Adaptive Mutual–Human Routing: A Unified Framework for Collaborative Multi-LLM Routing and Human-Assisted Routing with Context-Aware Mechanism Selection and Switching

The increasing diversity of large language models (LLMs) has created new opportunities for intelligent task allocation, model selection, distributed inference, and collaborative artificial intelligence. Existing LLM routing approaches primarily investigate how to select an appropriate model according to query characteristics, model capabilities, response quality, computational cost, or latency. Multi-agent routing further extends this problem by determining collaboration structures, agent roles, and model execution strategies. Meanwhile, human-in-the-loop approaches introduce human judgment into automated decision-making and escalation procedures. However, these research directions do not necessarily provide a unified, mechanism-independent architecture in which routing can adaptively alternate between collaborative routing among multiple LLMs and human-assisted routing as two distinct but combinable operational dimensions. This paper introduces Adaptive Mutual–Human Routing (AMHR), a theoretical framework developed as a specialized routing application of the Universal Mutual-Human Model (UMHM). AMHR establishes two foundational dimensions: Mutual Routing, in which multiple LLMs interact, negotiate, evaluate, coordinate, and collectively determine routing decisions; and Human-Assisted Routing, in which human individuals or groups contribute routing knowledge, judgment, intervention, correction, authorization, and decision-making. These dimensions generate three non-empty configurations: Mutual Routing, Human-Assisted Routing, and Mutual–Human Routing. The central contribution is an adaptive mechanism that allows a routing system to select, switch between, combine, and reconfigure these configurations according to task complexity, uncertainty, model disagreement, routing confidence, expected performance, operational risk, resource availability, and the value of human participation. Unlike architectures that permanently assign routing authority to an autonomous router or require human involvement according to a fixed escalation policy, AMHR conceptualizes routing as a dynamically configurable process in which the source and organization of routing decisions can change throughout execution. The paper develops a formal representation, an adaptive decision model, switching conditions, a unified architecture, a routing algorithm, theoretical propositions, application scenarios, governance principles, and an empirical evaluation methodology. It also distinguishes routing among models from routing of tasks, routing authority, and routing-policy adaptation. To the best of our knowledge, this paper is the first to formally articulate this particular UMHM-grounded integration of collaborative multi-LLM routing and human-assisted routing as independently operable, jointly configurable, and adaptively switchable routing mechanisms within one general architecture. The proposed framework provides a theoretical foundation for intelligent systems that can determine not only which model should execute a task, but also who or what should make that routing decision and when routing authority should move between artificial and human participants.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24362682
Primary Topic
Natural Language Processing Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Adaptive Mutual–Human Routing: A Unified Framework for Collaborative Multi-LLM Routing and Human-Assisted Routing with Context-Aware Mechanism Selection and Switching

Jincheng Zhang
Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
preprint

Adaptive Mutual–Human Routing: A Unified Framework for Collaborative Multi-LLM Routing and Human-Assisted Routing with Context-Aware Mechanism Selection and Switching

Jincheng Zhang
preprint en

Abstract

The increasing diversity of large language models (LLMs) has created new opportunities for intelligent task allocation, model selection, distributed inference, and collaborative artificial intelligence. Existing LLM routing approaches primarily investigate how to select an appropriate model according to query characteristics, model capabilities, response quality, computational cost, or latency. Multi-agent routing further extends this problem by determining collaboration structures, agent roles, and model execution strategies. Meanwhile, human-in-the-loop approaches introduce human judgment into automated decision-making and escalation procedures. However, these research directions do not necessarily provide a unified, mechanism-independent architecture in which routing can adaptively alternate between collaborative routing among multiple LLMs and human-assisted routing as two distinct but combinable operational dimensions. This paper introduces Adaptive Mutual–Human Routing (AMHR), a theoretical framework developed as a specialized routing application of the Universal Mutual-Human Model (UMHM). AMHR establishes two foundational dimensions: Mutual Routing, in which multiple LLMs interact, negotiate, evaluate, coordinate, and collectively determine routing decisions; and Human-Assisted Routing, in which human individuals or groups contribute routing knowledge, judgment, intervention, correction, authorization, and decision-making. These dimensions generate three non-empty configurations: Mutual Routing, Human-Assisted Routing, and Mutual–Human Routing. The central contribution is an adaptive mechanism that allows a routing system to select, switch between, combine, and reconfigure these configurations according to task complexity, uncertainty, model disagreement, routing confidence, expected performance, operational risk, resource availability, and the value of human participation. Unlike architectures that permanently assign routing authority to an autonomous router or require human involvement according to a fixed escalation policy, AMHR conceptualizes routing as a dynamically configurable process in which the source and organization of routing decisions can change throughout execution. The paper develops a formal representation, an adaptive decision model, switching conditions, a unified architecture, a routing algorithm, theoretical propositions, application scenarios, governance principles, and an empirical evaluation methodology. It also distinguishes routing among models from routing of tasks, routing authority, and routing-policy adaptation. To the best of our knowledge, this paper is the first to formally articulate this particular UMHM-grounded integration of collaborative multi-LLM routing and human-assisted routing as independently operable, jointly configurable, and adaptively switchable routing mechanisms within one general architecture. The proposed framework provides a theoretical foundation for intelligent systems that can determine not only which model should execute a task, but also who or what should make that routing decision and when routing authority should move between artificial and human participants.

Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon