LLM-as-a-Judge for Dynamic Resource Allocation in QoT-Aware Elastic Optical Networks

Elastic optical networks (EONs) require online routing, modulation, and spectrum assignment (RMSA) under spectrum-continuity and physical-layer constraints. This paper evaluates compact large language models (LLMs) as bounded arbiters that select among quality of transmission (QoT)-feasible candidate actions generated by classical heuristics. Candidate actions are decoded, QoT-checked, deduplicated, and summarized through physical and spectral descriptors. A frozen shortlist ranker then supplies up to three candidates to the LLM judge. On the \textit{nobel-eu} topology, three zero-shot deterministic LLM judges are compared with fixed heuristics and a calibrated shortlist ranker-only baseline from 100 to 300 Erlang. Blocking is rare at low loads, but under congestion, Llama-3.1 8B has the lowest observed blocking, with 77.8% and 42.0% reductions relative to the calibrated baseline at 250 and 300 Erlang, respectively. Diagnostics show that most calls contain multiple candidates and that LLM selections often differ from the ranker's top candidate.