skip to content

Step-Back Prompting

Ask for the general principle behind a question, then answer with that principle in hand. Interviewers ask when this two-step prompt beats reasoning straight through, and where it still fails.

part ofPrompt engineeringread it →
on this pageshow

questions

6

What is step-back prompting, and how do its abstraction and reasoning steps work?

level: middleimportance: should knowfreq 35%

answer

  1. two steps, general before specific
  2. a question one level higher
  3. state the principle, then apply it
  4. Zheng et al. 2023: abstraction, then reasoning

basics

~20 s

Step-back prompting first has a model answer a more general question about the concept or principle behind a task, then answer the original question grounded on those facts. The two steps are named abstraction and reasoning.

solid answer

~50 s

Step-back prompting, from Zheng et al. (2023), splits answering into two steps. In **abstraction**, few-shot demonstrations show the model how to derive a higher-level question or principle from the specific one and state the relevant facts: a question about an ideal gas with doubled temperature and eightfold volume steps back to "which principles apply?", answered by the ideal gas law. In **reasoning**, the model answers the original question grounded on those facts, so the pressure comes out at a quarter of its starting value. The premise is that the general question is usually easier and that grounding reduces errors in the intermediate steps. In that paper it raised accuracy on reasoning-heavy benchmarks, by about 7 and 11 points on high-school physics and chemistry for one large model. That is a finding on those tasks and models, and a later model may behave differently.

code

plaintext · 15 lines
plaintext
PROMPT 1, abstraction
You are given a physics problem. Extract the concepts and principles involved in solving it.

Question: <demonstration question>
Principles: <demonstration principles>

Question: What happens to the pressure of an ideal gas if the temperature is doubled and the volume is increased by a factor of 8?
Principles:

PROMPT 2, reasoning
You are given a physics problem and the principles involved in solving it. Solve it step by step by following the principles.

Question: What happens to the pressure of an ideal gas if the temperature is doubled and the volume is increased by a factor of 8?
Principles: Ideal gas law, PV = nRT, where P is pressure, V is volume, n is moles, R is the gas constant, T is temperature.
Answer:

go deeper

for a junior

Recall the order: a more general question first, then the original question answered with what came back. Name the two steps, abstraction and reasoning.

for a middle

Explain what abstraction means in the paper, why few-shot demonstrations teach it, and walk one small case through both prompts without notes.

for a senior

Say where the general facts come from, the model's memory or supplied context, and scope every accuracy figure to the paper's tasks and models.

for a principal

Treat it as an option with a price: an extra step on every question it touches. Tie adoption to measured question classes, not to the paper's headline numbers.

## What the technique is **Step-back prompting** is a prompting technique introduced by Zheng et al. (2023). Instead of answering a detailed question directly, the model first answers a more general one, then uses that answer to handle the detail. The paper defines a **step-back question** as a question derived from the original at a higher level of abstraction. **Abstraction** has a specific meaning here: moving from an instance with particular details to the high-level concept or first principle that covers it. It does not mean being vague. A date range, a name or a particular number falls away, and what remains is the general matter the instance belongs to. The paper's motivation is that many tasks hold so much detail that a model struggles to retrieve the relevant facts, and it can drift from the governing principle when it reasons straight at the question. ## The two steps 1. **Abstraction.** The model is prompted to ask a generic question about a higher-level concept or principle and to retrieve the relevant facts. It learns to do this from few-shot demonstrations, which are worked examples placed in the prompt. 2. **Reasoning.** Grounded on those facts, the model reasons to an answer to the original question. The paper calls this abstraction-grounded reasoning. Two details matter when you build it: - **The original question stays in the second prompt.** The general facts support the answer. They do not replace the task. - **The reasoning step still works through steps.** The paper's final-answer prompt for physics tells the model to solve the problem step by step by following the principles. ## A small worked case Take a question the paper uses: what happens to the pressure of an ideal gas if the temperature is doubled and the volume is increased by a factor of 8? ```plaintext Step 1, abstraction Step-back question: what physics principles are behind this question? Step-back answer: ideal gas law, PV = nRT Step 2, reasoning T becomes 2T, V becomes 8V P' = nR(2T) / (8V) = (1/4) * nRT / V Final answer: the pressure falls to a quarter of its value ``` In the paper's illustration, one model went wrong with a plain prompt and with chain of thought, and answered correctly once the law was stated first. That is one example chosen to show the method, not evidence on its own. The same shape fits a knowledge question. An internal benefits assistant asked whether a contractor who joined in March can claim a home-office allowance might first answer "what conditions decide who can claim this allowance?" from the policy text in its prompt, then apply those conditions to this contractor. The scenario is illustrative, and it shows a boundary: a model can recite a public law from what it knows, but a private policy has to be in its context, or the abstraction step may invent one. ## One prompt or two calls The paper runs the steps as separate prompts on physics, chemistry and knowledge questions: one prompt produces the principles or the step-back question, and a later prompt takes them and produces the answer. On arithmetic word problems it merges them, with a demonstration that states each principle beside its calculation, because the principles there are so simple. The paper reports no head-to-head comparison of the two forms on one task, so the choice is an engineering call. Two calls let you log and inspect the abstraction, and one prompt spares a round trip. ## What the paper measured | Benchmark, one large model | Plain prompt | Step-back alone | |---|---|---| | High-school physics | 66.4% | 73.2% | | High-school chemistry | 70.9% | 81.8% | | TimeQA, time-sensitive facts | 41.5% | 66% | | MuSiQue, multi-hop questions | 35.5% | 42.6% | These are that paper's figures, on 2023 models, and a later model may behave differently. The gain also varied by task and by model, which is why the figures describe those benchmarks and do not promise the same lift on yours. ## Where it does not apply The authors say abstraction is neither necessary nor possible in all scenarios: - **A simple lookup** has its answer readily available, so there is no need to step back. - **A question that already points at a first principle**, such as the speed of light, gains nothing, since abstraction would not make a difference. A broader question can also serve as a search query before retrieval. That form is covered with query transformation, and this technique's core is the answer side: abstract first, then reason from the principle.

  • When would you run the two steps in one prompt rather than two calls?
    One prompt fits when the principles are simple and a second round trip is costly. The paper merges them on arithmetic word problems because of how simple those principles are. Two calls let you log, inspect and test the abstraction on its own. The paper reports no head-to-head of the two forms, so this is engineering judgment to check on your own held-out set.
  • Where does the step-back answer come from when the model has no retrieval?
    From the model's own knowledge. The paper's physics demonstrations have the model recite relevant principles from what it knows, and its error classes count a wrong recited fact as a failure. That works best for public laws. For private facts such as an internal policy, the text has to be in the prompt, because the model has nothing reliable to recite.

Like writing the relevant formula at the top of an exam page before substituting the numbers: it points the working in the right direction, but it does not do the arithmetic for you.

  • Step-back means asking the model to reconsider its answer after it has replied.
  • The general answer is the final output, so the original question can be dropped.
  • Abstraction just means being vague so the model has room to think.
  • Any question gains from it, so even a simple lookup should step back first.
  • A private policy can be abstracted from the model's memory without putting it in context.
open as a page

In step-back prompting, how do you write the step-back question and the few-shot exemplars that teach the abstraction?

level: middleimportance: should knowfreq 20%

basics

~20 s

Strip the specifics from the question and ask about the concept covering them, so the answer still holds what the original needs. Exemplars pair an original question with its abstraction; one sufficed on the paper's tested tasks.

open as a page

How does step-back prompting differ from chain of thought and from decomposing a question into sub-questions?

level: seniorimportance: should knowfreq 30%

basics

~10 s

Chain of thought reasons forward from the question; step-back first climbs to a general principle, then reasons from it. Decomposition goes down into sub-questions, one-to-many, while many specific questions can share one abstraction, many-to-one.

open as a page

A step-back prompt still misses a fifth of your held-out questions: which step do you suspect first, and how do you confirm it?

level: seniorimportance: should knowfreq 25%

basics

~20 s

Suspect the reasoning step first. Zheng et al. (2023) place over 90% of step-back's physics errors there, with wrong principles a small fraction. Confirm by tagging failures per step and rerunning them with a known-correct principle.

open as a page

Across the tasks its paper tested, where did step-back prompting gain accuracy and where did it add little?

level: seniorimportance: should knowfreq 25%

basics

~20 s

In Zheng et al. (2023), one large model gained on physics, chemistry, time-sensitive facts and multi-hop questions. Step-back only tied chain of thought on arithmetic word problems and stayed flat without retrieval on a binary strategy benchmark.

open as a page

Should step-back prompting's extra abstraction call be a team's default, or earn its place only on a measured class of questions?

level: principalimportance: nice to knowfreq 15%

basics

~20 s

Prefer measured classes over a blanket default. The abstraction call adds latency and tokens to every request, published gains varied by task and model, and simple lookups need none of it. Gate it where a held-out set shows lift.

open as a page