Someone tells you:
Which tells you almost nothing. Here is what it actually pins down:
PEFT (LoRA, adapters) — one of six readings of “fine-tuned”.
48 techniques. Six dimensions. One coordinate each.
A six-dimensional taxonomy of post-training adaptation, built so a model change can be named precisely enough to document, compare, and audit.
Explore the taxonomy Classify a change Start with the concepts
Three problems this solves
Terminological ambiguity
The same word means different things. “Fine-tuning” covers three different scopes with three different validation burdens.
Disambiguate a term →Single-axis taxonomies
Is PEFT a fine-tuning method or an efficiency method? It is both, which is why one axis cannot hold it.
Filter on all six axes →Model-type conflation
Prompt engineering is meaningless for a random forest. Which techniques are even available depends on what you are adapting.
See the model-tier ladder →Training, or post-training?
Training and retraining are mechanistically identical — the same gradient updates, the same permanence, the same scope. What separates them is why you did it and what data you used.
| Pair | Identical on | Separated by |
|---|---|---|
| Training vs Retraining | D1, D4, D5, D6 | D2 build vs repair · D3 data regime |
| Retraining vs Full FT | D1, D4, D5 | D2 drift vs new task |
| Partial FT vs PEFT | D1, D3, D4, D6 | nothing outright — PEFT strictly extends it |
Why precision matters
Before
Update v2.1: the model was fine-tuned on recent hospital data to
improve performance and incorporate new clinical guidelines.
Does not say whether base weights changed, whether the change is permanent, or whether the guidelines were learned or retrieved.
After
Two layers, named separately — a parameter update to an adapter, and a retrieval corpus update. Each with its own six-dimensional profile.
See the full worked example →Overviews
Video overview
Fine-Tuning, RAG, or Prompting? The 6D Framework for AI Model Adaptation
SMAILE KI
Nothing loads from YouTube until you press play. Watch on YouTube instead
Audio overview
The 6D Taxonomy of AI Adaptation
SMAILE KI · 21 min
Download (19 MB) · nothing downloads until you press play
Cite
The preprint is arXiv:2608.06246 (placeholder until posted).
@misc{afdideh2026taxonomy,
title = {A Six-Dimensional Taxonomy of Post-Training Adaptation
Techniques with Applications in AI Governance},
author = {Afdideh, Fardin and Seoane, Fernando and Abtahi, Farhad},
year = {2026},
eprint = {2608.06246},
archivePrefix = {arXiv}
}