RLVR Framework for Efficient SLM-LLM Information Acquisition under API Budgets

This method treats SLM-LLM collaboration as an information acquisition problem rather than a task delegation problem. Using a three-stage RLVR framework, a small language model acts as the primary reasoner and selectively issues targeted queries to a black-box LLM advisor to minimize API costs.