LLM routing systems decide which model should handle a given request, often coordinating several heterogeneous models across different tasks. According to a new arXiv preprint, however, most existing approaches are built around a predefined execution paradigm—for example, they assume a single-round or multi-round interaction from the start. That separation between choosing a model and choosing how to run it may limit flexibility.

The paper introduces RouteWeaver, a framework that weaves mode selection and execution into unified LLM routing. Instead of treating the execution mode as a fixed input, RouteWeaver appears to make it part of the routing decision itself, allowing the system to select both the model and the interaction pattern together.

Because this is a single preprint and the abstract is truncated, the article cannot confirm how RouteWeaver is implemented or evaluated. No comparisons with other routing methods are available from the source, and no benchmark results are reported in the provided text. The contribution, as described, is conceptual: unifying two decisions that prior work handles separately.