What it is
A unified framework for routing across models and policies in text, generation, agents, and robotics, with theory plus a 25-benchmark omni-modal routing dataset. Instead of single-model or LM-only routing, it studies per-input routing across modalities with a unified quality-cost objective, a theory of routing under available information, and learned routers trained to predict each candidate's outcome rather than winner labels.
Key results
- Oracle routing beats best fixed by 5.4 pts across 24 benchmarks
- Learned routers gain 2.1 pts on video generation vs best fixed
- Robot routing reaches 71.6% vs 70.5% best single policy with 24% less time
Shows that much of the gain in general-purpose AI systems comes from input-level orchestration of specialized candidates, not from picking one globally best model.