Cost-Aware Mixture-of-Experts Coordination for Model Markets

arXiv:2610.11908 · cs.DB, cs.LG · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Cost-Aware Mixture-of-Experts Coordination for Model Markets".

Jane: The gist The proposed MoE-based model market framework lifts Mixture-of-Experts from a model-level learning architecture to a market-level coordination mechanism where brokers use gating networks to coordinate multiple heterogeneous experts…

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Let's get into the specifics of who wrote this. The authors are Yizhou Ma, Wenbo Wu, Xikun Jiang, Zhuoqin Yang, and Luis-Daniel Ibáñez from various universities including Southampton and Aalborg.

Jane: They’re researchers diving deep into how to structure these model markets so that they actually benefit from the differences between the experts involved.

Lu: The title itself tells you a lot about their goal: making coordination cost-aware, which means they aren't just looking at accuracy but also what it costs to get that accuracy.

Meng: So, we’re talking about a framework where brokers use gating networks to coordinate multiple heterogeneous experts and deliver one composite model service. That’s the core idea.

Tom: Exactly. It addresses the problem that existing marketplaces treat individual models like these indivisible commodities, which limits their ability to see how different models complement each other when you need something complex.

Jane: So, they want to lift Mixture-of-Experts from just being a model-level learning architecture to this market-level coordination mechanism where brokers manage the experts.

Lu: This shifts the problem definition from training a single unified system to designing a system that intelligently routes requests through several different specialized AI models simultaneously.

The paper's summary: Tom: So, what exactly is this framework they’ve put together in "Cost-Aware Mixture-of-Experts Coordination for Model Markets"? It describes the MoE Market as a tuple B, N, P, G, E.

Jane: That means we have buyers who make requests and brokers who coordinate everything—they manage the data providers and the set of heterogeneous experts available.

Meng: And the key piece here is that each broker operates a gating network G that routes those incoming requests to a specific subset of experts, producing a participation vector wk.

Lu: That routing decision is what makes this different; it's not just about which expert gives the best prediction for the input rk, but also which expert is economically worthwhile given its execution cost and expected compensation.

Tom: They formalize the workflow in six steps: request submission, data acquisition from providers Pk, then expert coordination using that gating network to get wk.

Jane: Then you get this composite service generation where those activated experts jointly produce an output through weighted aggregation based on the routing decisions.

Meng: And finally, the buyer pays a service price pi hat k to the broker, and what’s left over is this residual revenue pool Rk that gets allocated among the experts according to some cost-aware rule.

The paper's improvements: Tom: Now let's talk about what they actually improved. They introduce two major things here: a cost-aware gating mechanism and a cost-adjusted revenue allocation rule for the residual pool Rk.

Jane: The gating mechanism is designed to maximize market welfare by learning the gating parameters phi by minimizing a market-aware loss Lmarket.

Lu: That loss function rewards routing policies that produce high quality composite predictions while penalizing experts that are too costly to use, which is a really smart way to balance the two.

Meng: And then there's this revenue allocation rule where the residual revenue Rr is distributed among experts based on a score called si of r.

Tom: That score is defined as E over D r of h w i x exp minus lambda c i x, and lambda controls how much discount you give for higher costs.

Jane: So, the allocation rule means revenue isn't just split based on how many experts participated, but it’s adjusted by both their contribution to the prediction and their actual execution cost.

Lu: The authors show that this cost-adjusted allocation is budget-balanced, participation monotonic, and cost sensitive. It’s a solid mechanism for ensuring the system remains economically grounded while still prioritizing performance.

Conclusion: Tom: So to wrap up on "Cost-Aware Mixture-of-Experts Coordination for Model Markets," the main point is that this MoE framework successfully formalizes the whole structure—the participants, the workflow, and the welfare objective.

Jane: They show that by deriving these cost-aware gating and allocation mechanisms, you can balance predictive utility against measured execution cost in a way that previous methods didn't really address.

Meng: The result is that this framework demonstrates that MoE can function as a market-level coordination principle for collaborative, cost-aware, and economically grounded model marketplaces.

Lu: It suggests that we can move toward more sophisticated AI services where the routing isn't just about finding the best prediction but also about understanding the economic trade-offs involved in achieving it.

Tom: They confirm this by showing that MoE Market achieves the highest mean welfare on all fifteen datasets and a lower mean expected cost than Standard MoE in every case.

Jane: And they also showed that increasing lambda consistently shifts revenue toward experts who have lower execution costs, which is a clear economic signal for how to use the system effectively.

Meng: It’s efficient because it’s fast to compute compared to exact Shapley allocation, and it shows a systematic sensitivity to both expert participation and cost.

Lu: This opens the door for dynamic model markets where expert availability or costs might change over time, which is the next logical step for this kind of coordination.

Tom: That’s the picture with the Cost-Aware Mixture-of-Experts Coordination for Model Markets. We've seen how to ground these complex architectures in real economic terms.

Yizhou Ma, Wenbo Wu, Xikun Jiang, Zhuoqin Yang, Luis-Daniel Ibáñez

University of Southampton

cs.DB, cs.LG

Submitted: 2026-10-08

Updated: 2026-10-08

Code: https://github.com/MAYIZHOU/artifact-code_moe

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: The gist The proposed MoE-based model market framework lifts Mixture-of-Experts from a model-level learning architecture to a market-level coordination mechanism where brokers use gating networks to

Key concepts

MoE Market Framework
This is a structure where brokers act as intermediaries coordinating multiple specialized AI experts. Instead of trading individual models, the broker uses a gating network to decide which experts participate in generating a composite model service for a buyer's specific request. It treats experts as collaborative resources rather than independent commodities.
Gating Network
The gating network is the mechanism used by brokers to coordinate expert participation. It takes the request and context as input and outputs a vector indicating how much each expert should contribute to the final composite model. The broker learns this network to maximize overall market welfare, balancing prediction quality against expert usage costs.
Cost-Aware Gating Mechanism
This learning objective guides the gating network training. The broker optimizes the gating parameters by maximizing a welfare function that rewards high-quality predictions while penalizing the total expected cost of using various experts. This ensures that routing decisions favor combinations of experts that are both accurate and economically efficient.
Cost-Adjusted Revenue Allocation
After a service is delivered, the remaining revenue pool is distributed among participating experts. Instead of simple proportional sharing, this method uses a cost-adjusted score that discounts the revenue contribution based on an expert's individual execution cost. This ensures the allocation is budget-balanced and favors lower-cost experts when costs are high.

Terminology

Summary

The gist The proposed MoE-based model market framework lifts Mixture-of-Experts from a model-level learning architecture to a market-level coordination mechanism where brokers use gating networks to coordinate multiple heterogeneous experts and deliver a composite model service.

Problem Motivation

Existing model marketplaces typically trade and select individual models as indivisible units, limiting their ability to exploit complementarities among heterogeneous experts<ref:2610.11908#pg6>. This paper proposes an MoE-based model market framework that lifts Mixture-of-Experts from a model-level learning architecture to a marketlevel coordination mechanism<ref:2610.11908#pg4>. The motivation stems from the observation that existing marketplaces often treat models as indivisible commodities, failing to support scenarios where the best model is the result of combining complementary capabilities of multiple providers<ref:2610.11908#pg6>. This lack of cooperation limits providers' ability to jointly create higher-value composite model services<ref:2610.11908#pg4>.

Formal Framework and Workflow

The MoE Market is formally defined as a tuple M = (B, N, P, G, E), which includes the set of buyers B, brokers N, data providers P, gating network G operated by the broker, and heterogeneous experts E<ref:2610.11908#pg4>. The market workflow consists of six steps for each buyer request rk:

  1. Request Submission<ref:2610.11908#pg4>

  2. Data Acquisition, where the broker solicits taskrelevant datasets from a subset of data providers Pk ⊆ P and constructs the aggregated dataset Dk<ref:2610.11908#pg4>.

  3. Expert Coordination, where the broker invokes the gating network to produce an expert participation vector wk (x) = g(x, rk;φ)<ref:2610.11908#pg4>.

  4. Composite Service Generation, where activated experts jointly produce a composite model output through weighted aggregation<ref:2610.11908#pg4>.

  5. Service Delivery and Payment, where the buyer pays the negotiated service price πˆk to the broker, and the residual surplus Rk = πˆk − kappak − C(r) forms the allocatable revenue pool<ref:2610.11908#pg4>.

  6. Revenue Allocation, where the broker allocates Rk among participating experts according to a cost-aware allocation rule<ref:2610.11908#pg4>.

Cost and Utility Modeling

The paper formalizes the economic aspects by defining expert costs ci(x, r) for each expert Ei<ref:2610.11908#pg4>. The expected service cost under request r is defined as E[Cost r] = Ex∼Dr "∑ E i=1 wi(x, r)ci(x, r) <ref:2610.11908#pg4>. A budget constraint is imposed where the broker must satisfy E[Cost r] ≤ πr<ref:2610.11908#pg4>. The utility of the delivered composite model service is defined as U (r) = alpha · E(x,y)∼Dr exp − ltau (y,yˆ(x)) / lref, where α > 0 controls utility scaling<ref:2610.11908#pg4>. The market welfare objective is defined as Welfare(r) = U (r) − beta · Ex∼Dr h∑ E i=1 wi(x, r)ci(x, r)i, where beta ≥ 0 controls the trade-off between predictive quality and cost sensitivity<ref:2610.11908#pg4>.

Cost-Aware Gating Mechanism

The broker learns the gating parameters φ by maximizing the welfare under request r in Eq. (27), which is equivalent to minimizing the market-aware loss Lmarket<ref:2610.11908#pg4>. The learning objective is min φ Lmarket = − alpha · E(x,y)∼Dr exp − ltau (y,yˆ(x)) / lref + beta · Ex∼Dr "∑ E i=1 wi(x, r)ci(x, r) >, which rewards routing policies that produce high-quality composite predictions while penalizing costly expert usage<ref:2610.11908#pg4>.

Cost-Adjusted Revenue Allocation

The residual revenue pool Rr is distributed among participating experts according to the cost-adjusted contribution score si(r) = Ex∼Dr h wi(x, r) · exp− λ ci(x, r)i, where λ > 0 controls the strength of this discount<ref:2610.11908#pg4>. The revenue allocated to expert Ei is then defined as a i(r) = Rr · si(r) / Í E j=1 sⱼ (r) >, which distributes the residual revenue proportionally to cost-adjusted contribution scores rather than raw participation alone<ref:2610.11908#pg4>. This rule is shown to be budget-balanced, participation monotonic, and cost sensitive<ref:2610.11908#pg4>.

Experimental Results

Experiments over fifteen tabular and image benchmarks show that MoE Market achieves the highest mean welfare on all fifteen datasets and a lower mean expected cost than Standard MoE in every case<ref:2610.11908#pg4>. The allocation experiments further demonstrate systematic sensitivity to expert participation and cost, together with substantially lower computational overhead than exact Shapley allocation<ref:2610.11908#pg4>. Overall, cost-aware routing improves market efficiency while retaining competitive predictive performance<ref:2610.11908#pg4>. The results indicate that the proposed rule produces cost-sensitive revenue shares and is substantially faster to compute than exact Shapley allocation<ref:2610.11908#pg4>.

Conclusion

This paper presents an MoE-based model market that coordinates heterogeneous experts as a market-level coordination mechanism for collaborative AI service delivery<ref:2610.11908#pg4>. The framework successfully formalizes the market participants, workflow, cost structure, and welfare objective while developing a cost-aware gating mechanism and revenue allocation rule that balances predictive utility against measured execution cost<ref:2610.11908#pg4>. The proposed framework demonstrates that MoE can serve as a market-level coordination principle for collaborative, cost-aware, and economically grounded model marketplaces<ref:2610.11908#pg4>. The results suggest that MoE can serve as a market-level coordination principle for collaborative, cost-aware, and economically grounded model marketplaces<ref:2610.11908#pg4>. The paper concludes by showing that MoE Market achieves the highest mean welfare on every dataset and a lower mean expected cost than Standard MoE in every case<ref:2610.11908#pg4>. The allocation experiments further show that increasing λ consistently shifts revenue toward lower-cost experts<ref:2610.11908#pg4>. Future work will focus on dynamic model markets in which expert availability, execution costs, and deployment distributions evolve over time<ref:2610.11908#pg4>. The artifacts for reproducing our experiments are available at https://github.com/MAYIZHOU/artifact-code moe.git<ref:2610.11908#pg4>. The paper was partially funded by the Horizon Europe projects DATAPACT (No. 101189771) and RAISE Suite (No. 101188337). The references list provides further context on related work, including data marketplaces and Mixture of Experts. The paper was published on OpenProceedings.org under ISBN 978-3-89318-106-3. The paper is distributed under the terms of the Creative Commons license CC-by-nc-nd 4.0. The paper was published on OpenProceedings.org under ISBN 978-3-89318-106-3. The paper is distributed under the terms of the Creative Commons license CC-by-nc-nd 4.0. The paper was published on OpenProceedings.org under ISBN 978-3-89318-106-3. The paper is distributed under the terms of the Creative Commons license CC-by-nc-nd 4.0. The paper was published on OpenProceedings.

Improvements for AI systems

  1. A cost-aware gating mechanism that balances predictive quality and service cost can be designed to coordinate heterogeneous experts under explicit model quality–cost trade-offs. This allows the system to dynamically select which experts are most useful for a given input while factoring in their execution costs and compensation considerations.

  2. A cost-adjusted revenue allocation mechanism can be introduced, defined by the score si(r) = E[w i(x, r) · exp− λ c i(x, r)], which distributes residual revenue according to both realized contribution and execution cost. This ensures that experts are rewarded for their actual service delivery while explicitly discounting compensation when achieving that contribution requires higher costs.

  3. The improved system can achieve the highest mean welfare on all fifteen datasets and a lower mean expected cost than Standard MoE in every case, suggesting it can serve as a market-level coordination principle for collaborative, cost-aware, and economically grounded model marketplaces.

Abstract

Existing model marketplaces typically trade and select individual models as indivisible units, limiting their ability to exploit complementarities among heterogeneous experts. This paper proposes an MoE-based model market framework that lifts Mixture-of-Experts from a model-level learning architecture to a market-level coordination mechanism. In this framework, brokers use gating networks to coordinate multiple heterogeneous experts and deliver a composite model service. We formalize the market participants, service workflow, expert cost structure, and a welfare objective that combines predictive utility with heterogeneous execution costs. We then derive a cost-aware gating mechanism and market-aware training objective, and introduce a cost-adjusted revenue allocation rule that distributes residual revenue according to realized expert participation and execution cost. We also establish basic theoretical properties of the allocation rule, including budget balance, participation monotonicity, and cost sensitivity. Experiments over five random seeds on fifteen tabular and image benchmarks use independently trained and frozen neural and tree-based experts together with latency-derived execution costs. MoE Market achieves the highest mean welfare on all fifteen datasets and a lower mean expected cost than Standard MoE in every case, while maintaining competitive predictive performance. The allocation experiments further demonstrate systematic sensitivity to expert participation and cost, together with substantially lower computational overhead than exact Shapley allocation. These results suggest that MoE can serve as a market-level coordination principle for collaborative, cost-aware, and economically grounded model marketplaces.

Sources

Related papers