Vision-language-action models increasingly use a dual-system design: a fast policy runs at high frequency, while a slower, more expensive module is available for difficult cases. The central question is when to invoke that slow module. According to the arXiv abstract, published trigger mechanisms for this decision have no shared objective, no measured costs, and no standard for separating appropriate from inappropriate invocations.
The authors argue that this makes it difficult to compare systems or to know whether a given trigger is actually helping. Without a common measure of cost and benefit, a trigger that works in one setting may be wasteful in another.
The paper proposes a decision-theoretic account of slow-module scheduling, along with a measurement standard. This is intended to give researchers a principled way to decide when asking for help is worth the expense, and to evaluate triggers against a common baseline.