Event

Proposes the Stateful Harness framework and the WorldState-Bench benchmark

1 reports · 1 sources

The paper proposes the Stateful Harness framework, which separates world evolution from video rendering, using a frozen video model to render events and write back the world state, so as to support rule-based interaction with hidden states; it also builds the WorldState-Bench benchmark, containing 50 counterfactual twin pairs that share initial frames and actions but differ in hidden states and required outcomes, for evaluating rule-following capability. Evaluation shows that, under the condition of only rule prompts, the best of the 11 tested video backends produced both required outcomes on only 32% of the twin pairs. For MiniMax-H3, explicitly giving the required outcome raised this proportion from 32% to 64%, while the full harness using the same frozen model reached 92%.