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

Pengfei Zhou; Yixing Ma; Hexin Wang; Weidong Tang; Zhenglin Wan; Wangbo Zhao; Yang You

cs.CVNov 3, 2026

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

Authors: Pengfei Zhou, Yixing Ma, Hexin Wang, Weidong Tang, Zhenglin Wan, Wangbo Zhao, Yang You

Organizations: National University of Singapore, HPC-AI Lab · InfRec, Cardinal AI Lab · University of California, Berkeley · Hong Kong University of Science and Technology

Abstract

Artificial general intelligence (AGI) requires a method that can understand, generate, reason, and act across diverse domains and modalities. Rather than relying on a single model to master every capability, a promising path toward AGI now is to orchestrate a heterogeneous ecosystem of specialized models, tools, and policies (candidates) into a unified AI system. At the heart of this vision lies a fundamental question: how can an AI system dynamically select and coordinate the most capable candidate for each task while balancing performance and computational cost? In this work, we present Route Anything, a unified framework for studying model and policy routing across general-purpose AI systems. We develop a unified theoretical foundation that characterizes optimal routing under quality--cost trade-offs and establishes how the information available to a router determines its attainable performance. To systematically investigate routing across diverse capabilities, we first construct an omni-modal trajectory matrix spanning 25 benchmarks across language understanding, image, video, and speech generation, coding and workplace agents, and robotic manipulation, comprising 374,401 scored runs from 452 candidate configurations. Building on this framework, we train and test different routers to investigate the potential of omni-modal model routing, the capabilities of learned routers, and the conditions under which routing improves efficiency. Our results reveal substantial opportunities for capability orchestration: per-input oracle routing outperforms the best fixed candidate by 5.4 points, and learned routers outperform it by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation while using 24% less policy time. More importantly, routing raises efficiency by 8.82x on average compared with the reference model.

Figures & tables

Figure 1

Group	Benchmarks	Candidates	Inputs	Fixed	Best fixed in hindsight	Oracle	Gap
Text QA	7	3–4	9,482	83.27	83.27	87.56	4.28
Image-text QA	3	3	5,708	85.32	85.32	88.02	2.70
Image generation	5	4–6	1,515	82.93	82.99	87.92	4.99
Video generation	1	3	158	62.43	62.43	81.36	18.93
Speech generation	2	3	608	98.62	98.63	99.30	0.68
Code and workplace agents	5	4	343	69.17	69.78	77.29	8.11

Table 1: Per-input potential gain on the test split. Inputs on which every candidate of the pool is scored; values are benchmark metrics on a 0–100 scale, pooled over inputs. Fixed plays the best candidate per benchmark on training inputs, best fixed in hindsight the best candidate per benchmark on these test inputs, and the oracle the best candidate per input.

Figure 3

Study	Policies compared	Benchmarks	Inputs	Seeds	Fixed	Oracle
Potential gain (§ 3 )	fixed candidate, hindsight lookups	24	19,814	–	82.60	87.98
Rule-based tree (§ 3 , § 4.1 )	rule-based type router	18	5,034	–	77.03	87.44
One router (§ 4.1 )	capability-tree and untrained routers	16	3,444	1	79.68	84.80
Robots (§ 4.2 , § 4.3 )	outcome and winner-label routers	1	2,000	1	70.45	88.85
Targets (§ 4.3 , § 4.4 )	outcome and type routers	11	1,302	4	68.30	75.84
Winner labels (§ 4.3 )	12 winner-label recipes	16	3,211	1	79.44	85.77

Table 2: Evaluation panels. Each study uses the test inputs on which all its compared policies are scored; fixed-candidate and oracle scores (0–100) are comparable only within a row. The image-routing suite of Section 4.1 uses 158 GenEval2 and 193 Qwen-Image-Bench prompts, and the image price routes of Section 5.1 use the 1,515 image-generation inputs of Table 1 .

Group (benchmarks)	Inputs	Fixed	Router	Oracle	Router $-$ fixed	Share
Text QA (3)	588	65.14	64.97	71.43	$-0.17$	$-3\%$
Code and workplace agents (5)	519	69.24	70.01	74.78	$+0.77$	14%
Image generation (5)	1,559	82.96	83.21	87.92	$+0.25$	5%
Video generation (1)	161	62.17	64.28	80.22	$+2.12$	12%
Speech generation (2)	617	98.63	98.63	99.30	$0.00$	0%
All 16 benchmarks	3,444	79.68	79.98	84.80	$+0.30$	6%

Table 3: Learned routers against the fixed candidate and the oracle. Test inputs; one training seed. Share is the router’s gain over the fixed candidate divided by the oracle’s. The first block is the capability-tree router; the robot row is the outcome router of Section 4.2 .

Figure 6

Figure 7

Study	Policy	Inputs	Score	$-$ matched fixed [95% CI]
11 benchmarks, 4 seeds	Matched fixed candidate	1,302	68.30	–
	Outcome router	1,302	67.90	$-0.40$ $[-1.19,+0.33]$
	68-type classifier + table	1,302	66.98	$-1.32$ $[-3.05,+0.34]$
	Oracle	1,302	75.84	$+7.55$
16 benchmarks	Fixed candidate (best training mean)	3,211	79.44	–
	Winner-label router (generative)	3,211	78.87	$-0.58$

Table 4: Training targets and the fixed baseline. Paired 95% intervals over source groups (scenes for robots). The first block averages four training seeds on 1,302 test inputs of 11 text, agent, and image benchmarks; the fixed candidate is matched (Section 4.4 ).

Figure 9

Testbed (resource)	Reference	Route	$\Delta V$ [95% CI]	$\eta/\eta_{B}$
Code, CodeRouterBench ($)	Always Opus 4.6	Fitted, $\epsilon=0$	$-0.44$ $[-1.75,+0.97]$	$25.5\times$
		Certified, $\epsilon=0$	$\geq 0$ in 200 of 200	$19.4\times$
Text QA, RouterBench ($)	Always GPT-4	Fitted, $\epsilon=0$	$-0.40$ $[-0.84,+0.08]$	$7.1\times$
		Certified, $\epsilon=0$	$\geq 0$ in 200 of 200	$6.7\times$
Image generation ($)	Fixed generator	Fitted, $\epsilon=0.5$	$-0.47$ $[-0.87,-0.06]$	$1.35\times$
		Fitted, $\epsilon=2$	$-1.56$ $[-2.39,-0.73]$	$2.03\times$

Table 5: Price routes across testbeds. Test-set quality change

\Delta V=\mathbb{E}[V-Y_{B}]

in points (0–100) with paired 95% intervals, and efficiency

\eta/\eta_{B}

against the reference. Fitted routes meet the floor on fitting data; certified routes are medians over 200 random halvings of the fitting data (Proposition 5 ).

Figure 11

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Inputs	Fixed candidate	Fixed	Best fixed in hindsight	Oracle	Gap
Text QA
MMLU-Pro	2,403	Opus 5.5	92.55	92.55	95.21	2.66
MMLU-Redux	1,080	Opus 5.5	96.67	96.67	98.06	1.39
SuperGPQA	5,300	Opus 5.5	78.26	78.26	83.66	5.40
GPQA (extended)	106	Opus 5.5	90.57	90.57	92.45	1.89
AA-LCR	42	Opus 5.5	97.62	97.62	100.00	2.38

Appendix

Table 6: Per-benchmark potential gain on the test split (Table 1 ). The best fixed candidate in hindsight is named when it differs from the fixed candidate chosen on training inputs.

Benchmark	Inputs	Fixed candidate	Fixed	Router	Oracle
AA-LCR	42	Opus 5.5	97.62	95.24	100.00
AA-Omniscience	120	Opus 5.5	77.50	77.50	85.83
HLE (text)	426	Opus 5.5	58.45	58.45	64.55
SciCode	187	Opus 5.5	70.05	70.05	70.59
Terminal-Bench	36	Opus 5.5	69.44	77.78	86.11
AutomationBench	160	Opus 5.5	92.75	92.75	96.53

Appendix

Table 7: Per-benchmark results of the capability-tree router (first block of Table 3 ); one training seed. SciCode steps are aligned across models; 175 of its 187 inputs have one scored candidate.

Benchmark	Inputs	Fixed candidate	Fixed	Outcome router	Type router	Oracle
AA-LCR	42	Opus 5.5	97.62	95.24	97.62	100.00
AA-Omniscience	120	Opus 5.5	77.50	77.29	71.88	85.83
HLE (text)	431	Opus 5.5	58.24	57.95	56.61	64.27
AA-Briefcase	1	GPT-6.1 Sol	75.00	62.50	50.00	75.00
AutomationBench	160	Opus 5.5	92.75	92.70	92.54	96.53
GDP.pdf	100	GPT-6.1 Sol	21.00	20.50	20.00	36.00

Appendix

Table 8: Per-benchmark results of the 11-benchmark study (first block of Table 4 ), averaged over four training seeds; the fixed candidate is matched.

		Fixed generator			Half-point route
Benchmark	Test prompts	Candidate	Quality	m$	Candidate	Quality	m$
GenEval2	160	GPT Image 2.5	82.48	208.8	GPT Image 2.5	82.48	208.8
Qwen-Image-Bench	193	GPT Image 2.5	64.09	211.1	Grok Imagine 2.0	61.16	60.0
LongT2I	543	MAI-Image 2.6	86.14	39.4	MAI-Image 2.6	86.14	39.4
OneIG	427	FLUX 3	83.35	48.0	Qwen-Image 3	83.02	40.0
WISE	192	Qwen-Image 3	92.19	40.0	Qwen-Image 3	92.19	40.0

Appendix

Table 9: Image price routes: test quality (0–100) and price (m$ per image) of the fixed generator and of the half-point route of Table 5 , fitted on training prompts.

Policy	$\eta/\eta_{B}$ [CI]	Cost	$\Delta V$ [CI]
21 cells, unconstrained	28.187 [26.840, 29.659]	0.395	$-0.0201$ [ $-0.0343$ , $-0.0056$ ]
21 cells, $\epsilon=0$	25.516 [24.278, 26.903]	0.452	$-0.0044$ [ $-0.0175$ , $+0.0097$ ]
21 cells, $\epsilon=0.01$	27.612 [26.311, 29.072]	0.409	$-0.0140$ [ $-0.0272$ , $-0.0004$ ]
9 dimensions, unconstrained	28.024 [26.713, 29.452]	0.398	$-0.0192$ [ $-0.0332$ , $-0.0047$ ]
9 dimensions, $\epsilon=0$	22.733 [21.584, 23.954]	0.502	$-0.0090$ [ $-0.0227$ , $+0.0060$ ]
Fixed Qwen3-Max dispatch	20.097 [18.965, 21.297]	0.525	$-0.0416$ [ $-0.0549$ , $-0.0275$ ]

Appendix

Table 10: CodeRouterBench policies on the 2,919 test tasks. Dollar efficiency

\eta/\eta_{B}

, cost in m

per task, and

\Delta V=\mathbb{E}[V-Y_{B}]

, with paired 95% intervals. Floors are fitted on probing; at

\epsilon\geq 0.02

they are slack and give the unconstrained policy. Fixed dispatch plays the best probing model, which is also the cheapest (Qwen3-Max); the per-task oracle, an in-sample ceiling, gives each task the candidate maximizing

V-\eta^{\star}C$ on its realized outcomes.

Dataset, threshold	Splits	Count	Paired	Residual	Indep.	Saving	Marginal
CodeRouterBench, $\theta=1$	1	139	97.1%	86.3%	86.3%	95.7%	46.8%
CodeRouterBench, $\theta=0.5$	1	194	96.4%	94.3%	91.8%	93.3%	40.7%
RouterBench, $\theta=0.5$	10	5,539	63.9%	61.9%	35.9%	63.5%	14.6%
RouterBench, $\theta=1$	10	212	64.6%	53.8%	36.3%	48.1%	0.0%

Appendix

Table 13: Best-candidate selection: the share of comparisons (one portfolio with several candidate additions) in which a rule’s fitted choice has the highest test efficiency. Counts sum over splits, which share tasks. Paired ranks

G_{m}(S)

of Eq. ( 8 ); residual ranks

N_{m}(S)

; Indep. assumes independent failures; saving ranks

r_{m}b-d_{m}

Verifier	AUROC	$r(\theta=0.5)$ [95% CI]	$r(\theta=1)$ [95% CI]	$Q(\theta=0.5)$
Compiler (static gate)	0.766	4.09 [3.23, 5.45]	4.09 [3.23, 5.45]	$-0.024$
Visible examples	0.537	1.98 [1.72, 2.35]	1.06 [1.03, 1.10]	$-0.036$
Hidden tests	0.994	2.87 [2.38, 3.53]	2.37 [2.03, 2.80]	$+0.024$
LLM judge	0.824	1.09 [1.00, 1.21]	0.64 [0.61, 0.66]	$-0.012$
Released-score oracle	1.000	2.46 [2.09, 2.93]	2.46 [2.09, 2.93]	$+0.045$

Appendix

Table 14: Verifier choice changes the cascade’s efficiency and quality.

r=\eta/\eta_{B}

;

Q

is mean score change. Qwen3-Max primary; the judge is charged at the primary’s token price.

Signal source	Native [95% CI]	Threshold [95% CI]	Within dimension
Fine-tuned Qwen 27B	0.555 [0.490, 0.614]	0.555 [0.490, 0.614]	0.062
Fine-tuned Qwen 9B	0.546 [0.475, 0.608]	0.546 [0.475, 0.608]	0.078
Fine-tuned Qwen 2B	0.466 [0.397, 0.531]	0.417 [0.355, 0.479]	$-0.006$
Fine-tuned Qwen 0.8B	0.422 [0.348, 0.491]	0.082 [0.066, 0.098]	$-0.008$
Logistic regression	$-1.218$ [ $-1.590$ , $-0.869$ ]	0	$-0.002$
TF-IDF MLP	$-1.736$ [ $-2.178$ , $-1.342$ ]	0	$-0.012$

Appendix

Table 15: Cross-fitted capture of the nine released CodeRouterBench routers as signals

Z

(chose Qwen3-Max) for Qwen3-Max at

\theta=1

, with no signal charge. Native use attempts on

Z=1

, threshold use when

s_{Z}>a/b

; the last column applies the threshold rule given the benchmark dimension and divides the aggregated saving by the aggregated bound.

Adaptive Model Routing

Adaptive model routing selects models or specialized experts per input to balance prediction quality, computational cost, latency, and resource constraints.

See all 96 papers

Feb 23, 2026

SkillOrchestra: Learning to Route Agents via Skill Transfer

Compound AI systems promise capabilities beyond those of individual models, yet their success depends critically on effective orchestration. Existing routing approaches face two limitations: (1) input-level routers make coarse query-level decisions that ignore evolving task requirements; (2) RL-trained orchestrators are expensive to adapt and often suffer from routing collapse, repeatedly invoking one strong but costly option in multi-turn scenarios. We introduce SkillOrchestra, a framework for skill-aware orchestration. Instead of directly learning a routing policy end-to-end, SkillOrchestra learns fine-grained skills from execution experience and models agent-specific competence and cost under those skills. At deployment, the orchestrator infers the skill demands of the current interaction and selects agents that best satisfy them under an explicit performance-cost trade-off. Extensive experiments across ten benchmarks demonstrate that SkillOrchestra outperforms SoTA RL-based orchestrators by up to 22.5% with 700x and 300x learning cost reduction compared to Router-R1 and ToolOrchestra, respectively. These results show that explicit skill modeling enables scalable, interpretable, and sample-efficient orchestration, offering a principled alternative to data-intensive RL-based approaches. The code is available at: https://github.com/jiayuww/SkillOrchestra.

Jiayu Wang, Yifei Ming, Zixuan Ke +3

University of Wisconsin-Madison · Salesforce AI Research

LLM Agent Orchestration LLM Agent Routing LLM Agent Skill Learning

Ask PDF

Sep 26, 2026

HM-ROUTER: Joint Model and Harness Routing for Agentic Systems

Agent performance depends on both the underlying model and the harness that manages its tool use and execution. Selecting a suitable pair requires accounting for their compatibility, yet training samples may cover only a subset of the growing combination space. We introduce HM-Router, a routing method that jointly selects a model and harness for each query. It learns separate model and harness representations shared across routes, with an interaction term inspired by canonical polyadic (CP) tensor decomposition to capture how their compatibility varies with the query. This sharing allows training samples from observed pairs to inform predictions for unobserved combinations. We curate a benchmark from 12 public agent benchmarks, covering 293 routes, 73 models, and 25 harnesses. HM-Router exceeds the strongest evaluated learned baseline by 7.3 percentage points in mean routing accuracy and leads at all seven evaluated cost budgets on the six-benchmark subset. When 90% of routes have their training outcomes withheld, allowing unobserved combinations improves normalized accuracy by 15.8 points over restricting the same router to observed routes. HM-Router has also demonstrated training sample efficiency for new routes and components and generalization to unseen benchmarks. Our code and data are open-sourced at https://github.com/hmarkc/HM-Router.

Hao Mark Chen, Royson Lee, Yasuyuki Okoshi +3

Imperial College London · Samsung · Institute of Science Tokyo

LLM Agent Routing LLM Agent Harnesses

Ask PDF

May 13, 2026

Position: Agentic AI System Is a Foreseeable Pathway to AGI

Is monolithic scaling the only path to AGI? This paper challenges the dogma that purely scaling a single model is sufficient to achieve Artificial General Intelligence. Instead, we identify Agentic AI as a necessary paradigm for mastering the complex, heterogeneous distribution of real-world tasks. Through rigorous theoretical derivations, we contrast the optimization constraints of monolithic learners against the efficiency of Agentic systems, progressing from simple routing mechanisms to general Directed Acyclic Graph (DAG) topologies. We demonstrate that Agentic AI achieves exponentially superior generalization and sample efficiency. Finally, we discuss the connection to Mixture-of-Experts, reinterpret the instability of current multi-agent frameworks, and call for greater research focus on Agentic AI.

Junwei Liao, Shuai Li, Muning Wen +2

1Shanghai Jiao Tong University 2Shanghai Innovation Institute · 1Shanghai Jiao Tong University · University College London.

Multi-Agent Systems Agentic AI Artificial General Intelligence

Ask PDF

Abstract

Adaptive Model Routing

Explore similar work

SkillOrchestra: Learning to Route Agents via Skill Transfer

HM-ROUTER: Joint Model and Harness Routing for Agentic Systems

Position: Agentic AI System Is a Foreseeable Pathway to AGI