Efficient expert selection is a central challenge for Mixture-of-Experts (MoE) systems that combine heterogeneous large language models (LLMs). Existing solutions rely on jointly trained gating networks, prompt‑based predictors, or costly multi‑model inference, which limit flexibility or increase computation. ScoutRoute introduces a lightweight post‑hoc routing framework that extracts internal signals from a cheap scout model in a single forward pass to predict each expert's competence. The approach consists of three steps:
- Use hidden‑state representations of the scout model to train lightweight logistic probes that estimate cross‑model competence;
- Convert model logits into confidence scores and integrate them with domain‑aware routing logic;
- Perform expert selection without modifying any underlying language model.
The method was evaluated on five open‑weight LLMs across four benchmark datasets covering mathematical reasoning, code generation, commonsense reasoning, and multidisciplinary knowledge. Experiments show that ScoutRoute consistently outperforms individual routing signals while requiring only one scout‑model inference. Moreover, the competence probes achieve better calibration than conventional logit‑based confidence measures, demonstrating that internal model representations provide transferable information for accurate and efficient expert routing.
Review: ScoutRoute leverages internal model features to achieve fast, scalable MoE routing, offering a practical pathway for collaborative inference among heterogeneous LLMs.