TSRouter, developed by Fangxu Yu, is a graph-based dynamic routing framework that selects the optimal modality and model for time series reasoning queries. It constructs a heterogeneous graph of task, query, modality, and model nodes to evaluate modality-model pairs based on performance-cost preferences. Evaluations across four tasks show "16% to 46% relative improvements" over baselines, with robust zero-shot generalization to unseen models and reduced computational overhead.
No score is assigned. Sources and their independence are shown in the citation chain below.