VDAR-Router: Adaptive LLMs Routing via Verbalized Query Difficulty Analysis Retrieval
The router decides which LLM gets a query by first spelling out how hard the query is.
VDAR-Router generates an explicit difficulty analysis, retrieves past examples with similar difficulty profiles, then weighs candidate models by performance and cost. The authors say this avoids routing based mainly on surface semantics or embedding similarity. Experiments on three datasets found better cost-performance trade-offs than baseline routers. Case studies in the paper claim the verbalized analysis leads to more relevant retrieval and steadier routing decisions. Source: ArXiv · AI/CL/LG's note
VDAR-Router generates an explicit difficulty analysis, retrieves past examples with similar difficulty profiles, then weighs candidate models by performance and cost. The authors say this avoids routing based mainly on surface semantics or embedding similarity. Experiments on three datasets found better cost-performance trade-offs than baseline routers. Case studies in the paper claim the verbalized analysis leads to more relevant retrieval and steadier routing decisions. Source: ArXiv · AI/CL/LG's note
score 4