Large language models currently treat training and learning as largely identical processes: both arerealized through parameter optimization using large datasets. During inference, contextualknowledge is typically incorporated through prompts, retrieval systems, or external tool interfaceswithout modifying the model parameters themselves. This paper proposes a conceptual separation between training and learning in neural languagesystems. Training produces a stable base model through dataset-driven optimization. Learning, incontrast, is defined as the creation of context-indexed parameter overlays Wc that modify theeffective behavior of the base model without permanently altering its base parameters. Theseoverlays may arise through transfer or reinforcement signals and are activated selectivelydepending on the operational context. To support scalable contextual learning, the paper further proposes a standardized contextinterface inspired by Service Provider Interface (SPI) concepts, enabling context consumersto ingest context as structured payloads rather than interacting with persistent context services e. g. , Model Context Protocol. Such an approach reduces architectural complexity and enables modularcontext ecosystems. The framework outlines a layered architecture in which context acts as a transformation operatoron a base model, enabling reusable contextual specialization while avoiding repeated retraining orexcessively large prompt contexts.
Das et al. (Mon,) studied this question.