A Six-Dimensional Taxonomy Tries to Disentangle Fine-Tuning, RAG, Prompting, Editing, and Unlearning for Model-Change Tracking
arXiv 2608.06246·low signal
This survey (2608.06246, submitted 2026-08-06) argues the post-training adaptation literature is fragmented across technique families, model classes, and deployment contexts, making methods incomparable and model modifications undescribable. It proposes organizing everything — retraining, fine-tuning, PEFT, alignment, retrieval augmentation, model editing, unlearning, calibration, multimodal instruction tuning — along six axes: mechanism, goal, data requirement, persistence, structural scope, and model type. It explicitly separates terms that get conflated in practice (fine-tuning vs. retrieval augmentation vs. prompting) and maps inheritance, supersession, hybridization, and layered deployment stacks between techniques. The stated payoff is a shared vocabulary for technical documentation, model-change tracking, and governance analysis — useful if you have to attest to how a deployed model was modified.