Route Anything: Model and Policy Routing for General-Purpose AI Systems
What it is
An arXiv paper proposing a unified framework for studying model and policy routing, where an AI system dynamically selects among heterogeneous specialized models, tools, and policies (candidates) per task under quality-cost trade-offs. It develops a theoretical foundation for optimal routing and how router information limits attainable performance, and builds an omni-modal trajectory matrix of 374,401 scored runs from 452 candidate configurations across 25 benchmarks (language, image, video, speech, coding/workplace agents, robotic manipulation). It then trains and evaluates routers, reporting that per-input oracle routing beats the best fixed candidate by 5.4 points and that learned routers improve on the best fixed candidate by 0.3 points across 16 benchmarks, 2.1 points on video generation, and 1.1 points on robot manipulation while using 24% less policy time, with an average 8.82x efficiency gain versus the reference model.
The problem it solves
A single model cannot be best at every domain and modality, and always calling the strongest candidate is wasteful; there is no unified way to decide which candidate to invoke per input while balancing quality against computational cost.
What you could use it for
- Selecting the cheapest sufficient model per request in a multi-model serving stack
- Benchmarking router policies against oracle and best-fixed-candidate baselines
- Studying quality-cost trade-offs across omni-modal tasks (text, image, video, speech, robotics)
- Informing agent orchestration that picks tools or policies per task
Why it matters
If the reported gains hold, routing could let systems match or exceed a single strong model's quality while cutting compute substantially (reported 8.82x average efficiency gain), making heterogeneous model ecosystems a practical alternative to scaling one monolithic model.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.