Rephrase and respond (RaR): a complete guide

You and the model rarely read a question the same way. The phrasing that's obvious to you carries assumptions the model never picks up, so it confidently answers a slightly different question. Rephrase and Respond (RaR) fixes the mismatch from the model's side: tell it to restate and expand your question in its own words, then answer that. One appended sentence lifted GPT-4 on last-letter concatenation from 52.05% to 99.09% (Deng et al., 2023).

See it work

Take a question that looks trivial to a human but trips the model on its hidden assumption.

Prompt (plain):
Was Ada Lovelace born on an even day of the month?

Output:
Ada Lovelace was born on December 10, 1815. Yes, she was a remarkable
mathematician.  ← dodges the actual question

Now append the RaR instruction and let the model reframe before it commits.

Prompt (RaR):
Was Ada Lovelace born on an even day of the month?
Rephrase and expand the question, and respond.

Output:
Rephrased: On what day-of-month was Ada Lovelace born, and is that
number (1–31) divisible by 2 with no remainder?
Answer: She was born on the 10th. 10 is even, so yes.

Same model, same facts. The only change is that the model first pinned down what "even day" means, then answered the question it actually wrote down. That self-clarification is the whole technique.

The mental model

Think of a sharp junior engineer who reads a ticket back to you before touching the code: "So you want the export to skip archived rows and keep the header, right?" Half the time, saying it out loud surfaces the ambiguity, and they fix the misread before writing a line.

RaR makes the model read your question back to itself, in its own words, so it answers what you meant instead of what it first assumed.

How it works

The flow is one extra clause in your prompt. The model expands the question, then answers the expanded version in the same pass.

1. Take the original question exactly as the user wrote it, ambiguities and all.
2. Append the rephrase instruction, e.g. "Rephrase and expand the question, and respond."
3. The model rephrases and expands the question, making implicit constraints explicit and resolving vague terms.
4. The model answers its own rephrased version in the same response, so you see both the clarified question and the answer.

Why it works

The gain comes from forcing a clarification step the model normally skips. Ranked by how much they drive the effect:

Where it shines

RaR was tested across 10 tasks spanning knowledge classification, knowledge comparison, commonsense, and symbolic reasoning. The gains are largest where the plain question hides a precise sub-question. Selected GPT-4 results, baseline vs RaR (Deng et al., 2023):

The pattern is consistent: tasks the model already nails (age comparison, CSQA) move a little; tasks where it misreads the ask (symbolic concatenation, even-day parity) jump dozens of points. RaR also helped on a fairness probe: on StereoSet it scored a 97.73% language-modeling score against 84.09% for zero-shot chain-of-thought.

When to use it (and when not)

Reach for RaR when:

• The question has a term the model could read multiple ways ("even day", "last letter", "before").
• You're seeing answers that respond to a near-miss version of what you asked.
• The task is symbolic or parity-style, where one misread flips the result.
• You can afford a longer response (the rephrase costs tokens).

Skip it when:

• The prompt is already fully explicit and self-contained.
• Latency or output length is tightly budgeted.
• You need terse, machine-parseable output with no preamble.

Model fit: the gains scale with model strength. GPT-4 saw the largest, most consistent lifts; GPT-3.5 and Vicuna improved more modestly. Weaker models sometimes rephrase poorly, which is exactly what two-step RaR is for.

Variants and alternatives

One-step vs two-step RaR

One-step is the whole technique in a single sentence. Two-step splits the work so a stronger model can do the rephrasing for a weaker one.

One-step: append the instruction and let one model do both.

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

Variations work too: "Reword and elaborate on the inquiry, then provide an answer" performs comparably, so you don't have to match the wording exactly.

Two-step: a rephrasing model first restates the question, then a responding model answers using both versions.

• Step 1 prompt to the rephraser: "rephrase and expand it to help you do better answering. Maintain all information in the original question."
• Step 2 prompt to the responder: feed both the original and rephrased question, then "Use your answer for the rephrased question to answer the original question."

The two models can be identical or different. The payoff of using a different, stronger rephraser is transfer: GPT-4's rephrasings, handed to Vicuna-13B, lifted it on several tasks (for example, Sports understanding rose from 65.00% to 73.64%, and CSQA from 51.36% to 55.00%). Higher-capability models write clearer questions, and that clarity carries over to weaker responders.

Combining RaR with chain-of-thought

RaR and chain-of-thought (CoT) fix different failures, so they stack. RaR repairs how the model reads the question; CoT repairs how it reasons through the answer. Append "let's think step by step" to the RaR prompt to get both.

CoT isn't always safe on its own: on Chinese Idiom inference, zero-shot CoT actually dropped GPT-4 below baseline (31.43% vs 32.38%) because the reasoning steps hallucinated. RaR is also more robust to bad demonstrations. When few-shot examples carried wrong reasoning, 4-shot CoT on last-letter concatenation collapsed to 52.27%, while RaR combined with the same flawed 4-shot examples recovered to 95.45%.

Implementation notes

Getting RaR into production is mostly prompt plumbing, not new infrastructure.

• One-step: concatenate your question with the rephrase instruction. No second call, no parsing required unless you want to extract the rephrased question separately.
• Two-step: make a rephrase call, capture its output, then build the responder prompt from the original plus the rephrasing. Cache rephrasings if the same questions recur.
• Parsing: if you only want the final answer, instruct the model to label sections ("Rephrased:" / "Answer:") so you can split the response reliably.

Configuration that works in practice:

Do:

• Keep the original question verbatim so no detail is lost in your own paraphrase.
• Tell the rephraser to "maintain all information" so it expands without dropping constraints.
• Use a stronger model as the rephraser when the responder is weak.

Don't:

• Pre-rephrase the question yourself; the point is the model's own framing.
• Use RaR on already-explicit prompts where it only adds tokens.
• Assume tiny models will rephrase well; verify or hand them a strong model's rephrasing.

Debugging

Limitations

• It assumes the model can clarify. Weaker models may rephrase incorrectly, baking in a wrong reading instead of fixing it.
• Gains are uneven. Tasks that are already clear barely move; the payoff concentrates on ambiguous or symbolic questions.
• It costs tokens and, for two-step, an extra call. That overhead is wasted on well-specified prompts.
• The rephrase can drift. Without "maintain all information", expansion sometimes adds assumptions or loses a detail.
• Not a reasoning fix. If the failure is multi-step logic rather than comprehension, you need CoT, not RaR.

How to prove it on your task

Run a clean A/B: the same questions with and without the appended instruction, scored on your own metric. Watch for the RaR-specific failure mode where the rephrasing changes the question's meaning rather than clarifying it. If accuracy doesn't move, your prompts were probably already explicit, and RaR is just extra cost. If it jumps on a subset, those are the ambiguous questions RaR was built for.

Summary

• RaR appends one instruction so the model restates and expands your question, then answers its own version, closing the gap between what you asked and what it understood.
• Introduced by Deng, Zhang, Chen, and Gu (UCLA, 2023, arXiv 2311.04205) as a simple, general prompting method.
• One-step does it in a single prompt; two-step uses a strong model to rephrase for a weaker responder, and the clarity transfers across models.
• Biggest gains on ambiguous and symbolic tasks: GPT-4 went from 52.05% to 99.09% on last-letter concatenation and 54.29% to 93.33% on even-day parity; easy tasks barely move.
• It's complementary to chain-of-thought; rephrase first, then think step by step, to fix comprehension and reasoning together.
• Skip it on already-explicit prompts and tight token budgets, and verify that small models rephrase faithfully before trusting them.