Pretrain Once, Route Anywhere: Towards a Foundation Model for LLM Routing

The paper proposes RouteFM, a foundation model for LLM routing that learns a reusable routing capability generalizing across tasks, candidate models, and deployment conditions instead of requiring retraining. Rather than binding routing decisions to fixed model identities, RouteFM learns to characterize anonymous candidate models from behavioral context and infer their target-specific capabilities. Through episodic pretraining across heterogeneous routing environments, a frozen router can reuse this capability and adapt to new environments through context alone. Experiments on MMR-Bench show RouteFM outperforms the strongest baseline with minimal observations.

RouteFM characterizes anonymous candidate models from behavioral context and infers their target-specific capabilities.

Episodic pretraining across heterogeneous routing environments allows a frozen router to adapt through context alone.

On MMR-Bench, which is excluded from pretraining, RouteFM outperforms the strongest baseline by 2.23 quality points with only eight observations per candidate.