Key Aspects of the Study:

The study looks at how best to write a prompt for an LLM when your request is in one language (say, Spanish) and the information you need to work with is in another (say, English). The authors tested different strategies: keep everything in English, translate everything into the user's language, or translate only certain parts of the prompt (instructions, role, answer options).

Key Result: The best performance often comes from a hybrid approach: the instructions for the LLM are given in the user's language, while the main content and the data to analyze stay in the original language (English).

The Core of the Method, Explained:

The core of the method is "splitting language duties" within a single prompt to get the most out of an LLM's strengths. Most powerful models were trained predominantly on English data. Because of that, they "think", analyze and generate best in English. But they're also trained well enough to understand instructions in many other languages.

The practical technique that follows from the study looks like this:

1. Identify the "task language" and the "content language". For example, you want a summary of an English article in Spanish. The "task language" is Spanish; the "content language" is English.

2. Write the instructions in the "task language". Clearly and thoroughly describe the role, the task, the steps and the required answer format in your native language (Spanish). This helps the model understand exactly what's expected of it.

3. Provide the main content in the "content language". Paste the article text, the data to analyze or the list of options in their original language (English). This lets the model work with the information without the losses that machine translation inevitably introduces, and bring its strongest analytical abilities to bear.

4. State the output language clearly. In your "task language" instructions, spell out that the final answer must be in that language.

This "hybrid" prompt works better than trying to translate everything into one language because it minimizes translation errors and lets the LLM both understand your request precisely and process the source data well.

Practical Applicability:

• Direct applicability: Very high. Any user can start using this approach immediately. For example, when analyzing English-language reviews to prepare a report in Spanish, you can give the instructions in Spanish and paste the reviews into the prompt in English. No technical skills required.

• Conceptual value: Huge. The study helps you see that an LLM being "multilingual" doesn't mean it performs equally well in every language. It gives users a useful "mental model": an LLM can be a "polyglot listener" (understanding instructions in many languages) while still being an "English-speaking thinker" (analyzing and processing information best in English).

• Adaptation potential: The method adapts easily to a wide range of tasks beyond classification.

  • Summarization: Give the instruction "Write a short summary in Spanish" and paste the English text.
  • Data extraction: Give the instruction "Find all the revenue figures in this English report and present them as a table in Spanish" and paste the report.
  • Creative tasks: Give the instruction "Write a social media post in Spanish using the key ideas from this English article about trends" and paste the article. The adaptation mechanism is simple: separate the instruction (what to do) from the context (what to work with) and use the most suitable language for each part.

A Practical Example:

**Role:** You are an experienced HR analyst preparing a report for leadership who don't speak English. (In practice, write this instruction block in your own language, e.g., Spanish.)

**Task:**
Your task is to analyze anonymous exit feedback from former employees, written in English. Based on this analysis, identify the 3 main reasons people quit and briefly describe them in **Spanish**.

**Context for analysis (reviews in English):**

1. "The salary was stagnant for three years. No matter how well I performed, there was no path for financial growth. My manager was supportive, but company policy tied their hands."
2. "I felt my career hit a ceiling. There were no opportunities for promotion or to learn new skills. The work became monotonous, and I saw no future for myself here."
3. "Work-life balance is non-existent. Constant pressure to work late and on weekends without extra compensation. It led to a complete burnout."
4. "While my direct team was great, the cross-departmental collaboration was a nightmare. The overall company culture feels toxic and overly political."
5. "The compensation is significantly below market average. I got an offer from a competitor for 40% more for a similar role."


**Answer format instructions:**
1.  Present the result as a numbered list in **Spanish**.
2.  For each reason, give a short heading (2-3 words).
3.  After the heading, give a brief explanation (1-2 sentences).
4.  Don't add anything extra, just the list of reasons.

Why It Works:

This prompt is effective because it uses the hybrid language approach described in the study:

• Clear instructions in the user's language: The role, task and output format are given in the user's language. The model understands the goal precisely ("identify 3 reasons people quit", "report for leadership", "answer in Spanish"), which matches the strategy of translating the instructions (Rules) and identity (Identity) into the user's language.
• Analyzing the data in the original: The employee feedback (<REVIEWS_EN>) is provided in English. This lets the LLM analyze the source text without losing the nuance and meaning that machine translation could strip away. The model uses its most developed English-processing abilities to find the essence of the feedback.
• Separating the tasks: The prompt cleanly separates "understanding the task" (in the user's language) from "analyzing the content" (in English), making the model first form a plan of action and then apply it to undistorted data.

Another Practical Example

**Role:** You are a food blogger. You help your audience adapt interesting recipes from abroad. (In practice, write this instruction block in your own language, e.g., Spanish.)

**Task:**
Analyze the English-language tomato soup recipe. Based on it, write step-by-step cooking instructions in **Spanish**. Pay special attention to tips that will help beginners.

**Context for analysis (recipe in English):**

Title: Creamy Tomato Basil Soup

Ingredients:
- 2 tbsp olive oil
- 1 large yellow onion, chopped
- 4 cloves garlic, minced
- 2 cans (28 oz each) crushed tomatoes
- 1 cup vegetable broth
- 1 cup heavy cream
- 1/2 cup fresh basil leaves, chopped
- Salt and black pepper to taste
- Optional: 1 tsp sugar to cut acidity

Instructions:
1. Heat olive oil in a large pot over medium heat. Add onion and cook until soft, about 5-7 minutes. Add garlic and cook for another minute until fragrant.
2. Pour in the crushed tomatoes and vegetable broth. Bring to a simmer, then reduce heat and let it cook for 15 minutes.
3. Use an immersion blender to blend the soup until smooth. Alternatively, carefully transfer in batches to a regular blender.
4. Stir in the heavy cream and fresh basil. Season with salt and pepper. Add sugar if needed. Serve hot.


**Answer format instructions:**
1.  Write the title in **Spanish**.
2.  Create two sections: "Ingredients" and "Preparation".
3.  In the "Preparation" section, describe each step in detail, clearly enough for a beginner cook.
4.  Add 1-2 "Chef's Secrets" based on the recipe (for example, about the sugar or using the blender).
5.  The entire answer must be in **Spanish**.

Why This Example Works, Explained.

This example works on the same key principle as the previous one, illustrating the study's findings:

• A localized task: Instructions in the user's language ("food blogger", "adapt the recipe", "tips for beginners") let the model understand the end goal and the audience precisely. The model doesn't just translate; it carries out a creative adaptation task.
• An undistorted source: The English recipe (<RECIPE_EN>) serves as an accurate, reliable source of information. The model analyzes the original ingredients and steps, which prevents errors that could arise from translating it first (for example, "heavy cream" could be translated inaccurately).
• Synthesizing information: The model performs a complex process: it reads the instructions in the user's language to understand what to do, then analyzes the English text to understand how to do it, and finally synthesizes new content (an adapted recipe with tips) in Spanish, as the task required. It's a perfect example of the paper's "I&R" (Identity & Rules) strategy in action.

Usefulness Score: 85