Routing inputs to the right expert is a core problem in Mixture-of-Experts (MoE) systems, especially when the experts are a heterogeneous collection of independently trained large language models. The new arXiv paper "ScoutRoute: Efficient Mixture-of-Experts Routing Using Internal Language Model Signals" addresses this issue.
The paper proposes ScoutRoute, which leverages internal language model signals to guide routing. The abstract begins to contrast this with existing routing approaches, but the sentence cuts off before explaining what those approaches typically do or how ScoutRoute improves on them.
Because the abstract is truncated, the available text only establishes the problem and the general approach. It does not include experimental results, comparisons to prior methods, or specifics about the internal signals used.