online learning approach to learning-augmented k-median

We generate global text-based explanations using representative nodes (exemplars) in the embedding space. The exemplars are selected via coverage maximization, and their signatures are explained using natural language rules from a self-refining LLM.

We model stable matchings under group-dependent bias and correlated evaluations, characterize equilibrium thresholds, and show how evaluator alignment amplifies or mitigates fairness loss in decentralized systems.