SCRouter preprint tests security-aware model and token selection

Researchers propose SCRouter, a system that selects a language model and output token budget by weighing answer usefulness, security and cost. The preprint reports improvements over baseline routers of 13-29 points in-domain and 19-23 points out-of-domain. SCRouter estimates whether each model's response will finish within its token allowance, predicts utility and security separately, and uses a small probe to weight those objectives for each query. Its benchmark contains 38,869 utility and security queries spanning nine security dimensions, tested across 12 models and 12 budgets, plus 2,753 out-of-domain queries. The reported gains concern benchmark performance.