Routing Should Pay for Itself: Sparse Supervision for Economical LLM Routing
SaveRouter cuts the training-feedback bill before an LLM router ever starts saving money.
The paper argues that routing systems can spend too much upfront collecting quality feedback from multiple candidate models. SaveRouter selectively gathers feedback and shares capability signals across related queries while still refining decisions at the query level. In four benchmarks, it used about 33-41% of available training feedback while keeping competitive or better routing quality. The authors say it reduced break-even deployment volume by roughly 1.9-9.5x versus the fastest conventional router. HF Daily Papers' note
The paper argues that routing systems can spend too much upfront collecting quality feedback from multiple candidate models. SaveRouter selectively gathers feedback and shares capability signals across related queries while still refining decisions at the query level. In four benchmarks, it used about 33-41% of available training feedback while keeping competitive or better routing quality. The authors say it reduced break-even deployment volume by roughly 1.9-9.5x versus the fastest conventional router. HF Daily Papers' note
score 4