PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Learning for LLMs over Mobile Networks: Issues and Solutions in the RAN Transport

Emilio Paolini, Andrea Pinto, Flavio Esposito, Luca Valcarenghi

arXiv:2610.01304v1cs.NIcs.AI

TL;DR

Federated LLM fine-tuning produces asynchronous, deadline-sensitive UE updates that conventional transport treats as independent flows, despite their shared round and destination. The paper proposes gNB aggregation and scheduling to expose bounded transfers for packet and optical transport. INA reduces upstream learning traffic and makes its transfers shorter and more regular, while optical provisioning depends on aggregate size and setup cost.

  • Problem

    Asynchronous UE updates share a learning round and deadline, but conventional transport treats them as independent flows, obscuring their structure for resource provisioning.

    Source: §I · p. 1

  • Method

    The paper aggregates asynchronous UE updates at the gNB into bounded transfers and describes a Window Scheduler and APN Controller for coordinating optical provisioning.

    Source: §I · p. 1

  • Results

    INA reduces gNB-to-parameter-server learning traffic by a factor of four, forwarding one 42.7-MB aggregate instead of four updates; aggregation also shortens communication phases and rounds.

    Source: §V · pp. 4–5

  • Takeaways & Limitations

    APN/OCS is useful when the RAN reshapes learning traffic into schedulable aggregates; packet transport remains appropriate for small updates and fallback.

    Source: §VII · p. 6

    The appropriate optical-transfer interface and adaptation to wireless variability remain open questions, including whether to expose learning priorities, aggregation policies, or tolerance to delayed updates.

    Source: §VI · p. 5

Abstract

from arXiv · show

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.

I. INTRODUCTION

Federated LLM adaptation creates recurring, deadline-sensitive transfers whose asynchronous UE arrivals are poorly represented by conventional device-level flows. The article proposes gNB aggregation to expose bounded, schedulable demands and coordinate packet and optical transport.

I. INTRODUCTION

  • Wireless variability, mobility, and device heterogeneity make UE updates arrive asynchronously, while conventional transport hides their shared learning process, destination, and deadline.Asynchronous, dispersed UE updates can also be individually too small to justify optical reconfiguration.Source: §I · p. 1
  • In-network aggregation at the gNB combines updates into fewer transfers with bounded volume, known endpoints, readiness time, and latest useful delivery time.These properties let the network choose whether to send an aggregate immediately over packet transport, coordinate it with other aggregates, or use a temporary optical circuit.Source: §I · p. 1
  • The article evaluates aggregation's effects on learning progress, delivery, backhaul traffic, and workload timing, and describes a Window Scheduler and APN Controller for coordinating transfers.The scheduler coordinates ready aggregates; the controller considers their combined volume and deadline slack when deciding whether to provision optical capacity.Source: §I · p. 1
  • The transport abstraction also applies to federated-learning workloads beyond LLMs when edge-aggregated updates can be exposed as bounded, deadline-constrained transfers.LLM adaptation is a motivating case because PEFT reduces full-model fine-tuning requirements while still producing recurring transfers ranging from a few megabytes to hundreds of megabytes.Source: §I · p. 2
  • INA's broader contribution is exposing distributed-learning traffic structure for deadline-aware coordination, while optical provisioning remains workload-dependent and must account for readiness uncertainty, fairness, and missed-deadline risk.Large aggregates may justify circuits when ready; smaller ones may need multi-gNB coordination to amortize setup cost.Source: §I · p. 2

II. IMPACT OF LEARNING TRAFFIC ON THE RAN

Federated-learning updates are correlated within rounds, deadline-sensitive, and heterogeneous across UEs, but the RAN commonly forwards them as independent flows. Round timing therefore affects both communication performance and which clients contribute to the global model.

II. IMPACT OF LEARNING TRAFFIC ON THE RAN

  • Federated learning keeps raw data local while devices compute and send model updates to a parameter server, although privacy attacks on FL systems have been demonstrated.The parameter server waits for enough useful updates to construct the next global model.Source: §II · p. 2
  • Updates from one learning round share an aggregation point and deadline, so timely delivery—not merely eventual delivery—determines whether they can affect the current model.Late updates may wait for a future round or be discarded.Source: §II · p. 2
  • UEs differ in radio conditions, computation, mobility, energy, and data, creating a trade-off between waiting for broader participation and closing rounds sooner.Waiting may improve data diversity, while earlier closure reduces latency but may exclude slower clients and bias the model toward better-connected users.Source: §II · pp. 2–3
  • The gNB generally forwards updates as independent flows without using their shared round or learning objective, missing an opportunity to exploit information already present in federated learning.The workflow includes global-model distribution, local UE training, and staggered uploads through the gNB to a parameter server.Source: §II · pp. 2–3

III. THE GNB AS A LEARNING-AWARE AGGREGATION

The gNB is the first convergence point for learning updates, and the proposed shift is to use it for same-round aggregation before traffic reaches the transport network. This contains some UE-level variability and turns it into more structured gNB-to-server demand.

III. THE GNB AS A LEARNING-AWARE AGGREGATION

  • The gNB can combine updates from the same learning round and forward one aggregate instead of sending each UE's complete update upstream.Clients still train locally and the parameter server still coordinates the global model; the change is the RAN's role in organizing update traffic.Source: §III · p. 3
  • INA combines update fragments as they arrive and closes the aggregate when enough useful contributions arrive or the round deadline is reached.The gNB need not buffer a complete update from every UE or wait indefinitely for the slowest UE.Source: §III · p. 3
  • Aggregation changes the transport-facing demand from many asynchronous UE flows to a smaller number of gNB-level aggregate transfers released after an aggregation window.It contains some device-level variability at the gNB rather than eliminating late UEs or missed updates.Source: §III · p. 3
  • A gNB aggregate has an estimable size and a deadline inherited from its learning round, giving an APN/OCS controller information for optical reconfiguration decisions.FLAG is one example: it embeds aggregation in the 5G gNB user plane to combine updates from multiple UEs before upstream forwarding.Source: §III · p. 3

IV. OPTICAL-NATIVE RANS: A NEW TRANSPORT OPPORTUNITY

gNB aggregation turns federated-learning traffic into transport requests with known endpoints, approximate size, and a round-derived deadline. Those properties can support selective APN/OCS provisioning, while packet switching remains appropriate for small updates and other traffic.

IV. OPTICAL-NATIVE RANS: A NEW TRANSPORT OPPORTUNITY

  • Optical transport can exploit a learning transfer when its source, destination, traffic volume, and transmission window are known.A gNB aggregate exposes these attributes through its gNB source, parameter-server destination, estimable size, and learning-round deadline.Source: §IV · pp. 3–4
  • Packet switching remains suitable for control traffic, model distribution, small updates, and fallback, while optical provisioning is attractive only when aggregate size and expected transmission gain outweigh reconfiguration cost.INA enables schedulable optical demand; it does not replace packet transport.Source: §IV · p. 3
  • INA creates predictable, deadline-aware transfer requests that an optical transport network can schedule efficiently.The opportunity lies on the gNB-to-parameter-server path after updates have been aggregated.Source: §IV · pp. 3–4
  • A Window Scheduler uses aggregate readiness, deadlines, and resource availability to assign a shared active lightpath to the next gNB transfer.In the example, the lightpath moves from gNB 2 to gNB 1 as the latter's aggregate approaches its deadline, rather than dedicating a wavelength to each site.Source: §IV · p. 4, Figure 2 · p. 4

V. LEARNING AND TRANSPORT EVALUATION

The evaluation finds that gNB-side aggregation shortens learning rounds and reshapes asynchronous UE uploads into bounded transfers. Optical circuit switching becomes advantageous when aggregate volume offsets setup time, with smaller aggregates potentially grouped when deadlines permit.

Evaluation setup

  • The learning experiment models 60 UEs across six gNBs, 6.28-MB LoRA updates, and a 10-Mbps backhaul; optical analysis assumes one model-update-sized aggregate per gNB.The evaluation uses different UE counts and update sizes across its two experiments.Source: §V · pp. 4–5

Learning-level impact

  • INA reaches target model accuracy in less elapsed time than conventional forwarding by reducing communication time over the constrained backhaul.Local-training time remains comparable because INA does not modify client computation.Source: §V · p. 4, Figure 3 · p. 4

Transport-level impact

  • INA combines four asynchronous 42.7-MB UE updates into one gNB aggregate, reducing backhaul learning traffic fourfold and producing a shorter, more regular transfer.The aggregate is followed by an identifiable idle interval, exposing one bounded transfer opportunity per learning round.Source: §V · pp. 4–5, Figure 4 · p. 5

Deadline-aware optical provisioning

  • At the considered settings, OCS becomes advantageous from four active gNBs for medium adapters, while one large aggregate amortizes setup and small adapters favor packet transport.The packet–OCS break-even volume is approximately 156.25 MB; smaller aggregates can be grouped when deadlines allow.Source: §V · p. 5, Figure 5 · p. 5

VI. TOWARD DEPLOYABLE AI-NATIVE RANS

The section frames deployment of the proposed architecture as an open set of research questions for future AI-native RANs. It calls for RANs to expose distributed-AI workload semantics so networking decisions can account for learning efficiency alongside throughput and latency.

RQ1: How should learning intent be exposed to the transport network?

  • The transport interface must expose enough learning intent to describe aggregates and adapt when wireless variability disrupts planned transmission windows.Open questions include whether source, destination, size, readiness time, and delivery deadline suffice, or whether priorities, aggregation policies, and delay tolerance are also needed.Source: §VI · p. 5

RQ2: How can transport optimize learning without introducing bias?

  • Deadline-driven aggregation may favor UEs with better radio conditions and exclude slower participants, raising fairness and statistical-representativeness concerns.The section asks whether staleness-aware weighting, delayed aggregation, or partial contributions can preserve model quality while reducing communication latency.Source: §VI · p. 5

RQ3: When should optical transport be used?

  • Hybrid packet/optical transport must determine when traffic is sufficient to amortize optical-circuit setup costs as workloads and network conditions change.Packet transport remains the natural choice for control messages, model distribution, small updates, and fallback operation.Source: §VI · p. 6

RQ4: How should AI-native RANs be evaluated?

  • Future evaluations should assess the complete learning-network-transport loop rather than isolated networking or machine-learning metrics.The proposed scope includes UE variability, gNB aggregation, transport scheduling, deadline satisfaction, communication efficiency, and model convergence.Source: §VI · p. 6

AI-Native RANs

  • RAN architectures should expose distributed-AI workload semantics so networking decisions can optimize learning efficiency alongside throughput and latency.Source: §VI · p. 6

VII. CONCLUSION

The conclusion argues that optical circuits are useful when the RAN reshapes learning traffic into a schedulable form, not as a universal replacement for packet switching. It presents coordinated computation, aggregation, and optical provisioning as parts of one training system.

  • APN/OCS becomes useful when the RAN reshapes learning traffic into a form optical transport can schedule; optical circuits are not always better than packet switching.Source: §VII · p. 6
  • The design question shifts from backhaul capacity alone to which parts of the learning process transport should see and when.Source: §VII · p. 6
  • Future RANs should coordinate computation, aggregation, and optical provisioning as parts of the same training system.Source: §VII · p. 6
Loading 2610.01304v1…