New Protocol Evaluates World-Model Cascades for Adaptive Inference

This research introduces a novel evaluation protocol designed to assess the efficiency of cascaded world-action models, specifically focusing on whether a less complex "Medium" predictor can effectively decide when to invoke a more comprehensive "Full" predictor. The protocol uses paired exact-reset physical outcomes, ensuring that both predictors operate on the same initial state and candidate actions, with the routing target defined by their physical-loss difference. This rigorous setup aims to determine if the computational overhead of switching to a Full model is justified by improved task-specific decision loss. Applying this protocol to a PushT bank, the study demonstrated that a frozen prediction-interface router successfully reduced overhead-inclusive decision costs compared to using either the Medium or Full models alone, or a latency-advantaged task-only router. A subsequent confirmation on a new PushT dataset further supported these findings, showing a measurable reduction in priced physical decision cost. However, the research also notes that this sequential router is still slower than fixed policies and its advantages are limited to scenarios with low compute prices, indicating that while it provides incremental routing information, it doesn't guarantee overall compute savings or broad generality.