Rephrase and respond (RaR) is a prompting technique in which a large language model is asked to restate and expand the user’s question in its own words before it produces an answer. The rewritten question makes hidden assumptions explicit, resolves ambiguous wording and adds missing context, so the subsequent answer addresses what the user most likely meant.
In Prompting practice, RaR belongs to the family of techniques that change how the model frames a task rather than what it knows. It was described in 2023 research on the observation that questions which look clear to a human author are often ambiguous from the model’s point of view, and that the model itself can repair much of that gap.
How rephrase and respond works
RaR works by inserting an explicit clarification step between the question and the answer. The model first produces a more detailed version of the question, then answers the original question with that clarified version in view. Because language models generate text left to right, the rephrased question becomes part of the context that conditions every token of the answer.
The technique exists in two forms, summarized in the following table.
| Variant | How it runs | Model calls | Main advantage |
|---|---|---|---|
| One-step RaR | A single instruction asks the model to rephrase, expand and then answer | One | Simple and cheap to add to any prompt |
| Two-step RaR | One call produces the rephrased question; a second call answers using both versions | Two | The rephrasing can come from a stronger model |
A typical one-step instruction appended to a question reads: Rephrase and expand the question, and respond. In the two-step form, the rephrasing model and the responding model can differ, which allows a capable model to prepare clearer questions that a smaller or cheaper model then answers.
Why clarifying the question improves answers
Human questions are written from the author’s frame of reference, so they frequently omit definitions, scope and intended format. A question such as whether a historical figure was born in an even month leaves open whether “even” refers to the month number or to something else. RaR forces the model to commit to an interpretation in writing, which reduces silent misreadings.
The mechanism is related to, but distinct from, step-by-step reasoning. Zero-shot chain-of-thought elaborates the solution path, whereas RaR elaborates the problem statement itself. The two address different failure points and can be combined in one prompt, first clarifying the question and then reasoning through it.
The table below lists the kinds of ambiguity that RaR tends to resolve.
| Ambiguity type | Example in a question | What the rephrasing adds |
|---|---|---|
| Undefined criterion | Is this number large? | A threshold or comparison point |
| Unclear scope | List the main causes | Time period, region or domain |
| Implicit format | Summarize the report | Expected length and structure |
| Overloaded word | Was the coin flipped even times? | Precise meaning of the key term |
Rephrase and respond compared with related techniques
RaR sits among several prompting patterns that add an intermediate step before the final answer. Each targets a different weakness of direct prompting, so knowing the differences helps practitioners choose or combine them.
| Technique | Intermediate step | Primary target |
|---|---|---|
| Rephrase and respond | Restated, expanded question | Ambiguous or underspecified questions |
| Zero-shot chain-of-thought | Step-by-step reasoning | Multi-step logic and arithmetic |
| Step-back prompting | A more general principle question | Problems needing abstraction first |
| Self-ask prompting | Explicit follow-up subquestions | Compositional factual questions |
| Chain-of-verification | Verification questions after a draft | Factual errors in the draft answer |
Unlike the flipped interaction pattern, in which the model asks the user clarifying questions, RaR resolves ambiguity without a round trip to the user. That makes it suitable for automated pipelines, but it also means the model chooses the interpretation on its own.
Applying RaR in practice
RaR is most useful when questions come from many different authors and cannot be curated in advance, such as support chatbots, internal assistants and evaluation sets written by non-specialists. The following decision table maps common situations to a recommended use.
| Situation | Recommended approach | Reason |
|---|---|---|
| Short, ambiguous user questions | One-step RaR | Low cost, immediate clarity gain |
| Cheap model answering high volumes | Two-step RaR with a stronger rephraser | Clear questions help weaker models |
| Reasoning-heavy tasks | RaR plus chain-of-thought | Clarify first, then reason |
| Precise, well-specified prompts | Skip RaR | Little ambiguity to remove |
Good practice is to show the rephrased question in logs or, in user-facing tools, to display it briefly. Visible rephrasings let reviewers detect when the model has drifted from the user’s intent, and they produce useful data for improving prompt templates.
Limitations and failure modes
RaR improves results only when the model can infer the intended meaning better than it reads it. Reported gains are largest for strong models and smaller, sometimes negligible, for weaker ones, which can produce rephrasings that are themselves confused.
- Intent drift: the expanded question may add assumptions the user never made, so the answer is precise but addresses the wrong problem.
- Higher token cost: the rephrasing adds output tokens, and the two-step form doubles the number of calls.
- Longer latency: users wait for the clarification before the answer begins.
- No new knowledge: RaR cannot fix factual gaps; it only improves how the existing question is understood.
A practical safeguard is to constrain the rephrasing step, for example by instructing the model to keep every fact stated by the user and to mark any added assumption explicitly. When ambiguity is high and the stakes are serious, asking the user to confirm the interpretation remains the safer choice than silent rephrasing.
The idea in brief
Rephrase and respond means asking the AI to repeat the question in its own words first. Only then does it answer. People often leave out details because the meaning seems obvious to them.
When the AI restates the question, it spells out what it thinks was meant. That catches many misunderstandings early. It works like a good waiter who repeats an order back before sending it to the kitchen.
A practical scenario
A company runs an internal assistant that answers employee questions about travel and expense policy. Many questions are short, for example asking whether a hotel stay is covered. The assistant often gave generic answers that missed the point, because it did not know the trip type, the city or the expense limit being asked about.
The team adds a one-step RaR instruction. The model now restates each question, for instance as a question about whether a three-night business hotel stay above the standard nightly limit is reimbursable and what approval it needs, and then answers that version using the policy text.
In a review of several hundred logged conversations, reviewers judge noticeably more answers as relevant to what the employee asked. They also use the visible rephrasings to spot recurring ambiguities and add missing details to the policy documents.