ZAC-Inference: Collective Provisioning of Artificial Intelligence Inference Across Heterogeneous Agents
Abstract
Artificial intelligence inference is commonly deployed by placing a complete model, or a sufficiently large model partition, on an individual device or within a conventional computing infrastructure. Although distributed inference allows computation to be divided across multiple devices, existing deployment architectures still face a fundamental systems challenge: determining how a population of heterogeneous agents should collectively provide the computational stages required by an inference task when local hardware capacity, communication conditions, model compatibility, task deadlines, and resource availability vary over time.
This paper proposes ZAC-Inference, a research framework for collective provisioning of artificial intelligence inference under the broader Zero-Agent-Compute (ZAC) perspective. Instead of treating each agent as an independent inference unit, the framework treats the population as a potential provider of end-to-end inference capabilities. An inference request is represented as a dependency graph of computational stages, and the system jointly determines which stages are required, which agents can execute them, where model components should reside, how intermediate representations should be transmitted, and which local computing capabilities can be omitted without violating operational requirements.
The proposed architecture introduces five interconnected mechanisms: requirement-driven inference decomposition, joint stage-to-agent and resource allocation, communication-aware execution planning, dynamic collective provisioning, and requirement-preserving recovery. A constrained optimization formulation is developed to minimize system-wide inference cost while accounting for computation, communication, coordination, energy, deadlines, inference quality, and failure risk. The framework also distinguishes sequential model partitioning, parallel execution, modular model composition, and heterogeneous model cooperation, which have different technical requirements and should not be treated as interchangeable operations.
A research methodology is proposed to compare ZAC-Inference against local inference, conventional edge-cloud offloading, fixed model partitioning, and existing adaptive collaborative inference methods. No experimental performance gains are claimed in this paper. The principal research hypothesis is that jointly optimizing inference requirements, local provisioning decisions, and collective execution assignments can improve selected system-level outcomes under identifiable workload and network conditions. The contribution must ultimately be established through formal analysis, reproducible implementation, and empirical comparison with established baselines.
Related
- ZAC-Specialization: Collective Capability Provisioning Through Hardware–Role Co-Design in Heterogeneous Intelligent Agent Populations
- ZAC-Training: Collective Provisioning of Model Training and Updates in Heterogeneous Agent Populations
- ZLP-Inference: Zero-Local-Provisioning of Inference Capabilities in Heterogeneous Intelligent Systems
- ZAC-Memory: Collective Provisioning of Memory and World Models for Embodied Intelligence
- ZAC-Perception: Collective Provisioning of Perceptual Capabilities in Multi-Agent Embodied Intelligence
BibTeX
@misc{zhang2026zacinference,
title = {{ZAC-Inference: Collective Provisioning of Artificial Intelligence Inference Across Heterogeneous Agents}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23450937},
url = {https://zenodo.org/records/23450937},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - ZAC-Inference: Collective Provisioning of Artificial Intelligence Inference Across Heterogeneous Agents
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/18
PB - Zenodo
DO - 10.5281/zenodo.23450937
UR - https://zenodo.org/records/23450937
AB - Artificial intelligence inference is commonly deployed by placing a complete model, or a sufficiently large model partition, on an individual device or within a conventional computing infrastructure. Although distributed inference allows computation to be divided across multiple devices, existing deployment architectures still face a fundamental systems challenge: determining how a population of heterogeneous agents should collectively provide the computational stages required by an inference task when local hardware capacity, communication conditions, model compatibility, task deadlines, and resource availability vary over time. This paper proposes ZAC-Inference, a research framework for collective provisioning of artificial intelligence inference under the broader Zero-Agent-Compute (ZAC) perspective. Instead of treating each agent as an independent inference unit, the framework treats the population as a potential provider of end-to-end inference capabilities. An inference request is represented as a dependency graph of computational stages, and the system jointly determines which stages are required, which agents can execute them, where model components should reside, how intermediate representations should be transmitted, and which local computing capabilities can be omitted without violating operational requirements. The proposed architecture introduces five interconnected mechanisms: requirement-driven inference decomposition, joint stage-to-agent and resource allocation, communication-aware execution planning, dynamic collective provisioning, and requirement-preserving recovery. A constrained optimization formulation is developed to minimize system-wide inference cost while accounting for computation, communication, coordination, energy, deadlines, inference quality, and failure risk. The framework also distinguishes sequential model partitioning, parallel execution, modular model composition, and heterogeneous model cooperation, which have different technical requirements and should not be treated as interchangeable operations. A research methodology is proposed to compare ZAC-Inference against local inference, conventional edge-cloud offloading, fixed model partitioning, and existing adaptive collaborative inference methods. No experimental performance gains are claimed in this paper. The principal research hypothesis is that jointly optimizing inference requirements, local provisioning decisions, and collective execution assignments can improve selected system-level outcomes under identifiable workload and network conditions. The contribution must ultimately be established through formal analysis, reproducible implementation, and empirical comparison with established baselines.
ER -