RIVET preprint reports more solved tasks with 74.7% lower model-call cost

Researchers report that RIVET, a system for choosing language models during multi-turn tasks, solved more held-out MINT benchmark tasks than the strongest multi-turn baseline while using 74.7% less total call cost. The method uses responses recorded in past interactions to train a model-selection system. A training-only estimator observes the response each model produced, helping a separate estimator learn which model to select before a call. At deployment, routing uses only information available before the response arrives. The result is a preprint benchmark finding, with the reported cost comparison confined to MINT.