Large Language Models as Generalist Policies for Network Optimization

TL;DR

This paper introduces Trailblazer, a novel framework using large language models to create versatile, generalist network control policies that outperform traditional specialist methods in diverse tasks and environments.

Contribution

The paper presents the first systematic framework leveraging LLMs for generalist network policies, including a network alignment scheme and adaptive policy collaboration mechanism.

Findings

Trailblazer outperforms conventional policies in cross-task generalization.

It demonstrates strong real-world performance on Douyin.

The framework reduces the need for task-specific policy design.

Abstract

Designing control policies to ensure robust network services is essential to modern digital infrastructure. However, the dominant paradigm for network optimization relies on designing specialist policies based on handcrafted rules or deep learning models, leading to poor generalization across diverse tasks and environments. In contrast, large language models (LLMs), pretrained on Internet-scale corpora, provide a rich and unified knowledge base that encodes fundamental networking principles. Combined with their emergent abilities in generalization to unseen scenarios, LLMs offer a transformative foundation for generalist network policies that can generalize across diverse tasks and environments with minimal adaptation. In this paper, we present Trailblazer, the first systematic framework to realize such a generalist policy for networking. Trailblazer incorporates a network alignment scheme to ground the LLM in specific networking tasks, and an adaptive policy collaboration mechanism that offloads simple control cases from the LLM to a lightweight policy for computational efficiency. Through extensive simulations and large-scale real-world online evaluation on Douyin (the Chinese version of TikTok), Trailblazer, powered by a single LLM, demonstrates stronger cross-task and cross-environment generalization than conventional specialist policies. Our results validate LLMs as the foundation for generalist network policies, and position Trailblazer as the first step toward the generalist-driven paradigm that enables strong generalization with minimal efforts in policy design.