ZLP-Inference: Zero-Local-Provisioning of Inference Capabilities in Heterogeneous Intelligent Systems

@misc{zhang2026zlpinference,
  title = {{ZLP-Inference: Zero-Local-Provisioning of Inference Capabilities in Heterogeneous Intelligent Systems}},
  author = {Jincheng Zhang},
  howpublished = {Zenodo},
  year = {2026},
  month = oct,
  doi = {10.5281/zenodo.23451066},
  url = {https://zenodo.org/records/23451066},
  note = {CC BY 4.0},
}

TY  - UNPB
TI  - ZLP-Inference: Zero-Local-Provisioning of Inference Capabilities in Heterogeneous Intelligent Systems
AU  - Jincheng Zhang
PY  - 2026
DA  - 2026/10/18
PB  - Zenodo
DO  - 10.5281/zenodo.23451066
UR  - https://zenodo.org/records/23451066
AB  - Artificial intelligence systems are commonly designed under the assumption that each intelligent object must locally possess sufficient inference capabilities to perform its assigned tasks. This assumption can lead to duplicated model deployments, redundant accelerator hardware, excessive memory requirements, energy consumption, and lifecycle costs. However, an object's inference requirements do not necessarily imply that the corresponding inference capabilities must be provisioned locally. When compatible computing resources, inference services, model components, or alternative reasoning mechanisms are available elsewhere in the system, some or all local inference provisioning may be unnecessary.  This paper proposes Zero-Local-Provisioning of Inference Capabilities (ZLP-Inference), a domain-specific framework for investigating the collective provisioning of artificial intelligence inference capabilities across heterogeneous intelligent systems. Its central principle is that an individual object may have zero locally provisioned capacity for a specified inference function, or may operate without a locally deployed inference model, provided that the system can satisfy its mandatory inference requirements through feasible collective mechanisms. Unlike approaches that focus exclusively on accelerating, distributing, or routing existing inference workloads, ZLP-Inference explicitly treats the local provisioning of inference capabilities as a system-design decision.  The framework distinguishes locally provisioned inference capability from effective inference functionality available to an object. It formalizes objects, inference requirements, model configurations, local inference capacity, collective inference services, model-component compatibility, quality constraints, communication costs, reliability, and lifecycle expenditure. A unified optimization formulation jointly determines which inference capabilities should be provisioned locally, which inference requirements should be satisfied collectively, and which objects may operate without selected local inference capabilities. The formulation accommodates complete model execution, partitioned inference, shared model services, heterogeneous inference providers, and alternative inference workflows without assuming that these mechanisms are interchangeable in every application.  The paper derives basic feasibility conditions, presents a zero-permissive optimization model, proposes an implementation architecture, and develops a reproducible evaluation methodology. It also distinguishes zero local inference provisioning from zero computation, zero system-wide inference capacity, and unrestricted dependence on remote services. No empirical superiority or universal novelty is assumed. The principal research question is whether explicitly allowing local inference provisioning to equal zero, while jointly optimizing system-level inference satisfaction, provides a useful and generalizable design methodology beyond established distributed inference and model-deployment approaches.
ER  -