Huawei Technologies (France)

Artificial intelligence will be one of the key pillars of the next generation of mobile networks (6G), as it is expected to provide novel added-value services and improve network performance. In this context, large language models have the potential to revolutionize the telecom landscape through intent comprehension, intelligent knowledge retrieval, coding proficiency, and cross-domain orchestration capabilities. This paper presents Telco-oRAG, an open-source Retrieval-Augmented Generation (RAG) framework optimized for answering technical questions in the telecommunications domain, with a particular focus on 3GPP standards. Telco-oRAG introduces a hybrid retrieval strategy that combines 3GPP domain-specific retrieval with web search, supported by glossary-enhanced query refinement and a neural router for memory-efficient retrieval. Our results show that Telco-oRAG improves the accuracy in answering 3GPP-related questions by up to 17.6% and achieves a 10.6% improvement in lexicon queries compared to baselines. Furthermore, Telco-oRAG reduces memory usage by 45% through targeted retrieval of relevant 3GPP series compared to baseline RAG, and enables open-source LLMs to reach GPT-4-level accuracy on telecom benchmarks.

Large Language Models (LLMs) achieve near-human performance on general-purpose benchmarks such as MATH and GSM8K, yet their ability to solve <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">domain-specific numerical problems</i> remains underexplored. In telecommunications, many core tasks—from signal processing to network optimization—require precise quantitative reasoning, where errors directly impact engineering practice. Existing benchmarks evaluate LLMs on general mathematics or telecom-related language tasks, but none assess llM capability to solve numerical mathematical problems in the telecom domain. To address this gap, we introduce TeleMath, the first benchmark dataset specifically designed to evaluate LLM performance in solving mathematical problems with numerical solutions in the telecommunications domain. TeleMath is built through a novel data generation methodology that starts from a compact seed of expert-crafted problems and expands it via problem decomposition, code- and symbolic math–driven blueprint generation, parameterized synthesis, and semantic validation. The resulting benchmark comprises 500 expert-verified question– answer pairs spanning diverse topics in the telecom domain with different levels of difficulty, each requiring a numerical solution. We then conduct a systematic benchmarking study of leading open-source LLMs, comparing reasoning-oriented models against general-purpose ones. On TeleMath, the best reasoning model, Qwen3-32B, achieves 69.51% pass@1 accuracy, substantially outperforming the best non-reasoning baseline, Qwen2.5-Math-72B-Instruct, which achieves 39.99% average accuracy with 2× more parameters, underscoring the unique challenges of telecom mathematics. We release TeleMath and its evaluation framework publicly to foster progress in domain-specific reasoning for telecommunications and beyond.

We propose a holistic framework for evaluating training time, cost, and energy across distributed AI training configurations. With optimal DCI bandwidth, metro-distributed training(time) takes 7% longer than single-datacenter; long-haul, 37% longer at equivalent cost. Protocol advancements like Bifrost further reduce long-haul training time by 26%.

We propose a stable, fast and parameter-free method to compute power profiles for ultrawideband systems, suitable for arbitrary bidirectional signals and high-power backward Raman pumps. We demonstrate convergence for third-order Raman pumping and speed advantages in optimizing pre-emphasis for E+S+C+L transmission. ©2025 The Author(s)

We experimentally validate a nonlinearity-tailored carrier phase estimation (CPE) technique for subcarrier multiplexing systems. Leveraging data-aided modeling of nonlinear phase noise, the proposed method outperforms joint CPE by up to 0.2 dB after 3000 km transmission.

Integrated terrestrial and non-terrestrial network (TN-NTN) architectures offer a promising solution for expanding coverage and improving capacity for the network. While nonterrestrial networks (NTNs) are primarily exploited for these specific reasons, their role in alleviating terrestrial network (TN) load and enabling energy-efficient operation has received comparatively less attention. In light of growing concerns associated with the densification of terrestrial deployments, this work aims to explore the potential of NTNs in supporting a more sustainable network. In this paper, we propose a novel online optimisation framework for integrated TN-NTN architectures, built on a multi-armed bandit (MAB) formulation and leveraging the Bandit-feedback Constrained Online Mirror Descent (BCOMD) algorithm. Our approach adaptively optimises key system pa-rameters-including bandwidth allocation, user equipment (UE) association, and macro base station (MBS) shutdown-to balance network capacity and energy efficiency in real time. Extensive system-level simulations over a 24 -hour period show that our framework significantly reduces the proportion of unsatisfied UEs during peak hours and achieves up to 19 % throughput gains and 5 % energy savings in low-traffic periods, outperforming standard network settings following 3GPP recommendations.

This paper presents the development of several multilayer feed-forward artificial neural network (ANN) models aimed at the simulation and parametrization of semiconductor optical amplifiers (SOAs). In the direct approach, it is assumed that the SOA physical parameters are known, and the ANNs can reproduce the results of detailed SOA models up to 20 times faster. In the inverse approach, it is assumed that the SOA is a grey-box with a known model but unknown parameters, and the ANNs infer the SOA model parameters from the input-output data. These results provide efficient surrogate models to represent SOAs in future digital twins for on-line optimization of ultra-wide band transmission systems that employ the SOA as the repeater element.

We experimentally show that DPD nonlinear compensation can improve the power budget from 13.5 dB to 16.9 dB with an 800 Gbps 90 Gbaud signal thanks to enabling larger transmitted power while reducing nonlinear SNR degradation.

This paper investigates channel estimation for linear time-varying (LTV) wireless channels under <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">double sparsity</i>, i.e., sparsity in both the delay and Doppler domains. An on-grid approximation is first considered, enabling rigorous hierarchical-sparsity modeling and compressed sensing-based channel estimation. Guaranteed recovery conditions are provided for affine frequency division multiplexing (AFDM), orthogonal frequency division multiplexing (OFDM) and single-carrier modulation (SCM), highlighting the superiority of AFDM in terms of doubly sparse channel estimation. To address arbitrary Doppler shifts, a relaxed version of the on-grid model is introduced by utilizing multiple elementary Expansion Models (BEM) each based on Discrete Prolate Spheroidal Sequences (DPSS). Next, theoretical guarantees are provided for the precision of this off-grid model before further extending it to tackle channel prediction by exploiting the inherent DPSS extrapolation capability. Finally, numerical results are provided to both validate the proposed off-grid model for channel estimation and prediction purposes under the double sparsity assumption and to compare the corresponding mean squared error (MSE) and overhead performance when different wireless waveforms are used.