Paper deep dive
Federated Learning for LLMs over Mobile Networks: Issues and Solutions in the RAN Transport
Emilio Paolini, Andrea Pinto, Flavio Esposito, Luca Valcarenghi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/3/2026, 2:36:53 AM
Summary
This paper proposes a learning-aware transport architecture for Federated Learning (FL) of Large Language Models (LLMs) over mobile Radio Access Networks (RANs). It addresses the mismatch between asynchronous, heterogeneous UE model updates and the need for predictable, schedulable traffic in optical transport networks (OCS/APN). The solution involves using the gNB as an aggregation point to perform In-Network Aggregation (INA), transforming many small, asynchronous UE updates into fewer, bounded, deadline-sensitive aggregate transfers. This reshaping enables efficient coordination with optical circuit switching, allowing high-capacity paths to be provisioned dynamically during aggregate-transfer windows.
Entities (10)
Relation Signals (8)
gNB → performs → In-Network Aggregation
confidence 95% · Through in-network aggregation at the gNB, asynchronous UE updates can be transformed into fewer aggregate transfers
Federated Learning → uses → Large Language Models
confidence 95% · Federated LLM fine-tuning enables large models to be adapted using private and geographically distributed data
In-Network Aggregation → enables → Optical Circuit Switching
confidence 92% · Once shaped in this way, federated LLM traffic becomes a suitable candidate for selectively provisioned optical connectivity
Radio Access Network → connects → User Equipment
confidence 90% · This challenge is particularly relevant in mobile RANs, where wireless variability... cause model updates to arrive asynchronously
Radio Access Network → connects → gNB
confidence 90% · updates arrive asynchronously... at the gNB
gNB → connectsto → Parameter Server
confidence 90% · forwarding them upstream... to the parameter server
In-Network Aggregation → transforms → User Equipment
confidence 90% · asynchronous UE updates can be transformed into fewer aggregate transfers with bounded size and delivery requirements
Low-Rank Adaptation → reduces →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Federated LLM fine-tuning enables large models to be adapted using private and geographically distributed data at the network edge, creating recurring and deadline-sensitive communication workloads across access and transport networks. This challenge is particularly relevant in mobile RANs, where wireless variability, mobility, and device heterogeneity cause model updates to arrive asynchronously. Although these updates belong to the same learning round and share a common destination and deadline, conventional transport networks treat them as independent device-originated flows, hiding their underlying structure and limiting the ability to efficiently provision transport resources. This mismatch is particularly problematic for optical circuit switching and all-photonics transport, which benefit from predictable and schedulable traffic demands. We argue that future RANs should act as learning-aware traffic shapers by exposing the communication structure of distributed model adaptation to the transport layer. Through in-network aggregation at the gNB, asynchronous UE updates can be transformed into fewer aggregate transfers with bounded size and delivery requirements. Once shaped in this way, federated LLM traffic becomes a suitable candidate for selectively provisioned optical connectivity, where high-capacity paths can be established during aggregate-transfer windows and released between learning rounds. The resulting architecture combines the flexibility of packet-based mobile access with dynamically provisioned optical capacity, illustrating a broader approach for coordinating distributed AI workloads across programmable access and transport networks.
Tags
Links
- Source: https://arxiv.org/abs/2610.01304v1
- Canonical: https://arxiv.org/abs/2610.01304v1
Trouble viewing inline? Open PDF directly →
Full Text
35,042 characters extracted from source content.
Expand or collapse full text
1 Federated Learning for LLMs over Mobile Networks: Issues and Solutions in the RAN Transport Emilio Paolini 1 Andrea Pinto 2 Flavio Esposito 2 Luca Valcarenghi 1 1 Tecip Institute, Scuola Superiore Sant’Anna, Pisa, Italy 2 Computer Science Department, Saint Louis University, St. Louis, USA Abstract—Federated LLM fine-tuning enables large models to be adapted using private and geographically distributed data at the network edge, creating recurring and deadline-sensitive communication workloads across access and transport networks. This challenge is particularly relevant in mobile RANs, where wireless variability, mobility, and device heterogeneity cause model updates to arrive asynchronously. Although these updates belong to the same learning round and share a common desti- nation and deadline, conventional transport networks treat them as independent device-originated flows, hiding their underlying structure and limiting the ability to efficiently provision transport resources. This mismatch is particularly problematic for optical circuit switching and all-photonics transport, which benefit from predictable and schedulable traffic demands. We argue that future RANs should act as learning-aware traffic shapers by exposing the communication structure of distributed model adap- tation to the transport layer. Through in-network aggregation at the gNB, asynchronous UE updates can be transformed into fewer aggregate transfers with bounded size and delivery requirements. Once shaped in this way, federated LLM traffic becomes a suitable candidate for selectively provisioned optical connectivity, where high-capacity paths can be established during aggregate- transfer windows and released between learning rounds. The resulting architecture combines the flexibility of packet-based mobile access with dynamically provisioned optical capacity, illustrating a broader approach for coordinating distributed AI workloads across programmable access and transport networks. Index Terms—Federated learning, Radio Access Network, in- network aggregation, optical circuit switching, APN, LLM, fine- tuning. I. INTRODUCTION Foundation models, including Large Language Models (LLMs), are increasingly moving toward distributed adapta- tion over private data generated at the network edge [1]. Federated Learning (FL) supports this shift by allowing clients to exchange model updates rather than raw data [2], while parameter-efficient techniques such as Low-Rank Adap- tation (LoRA) reduce the number of trainable parameters communicated during adaptation [3]. Nevertheless, federated LLM adaptation still produces recurring and deadline-sensitive transfers whose value depends on whether updates arrive in time to contribute to the current learning round. This challenge is particularly relevant in mobile Radio Access Networks (RANs), where wireless variability, mobility, and device het- erogeneity cause updates to become ready asynchronously [4]. Conventional RAN transport does not expose their learning- level relationships: each user equipment (UE) forwards its This work has been sponsored by NSF Awards #2133407 and #2201536, and in part by the CHIPS-JU SMARTY project (GA101140087). update independently, while the transport network observes ordinary device-originated flows. This hides the fact that these transfers share a learning process, destination, and delivery deadline. Such traffic is also poorly matched to optical-native transport. Optical Circuit Switching (OCS) and All-Photonics Network (APN) architectures are most effective when demands are sufficiently predictable and schedulable [5]–[7]. UE-level updates instead arrive asynchronously and may be individually too small or dispersed to justify optical reconfiguration. Addi- tional capacity alone is therefore insufficient: the RAN must first reshape learning traffic into demands that the transport network can understand and schedule. This motivates a learning-aware role for the RAN. In- Network Aggregation (INA), originally explored in datacenter networks to combine model-update fragments in transit [8], can be applied at the gNB to aggregate updates belonging to the same learning process before forwarding them up- stream [9]. The transport network then observes fewer gNB- to-parameter-server aggregates rather than many asynchronous UE-originated transfers. The impact extends beyond band- width reduction. Each aggregate has a bounded volume, known endpoints, a readiness time, and a latest useful delivery time. INA therefore converts device-level variability into schedula- ble transport demands and creates a control point for deciding whether an aggregate should be sent immediately over the packet path, coordinated with other aggregates, or carried over a temporarily provisioned optical circuit. This reshaping is particularly attractive because distributed AI training alternates between computation and synchronization phases [10]. By absorbing part of the UE-level asynchrony, the gNB can create temporally localized aggregate transfers that are better suited to selectively provisioned optical connectivity, with quieter intervals emerging between learning rounds. Building on this observation, this article investigates how learning-aware traffic shaping at the gNB can bridge federated LLM adaptation and optical-native transport. We evaluate how gNB-based aggregation transforms asynchronous UE uploads into bounded gNB-to-parameter-server transfers and quantify its effects on learning progress, update delivery, backhaul traf- fic, and the temporal structure of the resulting workload. We further discuss a coordination framework in which the RAN exposes each aggregate through its volume, readiness time, source, destination, and latest useful delivery time. A Window Scheduler coordinates ready aggregates, while an APN Con- troller determines whether their combined traffic volume and deadline slack justify optical provisioning. Although federated LLM fine-tuning is the motivating workload of this article, arXiv:2610.01304v1 [cs.NI] 1 Oct 2026 2 0) Model broadcast (downlink) 1) Local training at UEs 2) Update arrivals at gNB (uplink) time . . . (a) (b) 0) Model broadcast (downlink) 1) Local training at UEs 2) Update arrivals at gNB (uplink) time Aggregation deadline . . . (d) gNB Parameter Server OCS candidate flow In-Network Aggregation FL flow ID Fragment-wise INA Deadline closure (c) gNB Round-correlated update traffic Parameter Server Global model (returned to UEs) 1) Local training 0) Model broadcast to selected UEs 2) Server aggregation Fig. 1: (Left) Federated LLM fine-tuning over the RAN: (a) learning workflow and (b) traffic timeline. (Right) Learning-aware gNB aggregation: (c) processing pipeline and (d) transport timeline. the proposed transport abstraction is not model-specific. It applies more broadly to federated-learning workloads whose updates can be aggregated at the network edge and exposed as bounded, deadline-constrained transfers. LLM adaptation is particularly relevant because PEFT substantially reduces the computation and communication requirements of full-model fine-tuning, while still generating recurring transfers that may range from a few megabytes to hundreds of megabytes. Such traffic volumes can materially affect packet-versus-OCS provisioning and therefore benefit from coordinated access and optical transport. This analysis shows that INA should not be interpreted only as a mechanism for reducing transmitted bytes. Its broader contribution is to expose the structure of distributed- learning traffic and enable deadline-aware coordination across radio, packet, and optical domains. The resulting transport decision is workload-dependent: large aggregates may justify an optical circuit as soon as they become ready, whereas smaller aggregates may require coordination across multiple gNBs to amortize the optical setup cost. Such coordination must also account for uncertain readiness times, heterogeneous wireless conditions, competing learning jobs, fairness, and the risk that waiting for additional traffic may cause ready updates to miss their delivery deadlines. Future programmable access networks can therefore better support distributed AI by transforming application-level communication into bounded, semantically meaningful, and schedulable aggregate transfers. The RAN becomes not only the point at which devices access the network, but also the point at which learning demands are shaped for coordinated use of the end-to-end communication infrastructure. I. IMPACT OF LEARNING TRAFFIC ON THE RAN While privacy attacks have been demonstrated on FL sys- tems, it is understood that FL preserves some level of privacy during training, as devices keep data locally, compute model updates, and send those updates to a parameter server [11]. When FL clients are connected through a mobile network, the same process acquires an additional meaning: model adap- tation generates a recurring transport workload. Figure 1(a) illustrates such learning workflow. Selected UEs receive the global model, train locally, and upload their updates through the gNB to the parameter server. Figure 1(b) shows the cor- responding traffic timeline: model distribution is followed by local computation and by staggered update arrivals constrained by a common aggregation deadline. The server does not simply receive unrelated application flows; it waits for enough useful updates to construct the next global model. The Main Insight: Structured Learning Traffic. A pro- grammable RAN should treat FL updates as a distributed computation rather than as independent network flows. Updates generated during the same training round share a common destination, a common aggregation deadline, and a common semantic objective, constructing the next global model. Exploiting this hidden structure enables the network to jointly optimize communication efficiency and learning performance. Unlike conventional mobile traffic, where each flow can be managed independently, federated learning traffic consists of multiple correlated updates that jointly contribute to the same training round. Consequently, the value of a model update depends not only on successful delivery, but also on timely delivery before the aggregation deadline. An update that arrives too late may still reach the server, yet it can no longer influence the current model and must wait for a future round or be discarded. The resulting communication workload exhibits three key characteristics. First, it is round-correlated: many updates are generated during the same training round and converge toward a common aggregation point. Second, it is deadline-sensitive: the parameter server must eventually close the round, even if some devices have not completed their uploads. Third, it is 3 heterogeneous: UEs differ in radio conditions, computational capabilities, mobility, energy availability, and local data dis- tributions [12]. As a result, the network influences not only communication latency but also which clients ultimately con- tribute to the global model. These characteristics create a trade- off between learning quality and communication efficiency. Waiting longer allows more devices to participate, improving data diversity, while closing the round earlier reduces latency but may exclude slower clients and bias the model toward users with better connectivity. Consequently, communication performance depends not only on traffic volume but also on the timing and coordination of model updates. Today’s RAN, however, is largely unaware of this structure. The gNB schedules and forwards packets as independent flows, without considering that many updates belong to the same learning round and share a common objective. As a result, the network currently misses the opportunity to exploit information already present in the federated learning process. I. THE GNB AS A LEARNING-AWARE AGGREGATION The previous section showed the limitation of the conven- tional transport view: the gNB is the first point where learning updates converge, but it normally forwards them as separate upstream flows. The paradigm shift is to use this convergence point as part of the communication phase of training. Instead of forwarding one complete update per UE toward a remote parameter server, the gNB can combine updates that belong to the same learning round and forward one aggregate update upstream. The clients still train locally, and the parameter server still coordinates the global model. What changes is the role of the RAN. The gNB is no longer only a radio scheduler and forwarding point; it becomes a learning-aware aggregation point. It uses the round structure of FL to organize update traffic before it reaches the transport network. Exploit the structure of FL traffic at the gNB. The proposed change is to move aggregation from the param- eter server toward the network edge. The gNB can use the shared round structure of FL to convert many UE- level updates into a small number of predictable, deadline- aware transfers before they enter the transport network. Figure 1(c) shows such a change in the learning workflow. On the UE-facing side, update arrivals remain variable because devices finish local training at different times and experience different radio conditions. However, these arrivals are no longer exposed one by one to the upstream transport network. They are absorbed at the gNB and combined into an aggregate that represents the contribution of that gNB for the current round. The box inside Fig. 1(c) shows the internal gNB pipeline that makes this transformation possible. First, FL flow identification separates model-update traffic from ordinary mobile traffic, for example through a dedicated QoS flow or a user-plane classification rule. Second, deadline closure determines the conditions under which the aggregate must be released: the gNB closes it when enough useful contributions have arrived or when the round deadline is reached, rather than waiting indefinitely for the slowest UE. Third, fragment-wise In-Network Aggregation (INA) combines update fragments as they arrive, avoiding the need to buffer a complete update from every UE before aggregation. The gNB cannot wait indefinitely for the slowest UE; it closes the aggregate when enough useful contributions have arrived or when the round deadline is reached. Figure 1(d) shows the corresponding transport timeline. Be- fore aggregation, the gNB-to-parameter-server link sees many UE-originated transfers spread over the update-arrival interval. With INA, the gNB first collects and combines updates during an aggregation window and then releases one aggregate trans- fer upstream. The transport-facing demand therefore changes from many device-level flows to a smaller number of gNB- level aggregate transfers. This does not require the wireless edge to become perfectly synchronized. Some UEs may still be late, and some updates may miss the current round. The role of INA is not to eliminate device-level variability, but to contain part of it at the gNB. The gNB becomes the boundary where variable UE-level arrivals are converted into structured gNB-to-server demand. This abstraction shift is important for optical transport. The output of the gNB is not just a smaller amount of traffic; it is a better-defined transport request. The aggregate consolidates multiple UE-originated flows into a single gNB-to-parameter- server transfer with an estimable size and a delivery deadline inherited from the learning round. These are precisely the properties that an APN/OCS controller needs to decide for performing the optical transport network reconfiguration. The FLAG architecture is one example of this principle [9]. FLAG embeds aggregation inside the 5G gNB user plane, allowing model updates from multiple UEs to be combined before being forwarded upstream. The goal is not to replace FL, but to make the network aware of its communication structure. Section V quantifies the resulting learning-level benefits, including the reductions in communication time and overall round duration. IV. OPTICAL-NATIVE RANS: A NEW TRANSPORT OPPORTUNITY Optical transport can exploit learning traffic when four pieces of information are known: source, destination, traffic volume, and transmission window. Raw UE-level FL traffic does not naturally provide this description because updates are distributed across many devices and affected by wireless variability. A gNB aggregate is different: its source is the gNB, its destination is the parameter server, its size can be estimated once the aggregation group is known, and its deadline is inher- ited from the learning round. This does not mean every update should use an optical circuit. Packet switching remains appro- priate for control traffic, model distribution, small updates, and fallback operation. Optical provisioning becomes attractive only when the aggregate is sufficiently large and the expected transmission gain outweighs the circuit reconfiguration cost. INA is therefore not a replacement for packet transport; it enables learning traffic to become schedulable optical demand. In an APN view, fronthaul, midhaul, and backhaul form a common optical substrate. For FL, the most immediate opportunity lies on the gNB-to-parameter-server path, where 4 APN/OCS Spine gNB1 gNB2gNB3 gNB4 Parameter Server Active APN/OCS lightpath Deadline APN/OCS Spine gNB1 gNB2gNB3 gNB4 Parameter Server Active APN/OCS lightpath Deadline Δupdate Window Scheduler RAN ControllerAPN Controller Fig. 2: Learning-aware APN/OCS scheduling of gNB aggre- gate transfers. updates have already been aggregated into transfers with a known source, destination, approximate size, and deadline. Predictable Optical Transfer Requests. The real benefit of gNB aggregation is not simply fewer packets, but the creation of predictable, deadline-aware transfer requests. This enables federated learning traffic to be scheduled efficiently by an optical transport network. This enables a learning-aware control loop between the RAN and the optical layer. As illustrated in Fig. 2, a Window Scheduler combines information from a RAN Controller, conceptually aligned with O-RAN RIC-based closed-loop control [13], and an APN Controller that tracks optical re- sources [14]. Based on aggregate readiness, delivery deadlines, and resource availability, the scheduler assigns the active lightpath to the gNB whose transfer should be served next. In the example, the lightpath is initially assigned to gNB 2 and later reconfigured to gNB 1 as its aggregate approaches the deadline. This allows multiple gNBs to share a reconfigurable APN/OCS spine without dedicating a wavelength to each site. Rather than provisioning optical circuits for every UE update, the network schedules high-capacity circuits only for aggregated bursts after the gNB has absorbed device-level variability. The RAN therefore remains packet-switched for ordinary mobile services, while optical circuits are allocated dynamically to accelerate federated learning transfers. V. LEARNING AND TRANSPORT EVALUATION The previous sections introduced a learning-aware transport architecture that exploits the structure of federated learning traffic. In this section we evaluate its impact on learning performance, transport behavior, and optical provisioning. In particular, we use two complementary experiments to illustrate how gNB-side INA affects both federated learning and the traffic offered to the transport network. The first experiment evaluates the effect of INA on round duration and model convergence. The second examines how INA reshapes asyn- chronous UE uploads into bounded gNB-to-parameter-server transfers. Because the two experiments target different aspects of the proposed architecture, they use different numbers of UEs and model-update sizes. Learning-level impact. The first experiment uses the FLAG 5G user-plane emulator, which models the SDAP, PDCP, and 02,0005,000 0 20 40 60 Time (s) Accuracy (%) Fwd. INA (a) Accuracy over elapsed time. Fwd. INA 0 20 40 Round (s) TrainUploadReceive (b) Round-time composition. Fig. 3: Effect of gNB in-network aggregation on model accuracy and round duration. GTP-U layers together with MTU, queueing, and wireless-loss effects [9]. The scenario considers 60 UEs evenly distributed across six gNBs. Each UE is assigned a 100-Mbps uplink, while the gNB-to-parameter-server backhaul is constrained to 10 Mbps. The federated LLM fine-tuning workload generates LoRA adapter updates of approximately 6.28 MB. We define the round time as the elapsed time from the beginning of local training until the participating UEs receive the global model for the following round. It therefore includes local training, update upload and aggregation, and global-model reception. Figure 3(a) shows that INA reaches the target model ac- curacy in less elapsed time than conventional forwarding. Figure 3(b) explains the source of this improvement. Local- training time remains comparable because INA does not mod- ify client computation. In contrast, the communication com- ponent of the round is reduced because the gNB combines UE updates before forwarding them over the constrained backhaul. Shorter communication phases consequently produce shorter rounds and faster model convergence over time. Transport-level impact. The same mechanisms also modify the traffic exposed to the backhaul. FL flow identification dis- tinguishes learning traffic from ordinary user traffic, fragment- wise aggregation combines UE update streams, and deadline closure bounds the release time of the resulting aggregate. Together, these functions consolidate multiple asynchronous UE-originated transfers into a single-source, single-destination gNB-to-parameter-server demand with an estimable volume and a delivery deadline inherited from the learning round. Without INA, each selected UE forwards its update inde- pendently. The traffic observed on the backhaul is therefore arrival-driven and depends on local-training completion times, radio scheduling, retransmissions, and packetization. INA in- stead absorbs this device-level variability at the gNB and re- leases one aggregate transfer per learning round. The resulting traffic is not necessarily higher rate than the superposition of UE flows, but it is more regular, bounded, and temporally localized. Figure 4 reports the throughput observed on the gNB-to- parameter-server link over three consecutive learning rounds. This experiment uses the OAI/FLAG emulation environment with a cohort of 12 UEs, of which four are selected in each round. Each selected UE transmits an update of approximately 42.7 MB, representative of a LoRA PEFT adapter. With con- ventional forwarding, the backhaul carries four independent updates per round, corresponding to approximately 170.8 MB 5 0246810 Time from Round Start (seconds) 0 250 500 750 1000 Uplink Throughput (Mbps) Conventional ForwardingIn-Network Aggregation Fig. 4: gNB-to-parameter-server traffic with conventional for- warding and in-network aggregation. of upstream learning traffic. Because the updates become available at different times, their transmissions extend over a relatively long and irregular interval, while partial over- lap among the flows produces throughput fluctuations. With INA, the four updates are combined at the gNB and only one aggregate, equal in size to a single 42.7-MB update, is forwarded to the parameter server. INA therefore reduces the gNB-to-parameter-server learning traffic by a factor of four. The aggregate appears as a shorter and more regular transfer approaching 1 Gbps, followed by a clearly identifiable idle interval. The transport network consequently observes one bounded aggregate-transfer opportunity per round rather than several asynchronous UE-originated flows. INA therefore pro- vides not only traffic reduction, but also the timing information needed to coordinate transport provisioning. Deadline-aware optical provisioning. Provisioning an op- tical circuit is worthwhile only when its higher transmission capacity offsets the time required to establish or reconfigure the lightpath. Once a gNB closes its aggregate, the Window Scheduler knows its source, destination, size, readiness time, and latest useful delivery time. The APN Controller comple- ments this information with the capacities of the available packet and optical paths and the optical setup time. The following analysis assumes that INA has already combined the updates from each gNB into a single model- update-sized aggregate. The resulting gNB-level traffic can then be transmitted either over the existing packet path or a temporarily provisioned optical circuit, isolating the packet- versus-OCS transport decision. Let N gNB denote the number of active gNBs sharing the same transport segment. If each gNB generates one aggregate of size S, the total traffic volume is B = N gNB S.(1) The corresponding completion times are T pkt = 8B C pkt ,T ocs = T setup + 8B C ocs ,(2) where C pkt and C ocs are the packet and optical capacities, and T setup is the optical setup time. OCS is beneficial when- ever T ocs < T pkt . Figure 5 illustrates this trade-off for PEFT adapters of 6.28 (Small), 42.7 (Medium), and 256 MB (Large), assuming 2- Gbps packet transport, 10-Gbps optical transport, and a 0.5- s setup time. Packet completion time increases linearly with 123456789101112 Number of Active gNBs 10 1 10 0 10 1 Transport Completion Time (s) Small Adapter (Packet) Small Adapter (OCS) Medium Adapter (Packet) Medium Adapter (OCS) Large Adapter (Packet) Large Adapter (OCS) Fig. 5: Packet-versus-OCS completion time as a function of the number of active gNBs for small, medium, and large PEFT adapter aggregates. traffic volume, whereas OCS incurs a fixed setup cost but benefits from higher capacity. Consequently, packet transport remains preferable for small adapters, OCS becomes advanta- geous from four active gNBs (170.8 MB) for medium adapters, and a single large aggregate already amortizes the setup delay. More generally, optical provisioning should depend on the traffic available within a transmission window rather than on the number of active gNBs alone. Under the considered parameters, the packet–OCS break-even volume is approxi- mately 156.25 MB, corresponding to one large aggregate, four medium aggregates, or about 25 small aggregates. This also highlights the role of the Window Scheduler: large aggregates can be transmitted immediately, whereas smaller ones may be grouped into a common transmission window whenever their delivery deadlines allow, amortizing optical setup costs without missing the learning deadline. VI. TOWARD DEPLOYABLE AI-NATIVE RANS The proposed architecture illustrates how learning-aware aggregation can expose federated learning traffic as structured transport demand. Several research questions remain before this vision can become part of future AI-native RANs. RQ1: How should learning intent be exposed to the transport network? The optical layer does not need to understand the details of the learning algorithm, but it does require a compact description of each aggregate transfer. What information should the RAN expose? Is the source, destina- tion, size, readiness time, and delivery deadline sufficient, or should future interfaces also communicate learning priorities, aggregation policies, or tolerance to delayed updates? More- over, how should transport adapt when wireless variability prevents the planned transmission window from being met? RQ2: How can transport optimize learning without introducing bias? Communication decisions influence which clients ultimately contribute to the global model. Deadline- driven aggregation may consistently favor UEs with better radio conditions while excluding slower participants. How should transport scheduling account for fairness and statis- tical representativeness? Can mechanisms such as staleness- aware weighting, delayed aggregation, or partial contributions preserve model quality while still reducing communication latency? 6 RQ3: When should optical transport be used? Packet transport remains the natural choice for control messages, model distribution, small updates, and fallback operation, whereas optical circuits become attractive only when sufficient traffic is available to amortize setup costs. How should a hybrid packet/optical infrastructure dynamically determine the operating point as workloads, network conditions, and optical resources evolve? RQ4: How should AI-native RANs be evaluated? The proposed architecture spans the learning system, the wireless network, and the transport infrastructure. Future evaluations should therefore move beyond isolated networking or machine learning metrics and instead consider the complete closed loop, including UE variability, gNB aggregation, transport scheduling, deadline satisfaction, communication efficiency, and model convergence. AI-Native RANs. Ultimately, RAN Architectures should evolve from communication infrastructures into learning- aware systems that expose the semantics of distributed AI workloads to the transport network. Doing so enables networking decisions to optimize not only throughput and latency, but also the efficiency of the (distributed) learning process. VII. CONCLUSION The main message of this article is not that optical circuits are always better than packet switching. Rather, APN/OCS becomes useful when the RAN first reshapes learning traf- fic into something that optical transport can schedule. This shifts the design question from “how much capacity does the backhaul need?” to “which part of the learning process should the transport fabric see, and when?” Future RANs that support distributed model adaptation will need to become more than passive carriers of model updates. They will need to act as learning-aware traffic shapers, exposing the right abstractions to packet and optical transport. The opportunity is to build a RAN in which computation, aggregation, and optical provisioning are coordinated parts of the same training system. REFERENCES [1] D. Cai, Y. Wu, S. Wang, F. X. Lin, and M. Xu, “Efficient federated learn- ing for modern NLP,” in Proceedings of the 29th annual international conference on mobile computing and networking, 2023, p. 1–16. [2] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics.Pmlr, 2017, p. 1273– 1282. [3] K. Kuo, A. Raje, K. Rajesh, and V. Smith, “FLASC: Federated LoRA with Sparse Communication,” in Proceedings of the ACM Conference on AI and Agentic Systems, 2026, p. 990–1008. [4] M. M. Amiri and D. G ̈ und ̈ uz, “Federated learning over wireless fading channels,” IEEE transactions on wireless communications, vol. 19, no. 5, p. 3546–3557, 2020. [5] F. Saliou, P. Chanclou, G. Simon, J. Potet, G. Gaillard, J. Zandueta, and D. Chevalier, “Optical access networks to support future 5G and 6G mobile networks,” Journal of Optical Communications and Networking, vol. 17, no. 7, p. C22–C29, 2025. [6] M. Nakagawa, H. Minami, T. Fukatani, T. Seki, R. Hayashi, and T. Kuwahara, “Multi-band optical networking: Recent advances toward iown all-photonics network,” IEICE Transactions on Communications, vol. 107, no. 12, p. 842–851, 2024. [7] C. Avin and S. Schmid, “Revolutionizing datacenter networks via reconfigurable topologies,” Communications of the ACM, vol. 68, no. 6, p. 44–53, 2025. [8] A. Sapio, M. Canini, C.-Y. Ho, J. Nelson, P. Kalnis, C. Kim, A. Krish- namurthy, M. Moshref, D. Ports, and P. Richt ́ arik, “Scaling distributed machine learning with in-network aggregation,” in 18th USENIX Sym- posium on Networked Systems Design and Implementation (NSDI 21), 2021, p. 785–808. [9] E. Paolini, A. Pinto, L. Valcarenghi, and F. Esposito, “Programmable in- network aggregation for communication-aware federated learning in 5g rans,” IEEE Transactions on Network and Service Management, vol. 23, p. 5885–5899, 2026. [10] S. Rajasekaran, S. Narang, A. A. Zabreyko, and M. Ghobadi, “MLTCP: A distributed technique to approximate centralized flow scheduling for machine learning,” in Proceedings of the 23rd ACM Workshop on Hot Topics in Networks, 2024, p. 167–176. [11] Z. Zhang, A. Pinto, V. Turina, F. Esposito, and I. Matta, “Privacy and efficiency of communications in federated split learning,” IEEE Transactions on Big Data, p. 1–12, May 2023. [12] P. Kairouz and H. B. McMahan, “Advances and open problems in federated learning,” Foundations and trends in machine learning, vol. 14, no. 1-2, p. 1–210, 2021. [13] M. Polese, L. Bonati, S. D’oro, S. Basagni, and T. Melodia, “Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,” IEEE Communications Surveys & Tutorials, vol. 25, no. 2, p. 1376–1411, 2023. [14] K. Miyamoto, Y. Sugai, and R. Inohara, “Progress of an all-photonic network for a 6G mobile network in the IOWN Global Forum,” Journal of Optical Communications and Networking, vol. 17, no. 11, p. E9– E16, 2025. Emilio Paolini received his Ph.D. degree in Emerging Digital Technologies, cum laude, from Scuola Superiore Sant’Anna, Pisa, Italy, in 2024. He is currently an Assistant Professor with the TeCIP Institute, Scuola Superiore Sant’Anna, Pisa, Italy. His research interests include AI-enabled NextG networks, programmable data planes, and federated learning. He received the GTTI Ph.D. Thesis Award for the Best Italian Ph.D. in Communication Technologies in 2024. Andrea Pinto received his BS and MS in computer engineering from the University of Naples Federico I. He graduated with a Ph.D. in Computer Science from Saint Louis University in 2026, specializing in computer networks, distributed systems, and scalable machine learning systems. His research focuses on distributed training and bandwidth-aware orchestration to alleviate network bottlenecks. Flavio Esposito is an Associate Professor and Graduate Coordinator in the Department of Computer Science at Saint Louis University, where he also serves as a Fellow of the Research Institute. He received his Ph.D. in Computer Science from Boston University. His research focuses on edge computing, programmable and virtualized networks, network security, CPS, and applied artificial intelligence. Luca Valcarenghi is a Full Professor at the Scuola Superiore Sant’Anna of Pisa, Italy, since 2024. He received the Ph.D. from UTD in 2001. Dr. Valcarenghi received a Fulbright Research Scholar Fellowship in 2009 and a JSPS ”Invitation Fellowship Program for Research in Japan (Long Term)” in 2013. He coordinated, as a PI or local PI, several National and International projects. His main research interests are optical networks design, analysis, and optimization; energy efficiency in communications networks; optical access networks; zero touch network and service management; 5G technologies and beyond.