Researchers have developed a framework called SCOPE for improving video world models used in planning and decision-making. These models are simulators that mimic real-world environments to help robots or agents make decisions. However, it's challenging to evaluate the improvements made to these models during inference time because multiple components of the model may change together. SCOPE addresses this issue by introducing a typed state that represents external controls and updates it through bounded changes supported by evidence. The framework improves the performance of frozen video world models on the Physics-IQ benchmark, but its benefits do not transfer uniformly across different models and settings. The study suggests that reliable inference-time adaptation requires not only better proposals but also a principled mechanism for deciding which updates should be part of the deployed system. --- Why it matters: This research is important because it tackles the challenge of evaluating improvements to video world models during inference time, which is crucial for developing reliable and efficient decision-making systems. The findings have implications for the development of autonomous robots or agents that rely on these models. Source: https://arxiv.org/abs/2608.15043