Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.24348

Blackbox Performance Analysis of LLM-based Network Control

Mahdi Alizadeh·Pooria Namyar·Arvin Ghavidel·Ramesh Govindan
View original source ↗Read source document ↗Export metadata ↓

Abstract

LLMs are becoming an indispensable part of network control systems due to their impressive reasoning. Before such LLM-enabled systems can be deployed, it is imperative to develop tools that can audit their performance on previously unseen inputs. However, auditing these systems is challenging due to the inherent opacity of LLMs. LENS is the first step towards black-box performance analysis of LLM-enabled network control systems. It uses novel methods to extend deep reinforcement learning to guide the search for performance gaps and inputs that make LLMs struggle. LENS finds large gaps for LLM-based traffic engineering and GPU allocation systems, and even matches existing white-box analyzers designed for hand-crafted TE heuristics.

Explore related subjects

cs.NIcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Mahdi Alizadeh, Pooria Namyar, Arvin Ghavidel, Ramesh Govindan. 2026-10-17. Blackbox Performance Analysis of LLM-based Network Control. https://arxiv.org/abs/2610.24348

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Decentralized Decoding with Adaptive Local Combining for Cell-Free Massive MIMO

In existing literature, a two-stage architecture comprising of large-scale fading decoding (LSFD) at the central processing unit (CPU) and local combining at the access points (APs) is commonly deployed for uplink processing in scalable cell-free massive multiple-input multiple-output (CF-mMIMO) systems. However, centralized LSFD (c-LSFD) requires cross-AP channel statistics at the CPU and local combining uses schemes that either do not suppress interference optimally or result in analytically intractable solutions. We provide novel schemes to address both concerns. First, we propose the decentralized LSFD (d-LSFD) scheme that maximizes a surrogate signal-to-interference-plus-noise ratio built only from local interference covariance statistics and identify covariance-action regime where it is equivalent to c-LSFD. Second, based on the fact that the antennas at each AP provide spatial degrees of freedom (DoFs) for both interference suppression and coherent array gain, we interpret every combining scheme as solving the optimal DoF allocation problem over available DoFs. Selective zero-forcing (ZF) schemes solve this problem over both the dimension and composition of the local ZF subspace. We then provide the DoF-adaptive ZF-subspace (DAZS) combining scheme, which lets each AP select the optimal directions in the ZF subspace by maximizing its local sum spectral efficiency. Our numerical results establish the efficacy and efficiency of the proposed schemes compared to the corresponding existing schemes.

cs.NI↗
arXiv2026

Semantic Routing Framework for Multi-Domain Telecom Intent-Based Networking

Intent-Based Networking (IBN) is a key enabler of autonomous network management in next-generation networks, as it allows operators to drive the network through high-level intents instead of low-level configuration commands. In this paper, we propose a semantic routing framework for telecom intent classification and translation across the radio access networks (RANs), transport, 5G core, and cloud/edge domains. This framework operates locally to preserve intent privacy and avoid the latency and cost of cloud-based Large Language Model (LLM) solutions. The proposed framework is composed of intent classification and translation. To accurately classify the intent, we propose fine-tuning sentence-transformer models using contrastive learning over a 3GPP TS 28.312-aligned taxonomy of five top-level routes and 22 branches, supported by a telecom-specific codebook of 1,100 intents. In addition, we develop a branchaware, template-driven rule-based module for intent translation that enforces deterministic, hallucination-free parameter extraction and explicit missing-field detection, ensuring a structured and verifiable output pipeline. The numerical results show that the best fine-tuned encoder achieves up to 99.02%/94.61% route/branch classification accuracy, outperforming both base encoders and a commercial embedding baseline. The proposed translator achieves up to 96.24% translation accuracy, compared to 85.37% for a cloud-based LLM benchmark. Moreover, since the proposed pipeline runs entirely on local hardware, an average translation latency of 4.81 ms is recorded against 4,163.58 ms for the cloud-based benchmark, i.e., a speedup of approximately 866x.

cs.NI↗
arXiv2026

Task Offloading and Resource Allocation for Deterministic Service Provisioning in the IoT-Edge-Cloud Continuum

Future networks will seamlessly integrate computing and communication resources within a unified ecosystem of IoT devices, edge nodes, and cloud servers, forming an IoT-edge-cloud continuum. Within such an ecosystem, tasks can be dynamically offloaded to leverage resources across the continuum and efficiently scale with demand. Task offloading has traditionally focused on minimizing task execution latency or energy consumption. In contrast, we propose a deterministic task offloading and resource allocation scheme for the joint management of communication and computing resources to support services with deterministic service-level requirements. The proposed scheme aims to maximize the number of tasks completed on time, rather than minimizing the latency of individual tasks. To do so, it exploits variable task latency requirements to distribute the workload across the continuum. It can also prioritize high-priority tasks under resource scarcity. Numerical evaluations demonstrate that a deterministic task offloading strategy scales more effectively and can better guarantee timely task completion by balancing resource utilization across the continuum.

cs.NI↗