Cross-lingual prompting: choosing the language of instructions, data, reasoning and reply

What cross-lingual prompting is

Cross-lingual prompting means choosing a language for each part of the prompt separately: the instructions, the source data, the model's reasoning and the reply. They do not have to match. You can write instructions in Spanish, paste an English report as is, ask the model to reason in English, and get the answer in Spanish.

The technique has four levers:

• Reasoning language: where the model does its step-by-step work before answering.
• Data language: whether source text stays in the original or is translated first.
• Anchors: English names or terms added next to key concepts in a non-English prompt.
• Reply language: stated explicitly instead of left for the model to infer.

The studies agree on the problem: models are less accurate and less consistent outside English. They disagree on whether to translate.

Why the language choice matters

Most large models are trained mostly on English text. The XCoT authors found that multilingual models hallucinate far more often in lower-resource languages, where both their knowledge and their reasoning are thinner. Their remedy is to have the model reason step by step in English and then write the answer in the target language. At that last step it restates conclusions it has already built instead of generating them in a weaker language. The reported reduction in hallucinations of up to 62% came from combining this prompt with fine-tuning, so a chat user gets only part of that effect.

A second mechanism is concept alignment. The entity-alignment study tested 12 models on 2,619 facts in 17 languages. Models gave consistent answers across languages almost only when they had correctly linked the local name of an entity to its English counterpart: 99% of consistent answers were such cases. Adding the English name in parentheses (SUBINJ) or substituting it (SUBSUB) raised factual accuracy by 2-32% and cross-lingual consistency by 5-44%. The largest gains were for English-centric models and non-Latin scripts.

The third mechanism is about the reply. Models infer the expected reply language from surface signals: the share of characters from each language and, above all, the language of the last word. Mixed-language prompts therefore often get an answer in the wrong language. The OLA benchmark measured 21-75% correct-language replies without an instruction and 96-100% with an explicit "Reply in English" or "Responde en español". Chain-of-thought did not fix it and in some cases made it worse, by up to -50%.

Variants that tested well

• Think in English, answer in your language. A two-stage prompt from XCoT: reason in English with explicit action steps, then write the final answer in the target language. Best suited to factual questions in languages where the model has little data.
• Structure first, then reason. Structured-of-Thought adds three steps to the English restatement: extract entities, numbers and relationships, explain language-specific expressions (a Chinese "four-five discount" means paying 45% of the price), then answer in the original language. It reached 68.3% on MGSM and 76.5% on MSVAMP without fine-tuning, ahead of the training-free baselines it was compared with. The ablation is telling: structuring and idiom explanation together gave 61.6% against 62.8% for the full method, while switching the thinking to English added only 1.2%.
• English anchors for key terms. Write the concept in your language and the English name in parentheses: "Organización Mundial de la Salud (World Health Organization)". This is the SUBINJ variant. It suits names, places and technical terms, not creative text.
• Hybrid prompt. Instructions in your language, data in the original: set the task in the language you want the answer in, paste the English source untranslated, and state the output language. This avoids two rounds of translation loss.
• Explicit reply language. End every mixed-language prompt with a short reply-language line. In a long prompt, repeat it in the middle and at the end (OLA).
• Language as a perspective. Two less common uses: thinking in a distant language such as Hebrew gave +12.78% answer diversity over thinking in English, and cultural debate between personas raised cultural neutrality to 86% from 69%. The second study also found that simply prompting in Chinese shifted the bias toward East Asian defaults instead of removing it.

Translate or not: where the studies disagree

The main disagreement is about translation. XCoT says reason in English. The Dutch clinical-extraction study found the opposite for input: machine-translating Dutch medical documents into English before extraction consistently and significantly lowered accuracy. Mid-sized open models from 14B parameters worked well on the original text. These results can coexist, because they test different things. Translating the data with an external tool loses domain vocabulary before the model sees it. Asking the model to reason in English, or to add an English anchor next to a term, keeps the original in the context. The SoT ablation points the same way: the English step was the smallest contributor, and structure did most of the work.

A practical rule that fits all the evidence:

• Keep source documents in their original language. Do not pre-translate them.
• For hard factual or multi-step questions in a lower-resource language, let the model reason in English, or at least extract the structure first.
• Add English anchors to names and technical terms when cross-language consistency matters.
• Always state the reply language when the prompt mixes languages.

Limits: anchors help only when the model knows the fact in some language, and a wrong English term can make things worse. Multilingual models such as Gemma and Qwen gained only 2-10% from anchors (anchor study). An explicit reply language fixes which language you get, but in the OLA tests it did not stop stray characters from a third language, and the text could still read like a translation. Simple everyday requests do not need any of this. For the reasoning step itself, see chain-of-thought prompting.

Role: experienced dietitian. Task: explain the long-term risks of the keto diet to someone without a medical background.

Step 1 — Reason in English:
Action: Define ketogenic diet and how ketosis works.
Action: List long-term health risks: cardiovascular, kidney stones, nutrient deficiencies, gut microbiome, bone density.
Action: For each risk explain the mechanism in simple terms.
Action: Structure the answer with clear categories.

Step 2 — Give the answer in Spanish based on your reasoning. Use subheadings, avoid medical jargon.

I'll give you a problem in {source_language}. Apply the SoT (Structured-of-Thought) method:
Step 1: Restate the problem in English and think in English
Step 2: Extract all entities (numbers, units, objects) and their relationships in structured form
Step 3: Explain language-specific expressions (idioms, cultural conventions)
Step 4: Solve the problem and give the final answer in {source_language}
Problem: {your_problem}

¿En qué año se fundó la Organización Mundial de la Salud (World Health Organization)?
¿Dónde está la sede del Banco Central Europeo (European Central Bank)?
¿Cuántos empleados tiene la Agencia Espacial Europea (European Space Agency)?