RE2 (Re-Reading)
The simplest technique that shouldn’t work — but does. Include the question twice in your prompt and watch accuracy climb. RE2 forces the model to process your question on a second pass, catching nuances, constraints, and details that a single reading misses.
Introduced: RE2 (Re-Reading) was published in 2023 by Xu et al. The researchers discovered that a deceptively simple input modification — repeating the question with a “Read the question again:” prefix — produced consistent accuracy improvements across multiple reasoning benchmarks. Unlike most prompting techniques that modify the output side (how the model responds), RE2 operates entirely on the input side, changing how the model processes and attends to the question itself.
Modern LLM Status: RE2 remains a still active technique in prompt engineering, particularly valuable for complex multi-constraint questions where models frequently miss buried details. The technique is especially effective with questions containing multiple conditions, subtle negations, or layered requirements. Its zero-cost simplicity — requiring no special formatting, no examples, and no additional instructions — makes it one of the most accessible comprehension-enhancing methods available.
Read It Again, Understand It Better
When humans encounter a complex passage — a legal clause, a multi-step math problem, or an intricate set of instructions — the natural instinct is to read it again. The second reading isn’t redundant. It builds on the comprehension scaffold from the first pass, allowing deeper processing of relationships, constraints, and implications that were only partially absorbed initially.
Language models benefit from the same pattern. When a question appears twice in the input, the model’s attention mechanism processes it with different contextual weights on the second encounter. The first reading establishes the general topic and structure; the second reading — now informed by that initial understanding — catches the specific details, edge conditions, and nuances that determine whether the answer is correct or merely plausible.
Think of it like proofreading: you almost never catch every error on the first read. The second pass, armed with the gist from the first, focuses on what matters most. RE2 gives the model that same second chance at comprehension.
Reinforced Attention: The repeated question receives stronger attention weights in the model’s transformer layers, ensuring critical details aren’t diluted by surrounding context.
Constraint Recognition: On the second pass, the model is better positioned to identify all constraints simultaneously rather than processing them sequentially and losing track of earlier ones.
Detail Anchoring: Specific numbers, negations, and qualifying phrases that might be glossed over in a single read become anchored reference points when encountered a second time.
The RE2 Process
Three steps to deeper question comprehension
Present the Original Question
Start with the question exactly as you would normally ask it. This serves as the model’s first encounter with the problem — establishing the topic, scope, and general structure. No special formatting is needed; just write the question naturally with all its constraints and details intact.
“A store sells apples in bags of 6 and oranges in bags of 8. If Maria wants to buy exactly 30 fruits with at least 2 bags of each type, and she wants more oranges than apples, how many bags of each should she buy?”
Add the Re-Reading Directive
After the original question, insert the phrase “Read the question again:” followed by a complete repetition of the question. This triggers a second processing pass where the model’s attention mechanism re-engages with every constraint, number, and condition — this time with the benefit of the initial comprehension context.
“Read the question again: A store sells apples in bags of 6 and oranges in bags of 8. If Maria wants to buy exactly 30 fruits with at least 2 bags of each type, and she wants more oranges than apples, how many bags of each should she buy?”
Request the Answer
The model now generates its response with a more thorough understanding of the question. Having processed all constraints twice, it is significantly less likely to overlook a condition, misinterpret a negation, or produce an answer that satisfies most — but not all — of the requirements. The result is measurably higher accuracy, especially on multi-constraint problems.
The model explicitly lists all four constraints (exactly 30 fruits, at least 2 bags each, bags of 6 and 8, more oranges than apples) and systematically tests combinations against every condition before committing to an answer — rather than latching onto the first combination that totals 30.
See the Difference
How a second reading catches what the first one misses
Single-Read Prompt
A company has 120 employees. They need to form teams of either 5 or 8 people. Every employee must be on exactly one team. They need at least 3 teams of each size, and the number of small teams must exceed the number of large teams. How many teams of each size should they form?
Let me solve this. If we use 8 teams of 5 (40 people) and 10 teams of 8 (80 people), that’s 120 people total. So 8 small teams and 10 large teams.
[Misses the constraint that small teams must exceed large teams — 8 is less than 10]
RE2 Prompt
[Same question] Read the question again: [Same question repeated]
Let me list all constraints: (1) 120 total, (2) teams of 5 or 8, (3) every employee on one team, (4) at least 3 of each size, (5) more small teams than large teams. Testing: 16 teams of 5 (80) + 5 teams of 8 (40) = 120. Check: 16 > 5 (small exceeds large), both above 3 minimum. Valid solution.
Practice Responsible AI
Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.
48 US states now require AI transparency in key areas. Critical thinking remains your strongest tool against misinformation.
RE2 in Action
See how re-reading improves accuracy across different question types
“A baker makes cookies in batches of 12 and brownies in batches of 9. She needs at least 100 items for a bake sale, but no more than 120. She must make at least 3 batches of cookies and at least 2 batches of brownies, and she wants more cookies than brownies. How many batches of each should she make?”
“A baker makes cookies in batches of 12 and brownies in batches of 9. She needs at least 100 items for a bake sale, but no more than 120. She must make at least 3 batches of cookies and at least 2 batches of brownies, and she wants more cookies than brownies. How many batches of each should she make?
Read the question again: A baker makes cookies in batches of 12 and brownies in batches of 9. She needs at least 100 items for a bake sale, but no more than 120. She must make at least 3 batches of cookies and at least 2 batches of brownies, and she wants more cookies than brownies. How many batches of each should she make?”
Result: The model systematically enumerates constraints — batch sizes, range (100–120), minimums per type, and the comparison requirement — then tests combinations. Without RE2, models frequently satisfy the total range but overlook the “more cookies than brownies” condition.
“In a survey of 200 employees, everyone speaks at least one language. 120 speak English, 90 speak Spanish, and 50 speak French. 40 speak both English and Spanish, 20 speak both English and French, and 15 speak both Spanish and French. How many employees speak all three languages?”
Add “Read the question again:” followed by the complete question repeated verbatim.
Result: The re-read helps the model carefully track every number in the inclusion-exclusion formula. On a single read, models often transpose the “both” values or confuse which pairs overlap. The second pass anchors each specific number to its correct pair, leading to the accurate application: 200 = 120 + 90 + 50 − 40 − 20 − 15 + x, so x = 15 employees speak all three.
“Five friends — Alex, Blake, Casey, Drew, and Ellis — sit in a row of five chairs. Alex cannot sit next to Blake. Casey must sit in one of the two end chairs. Drew must sit somewhere to the left of Ellis but not immediately next to Ellis. How many valid seating arrangements exist?”
Add “Read the question again:” followed by the complete question repeated verbatim.
Result: The hidden complexity here is the “to the left of Ellis but NOT immediately next to Ellis” constraint — requiring at least one chair between Drew and Ellis. On a single read, models often process “to the left of” but skip the “not immediately next to” qualifier. With RE2, the model reliably catches both parts of the compound constraint, correctly reducing the solution space and producing the accurate count.
When to Use RE2
Maximum impact for complex questions, minimal overhead
Perfect For
Math problems, logic puzzles, and scenarios with three or more simultaneous constraints — exactly where single-read comprehension breaks down most often.
When important conditions are embedded in the middle of a lengthy prompt, re-reading ensures they receive proper attention rather than being overshadowed by the beginning and end.
Tasks requiring exact adherence to formatting rules, numerical limits, or specific output structures — the second pass reinforces every requirement.
SAT, GRE, and similar exam questions often contain carefully worded traps and qualifiers that benefit significantly from a second processing pass.
Skip It When
“What is the capital of France?” has no hidden constraints or subtle details — repeating it wastes tokens without improving accuracy.
RE2 doubles the question length. When operating near token limits or paying per-token costs on long prompts, the overhead may not justify the accuracy gain.
If your prompt is short, well-structured, and contains a single clear constraint, the model already processes it thoroughly on the first read — repetition adds no value.
Use Cases
Where RE2’s second-pass comprehension delivers the most value
Exam Preparation
Practice standardized test questions with higher accuracy by ensuring the model catches every qualifier, negation, and condition in complex exam problems.
Contract Review
When asking AI to analyze contract clauses with multiple conditions, RE2 ensures every obligation, exception, and deadline is identified rather than just the most prominent terms.
Requirements Analysis
Software requirements often contain interrelated constraints. RE2 helps the model catch dependency chains and conflicting requirements that a single pass might overlook.
Data Analysis Queries
Complex analytical questions with multiple filters, groupings, and conditions benefit from re-reading to ensure every data constraint makes it into the query or analysis plan.
Compliance Checking
Regulatory questions with layered conditions, exemptions, and thresholds are prime candidates for RE2 — missing a single qualifier can mean the difference between compliant and non-compliant.
Technical Specifications
When asking AI to generate or validate against technical specs with precise tolerances, ranges, and interdependent parameters, the second read prevents specification drift.
Where RE2 Fits
RE2 anchors the input-side comprehension family of techniques
RE2 is the lightest-weight technique in the comprehension family. While RaR asks the model to rephrase (adding output overhead), S2A requires context filtering, and Self-Ask demands decomposition — RE2 achieves its gains through pure repetition with zero additional instructions. This makes it the ideal first-line technique: try RE2 first, and escalate to more complex methods only if accuracy remains insufficient.
Related Techniques
Explore complementary approaches to comprehension and accuracy
Read It Twice
Try adding RE2’s re-reading technique to your next complex question, or explore other comprehension-enhancing methods in the Praxis Library.