Beyond Task-Level Collaboration: Workload-Level Adaptive Allocation of Multi-Model Cooperation in Large Language Model Systems

Multi-model and multi-agent large language model (LLM) systems increasingly rely on collaboration to solve complex tasks, combine complementary capabilities, verify intermediate outputs, and improve decision quality. However, collaboration is not a cost-free operation. Additional model invocations, inter-model communication, context transfer, synchronization, result aggregation, and repeated inference consume computational resources and may increase latency. Existing adaptive collaboration approaches primarily address whether an individual task should be executed independently or collaboratively. Yet, in practical LLM services, multiple tasks frequently arrive concurrently and compete for limited GPU capacity, inference budgets, model availability, communication bandwidth, and response-time guarantees. Consequently, making an isolated collaboration decision for each task may produce globally inefficient resource allocation, even when each local decision appears reasonable. This paper introduces Workload-Level Adaptive Collaboration Allocation (WLACA), a conceptual and algorithmic framework for jointly managing collaboration demand across multiple concurrent tasks in heterogeneous LLM systems. Rather than treating collaboration as a binary task-level decision, WLACA models collaboration as a dynamically allocatable system resource. It jointly considers task admission, collaboration eligibility, collaboration intensity, model-team composition, resource budgeting, execution scheduling, adaptive scaling, termination, and cross-task budget reallocation. The framework introduces a workload-level net collaboration value function, a resource-constrained allocation model, a marginal collaboration value mechanism, and a feedback-driven controller that continuously revises allocation decisions as task states and system conditions change. It supports independent execution, selective consultation, limited collaboration, intensive collaboration, deferred collaboration, and collaboration termination. The central contribution is a shift from isolated collaboration selection to coordinated allocation of collaboration opportunities across an entire workload. To the best of our knowledge, this paper proposes one of the first unified formulations specifically centered on workload-level adaptive allocation of collaboration demand, integrating task admission, collaboration intensity, shared-budget competition, dynamic resource reallocation, and termination within one decision framework. This priority claim is deliberately limited to the specific integrated formulation and does not claim that adaptive collaboration, multi-agent scheduling, or resource allocation has never previously been studied. The paper develops the theoretical foundations, formal problem definition, algorithmic architecture, optimization objectives, evaluation methodology, and research hypotheses required for subsequent empirical validation.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24358616
Primary Topic
Cloud Computing and Resource Management
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Beyond Task-Level Collaboration: Workload-Level Adaptive Allocation of Multi-Model Cooperation in Large Language Model Systems

Jincheng Zhang
Zenodo (CERN European Organization for Nuclear Research)
Cloud Computing and Resource Management
preprint

Beyond Task-Level Collaboration: Workload-Level Adaptive Allocation of Multi-Model Cooperation in Large Language Model Systems

Jincheng Zhang
preprint en

Abstract

Multi-model and multi-agent large language model (LLM) systems increasingly rely on collaboration to solve complex tasks, combine complementary capabilities, verify intermediate outputs, and improve decision quality. However, collaboration is not a cost-free operation. Additional model invocations, inter-model communication, context transfer, synchronization, result aggregation, and repeated inference consume computational resources and may increase latency. Existing adaptive collaboration approaches primarily address whether an individual task should be executed independently or collaboratively. Yet, in practical LLM services, multiple tasks frequently arrive concurrently and compete for limited GPU capacity, inference budgets, model availability, communication bandwidth, and response-time guarantees. Consequently, making an isolated collaboration decision for each task may produce globally inefficient resource allocation, even when each local decision appears reasonable. This paper introduces Workload-Level Adaptive Collaboration Allocation (WLACA), a conceptual and algorithmic framework for jointly managing collaboration demand across multiple concurrent tasks in heterogeneous LLM systems. Rather than treating collaboration as a binary task-level decision, WLACA models collaboration as a dynamically allocatable system resource. It jointly considers task admission, collaboration eligibility, collaboration intensity, model-team composition, resource budgeting, execution scheduling, adaptive scaling, termination, and cross-task budget reallocation. The framework introduces a workload-level net collaboration value function, a resource-constrained allocation model, a marginal collaboration value mechanism, and a feedback-driven controller that continuously revises allocation decisions as task states and system conditions change. It supports independent execution, selective consultation, limited collaboration, intensive collaboration, deferred collaboration, and collaboration termination. The central contribution is a shift from isolated collaboration selection to coordinated allocation of collaboration opportunities across an entire workload. To the best of our knowledge, this paper proposes one of the first unified formulations specifically centered on workload-level adaptive allocation of collaboration demand, integrating task admission, collaboration intensity, shared-budget competition, dynamic resource reallocation, and termination within one decision framework. This priority claim is deliberately limited to the specific integrated formulation and does not claim that adaptive collaboration, multi-agent scheduling, or resource allocation has never previously been studied. The paper develops the theoretical foundations, formal problem definition, algorithmic architecture, optimization objectives, evaluation methodology, and research hypotheses required for subsequent empirical validation.

Zenodo (CERN European Organization for Nuclear Research)
Cloud Computing and Resource Management
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