Beyond Fixed Model Selection: A Universal Adaptive Routing, Collaborative Routing, and Human-Assisted Routing Framework for Large Language Models
Large language models (LLMs) increasingly operate within heterogeneous computational environments containing models with different capabilities, architectures, parameter scales, domain expertise, reasoning abilities, inference costs, and operational constraints. Existing model-routing approaches primarily investigate how to select an appropriate model or expert according to input characteristics, predicted task difficulty, model performance, uncertainty, or computational budgets. Other approaches explore collaborative multi-agent systems, mixture-of-experts architectures, and human-in-the-loop decision-making. However, a general theoretical architecture that integrates autonomous adaptive routing, collaborative routing in which multiple models participate in determining routing decisions, and human-assisted routing within a unified dynamically configurable mechanism remains insufficiently established. This paper proposes the Universal Adaptive Routing and Human Assistance Framework (UARHA), grounded in the Universal Self-Mutual-Human Model (USMHM). The framework establishes three fundamental routing dimensions: Adaptive Routing, representing autonomous routing decisions made by an individual routing system; Collaborative Routing, representing coordinated routing decisions made by multiple artificial intelligence models or agents; and Human-Assisted Routing, representing routing decisions involving human judgment, expertise, authorization, or intervention. Its central principle is that routing should not be permanently restricted to one decision-making configuration. Instead, an intelligent system should be capable of determining whether routing should be performed autonomously, collaboratively, with human assistance, or through hybrid configurations. The proposed architecture introduces Universal Adaptive Routing Selection, Adaptive Routing Switching, Adaptive Routing Combination, and Adaptive Routing Reconfiguration. It further distinguishes the selection of a destination model from the selection of a routing decision mechanism, and distinguishes the models responsible for determining a route from the models or experts ultimately executing the task. A formal mathematical representation, multi-level routing architecture, decision lifecycle, switching conditions, coordination principles, human-authority model, and evaluation framework are developed. The framework accommodates single-model routing, multi-model routing, expert routing, distributed routing, hierarchical routing, sequential routing, parallel routing, and heterogeneous human–AI routing. To the best of our knowledge, this paper presents the first unified theoretical formulation specifically integrating autonomous adaptive routing, collaborative routing, and human-assisted routing with mechanism-independent adaptive selection, switching, combination, and reconfiguration within the USMHM conceptual architecture. Rather than proposing another isolated routing algorithm, this work establishes a general theoretical foundation for designing dynamically adaptive routing systems capable of changing not only where tasks are routed, but also who determines the routing decisions and how those decisions are produced.
Authors
- Jincheng Zhang (ORCID: https://orcid.org/0009-0005-1860-0009)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24357654
- Primary Topic
- Natural Language Processing Techniques
- Type
- preprint