JOVE preprint reports lower reasoning costs and latency through model allocation

A JOVE preprint reports competitive accuracy against standard inference baselines across four reasoning benchmarks while reducing average cost and latency by at least 3.17 times. The framework divides complex reasoning queries into dependent subtasks and assigns them to different language models under a long-term budget and a per-query latency constraint. JOVE also selects intermediate outputs for paid verification, which runs asynchronously. That feedback updates estimates of each model's quality for particular tasks and informs future assignments. Its allocation decisions balance current execution spending against the value of learning which models perform well. The reported gains come from benchmark evaluations.