The paper proposes treating LLM routing, the task of assigning incoming queries to the best-fit model from a heterogeneous pool, as a foundation model problem rather than an environment-specific optimization. The resulting system, RouteFM, learns to characterize anonymous candidate models from behavioral context and infer their capabilities without binding decisions to fixed model identities, using episodic pretraining across many diverse routing environments so a single frozen router can adapt to new settings purely through context. On MMR-Bench, a benchmark held out from pretraining, RouteFM outperformed the strongest baseline by roughly 2.23 quality points while using only eight behavioral observations per candidate model. The authors frame this as a shift from repeated local router optimization toward a reusable pretrain-once, route-anywhere paradigm.
