Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems

Route Anything presents a unified theory and omni-modal trajectory matrix for routing among models, tools, and policies, finding a 5.4-point per-input oracle gap, modest learned-router gains, and 8.82× average efficiency from cheaper routes.

General-purpose AI systems need to understand, generate, reason, and act across domains and modalities, so a system may orchestrate specialized models, tools, and policies rather than rely on one model. Routing research has mostly studied language models, leaving open whether routing helps across modalities. The paper asks how much routing can gain, how well learned routers work, and when a cheaper route is good enough.

The paper develops a unified quality–cost routing theory whose value ladder runs from the best fixed candidate to the per-input oracle, with a reference policy and quality tolerance. It constructs an omni-modal trajectory matrix of 25 benchmarks, 374,401 scored runs from 452 candidate configurations, and 103,307 tasks. Routers fine-tune Qwen3.5-4B on task text and output requirements, and for robots also on three camera frames and robot state. It trains outcome routers, type routers, and winner-label routers, and adds code and text testbeds with recorded prices for cost studies.

• On the test split, the per-input oracle beats the best fixed candidate per benchmark by 5.38 points over 24 benchmarks and 19,814 inputs; group gaps range from 0.68 for speech generation to 18.93 for video generation and 18.40 for robot manipulation. §3
• A capability-tree router over 16 benchmarks scores 79.98 against 79.68 for the fixed candidate and 84.80 for the oracle, a 0.3-point gain over the fixed candidate. §1
• On robots, a router using three camera frames sends clean scenes to PatchWAM and randomized ones to ME-Dex, reaching 71.6% success against 70.5% for the best single policy with 24% less policy time. §1
• With the same inputs and backbone, regressing each policy’s success reaches 71.6% robot success, while classifying the winner reaches 42.8%, because a latency tie-break decides 94% of the winner labels. §1
• Routing raises efficiency by 8.82× on average over four testbeds, including 25.5× on coding tasks and 7.1× on text question answering. §1
• On CodeRouterBench, a free compiler gate gives 4.09× dollar efficiency, while a list-price LLM judge gives 0.30×. §1

• The paper states that its theory takes the law of states as given and does not cover states that depend on the router’s own earlier choices, such as SciCode steps containing a model’s earlier code (§D).
• Certification needs the reference’s outcome on every certification input, and step-level routing for long-horizon agentic tasks deserves deeper exploration (§D).
• The paper states that most of the trajectory matrix records no costs, so Sections 3 and 4 mostly use λ=0 (§2).
• The experiments leave image routing open: none of 124 router configurations on GenEval2 and Qwen-Image-Bench beats the fixed generator significantly, and the fitted zero-loss route loses 0.33 on test at 1.34× efficiency (§1, §B.4).

• Matrix: 25 benchmarks, 374,401 scored runs, 452 candidate configurations, and 103,307 tasks. §B.1
• Oracle gap: 5.4 points over 19,814 inputs and 24 benchmarks. §3
• Learned router: 79.98 vs 79.68 fixed and 84.80 oracle over 16 benchmarks. §1
• Robot router: 71.6% vs 70.5% success with 24% less policy time. §1
• Outcome regression: 71.6% vs 42.8% for winner classification; 94% of winner labels are decided by a latency tie-break. §1
• Matched comparisons: dropping Opus 5.5 for 11 missing HLE scores turned a 0.4-point deficit into a 2.4-point gain. §1
• Efficiency: 8.82× average, 25.5× coding, and 7.1× text QA; CodeRouterBench’s zero-loss floor holds in all 200 splits at a median 19.4×. §1
• Verifier choice changes efficiency: compiler gate 4.09×, list-price LLM judge 0.30×. §1