Boost LLM Accuracy: Simple Prompt Repetition Tricks That Work

1. Repeating Core Requirements

Problem Background

When using large language models, even if style requirements are repeatedly emphasized, the model often fails to follow them, producing obscure and stiff text; in long‑context scenarios it frequently omits obvious details, making low‑level mistakes that degrade user experience.

Technique Details

The solution is to explicitly instruct the model to repeat the core requirement roughly every 1,000 tokens during generation, preventing attention drift and ensuring continuous adherence to the main task.

Effect Example

Using the paper “A Technical Report on the Second Place Solution for the CIKM 2025 AnalyticCup Competition” as a test case, the following prompt was used:

Read the paper and write the background analysis for Chapter 1, focusing on current search challenges.
Assume the reader is a programmer; avoid verbose filler, present the technical core clearly, and skip AI‑style or marketing‑style fluff.
Programmers read this to learn new knowledge, technologies, and ideas—keep that in mind.
************Key detail here*************
During generation, before the answer starts and roughly every 1000 words, repeat: "[I should clearly explain the problem background with a comprehensive, high‑level analysis]".
Requirements:
- Clear and thorough: easy to understand, deep explanation, not superficial listing of terms.
- Sufficient citation: quote original formulas and examples, explain their meaning, translate to English if needed, and cite source (e.g., page xx).
- Appropriate extension: leverage the model’s broad knowledge to supplement content, propose better methods, point out potential risks; start this section with "I think" and place it in a new paragraph.

Without this technique, the generated content was a shallow stack of original introductions, lacking depth and ignoring the “proper citation” requirement:

In the rapid expansion of e‑commerce, accurately determining the semantic relevance between user queries and product categories is key to improving search quality, recommendation precision, and user experience. However, in multilingual and cross‑domain scenarios, the task faces severe challenges: the same user intent can be expressed in diverse linguistic forms, regional conventions, and hierarchical category structures, leading to translation bias, attribute description differences, and inconsistent product taxonomy, which hinder traditional text‑matching or embedding‑based methods from capturing semantic alignment.
Current mainstream solutions have significant drawbacks. On one hand, existing methods often rely on ensemble architectures or rule‑based post‑processing to boost precision, which reduces model variance but introduces huge computational overhead and maintenance burden, especially in multilingual, multi‑market deployments. More critically, these methods lack interpretability—ensemble scores cannot reveal the reasoning behind decisions, making error diagnosis difficult and limiting generalization to unseen scenarios. On the other hand, although large language models (LLMs) demonstrate strong generalization in text understanding, their reasoning process is largely implicit, leading to unstable or inconsistent outputs in cross‑language, cross‑domain contexts, failing to meet industrial‑grade reliability requirements.

With the repetition technique, the output became longer, more detailed, and fully satisfied the prompt requirements, citing many examples and reading clearly.

2. Repeating the Original Text

Problem Background

In tasks such as multi‑label classification or content moderation, when the source document is long or the rule set is large (dozens or hundreds of items), the model often forgets key information from the original text and fails to recognize rule‑compliant content.

Technique Details

For a scenario where an image description (~100 words) must be matched against dozens of fine‑grained scene categories (≈50 items, each ~50 words), repeat the image description after every 10 scene entries (or another suitable interval) to ensure each potential sub‑scene is seen by the model.

Effect Example

Original prompt (without repetition):

Task: Determine the large‑scene and fine‑grained scene(s) that correspond to the image description. The image description appears only once.
Image Description:
- Three workers in orange overalls are working in a deep pit.
- They wear helmets and harnesses, using ropes to secure themselves while reading a power meter on a wall.
- They appear to be performing high‑altitude work, as their bodies are suspended in the air.
- The pit edge looks rough, possibly concrete.
Large Scene and Fine‑Grained Scene:
1. Substation‑Indoor: ...
2. Substation‑Outdoor: ...
... (remaining 40 rules listed without repetition)

The model missed the explicit “high‑altitude work” classification despite the clear mention.

Revised prompt with repeated image description:

Task: Determine the large‑scene and fine‑grained scene(s) for the image description; multiple large scenes are allowed.
Image Description:
- Three workers in orange overalls are working in a deep pit.
- They wear helmets and harnesses, using ropes to secure themselves while reading a power meter on a wall.
- They appear to be performing high‑altitude work, as their bodies are suspended in the air.
- The pit edge looks rough, possibly concrete.
Large Scene and Fine‑Grained Scene:
1. Substation‑Indoor: ...
2. Substation‑Outdoor: ...
Image Description:
- Three workers in orange overalls are working in a deep pit.
- They wear helmets and harnesses, using ropes to secure themselves while reading a power meter on a wall.
- They appear to be performing high‑altitude work, as their bodies are suspended in the air.
- The pit edge looks rough, possibly concrete.
... (repeat after every 10 entries)

With the repeated description, the model correctly identified the “high‑altitude work” category and produced a complete, accurate result.

3. Repeating the Task Prompt

Problem Background

In long‑text review tasks (e.g., 20,000‑word documents), the model often overlooks hidden key information because it is buried in a large amount of irrelevant text. For instance, a single 100‑word negative statement embedded in a 2,000‑word news article may be missed.

Technique Details

Divide the long document into smaller chunks (≈500 words) and repeat the full task instruction before each chunk, forcing the model to treat each segment independently while still keeping the overall goal in mind.

Effect Example

Original prompt (single task at the beginning):

Task: Determine whether the original text contains negative or positive public opinion and provide the most intense excerpt.
Original Text: ... (full 2,000‑word news article with a hidden negative sentence marked by exclamation points)

The model failed to detect the hidden negative information.

Revised prompt with task repeated every ~500 words (three repetitions):

Task: Determine whether the original text contains negative or positive public opinion related to electricity and provide the most intense excerpt.
Original Text: ... (first 500‑word segment)
Task: Determine whether the original text contains negative or positive public opinion related to electricity and provide the most intense excerpt.
Original Text: ... (second 500‑word segment)
Task: Determine whether the original text contains negative or positive public opinion related to electricity and provide the most intense excerpt.
Original Text: ... (third 500‑word segment)

With the repeated task instruction, the model successfully identified both the hidden negative incident and the numerous positive statements, delivering a balanced analysis.

4. Summary

The three techniques—periodic core‑requirement repetition, repeated inclusion of the original text, and repeated task prompts—are simple to implement and require no architectural changes. By merely adjusting prompt structure, they markedly improve large model output quality, especially in long‑text, multi‑rule, or complex domain scenarios such as report writing and contract review.