Rephrase and Respond Prompting: The Real 2023 Paper Behind RaR

TL;DR: Rephrase and Respond (RaR) has the model restate your question in its own words, then answer, in a single reply. It comes from a real 2023 paper by four UCLA researchers, not from prompt-engineering folklore, and the paper's own tests moved some question types from roughly half-right to nearly always right.

What is Rephrase and Respond (RaR) prompting?

Rephrase and Respond, or RaR, is a one-line addition to any prompt: you ask the model to restate and expand your question before it answers, inside the same reply. Nothing else changes about how you use it. You keep asking your question exactly as you would have anyway, and you append one instruction.

The idea sounds almost too simple to name, which is exactly why it is worth sourcing properly before repeating it. A lot of what circulates under prompting-technique names is folk wisdom that got a catchy label after the fact. RaR is the opposite case: it has an actual paper, actual authors, and actual numbers behind it, and most of what is written about it online paraphrases that paper without linking to it.

Where does Rephrase and Respond actually come from?

RaR is introduced in "Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves," by Yihe Deng, Weitong Zhang, Zixiang Chen and Quanquan Gu, all in the Department of Computer Science at the University of California, Los Angeles. It was posted to arXiv on November 7, 2023, as arXiv:2311.04205, and revised on April 18, 2024. Its arXiv listing carries no separate journal or conference reference, so it should be cited as a preprint rather than as a peer-reviewed publication.

The paper opens from a specific, well-argued premise. Its abstract states plainly: "Misunderstandings arise not only in interpersonal communication but also between humans and Large Language Models (LLMs)." The authors' point is that a question a human reads as unambiguous can still be read multiple ways by a model, and the model has no way to signal that it picked a different reading than the one you intended.

Their own motivating example is a good illustration of the failure mode RaR targets. Asked whether Ludwig van Beethoven was born in an even month, GPT-4 answered no, reasoning that December "is often categorized as an odd month because the number of its days is 31, which is an odd number." The model wasn't wrong about arithmetic. It had silently substituted a different definition of "even month" than the one the question intended, and nothing in a single exchange revealed that substitution before the wrong answer arrived.

The paper's stated fix, in its own words: "we present a method named ‘Rephrase and Respond’ (RaR), which allows LLMs to rephrase and expand questions posed by humans and provide responses in a single prompt." Having the model restate the question in its own words, before answering, surfaces exactly this kind of silent misreading, because the rephrased version shows you which definition the model actually picked.

What is the exact Rephrase and Respond prompt?

This is the one-step version, quoted directly from the paper's Section 3:

"{question}"
Rephrase and expand the question, and respond.

You append that single instruction after your actual question. The model rephrases it, then answers, in one reply. The paper is explicit that wording flexibility is part of the method, and lists four tested variants that all worked:

Reword and elaborate on the inquiry, then provide an answer.

Reframe the question with additional context and detail, then
provide an answer.

Modify the original question for clarity and detail, then offer an
answer.

Restate and elaborate on the inquiry before proceeding with a
response.

Any of the five (the original plus these four) does the job. Pick whichever reads most naturally next to your own question.

Does Rephrase and Respond actually improve accuracy?

On the tasks the paper tested, yes, and by a wide margin on some of them. Table 1 of the paper compares GPT-4's zero-shot accuracy on the original questions against RaR, across ten tasks:

Those numbers describe one model (GPT-4, tested by the authors between October 1 and 30, 2023) on ten specific benchmark tasks, several of which were deliberately chosen because GPT-4 answered them badly in a prior study. Treat the exact percentages as evidence that the mechanism works, not as a promise of what any current model will do on your own questions. The paper itself frames the result the same way: "In summary of the results, RaR provides a universal, plug-and-play zero-shot prompt that allows for efficient and effective performance improvement of LLMs on general tasks."

The tasks where RaR helped most share a pattern: they hinge on the model correctly parsing what the question is actually asking (an even day versus an even month versus an even year, or which letters to concatenate from which names), not on world knowledge it lacks. That is exactly the class of failure the technique is built to catch.

What is Two-step RaR, and when does it help more than the one-step version?

The one-step version above has the same model rephrase and answer in one call. Two-step RaR splits that into two calls, which the paper introduces this way: "a rephrasing LLM first rephrases the question and then passes the original and rephrased questions together to a different responding LLM." The rephrasing prompt and the responding prompt are separate:

Rephrasing prompt:
"{question}"
Given the above question, rephrase and expand it to help you do
better answering. Maintain all information in the original question.

Responding prompt:
(original) {question}
(rephrased) {rephrased question}
Use your answer for the rephrased question to answer the original question.

The reason to bother with two calls instead of one: the rephrasing and responding models don't have to be the same model. The paper tested whether a stronger model's rephrasing could help a weaker model answer better, using GPT-4 to rephrase questions for Vicuna-13b-v1.5. On last-letter concatenation (2 names), Vicuna answered 5.45% correctly using its own self-rephrased question, versus 10.45% using GPT-4's rephrasing of the same question. On the Sports task, Vicuna went from 65.00% with its own rephrasing to 73.64% with GPT-4's. Both gains are real but modest, which matches the paper's own framing: rephrasing quality is transferable across models, but a weaker responding model still caps how much any rephrasing can help it.

Two-step RaR is the shape you want if you're already running a pipeline where a stronger, more expensive model is available for one step and a cheaper model handles volume. One-step RaR is the shape for everything else.

How is RaR different from Chain-of-Thought prompting?

They target different failure modes, and the paper is direct about this. Chain-of-Thought, or CoT, assumes the model already understood the question correctly and asks it to show its reasoning toward the answer, typically by appending "Let’s think step by step." RaR instead assumes the question itself might be ambiguous to the model, and has it restate what it thinks was asked before reasoning about anything.

The paper's own comparison names this distinction and then goes further, arguing the two aren't competitors. It opens the section plainly: ‘It is widely known that zero-shot CoT, by appending the instruction “Let’s think step by step.” to queries, can effectively improve the performance of LLMs on reasoning tasks.’ It also reports cases where CoT alone made things worse (a Chinese-idiom completion task, where the added reasoning chain increased hallucination), while RaR held up. Its conclusion: "We show that RaR is complementary to CoT and can be combined with CoT to achieve even better performance." Combining them is as simple as appending both instructions to the same prompt. For a full walkthrough of CoT on its own, see Chain-of-Thought Prompting: Examples and When to Use It.

How is RaR different from asking the model to ask you clarifying questions first?

They solve overlapping problems with an opposite mechanism. Getting ChatGPT to ask you questions first is a two-turn exchange: you ask a question, the model asks you something back, you answer, and only then does it respond. That only works when a person is present in the loop to answer.

RaR never asks you anything. The model rephrases the question itself, in its own words, and answers in the same reply, with zero additional turns. That makes RaR the right tool for anything unattended: a batch script running through a spreadsheet of questions, an API call inside a larger pipeline, or any place a two-turn clarifying exchange simply isn't possible because nobody is there to answer.

If you genuinely need the model's help articulating an ambiguous task and a human is available to weigh in, the clarifying-questions approach will get you further, because it resolves the actual ambiguity rather than the model's best guess at it. RaR is the fallback for when that conversation can't happen.

When should you actually use Rephrase and Respond?

Reach for it in a few specific situations, based on where the paper's own tested tasks clustered:

• Symbolic or lexical tasks with a precise, easy-to-misread definition — letter concatenation, date and calendar reasoning, anything where the question hinges on one specific interpretation of an ordinary word.
• Unattended pipelines — batch jobs, API calls, anywhere a clarifying question can't be answered because no human is in the loop.
• Short, single-line questions — the paper's test set skews toward short factual or symbolic prompts. A question that's already several paragraphs of context probably has less ambiguity left for a rephrase to surface.

How do you add Rephrase and Respond to your own prompts?

Take any question you'd normally ask directly, and append the one line from Section 3:

[Your actual question here]
Rephrase and expand the question, and respond.

That's the entire change. If you want to see whether the rephrase surfaced a misreading before you trust the answer, read the rephrased sentence the model produces before you read anything after it. A rephrase that quietly changes what's being asked is a signal worth acting on, whether or not the final answer looks reasonable.

If you're building a saved template rather than typing this by hand each time, Prompt Architects' generator can slot the RaR instruction in as a fixed suffix on any prompt you save, so it runs every time you reuse it without you having to remember to add it.

What are the real limits of Rephrase and Respond?

Three, honestly stated. First, it's an arXiv preprint with no confirmed peer-reviewed venue attached, tested on GPT-4, GPT-3.5 and Vicuna-13b snapshots from late 2023, not on any 2026 model. Second, it costs extra output tokens and latency on every call, since the model now writes a rephrase before it writes an answer, and the paper reports no figure for how much that adds. Third, it fixes misreadings, not ignorance: a rephrased question can still get a wrong answer if the model's actual knowledge on the subject is wrong, which is a different problem that RaR was never built to solve.

RaR shapes the question. If what you actually need is control over the shape of the answer (forcing JSON, skipping a preamble), that's a different lever entirely; see Output Priming: Start the Answer for the Model for the vendor-side mechanics of that one, and for a rare honest case where a similarly-named technique has no paper behind it at all.