Rephrase And Respond: Definition and Examples

Rephrase And Respond (RaR) is a prompting method introduced by UCLA researchers in 2023. The principle is simple: instead of letting the model answer a question directly, you first ask it to rephrase that question before providing its answer. This intermediate step forces the model to clarify ambiguities and better grasp the intent behind the query.

This technique is based on a fundamental observation: human questions are often ambiguous, imprecise, or phrased in ways that can mislead the model. By rephrasing, the LLM performs active interpretation work that reduces misunderstandings and aligns its understanding with the user's true intent. This is comparable to what a human expert naturally does when rephrasing a question before answering.

RaR exists in two main variants. The "one-step" variant incorporates the rephrasing instruction directly into the initial prompt (e.g., "Rephrase this question then answer it"). The "two-step" variant separates the process into two distinct calls: first to obtain the rephrasing, then a second that uses the rephrased question to generate the answer. The two-step variant generally yields better results because it allows the model to fully focus on each step.

Experiments show that RaR significantly improves performance on various tasks, including multiple-choice questions, logical reasoning, and math problems. A major advantage is its compatibility with other techniques like Chain-of-Thought: you can combine rephrasing and step-by-step reasoning to get even better results.