Table of contents
In this article
Technical page · Models
The adaptation ladder
Prompting, retrieval, fine-tuning, then preference tuning
- Published
- 8 November 2026
- Last reviewed
- 8 November 2026
- Re-checked every
- 180 days
- Reviewed by
- Ales Analytics engineering
In one paragraph
The adaptation ladder is the order in which to try making a model fit a task: better prompts and context first, then retrieval, then supervised fine-tuning, then preference tuning or reinforcement learning. Each rung costs more than the last, so a rung is climbed only when evaluation shows the one below falls short. Retrieval adds knowledge; fine-tuning changes behaviour.
How it works
| Rung | Fixes | Costs |
|---|---|---|
| Prompt and context engineering | Unclear instructions, wrong format, missing examples | Hours; no training |
| Retrieval | Missing or changing knowledge | An index and a pipeline to keep it current |
| Supervised fine-tuning | Consistent behaviour the prompt cannot produce | Labelled examples, training runs, a model to version |
| Preference tuning or RL | Choosing between acceptable and better outputs; passing computable checks | Preference pairs or reward functions, more careful evaluation |
The evaluation set is built before the first rung. Without it there is no way to tell whether a rung helped.
Design choices and trade-offs
- Knowledge in weights or in retrieval. Knowledge that changes, or that must be cited, belongs in retrieval. Fine-tuning facts into weights makes them hard to update and impossible to cite.
- How many examples. A few hundred well-chosen examples often move behaviour noticeably with parameter-efficient methods; noisy examples move it the wrong way.
- Forgetting. Fine-tuning can degrade abilities the task did not exercise; parameter-efficient methods tend to forget less.
When not to use it
When the base model already passes the acceptance tests with a good prompt, stop at the first rung.
How we evaluate it
Each rung is compared against the previous one on the same frozen test set, with a separate check on general abilities the system still relies on, so a gain on the task is not paid for elsewhere.
Sources
- Biderman et al. (2024), LoRA Learns Less and Forgets Less (opens in a new tab)
- Model adaptation and evaluation
- Explainer: Tuning and routing
Should you have any questions regarding this page, please contact us at solutions@alesanalytics.com.