LMTE: Putting the "Reasoning" into WAN Traffic Engineering with Language Models

TL;DR

This paper addresses the challenge of WAN TE by introducing LMTE, the first framework that leverages pre-trained large language models to perform TE through reasoning-based, automata-inspired planning. It provides theoretical results showing LMs can express TE transitions and compute TE states with logarithmic-depth reasoning, and couples this with a practical architecture that aligns multimodal topology and traffic inputs to the LM via embedding-to-language alignment and task-specific prompting. Empirically, LMTE delivers consistent improvements in maximum link utilization and orders-of-magnitude speedups over LP solvers across five WAN topologies, with strong robustness to failures, distribution drift, and demand variations. The approach offers scalable, generalizable TE with interpretable reasoning and requires only lightweight adaptation, making it practical for real-world deployment and future network optimization research.

Abstract

The rapid expansion of modern wide-area networks (WANs) has made traffic engineering (TE) increasingly challenging, as traditional solvers struggle to keep pace. Although existing offline ML-driven approaches accelerate TE optimization with deep neural networks (DNNs), they often lack sufficient expressiveness and generalization on unseen traffic patterns or topologies, limiting their practicality. Inspired by the success of large language models (LMs), for the first time, this paper investigates their potential as general-purpose traffic planners. Our contributions are two-fold: (i) Theoretically, we show that pre-trained LMs can simulate the sequential decision processes underlying TE and, crucially, exhibit parallel reasoning capabilities, making them well-suited for the task; (ii) Practically, we present LMTE, a novel LM-driven TE framework that embraces these insights through efficient multimodal alignment and lightweight configuration generation, all while preserving the model's original abilities. Extensive experiments demonstrate that fold matches top-tier performance on five datasets, achieving up to 15\% better maximum link utilization (MLU) and consistently lower performance degradation across diverse scenarios, e.g., less than 5\% with high traffic dynamics and link failures. Moreover, it achieves 10 to 100 times speedups over traditional TE solvers. To aid future works, our codebase is available at https://github.com/Y-debug-sys/LMTE.

Figures (16)

• Figure 1: Our ideal LM-driven TE system design for software-defined WANs.
• Figure 2: Comparison of adaptation costs on a real-world topology: full-parameter fine-tuning of LLaMA-7B incurs high overhead, while Lmte is even more lightweight than LoRA by keeping the LM's backbone intact.
• Figure 3: Illustrative network traffic engineering (TE) exampleperry2023dote.
• Figure 4: Theoretical intuition and foundations for our LM-driven TE approach. (a) Simple TE example from a solvable automaton perspective (Lemma \ref{['lemma:1']}). (b) A pretrained LM can approximate any optimal state from an initial state (Theorem \ref{['theorem:2']}), and accelerate automaton simulation with logarithmic-depth (Theorem \ref{['theorem:3']}).
• Figure 5: An overview of the proposed LMTE, the first LM–driven WAN TE framework. Left: input embedding alignment to bridge the modality gap. Middle: a frozen LM backbone, reused without any modification. Right: a lightweight LM head that jointly outputs TE configurations.

Theorems & Definitions (16)

• Lemma 1
• Theorem 2
• Theorem 3
• Definition A.0.1
• Definition A.0.2
• Lemma 4
• proof
• Lemma 5
• proof
• Lemma 6