Integrated Prompting: Solve a Similar Problem First, Then Yours

Everything in ONE prompt:

STEP 1: The model invents a similar problem and solves it
        → we see on-topic reasoning

STEP 2: The model immediately solves your problem
        → using the warmed-up thinking from step 1

Both steps are one request. Don't copy the example into a separate prompt. That's a different method, and it doesn't work.

Task: Investor Mark Davis is listening to a pitch: "We want to launch a subscription service for legal templates aimed at freelancers and small businesses in the US, $19/month." He needs a quick read: does this business model have structural weaknesses that aren't obvious at first glance?

Prompt:

First, analyze a similar case: a subscription service for HR document templates 
for small businesses, $15/month. What structural weaknesses does a 
business model like this have? Go step by step: unit economics, churn, 
competition, barriers.

Now use the same approach to analyze my case: a subscription service 
for legal templates for freelancers and small businesses, $19/month. 
What could go wrong here?

Result: The model runs two rounds of analysis back to back. In the first, it builds a line of reasoning about the similar business model: churn, CAC, barriers to entry. In the second, it applies that same structure to your case, but now it's "warmed up" with ready-made categories of analysis. The answer to your problem will be more structured and deeper than if you'd asked directly.

An LLM doesn't think linearly: it generates token after token, and every next token depends on all the preceding context. When the prompt contains only the question, the model takes the shortcut to the answer.

When the model first solves a similar problem, it does something important: it builds a reasoning pattern in context, right now. That freshly built pattern sits "close by" in memory, and the model actively leans on it. Attention analysis showed that in Integrated mode the model looks at its own example 2–3 times more often than in Decoupled mode.

Why doesn't a ready-made example (Decoupled) give the same effect? Because a ready-made example is someone else's text in the context. It can be ignored. But an example you just created and solved is your own recent reasoning. The model relies on it automatically.

Control levers: - Add "intent": say explicitly why the example is being created: "Come up with a similar problem so you can better understand the structure of mine". This boosts accuracy even further - Make the example harder: if your task is big, ask for a more complex analog, not a simplified one - Specify the domain: the closer the similar problem is in topic, the closer the warm-up is to your actual task

First, come up with a problem similar to mine on the topic "{topic}" and solve it 
step by step. Show your reasoning.

Now use the same approach and solve my problem:
{your problem}

Placeholders: - {topic}: the field or type of task, e.g. "risk analysis", "building an argument", "reviewing contract terms" - {your problem}: your specific question or situation, in full

🚀 Quick start: paste this into the chat:

Here is an Integrated Prompting template. Adapt it to my task: [your task].
Ask me questions to fill in the fields.

[paste the template above]

The LLM will ask about the topic and the specific task, because it needs to know what similar problem to generate as a warm-up. It will take the pattern from the template and adapt it to your context.

⚠️ Works worse on fuzzy tasks: If the task is subjective ("make it sound nice", "come up with something interesting"), the warm-up will be fuzzy too. The method shines where there's structure: analysis, arguments, calculations, plans.

⚠️ Chain-of-Thought partly overlaps the effect: If you're already asking the model to "think step by step" (CoT), the gap between Integrated and a regular prompt narrows. The method still adds a boost, just a smaller one.

⚠️ The original task must be non-trivial: On simple questions the effect is minimal. The harder the task, the bigger the difference.

⚠️ More tokens: You add a whole block of similar-problem generation. Fine for short sessions; keep it in mind for very long contexts.

The idea was simple: figure out what actually helps, the example as text or the process of creating it. A team from KAIST compared three prompting variants on math problems (MATH and GSM8K): no examples, an example created on the spot (Integrated), and the same example pasted in separately (Decoupled). If Integrated = Decoupled, the example is what helps. If Integrated > Decoupled ≈ Zero-shot, the process is what helps.

The result was unambiguous: five different models, two datasets, the same picture everywhere. GPT-4O with Integrated scored 54.6% on GSM8K versus 50.5% for Decoupled and 49.9% for Zero-shot. In other words, Decoupled is practically no better than nothing, even though it uses the exact same example.

The researchers went further and swapped the self-generated example for random characters ("dummy"). Even with garbage content, Integrated edged out Zero-shot slightly, meaning part of the effect comes from the simple act of generating text before the task. But a meaningful example gives a noticeably bigger boost. They also tried explicitly repeating the example right in the prompt before the task (Repeat): better than Decoupled, but worse than Integrated. Where the example sits matters, but the process of creating it matters more.

The most surprising part was the attention analysis on the open LLaMA model. It turned out that with Integrated, the model actively looks at its own example in the lower layers (0–12, where lexical and syntactic processing happens) and in the upper layers (23–31, where decisions get made). In the middle layers the difference disappears. The upper layers are the ones responsible for the final reasoning, and that's exactly where Integrated loads the example as an anchor point.

The authors moved the full prompts to Appendix B of the original paper (not included in the publication). Based on the description in the text, the Integrated structure looks like this:

[User]: Generate a related math problem about {topic} and solve it step by step.
Then solve the following test question: {test question}

[Assistant]: [generates related problem] → [solves it] → [solves test question]

Context: The authors tested on competition math problems (MATH) and grade-school arithmetic word problems (GSM8K). The key point: Integrated = one request, both steps in a single model response.

💡 Adaptation: add generation "intent"

From Table 5 of the paper: if you explicitly say why the similar example is being created, accuracy jumps sharply. Without intent, Integrated gives 35.19%; with intent, 45.50% (+10 pp on the same dataset).

Come up with a problem similar to mine on the topic "{topic}", so you can understand 
the structure of a similar problem and work out a method for solving it. Solve it 
step by step and show your reasoning.

Now, building on that method, solve my problem:
{your problem}

🔧 Technique: several similar problems → a deeper warm-up

From the multi-shot table: with 2 examples in Integrated, accuracy goes up (35.19% → 38.93%). With 4, it's also above baseline (38.24%). With Decoupled it's the opposite: adding examples doesn't help.

Come up with two problems similar to mine on the topic "{topic}". 
Solve each one step by step.

Now solve my problem:
{your problem}

When to use it: the task is especially hard or unusual, and a single warm-up may not be enough.

🔧 Technique: Integrated + Chain-of-Thought

The authors tested CoT separately: combined with Integrated, the boost holds up, though it's smaller. You can stack them:

Come up with a problem similar to mine on the topic "{topic}". 
Solve it step by step, explaining each decision.

Now solve my problem using the same method, step by step:
{your problem}

Paper: Not the Example, but the Process: How Self-Generated Examples Enhance LLM Reasoning (2025)

Authors: Daehoon Gwak, Minseo Jung, Junwoo Park, Minho Park, ChaeHun Park, Junha Hyung, Jaegul Choo

Organizations: KAIST AI, Applied Artificial Intelligence — Sungkyunkwan University

Contact: [email protected], [email protected]