English as the Universal Reasoning Lingua Franca

This direction is split into other directions: cross-lingual instruction bias. Its last version stays readable and citable here.

Version 1 of 1125 papers153 conceptsfrozen 2026-10-04

Large Language Models (LLMs) are remarkably polyglot, capable of translating, summarizing, and chatting in dozens of languages. However, a growing body of research suggests that while these models can speak many languages, they primarily reason in one: English.

Even when a model is prompted in a high-resource language like Spanish or a low-resource one like Bengali, its internal logical "circuits" often function most effectively when mediated through English. This phenomenon, often called "thinking in English," reveals a fundamental asymmetry in how multilingual intelligence is structured within neural networks.

The English Centricity of Logic

The intuition behind this research direction is rooted in the distribution of pre-training data. Because the vast majority of high-quality reasoning data—such as source code, mathematical proofs, and philosophical discourse—is written in English, the model's internal representation of "logic" becomes disproportionately anchored to English linguistic patterns.

Aitor Soroa, Mikel Artetxe, and their colleagues demonstrated this in Do Multilingual Language Models Think Better in English?. They found that models possess significant reasoning capabilities that remain "locked" when prompted in non-English languages. By simply asking the model to translate the input into English before solving it—a "self-translate" strategy—they achieved consistent performance gains. This suggests that the model's "intelligence" is not evenly distributed across its vocabulary; rather, English acts as a high-performance substrate for complex thought.

English as a Thinking Pivot

If a model reasons better in English, can we explicitly use English as a bridge to improve performance in other languages? This is the core idea behind Cross-Lingual-Thought (XLT) prompting, introduced by Haoyang Huang and Dongdong Zhang in Not All Languages Are Created Equal in LLMs.

XLT prompting encourages the model to use English as a "thinking pivot." Instead of solving a math problem directly in Chinese, the model is prompted to translate the problem to English, perform the Chain-of-Thought reasoning in English, and then translate the final answer back. This technique has yielded massive performance boosts, particularly in arithmetic and symbolic reasoning, where the model's internal English logic is most robust.

The Reasoning Tax and Linguistic Inertia

The emergence of Large Reasoning Models (LRMs), such as DeepSeek-R1, has made this internal process visible through "thinking traces." These models often exhibit linguistic inertia: even when asked a question in a non-English language, they frequently default to English for their internal scratchpad.

Research by Jirui Qi and Arianna Bisazza in When Models Reason in Your Language identifies a significant "reasoning tax" associated with forcing language consistency. When researchers used prompt engineering or fine-tuning to force a model to reason in the user's native language, the accuracy of the final answer often dropped. Forcing the model to abandon its English "lingua franca" for internal thought appears to degrade its ability to access sophisticated cognitive strategies like backtracking and sub-goal setting.

The Double-Edged Sword

While English-centric reasoning provides a performance floor for multilingual tasks, it introduces new failure modes. Alan Saji and Raj Dabre describe this as a Double-Edged Sword. The primary risk is "Lost in Translation" errors: a model might reason perfectly in English but fail because it misinterpreted a subtle cultural nuance or idiomatic expression during the initial translation from the source language.

Furthermore, this English-centricity creates a performance gap for low-resource languages. As noted by Krishno Dey and Usman Naseem in their study of South Asian languages, it is almost always "better to ask in English." This suggests that current multilingual models do not truly achieve cross-lingual equity; instead, they operate as English-centric engines with a sophisticated translation layer on top.

Beyond the English Pivot

The research community is now exploring ways to break this dependence. While English remains the most effective anchor for instruction-following and logic today, projects like ArabicMMLU by Fajri Koto and NusaCrowd by Samuel Cahyawijaya emphasize the need for native-language benchmarks and data. The goal is to move toward models where reasoning is truly language-agnostic, or where models can "think" natively in the cultural and linguistic context of the user without paying a tax on their intelligence.

Papers in this version

Showing 60 of 125 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_english_as_the_universal_reasoning_lingu_a9d4270d,
  title = {English as the Universal Reasoning Lingua Franca},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/english-as-the-universal-reasoning-lingua-franca/rdr_a9d4270d0e6840d884abca0bbc973ce3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_english_as_the_universal_reasoning_lingu_a9d4270d_v1,
  title = {English as the Universal Reasoning Lingua Franca},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/english-as-the-universal-reasoning-lingua-franca/rdr_a9d4270d0e6840d884abca0bbc973ce3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}