Ask a friend "was Tom Cruise born in an even month?" and they answer without blinking. They quietly decided you meant an even-numbered month (February, April, and so on), not a month with an even number of days. That silent disambiguation is so automatic you never notice it happening.
A language model doesn't get to do that. It can't stop and ask what you meant. It just continues the most likely text, which means it picks one reading of your question — often the most common one in its training data, not the one you had in mind — and answers with total confidence. There's no signal that a choice was even made. The model isn't missing knowledge; it's answering the wrong question.
Rephrase-and-Respond (RaR) fixes this with almost no effort. Before answering, you have the model restate and expand your question in its own words, and only then answer that restatement. That one extra move drags the ambiguity into the open and forces the model to commit to a single interpretation on the record.
Watch it flip
Take that Tom Cruise question. Answer it directly and the model shrugs out a confident "Yes" — it never pinned down what "even month" means, and it's wrong. Now add RaR:
Restating your question: you're asking whether the numeric position of Tom Cruise's birth month (Jan = 1 … Dec = 12) is even. He was born July 3, 1962. July is the 7th month.
Answer: No — 7 is odd.
Same model, same facts it knew all along. The only thing that changed is that it said out loud which question it was answering before answering. The restatement lands in the context the answer is generated from, so every following token is conditioned on "the numeric month is 7." It can't quietly drift onto a different reading halfway through.
The same trick rescues a pile of everyday ambiguities:
- Bare dates. "Is 02/03 before or after Valentine's Day?" Month-first it's February 3 (before); day-first it's March 2 (after). Answer directly and the model silently assumes one format. RaR surfaces both and says: it depends on the format.
- Missing units. "My oven is set to 200 — hot enough for bread?" 200 C is nearly ideal; 200 F (about 93 C) is far too cool. The direct answer guesses a scale. RaR states the assumption explicitly.
- Self-referential prompts. "How many words are in your response?" The direct answer blurts a number that doesn't match its own reply. RaR makes the model fix one sentence, then count it.
One-step vs two-step
There are two ways to run RaR.
One-step folds everything into a single prompt: "Rephrase and expand this question in your own words, making any ambiguous part explicit, then answer it." The model emits the restatement and the answer in one response. Cheapest and simplest.
Two-step uses two calls. The first returns only the rephrased question; the second answers that clean, disambiguated version. It costs an extra round-trip, but the rephrasing becomes an inspectable artifact — you can log it, validate it, show it to the user for confirmation, or route on it. Both variants reach the same corrected answer; you're trading a call for control.
It isn't Chain-of-Thought
RaR looks like Chain-of-Thought because both add text before the answer, but they guard opposite ends of the problem. CoT reasons toward the answer — it breaks the solution into steps once the question is understood. RaR clarifies the question — it fixes which problem is being solved before any reasoning starts. A flawless chain of reasoning about the wrong interpretation is still wrong.
Because they fix different failures, they stack into a strong default: rephrase first, reason second. Step 1 restates and disambiguates; step 2 walks through the solution for that pinned-down version; step 3 answers. Doing the rephrase first anchors the whole chain to the correct reading from its opening token.
The cost
RaR isn't free. Every answer now carries a restatement you didn't ask for, so you pay extra tokens and latency, and in a chat UI the rephrasing can read as padding — hide it or cap it at one line. Watch for a subtler failure too: a rephrase that quietly narrows or shifts the question, so the model then answers something slightly different, precisely. Guard against it by checking the restatement still contains the question's key terms, and fall back to a direct answer if it drifted.
The rule of thumb: reach for RaR when there's more than one way to read the question — ambiguous phrasing, dates and times, numbers without units, dangling pronouns. On crisp, single-reading questions it just adds words.
I built an interactive demo — pick an ambiguous question and watch the direct answer misread it while RaR restates and gets it right, with a toggle for one-step vs two-step:
https://dev48v.infy.uk/prompt/day28-rephrase-and-respond.html
Top comments (0)