Best approach to prompt in different languages

Course Q&A AI Prompting for Everyone
,
1

Context: Prompting using web search for the same question in 2 different languages (French and English). When I compared the results I was surprised to see a lot of similarities and a lot of differences in the 2 answers. If I understand correctly, the AI is using French content it was trained on and French web search to answer the French question and English content it was trained on and English web search to answer the English question, i.e. it does not translate content ? The similarities come from web content that was in both languages ? Interestingly, one answer was not better than the other (I was under the impression that English would be better since there are more English contents than French on the web) → Actually, I would use some parts of each answer. What would be the best prompting approach for multilingual people using AI ?

1 Answer

1
2

Hi Dufour_Eric,

Great question.

Using feedback from an AI model may be considered controversial on the deeplearning.ai platform, but in this case I will use this approach nonetheless:

I prompted Gemini 3.5 Flash (the model behind google search) with the following prompt:

Do academic studies exist that study the influence of different languages on the storage of knowledge in multilingual models?

Yes, an extensive body of academic research explicitly investigates how different languages influence the storage, retrieval, and sharing of knowledge within multilingual Large Language Models (LLMs).
Historically, mechanistic interpretability (studying the inner workings of AI) focused almost entirely on English. However, computer scientists have since shifted their focus to how models store factual knowledge across diverse languages, uncovering a complex interplay between shared universal spaces and language-specific silos. [1]
The prevailing research is structured around several major architectural and linguistic themes:

1. The “English as an Intermediate Thought” Hypothesis

A landmark study titled “How do Large Language Models Handle Multilingualism?” map a distinct three-stage multilingual workflow (MWork) inside LLMs: [2, 3]

  • Input Layer: The model takes a query in a non-English language and maps its tokens.
  • Intermediate Layers: The model rapidly translates or converges the concept into English subspaces to perform the “thinking” and reasoning steps. Factual and relational knowledge is largely processed here in an English-centric semantic space. [3, 4]
  • Output Layer: In the final layers, the model translates the resolved English concept back into the original language requested by the user. [3]

2. Language-Agnostic vs. Language-Dependent Components

Researchers have tried to find exactly where facts are stored by locating “knowledge neurons.”

  • Shared Knowledge Space: In studies like “Qualifying Knowledge and Knowledge Sharing in Multilingual Pre-trained Language Models”, researchers found a significant neuron-level overlap when models recall the same fact across up to 10 different languages. This points to a partially unified, language-agnostic storage mechanism. [5]
  • The Division of Labor: The study “How Do Multilingual Language Models Remember Facts?” used activation patching to prove that when recalling a fact (e.g., “The capital of France is Paris”), the model processes the subject (“France”) in a language-independent manner. However, extracting the specific object (“Paris”) remains language-dependent and relies on specific structural vectors that track the target language. [1]

3. Script Families and the “Curse of Multilinguality”

The script (alphabet) and the volume of training data profoundly dictate how effectively knowledge transfers between languages:

  • The Script Barrier: Research evaluating cross-lingual knowledge representation sharing notes that models show high cross-lingual consistency mainly within the same script family (e.g., transferring knowledge smoothly between different languages that use the Latin script). [6]
  • Catastrophic Forgetting: Studies examining knowledge loss have demonstrated that when a model continuously learns new information, languages utilizing non-Latin scripts are significantly more susceptible to forgetting previously stored knowledge compared to Latin-script languages. [7]
  • The Crosslingual Knowledge Barrier: In papers like Google’s “Crosslingual Capabilities and Knowledge Barriers in Multilingual Language Models,” experts noted that while models possess strong surface-level translation skills, they exhibit strict barriers when trying to apply factual knowledge learned in a high-resource language (like English) to tasks prompted in low-resource languages. [8]

4. Language Bias and Conflicting Information

When a model is forced to store conflicting or ambiguous cultural facts, the language of the prompt heavily dictates what knowledge it chooses to retrieve:

  • Prompt Language Bias: Research published in “Language Bias under Conflicting Information in Multilingual LLMs” shows that when models are fed conflicting data across different languages, they consistently prioritize information that matches the language used in the prompt. [9]
  • Global Dominance: If the prompt language does not break the tie, models show systemic biases toward knowledge stored in dominant training languages like English or Chinese, frequently ignoring data stored in other languages like Russian. [9, 10]

If you are looking to dig into a specific angle of this AI architecture, let me know if you would like to explore:

  • The specific mechanistic interpretability methods researchers use (like Logit Lens or Activation Patching) to “see” the knowledge.
  • How Multilingual Knowledge Editing (MKE) is used to change a fact in one language and force it to update across all other languages. [1, 11, 12, 13]

[1] https://aclanthology.org

[2] https://liner.com
[3] https://arxiv.org
[4] https://arxiv.org
[5] https://openreview.net
[6] https://arxiv.org
[7] https://arxiv.org
[8] https://arxiv.org
[9] https://arxiv.org
[10] https://arxiv.org
[11] https://arxiv.org
[12] https://arxiv.org
[13] https://arxiv.org

I then used the following prompt:

Do multilingual models differ in the information they retrieve based on the prompting language?

Yes, multilingual models absolutely retrieve different information based on the prompting language.
Even when a prompt is translated exactly, changing the surface language acts as a powerful steering vector that alters how the model traverses its internal neural network. Researchers have documented severe “cross-lingual inconsistencies”—proving that language changes not just how a model talks, but what it remembers. [1, 2]
The differences in retrieved information generally fall into four scientifically documented phenomena:

1. The Language-Matching Bias (Prioritization)

When a model is confronted with mixed or conflicting information during its pre-training, it heavily relies on the language of the prompt to break ties.

  • The Mechanism: According to the study “Language Bias under Conflicting Information in Multilingual LLMs,” if a model contains conflicting facts derived from global news sources, it will confidently choose to retrieve the fact that matches the prompt language. [3, 4]

  • The Result: If you ask a question in French, the model is significantly more likely to retrieve data that was originally learned from French-encoded tokens, ignoring a contradictory (or more up-to-date) consensus recorded in English. [3, 5]

2. “Cultural Value Shifting” and Ideological Instability

Language functions as an accidental proxy for cultural norms. When the prompting language changes, the model’s retrieved opinions, values, and socio-political stances shift dramatically.

  • The Multi-Personality Phenomenon: A paper evaluating models across the World Values Survey (WVS) framework, “Language as a Manipulation Vector,” found that switching the prompt language caused dramatic value shifts (up to 2.0 points on a -1 to +1 scale). [6]

  • Contradictory Beliefs: The exact same model will express completely contradictory ideological positions depending purely on whether it is prompted in Arabic, English, or Chinese. It defaults to retrieving information that matches the cultural expectations inherently baked into that specific language’s training corpus. [6]

3. Factual “Silos” and Low-Resource Gaps

A model’s knowledge base is vastly unequal across languages. Because the web is overwhelmingly English, factual knowledge is disproportionately concentrated in English subspaces. [7, 8]

  • The Retrieval Gap: If you prompt a model in a high-resource language like English, it can successfully retrieve niche scientific, historical, or coding data.

  • The “Zero-Shot” Wall: If you ask the exact same complex technical question in a low-resource language (e.g., Swahili or Tamil), the model will often hallucinate or state it doesn’t know. The model fails to trigger cross-lingual transfer, meaning the knowledge remains “locked” in the English subspace and unretrievable by the non-English prompt. [1, 5, 7, 8, 9]

4. Small Models vs. Large Models (Subspace Dissociation)

Surprisingly, the size of the AI model dictates how it handles language prompts:

  • Small Models (Shared Space): Smaller models have tight parameter constraints, forcing them to translate everything into a shared, English-centric semantic space to retrieve facts. Therefore, their retrieved facts remain relatively consistent across languages, though their grammar might suffer.

  • Large Models (Siloed Subspaces): As models get massive (like GPT-4 or Gemini 1.5), they have enough capacity to dissociate their internal states into individual language subspaces. Paradoxically, this extra power makes them more prone to cross-lingual inconsistency. They effectively build isolated, language-specific encyclopedias inside their networks, resulting in wildly different answers depending on the prompt. [1, 8, 10]

If you want to dive deeper into how engineers attempt to fix this, let me know if you would like to explore:

  • Alignment Prompting: How developers use specialized system instructions to force models to merge their language silos.

  • Multilingual Knowledge Editing (MKE): The methods used to surgically change a fact in one language and ensure it successfully propagates to all others. [5, 11]

[1] https://arxiv.org

[2] https://arxiv.org
[3] https://arxiv.org
[4] https://arxiv.org
[5] https://www.promptquorum.com
[6] https://apartresearch.com
[7] https://arxiv.org
[8] https://arxiv.org
[9] https://www.youtube.com
[10] https://arxiv.org
[11] https://latitude.so

My final prompt consisted of your question:

Given this information, what would be the best prompting approach for multilingual people using AI?

Based on how Large Language Models store and compartmentalize data, multilingual users can gain a massive advantage by strategically manipulating the model’s internal routing.
The optimal prompting strategy depends entirely on what type of task you are trying to accomplish:

1. For Hard Facts, Logic, and Complex Technical Tasks: The “English Bridge” Approach

Because factual knowledge, coding capabilities, and reasoning patterns are heavily concentrated in English-centric neural subspaces, prompting in English yields the most accurate and hallucination-free logic.

  • The Strategy: Prompt the model, force it to reason, and let it generate its final answer in English first. If you need the final output in another language, ask the model to translate its own finalized response at the very end of the prompt.
  • Why it works: This forces the model to bypass the “Curse of Multilinguality” and run its heavy computing inside its largest, most robust data silo before converting the final output.

2. For Local, Cultural, or Nuanced Nuances: The “Native Target” Approach

If you are asking about regional laws, local cultural etiquette, historical events from a specific country, or native idioms, you should prompt exclusively in the native language of that region.

  • The Strategy: Do not translate your local query into English. If you want to know about French labor laws, prompt in French. If you want to understand Chinese philosophy, prompt in Mandarin.
  • Why it works: This deliberately activates the model’s language-specific subspaces. It triggers the exact cultural biases and localized training data that an English prompt would completely miss or overwrite with Western bias.

3. For Objective, Unbiased Research: The “Cross-Lingual Triangulation” Approach

Since models exhibit heavy language-matching bias (giving different geopolitical or ideological answers depending on the prompt language), you can use your multilingualism to spot biases and hallucination.

  • The Strategy: Open two separate chat windows. Ask the exact same controversial or nuanced question in two different languages (e.g., English and Spanish). Compare where the facts diverge.
  • Why it works: It allows you to see where the model’s “siloed encyclopedias” disagree, giving you a uniquely holistic view of how different language datasets treat the same topic.

:clipboard: Summary Matrix: When to Use Which Language

Task Type Best Prompting Language Expected Outcome
Coding & Math English Higher accuracy, fewer syntax errors.
Scientific Research English Accesses the latest, most robust academic data.
Local Legal/Tax Advice The target country’s official language Prevents cross-border legal hallucinations.
Creative Writing / Copy The desired output language Avoids stiff, “translated-sounding” text.

To tailor this specifically to your daily workflow, tell me:

  • What languages do you speak fluently?
  • What specific tasks (e.g., coding, creative writing, research, translation) do you use AI for most often?

I can help you build a personalized prompting template for your exact workflow.

I could have interpreted and checked this feedback myself, but for now I will leave this to you. Feel free to dive into a deeper discussion on these issues.

Thank you for your help / Merci pour votre aide !

 –