Researchers introduce SeLMRoute, a framework for routing queries to the most suitable large language model among several candidates. Instead of learning routing decisions directly from query embeddings or preference data, the system first evaluates interpretable questions about a query’s reasoning requirements and need for external knowledge, retaining each judgment as a probability distribution. A lightweight supervised router then uses this semantic evidence to estimate how each candidate model would perform, allowing the same representation to support both performance-oriented and cost-aware routing objectives. On LLMRouterBench, a benchmark spanning 15 datasets, 20 candidate models, and 11,481 queries, SeLMRoute reached 72.08% average accuracy, compared with 69.23% for the strongest single fixed model.
