SUBSUB and SUBINJ: An English Anchor for Accurate Facts in Any Language

SUBSUB — substitution:

Original in Spanish: "¿Dónde está la capital de Francia?"
↓
With SUBSUB: "¿Dónde está la capital de France?"

SUBINJ — injection:

Original in Spanish: "¿Dónde está la capital de Francia?"
↓
With SUBINJ: "¿Dónde está la capital de Francia (France)?"

Both techniques are applied within a single prompt, in an ordinary chat.

Task: You're writing a briefing in Spanish about international institutions. You need accurate facts — founding year, headquarters location, headcount. You work in Spanish, but the LLM keeps mixing up facts or giving different answers in different chats.

Prompt with SUBINJ:

¿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)?

Result: The model finds facts more accurately, because the English names are a direct path to how these organizations are represented in its training data. Instead of "Organización Mundial de la Salud" → internal representation → fact lookup, the model goes straight from "World Health Organization" → fact. Fewer steps = fewer errors.

The LLM's weakness: The model doesn't always map concepts correctly between languages in its internal space. "Francia" in Spanish and "フランス" in Japanese should land on the same concept, but often they don't — especially for non-Latin scripts and lower-resource languages.

The LLM's strength: The model works well with English as its base language. Nearly all multilingual LLMs are trained mostly on English data and use English as a pivot language in their concept space. When the model sees an English concept, it lands in the right place immediately.

How the method exploits this: Instead of relying on an unreliable translation (your language → internal space), we give the model a direct anchor — the English name. The model skips the translation step and works straight in its "native" English space. It's like giving GPS coordinates instead of an address in an unfamiliar language.

Control knobs:

• SUBSUB vs. SUBINJ: SUBSUB is more aggressive (full replacement) — use it for maximum accuracy, but it can make the text sound unnatural. SUBINJ is gentler (addition) — it keeps things readable and is better for final copy.
• Which concepts to translate: Translate the key subjects (names of companies, people, places) and domain-specific terms. Leave function words alone.
• When you don't need it: If you're writing creative copy or need natural-sounding speech, skip it. The method is for facts, not style.

SUBINJ (recommended for most cases):

{question containing the key concept} ({English translation of concept})

Example: ¿Cuál es la capital de Francia (France)?

SUBSUB (for maximum accuracy):

{question with the key concept replaced by its English translation}

Example: ¿Cuál es la capital de France?

What to plug in:

• {key concept} — a name, company, place or specialized term in your language
• {English translation} — the exact English translation of that concept

When to use it:

• Factual queries (dates, figures, names)
• Questions about well-known brands, companies, people
• Technical terms with an established English name
• Whenever accuracy and consistency matter

When NOT to use it:

• Creative writing
• Literary translation
• When natural-sounding text matters more than accuracy
• Casual conversation

The study tested the hypothesis on 12 models (LLaMA, Qwen, Gemma, OLMo) of different sizes and 17 languages (from English to Japanese, from Spanish to Arabic). The researchers took 2,619 facts and measured how well the models:

1. Translate names between languages (e.g., "Francia" → "France")
2. Answer consistently to the same question in different languages

The main finding: Consistency strongly correlates with the ability to translate concepts (r > 0.7-0.8 for most models). If a model can't correctly translate "Francia" into "France", it almost never gives a consistent answer to a question about France across languages.

A detailed analysis showed that 99% of consistent answers are cases where the model correctly aligned the concepts. And conversely: when the model failed to align the concepts, consistent answers practically vanish (under 10%).

Why do SUBSUB and SUBINJ help? A mechanistic analysis (Logit Lens) showed that when we add the English concept, the representations of that concept across all languages move closer together in the model's hidden space. The model stops "tripping" over translation and works in a single shared concept space.

The effect is especially strong for:

• English-centric models (OLMo): up to +44% consistency and +32% accuracy
• Non-Latin scripts (Cyrillic, Arabic, Chinese): there's no surface-level token overlap with English, so an explicit translation is critical
• Small models: their internal concept alignment is weaker, and an external anchor makes up for it

⚠️ Requires English: You need to know the correct English translation of the concept. Get it wrong and you can make the problem worse.

⚠️ Not for every task: It works for facts and specific concepts. For creative work, style or casual speech it's overkill, or it hurts naturalness.

⚠️ The effect depends on the model: Multilingual models (Gemma, Qwen) already align concepts well on their own, so the gain is smaller (2-10%). English-centric ones (OLMo) get a huge boost (25-44%).

⚠️ Not magic: If the model doesn't know the fact in any language, the method won't help. It only helps when the knowledge is there but gets lost in translation between languages.

The researchers used KLAR — a dataset of 2,619 facts in 17 languages. Each fact is a triple (subject, relation, object): for example, (France, capital, Paris). Every fact has translations into all 17 languages.

The experimental design was simple but clever: they tested two tasks in parallel:

Task 1 — Concept translation: Can the model translate a fact's subject or object from language 1 to language 2? For example: "日本語: フランス - English: ?" → should be "France". They tested all 136 language pairs (17 languages × 16).

Task 2 — Factual consistency: Does the model give the same (and correct!) answer to the same question in two languages? For example: "¿Dónde está la capital de Francia?" in Spanish → "París", and "Where is France's capital?" in English → "Paris".

Why this design is valuable: The authors didn't just measure the problem (models are inconsistent), they found its mechanical cause (poor concept alignment). And they proved the link: when alignment is good → consistency is high; when it's poor → consistency collapses.

The mechanistic analysis (Logit Lens) showed exactly HOW SUBSUB/SUBINJ work: they align the internal representations of concepts across all languages. When the prompt contains an English anchor, the representation of "Francia" in Spanish moves closer to the representation of "France" in English in the model's hidden space. The model stops "guessing" how to translate the concept — it already has the right answer in hand.

Scale of the evaluation: 12 models of different sizes (from 1B to 13B parameters), 4 families (LLaMA, Qwen, Gemma, OLMo), 17 languages, 8 scripts, 2,619 × 136 = 356,184 evaluated facts per model. That's solid.

The authors used 3-shot prompting for concept translation. Here's an example from the paper (translating a subject from Japanese into English):

日本語: フランス - English: France
日本語: セルビア - English: Serbia
日本語: イタリア - English: Italy
日本語: イギリス - English:

The model should continue with: "United Kingdom"

Context: The researchers tested whether the model can translate entity names between languages. This is a proxy task for measuring "alignment" — the model's ability to understand that different words in different languages refer to the same concept.

For factual queries they used a similar 3-shot format with question-answer examples in the target language.

💡 Adaptation for technical terms:

If you work in a specialized field (medicine, law, engineering) with established English terminology, always add the English term in parentheses.

Example (medical content):

Describe los síntomas de la hipertensión arterial (hypertension) y sus métodos de diagnóstico.
¿Qué medicamentos se usan para tratar la diabetes (diabetes)?

Why it works: Medical literature is mostly in English, and the model was trained on it. The English term = direct access to that knowledge.

🔧 Technique: Multiple concepts → batch anchoring

If your query has several key concepts, add translations for all of them:

Example (business analysis):

Compara las estrategias de Tesla (Tesla) y BYD (BYD) en el mercado de vehículos eléctricos (electric vehicles) en China (China).

Every key concept gets an English anchor. This amplifies the effect — the model works entirely in its English concept space.

💡 Adaptation for names in non-Latin scripts:

Companies, products and places written in Japanese, Chinese, Arabic and other scripts often have an official English spelling. Use it:

Example (company analysis):

トヨタ (Toyota) の2024年の売上高はいくらですか？
ソフトバンク (SoftBank) の契約者数は何人ですか？
東京 (Tokyo) にある任天堂 (Nintendo) の本社はどこですか？

🔧 Technique: Not sure about the translation → ask the model

Add a meta-question BEFORE the main one:

¿Cómo se dice "hipertensión arterial" en terminología médica en inglés?

[you'll get: "hypertension"]

Describe los síntomas de la hipertensión arterial (hypertension) y sus métodos de diagnóstico.

The model gives you the right term, and you use it. A two-step approach, but a reliable one.

On the Entity-Level Alignment in Crosslingual Consistency
Yihong Liu, Mingyang Wang, François Yvon, Hinrich Schütze
Center for Information and Language Processing, LMU Munich; Munich Center for Machine Learning (MCML); Sorbonne Université, CNRS, ISIR, France

Dataset: KLAR (Wang et al., 2025)
Models: LLaMA-3.x, Qwen3-Base, Gemma-3-pt, OLMo-2