Leveraging Generative AI for large-scale prediction-based networking

2025-10-07

AI Summary

To address throughput limitations and unpredictable latency in large-scale multimodal networks, this paper proposes a generative AI (GenAI)-driven predictive network layer architecture. The method embeds GenAI directly into the network protocol stack, employs a lightweight initialization protocol to optimize prompt size, and constructs an end-to-end predictive model for proactive cross-modal content delivery (e.g., images, text, time-series data). It further replaces conventional congestion control with a TCP-inspired flow control mechanism. Crucially, the approach tightly integrates generative modeling with network scheduling while preserving backward compatibility with existing protocols. Experimental evaluation demonstrates over 100% improvement in end-to-end delivered traffic for image transmission, validating both the feasibility and substantial performance gains of GenAI-enabled networking at scale—specifically under ten-thousand-user workloads.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Generation

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

The traditional role of the network layer is to create an end-to-end route, through which the intermediate nodes replicate and forward the packets towards the destination. This role can be radically redefined by exploiting the power of Generative AI (GenAI) to pivot towards a prediction-based network layer, which addresses the problems of throughput limits and uncontrollable latency. In the context of real-time delivery of image content, the use of GenAI-aided network nodes has been shown to improve the flow arriving at the destination by more than 100%. However, to successfully exploit GenAI nodes and achieve such transition, we must provide solutions for the problems which arise as we scale the networks to include large amounts of users and multiple data modalities other than images. We present three directions that play a significant role in enabling the use of GenAI as a network layer tool at a large scale. In terms of design, we emphasize the need for initialization protocols to select the prompt size efficiently. Next, we consider the use case of GenAI as a tool to ensure timely delivery of data, as well as an alternative to traditional TCP congestion control algorithms.

Problem

• Redefining network layer with Generative AI prediction
• Solving throughput limits and uncontrollable latency issues
• Scaling GenAI networks for multiple data modalities

Innovation

• Using Generative AI for prediction-based networking
• Implementing initialization protocols for prompt size selection
• Applying GenAI as alternative to TCP congestion control