Definition
Generation-time specialization adapts a model’s behavior on the fly — through prompting, retrieval, scaffolding, or lightweight adapters — rather than baking the specialization into the weights through additional training. It trades some accuracy for the ability to ship many specializations from a single base model.