Routing is the practice of sending each query to the most capable language model while keeping costs down. A common approach is to train a specialized router, but such routers can become stale when models or tasks shift. The COMPASS paper, posted on arXiv, proposes an alternative: an agentic router in which an LLM itself follows a routing recipe to decide which model to use.
COMPASS stands for Capability-Oriented Model Profiling with Adaptive Search, Selection, and Tool Use. Rather than relying on a fixed classifier, the router gathers performance information on the fly and uses adaptive search and selection to match queries to models. The abstract indicates that the router can also use tools as part of its process.
Because the routing logic is driven by an LLM following a recipe, the system may adapt to new models and changing task demands without retraining a dedicated router. The paper's abstract does not provide experimental results in the available excerpt, so the claimed benefits are not yet verified in this summary.