Networked AI: Architecture, Theory, and Applications

Liang Song, Jiayue Jin, Peng Sun, Di Li, Yang Liu, Athanassios Skodras, Dimitris Ampeliotis, Yong Yuan, Amit Dvir, Jiangchuan Liu, Victor C. M. Leung, Qi BiarXiv 2026 · arXiv

TL;DR

Networked AI (NetAI) is a new paradigm that shifts artificial intelligence from monolithic, offline-trained "black boxes" to a decentralized, hierarchical ecosystem of interacting agents. By synthesizing signal processing and control theory, NetAI achieves near-linear resource scaling and continuous online evolution, overcoming the exponential data and compute bottlenecks of current foundation models.

What Problem Does This Paper Address?

The current AI era is defined by "scaling laws": to get better performance, we build bigger models (LLMs, VLMs) trained on increasingly massive, static datasets. However, this approach is hitting a wall. Monolithic models require exponential growth in compute and high-quality data, which is becoming a scarce resource.

Furthermore, these models are often "static" once deployed; they struggle to adapt to "out-of-distribution" novelties in dynamic, real-world environments. They are also notoriously difficult to interpret or regulate. The authors argue that the next generation of AI must move toward system-level intelligence—systems that can continuously generate high-quality data through interaction, adapt in real-time, and maintain global stability.

Methods

The authors propose a three-layer hierarchical architecture for Networked AI, integrating foundations from data science, signal processing, and system control:

  1. Agent Layer: This layer hosts a community of heterogeneous agents, ranging from massive foundation models to lightweight edge modules. These are organized into a scalable, hierarchical topology where higher-level agents represent functional abstractions of lower-level ones.
  2. Interaction Layer: Instead of black-box communication, this layer uses adaptive signal processing (e.g., Kalman filters, Graph Signal Processing) to guide online data extraction. Agents monitor the "discrepancy" between their internal predictions and network-derived signals to trigger targeted, localized updates.
  3. Regulation Layer: To prevent the system from diverging during continuous evolution, this layer applies system control theory. Using Lyapunov stability analysis, it injects regulatory control inputs that constrain local updates to ensure global system-wide stability and convergence.
The three-layer architecture of Networked AI: Agent, Interaction, and Regulation.

Datasets

Rather than using a standard static benchmark, the authors developed a configurable grid-world environment to test the theoretical scaling laws of NetAI.

  • Setup: Multiple heterogeneous agents (perception and control) navigate a space with known and dynamically introduced unknown obstacles.
  • Goal: Measure the total computational cost (FLOPs) required to reach policy convergence as model scale, application scale, and task complexity ($N$ agents) increase.
  • Comparison: They compared a "standalone" end-to-end configuration against the "Networked AI" configuration to verify resource-scaling trends.

Key Insights and Ideas

  • Exponential to Near-Linear Scaling: This is the "aha moment." In standalone systems, task complexity leads to an exponential training-data requirement ($\Omega(c^N)$). By using hierarchical factorization and localized updates, NetAI reduces this to near-linear scaling ($O(N \cdot c^{\log N})$).
  • Online Evolutive Learning (OEL): NetAI doesn't just "infer" after deployment; it continuously evolves. The Interaction Layer acts as an "online data factory," extracting high-quality samples only when environmental novelties are detected.
  • From Black-Box to Gray-Box: By explicitly defining interaction and regulation layers through signal processing and control theory, the system becomes a "gray-box." This allows for formal stability guarantees and layer-wise explanations of decisions.
  • Cross-Layer Resilience: Security is treated as an architectural property. The layered design allows for multi-level defense, from individual-level protection to collective-level immunity and system-level resilience.

Key Results

The experimental results (visualized in Figure 2) strongly support the theoretical near-linear scaling:

  • Model Scale: As the number of parameters ($P$) increased, the computational cost for standalone AI grew super-linearly ($P^{1.21}$), while Networked AI remained nearly linear ($P^{1.04}$).
  • Application Scale: For increasing application complexity ($E$), standalone AI showed a high exponent ($E^{2.55}$), whereas NetAI was significantly more efficient ($E^{1.29}$).
  • Overall Task Complexity: When increasing the number of heterogeneous agents ($N$), standalone AI exhibited exponential growth, while Networked AI followed a linear trend. For example, moving from 2 to 6 agents increased costs by ~8x for standalone AI, but only ~1.37x for NetAI.
Fig. 2. Computational cost to convergence under different model, application and overall task complexity scales. The box plots summarize the variability across heterogeneous experimental configurations, while the solid and dashed curves represent the global power-law or exponential fits for standalone AI and Networked AI.

Limitations and Future Work

While promising, the authors identify several open challenges:

  • Adaptive Topologies: Designing interaction graphs that can evolve dynamically based on task requirements and resource availability remains difficult.
  • Catastrophic Forgetting: Ensuring that continuous local adaptation doesn't lead to the loss of previously learned global knowledge is a key theoretical hurdle.
  • Incentive and Coordination: In large-scale decentralized systems, balancing local specialization with global consistency requires sophisticated coordination mechanisms that are still in early development.
  • 6G Integration: The authors see NetAI as a native capability for future 6G networks, but standardizing these "AI as a Networking System" (NSAI) protocols is a massive industrial undertaking.

Summary updated

Abstract

Recent advances in Artificial Intelligence (AI), particularly large-scale deep learning, have been driven by monolithic or offline training on massive curated datasets. While successful, these approaches face structural limitations: escalating data and computational demands, limited online adaptability, and a lack of system-level intelligence in open, non-stationary environments. This paper presents Networked AI - a networked intelligent system composed of interacting, heterogeneous, and continuously evolving AI agents. We establish a hierarchical layered architecture comprising an Agent Layer for heterogeneous capability construction, an Interaction Layer for signal-processing-guided online data extraction, and a Regulation Layer for stability-constrained collective evolution. Foundations in data science, signal processing, and system control are hereby synthesized into a unified theoretical framework of Networked AI, providing superior data efficiency, intelligence efficiency, interpretability, and scalability. Theoretical and experimental analysis demonstrates that Networked AI can approach near-linear scaling in resource demand, in contrast to the exponential increase in standalone end-to-end systems. It provides not only online adaptation to out-of-distribution environmental novelties, but also the autonomous evolution towards application-specific intelligence. Application frontiers are discussed for IoT spaces, embodied, spatial, and industrial intelligences, as well as the cellular infrastructure and beyond. Key open challenges are identified towards realizing locally autonomous, collectively intelligent, continuously evolving, globally regulated Networked AI.
BibTeX
@misc{song2026networked,
  author = {Liang Song and Jiayue Jin and Peng Sun and Di Li and Yang Liu and Athanassios Skodras and Dimitris Ampeliotis and Yong Yuan and Amit Dvir and Jiangchuan Liu and Victor C. M. Leung and Qi Bi},
  title = {{Networked AI: Architecture, Theory, and Applications}},
  year = {2026},
  eprint = {2611.03063},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.03063},
}
Figure