rephrase and respond
**Rephrase and Respond (RaR)** is the **prompting strategy that instructs a language model to first restate the input question in its own words before generating an answer — forcing deeper comprehension of the query intent, resolving ambiguities, and activating more relevant knowledge pathways** — a deceptively simple technique that consistently improves performance on knowledge-intensive and reasoning tasks with near-zero implementation overhead.
**What Is Rephrase and Respond?**
- **Definition**: A two-step prompting approach where the model first rephrases or restates the question, then answers the rephrased version — separating comprehension from response generation.
- **Mechanism**: Rephrasing forces the model to process the question semantically rather than pattern-matching on surface features — similar to how humans benefit from restating a problem before solving it.
- **Self-Clarification**: Ambiguous, abbreviated, or poorly formulated questions are expanded and clarified during rephrasing — the model interprets what was meant rather than what was literally asked.
- **Knowledge Activation**: The rephrasing step often introduces domain-specific terminology present in the model's training data, improving retrieval of relevant knowledge during the answer generation phase.
**Why RaR Matters**
- **Consistent Improvement**: Studies show 3–10% accuracy gains across knowledge QA, commonsense reasoning, and mathematical reasoning benchmarks — meaningful gains for zero overhead.
- **Zero Computational Cost**: No additional model calls, retrieval, or multi-step orchestration — just a single instruction to rephrase before answering.
- **Handles Ambiguity**: Real-world queries are often vague, abbreviated, or context-dependent — rephrasing resolves ambiguities that would otherwise cause incorrect responses.
- **Combines With Other Techniques**: RaR stacks effectively with Chain-of-Thought, few-shot examples, and other prompting strategies without interference.
- **Reduces Misinterpretation**: 15–25% of LLM errors on benchmark tasks stem from misunderstanding the question rather than lacking knowledge — rephrasing directly addresses this error source.
**RaR Implementation**
**Basic RaR Prompt**:
- Append to system or user prompt: "Before answering, first rephrase the question in your own words to ensure you understand it correctly, then provide your answer."
- The model outputs a rephrased question followed by the answer — the rephrasing serves as an implicit comprehension check.
**Constrained Rephrasing**:
- "Rephrase the question identifying all key entities, constraints, and the specific information requested, then answer."
- More structured rephrasing produces more thorough comprehension — particularly effective for multi-part questions.
**Decomposition Variant**:
- "Restate this question by breaking it into its component sub-questions, then answer each sub-question before synthesizing a final answer."
- Extends RaR into implicit question decomposition — combines comprehension benefits with structured reasoning.
**RaR Performance Across Task Types**
| Task Category | Baseline Accuracy | With RaR | Improvement |
|---------------|------------------|----------|-------------|
| **Knowledge QA** | 74.2% | 79.8% | +5.6% |
| **Commonsense Reasoning** | 68.5% | 73.1% | +4.6% |
| **Math Word Problems** | 61.3% | 67.9% | +6.6% |
| **Ambiguous Queries** | 52.1% | 64.3% | +12.2% |
| **Multi-Part Questions** | 58.7% | 66.4% | +7.7% |
Rephrase and Respond is **the highest-ROI prompting technique available** — delivering measurable accuracy improvements across diverse tasks through the simple act of asking the model to understand the question before answering it, embodying the principle that comprehension precedes correct reasoning in both human and artificial intelligence.