Pretrain Once, Route Anywhere: Towards a Foundation Model for LLM Routing
RouteFM is pitched as a reusable router that adapts to new model pools from behavioral context instead of retraining.
The paper frames LLM routing as a general capability, not a local fit to one workload and candidate set. RouteFM learns from anonymous candidate-model behavior and infers task-specific strengths without relying on fixed model identities. In tests, the authors report transfer across domains, modalities, candidate pools, and context budgets, with the biggest gains when little behavioral evidence is available. On MMR-Bench, which was not used in pretraining, it beats the strongest baseline by 2.23 quality points using eight observations per candidate. HF Daily Papers' note
The paper frames LLM routing as a general capability, not a local fit to one workload and candidate set. RouteFM learns from anonymous candidate-model behavior and infers task-specific strengths without relying on fixed model identities. In tests, the authors report transfer across domains, modalities, candidate pools, and context budgets, with the biggest gains when little behavioral evidence is available. On MMR-Bench, which was not used in pretraining, it beats the strongest baseline by 2.23 quality points using eight observations per candidate. HF Daily Papers' note
score 4