What is Step-Back Prompting?
Step-Back Prompting was introduced in the paper Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models by Huaixiu Steven Zheng and colleagues at Google DeepMind, published at ICLR 2024. The idea mirrors how experts approach unfamiliar problems: before calculating anything, a physicist asks which law applies, and a historian first establishes the timeline.
The technique has two stages:
- Abstraction: the model answers a higher-level "step-back" question about the concept, principle or background behind the original question.
- Reasoning: the model answers the original question, grounded in the principle or facts it just retrieved.
Why it works: specific questions often contain distracting details that pull the model towards a plausible but wrong path. Naming the principle first narrows the search to the right knowledge. In the paper, the approach improved PaLM-2L on MMLU Physics by 7%, MMLU Chemistry by 11%, TimeQA by 27% and MuSiQue by 7%.
When to Use Step-Back Prompting
STEM Questions
Physics, chemistry and engineering problems where choosing the right law or formula is most of the work.
Knowledge-Intensive QA
Questions about people, events or organisations where the model should first recall the relevant history.
Multi-Hop Questions
Questions that need several facts combined, where recalling the broader context first avoids missing a hop.
Debugging and Root Cause
Ask what class of problem this is (race condition, caching, encoding) before looking at the specific stack trace.
Policy and Compliance
Identify the governing rule or regulation first, then apply it to the specific case.
Teaching and Explanations
Explaining the concept before the worked answer produces better tutoring responses.
How to Use Step-Back Prompting
- 1
Write the step-back question
Ask about the principle, category or background behind the specific question. Keep it general enough that the answer applies beyond this one case.
- 2
Let the model answer it first
Either as a separate call or as a labelled "Principle" section in the same response. Ask for concise, factual output.
- 3
Answer the original question using the principle
Explicitly instruct the model to base its reasoning on the principle it stated, and to reason step by step if the problem is multi-step.
- 4
Check for consistency
If the final answer contradicts the stated principle, that is a strong signal of an error worth a second pass.
Prompt Examples
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? Step 1 (step back): What physics principle relates the pressure, temperature and volume of an ideal gas? State it as a formula. Step 2: Using that principle, answer the question step by step. Final answer on the last line.
Our checkout service returns stale prices for about 5 minutes after a price change, but only for logged-in users. First, step back: which general classes of problems cause data to be stale for a fixed period for one group of users but not another? List the top 3 with a one-line explanation each. Then, using those categories, rank the most likely causes in our case and propose one diagnostic check per cause.
Pros and Cons
| 🟢 Pros | 🔴 Cons |
|---|---|
| Large reported gains on knowledge and STEM tasks (7–27%) | Extra stage adds tokens and latency |
| Simple to learn and to add to existing prompts | Little benefit for simple lookups or writing tasks |
| Stated principle makes errors easier to spot | A wrong principle in step 1 leads the whole answer astray |
| Combines naturally with Chain-of-Thought and RAG | Choosing the right level of abstraction takes practice |
Step-Back Prompting Prompt Template
Copy this template and replace each {{ variable }}. It uses Prompt Edit's variable syntax, so you can paste it into the app and fill the fields from a form every time you reuse it.
Question: {{ question }}
Step 1 (step back): What general principle, concept or background knowledge is needed to answer this? {{ step_back_question }}
Answer it concisely.
Step 2: Using only that principle, answer the original question step by step.
Final answer: one line.Frequently Asked Questions
What is Step-Back Prompting?
Step-Back Prompting is a two-step reasoning technique from Google DeepMind (Zheng et al., 2023). First the model is asked a more general 'step-back' question about the concept or principle behind the problem. Then it answers the original question using that principle. Abstracting first helps the model retrieve the right knowledge before it gets lost in the specific details.
How much does Step-Back Prompting improve accuracy?
In the paper, Step-Back Prompting improved PaLM-2L performance on MMLU Physics by 7% and Chemistry by 11%, on TimeQA by 27% and on MuSiQue multi-hop questions by 7%. Gains were also reported with GPT-4 and Llama 2 70B.
What is a good step-back question?
A good step-back question asks about the general principle, category or history behind the specific question. For 'What happens to the pressure of an ideal gas if the temperature doubles and the volume increases eight times?' the step-back question is 'What physics principle relates pressure, temperature and volume?' (the ideal gas law). For 'Which school did Estella Leopold attend between August 1954 and November 1954?' it is 'What was Estella Leopold's education history?'
What is the difference between Step-Back Prompting and Chain-of-Thought?
Chain-of-Thought reasons forward through the specific problem step by step. Step-Back first moves up a level of abstraction to recall the governing concept, then reasons down to the answer. They combine well: step back to get the principle, then apply it with Chain-of-Thought.
Can I do Step-Back Prompting in one prompt?
Yes. Ask the model to first state the underlying principle or background in a labelled section, then answer the original question using it. The paper used separate abstraction and reasoning steps, but a single structured prompt works well in chat.
When does Step-Back Prompting not help?
It adds little for simple lookups or tasks with no underlying principle to recall, such as rewriting or formatting. It helps most on knowledge-intensive and multi-hop questions where the model needs to recall the right facts or rules first.
Origin & Sources
- Zheng et al. (Google DeepMind) (2023). Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models. arXiv / ICLR 2024.