering for world interactions with hidden state." In pairwise human evaluation the harness wins 98% of comparisons against baselines, including a 100% rate against both AlayaWorld and Echo-WM Flash.

The work lands amid a run of AI video papers probing the limits of pure pixel-space world models. The benchmark skews toward turn-based scenarios — dungeons, dating sims, card games, escape rooms — and demo episodes run between 11 and 29 turns. The authors do not report runtime overhead for the 3D preview pipeline that sits alongside the frozen generator.