Glossary · Technique
Step-Back Prompting
Also known as: Abstract-first prompting
Before answering the specific question, derive the higher-level principle. Then apply it. Better generalization, fewer hallucinations.
When to use it
- Specific questions where the underlying principle is more reliable than the surface facts.
- Physics, chemistry, math, finance, law — anywhere principles compose.
- When direct factual recall is brittle but the meta-rule is well-known.
- Pre-step before Chain-of-Thought to avoid early commitment to a wrong frame.
When not to use it
- Pure factual lookups with no general principle.
- Tasks where the specific instance has no broader pattern.
- Time-sensitive prompts — adds reasoning overhead.
How it works
- 1Step 1: Ask the model to identify the general principle / rule / concept behind the specific question.
- 2Step 2: Confirm the principle is correct (sanity check).
- 3Step 3: Apply the principle to the specific case to derive the answer.
- 4Step 4: State the answer with the principle as justification.
Example
Lazy prompt
If a car accelerates from 30 m/s to 50 m/s in 4 seconds, what's the acceleration?
Using the technique
First, state the general principle for computing acceleration from velocity change over time. Then apply it to: v_initial=30 m/s, v_final=50 m/s, t=4s. Show the work.
Common pitfalls
- If the abstracted principle is wrong, the answer compounds the error.
- Models can over-abstract — pull in irrelevant principles that don't actually apply.
- Adds tokens; not worth it for trivial questions.
Where this came from
Zheng et al., 2023 — "Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models".