Can a digital twin-aided multi-agent LLM autonomously manage optical networks?

The system closes a soft-failure loop without manual operation

The paper's question is whether a multi-agent LLM framework with SDN control can autonomously optimize services in an optical network. The authors demonstrate this on a physical-layer testbed: the system detects soft failures before they escalate and uses digital-twin-assisted BER predictions to keep marginless connections running automatically over a 630 km path under emulated amplifier and fiber ageing.

The claimed capability is narrow and specific. It is not general network autonomy; it is autonomous soft-failure detection and correction, which the authors frame as essential for marginless operation, where a network is pushed toward its limits and needs proper performance observation and reaction when problems appear.

Three agents split planning, verification, and execution

The architecture instantiates a plan–verify–execute paradigm. A Supervisor Agent, a fine-tuned LLaMA3.1-8B, parses natural-language intent and delegates tasks. A Toolbox acts as a centralized orchestration bus that ingests real-time telemetry from a NETCONF/YANG-polled Telemetry Database, so downstream actions are grounded in current network state.

Two specialized agents then act. The Config Agent translates high-level intent into vendor-neutral NETCONF remote procedure calls for amplifier gain, variable optical attenuator, and wavelength selective switch settings. The QoT Agent performs quality-of-transmission estimation and parameter optimization based on detected impairments. Both are fine-tuned LLaMA3.1-8B models that query a shared Knowledge Base containing standards, vendor specifications, and physical-layer constraints via retrieval-augmented generation.

LLaMA3.1-8B fine-tuned won on both evaluation axes

Three models — LLaMA2-7B, LLaMA3.1-8B, and Mistral-7B — were fine-tuned via QLoRA on testbed telemetry and ITU-T standards documentation, then evaluated on standard LLM benchmarks (MMLU, ARC, and perplexity) and on OpED, the Optical Network Evaluation Dataset developed in this work, which covers optical hardware, DSP, and SDN/NFV automation and applies an honesty penalty for hallucinated responses.

LLaMA3.1-8B fine-tuned achieved the strongest results across both axes: the largest perplexity reduction at 36.70% on standard benchmarks, and the highest OpED accuracy at 96.1%. Its low honesty rate of 13.5% reflects a small number of abstained questions, the majority of which were answered correctly — unlike base models, which attempted all questions regardless of confidence and yielded no honesty signal. Given this overall superiority, LLaMA3.1-8B fine-tuned was selected for deployment across all agent roles.

The soft-failure threshold is statistical, not a fixed BER

The QoT Agent monitors pre-FEC BER against a soft-failure threshold defined as the mean BER plus four standard deviations, computed from a 9-sample window with the standard deviation taken from live measurements. Adaptation triggers when BER reaches about 20% of the FEC margin.

A failure detection algorithm then selects a lower-rate candidate mode and evaluates feasibility using three inputs: FEC headroom, expressed as log10 of FEC limit divided by BER; BER trend; and BER prediction using calibrated fiber attenuation coefficients. This assessment plus cross-channel context is passed to the LLM for an APPROVE, DEFER, or REJECT decision. If approved, the Config Agent applies the new mode via the ONOS SDN controller.

Gold and bronze service classes trade spectrum differently

The testbed has 8 commercial-grade reconfigurable add/drop multiplexer nodes in a mesh topology, carrying two channels from two generations of commercial coherent transceivers under two service class agreements. Gold class is high priority and must maintain constant net rate with a high margin. Bronze class is low priority and is allowed to operate at zero margin, with the caveat that net rate can be sacrificed under strenuous conditions.

To emulate fiber ageing, the authors increase VOA attenuation at ROADM-3 from 0 dB to 12 dB in steps of 0.2 dB, holding each level for 5–10 minutes to ensure steady-state conditions. Attenuation starts at 0 dB so the agents capture the baseline BER distribution and compute the threshold. As loss increases, the Q-factor derived from pre-FEC BER drops and the available FEC margin, defined as Qmeasured² minus QFEC² in dB, narrows for both services.

Mode transitions follow a verified sequence, not a single switch

The recorded transitions show a staged response. With increasing attenuation, the Gold service initially steps from G1 to G2 by forcing the Bronze service to downgrade from B1 to B2. With further attenuation, the digital twin calculates an 85% failure probability for the Gold service, which forces another change from G2 to G3 while B2 remains unchanged.

By reducing the symbol rate of the Bronze service, spectrum is freed for the Gold service to maintain the same net rate, under the assumption that both channels co-propagate through the same nodes until reception. The Configuration Agent verifies the proposed configuration on the digital twin, which emulates statistical network performance under the proposed trend failure and returns the expected Q² with failure probability when Q² is above the FEC margin, built from past measurements.

The digital twin is a verification layer against hallucination

The paper positions the digital twin as a verification layer that mitigates LLM hallucinations. Proposed outputs are intercepted before reaching live hardware, and the Configuration Agent checks the proposed configuration on the digital twin before deployment. This matters because the LLM decision layer is not trusted to act directly on the physical network.

The ONOS log records the event sequence — SOFT_FAILURE, CONFIG_INTENT, and CONFIG_APPLIED — during testbed configuration, and the LLM assessment includes decision rationale, confidence level, and an operator alert. The authors conclude that they experimentally demonstrated consistent service optimization through timely mode adaptation that autonomously maintains service class requirements under progressive degradation without manual operation.

Where the demonstration stops

The evidence is a testbed demonstration, not a deployed network trial. The setup uses 8 ROADM nodes, two channels, two service classes, and two generations of transceivers from the same vendor, with ageing emulated by VOA attenuation rather than observed in service. The 630 km path and the 0 dB to 12 dB attenuation sweep define the operating envelope the results cover.

The most decision-relevant result is the service-class split: the system preserves the Gold service's constant net rate by sacrificing Bronze net rate, and it does so only after the digital twin returns an acceptable failure probability. Readers should treat the 85% failure probability and the G1→G2→G3 sequence as the paper's specific threshold behavior under emulated ageing, not as a general rule for other topologies, vendor mixes, or traffic patterns.

Sources used in this answer