TL;DR
Language of Thought is a technique that controls the language of the model's internal reasoning, not of its final answer. The gist: the model thinks in one language (say, Hebrew or Turkish) but answers in another (English). That shifts the model into a different region of its "thinking space" and produces more diverse answers.
When a model answers the same question many times while thinking in English, the answers are often similar to one another: the model gets "stuck" in one region of its space. The researchers visualized hidden states (the model's internal representations) and found that different languages occupy different regions of the thinking space. Chinese and French sit close to English, while Hebrew, Norwegian and Bulgarian are far away. The farther a language is from English, the more diverse the answers (correlation 0.72-0.88).
There are two ways to apply it. Single-Language Sampling: every repeated request uses the same non-English thinking language (for example, all 15 requests think in Hebrew and answer in English). Mixed-Language Sampling: each request thinks in a different language (Portuguese, Hebrew, Chinese, Turkish...), but all answer in English. That gives a compositional effect and another +2-8% in diversity.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
Single-Language Sampling:
STEP 1: Request → "Think in {language}, answer in English"
STEP 2: The model thinks in the target language (internal process)
STEP 3: The model answers in English
STEP 4: Repeat 10-15 times with the same thinking language
STEP 5: Collect all the answers → assess the diversity
Mixed-Language Sampling:
STEP 1: Request 1 → "Think in English, answer in English"
STEP 2: Request 2 → "Think in Hebrew, answer in English"
STEP 3: Request 3 → "Think in Chinese, answer in English"
...
STEP 15: Request 15 → "Think in Turkish, answer in English"
STEP 16: Collect all the answers → maximum diversity
Each step is a separate request in the chat.
Example in Practice
Task: Come up with 10 offer headlines for the landing page of an online AI school. You need genuinely different angles, not clones of "Master ChatGPT in 3 Days".
Prompt 1 (thinking in English):
Suggest 5 offer headlines for the landing page of an online AI school (target audience: marketers, designers, product managers).
Prompt 2 (thinking in Hebrew):
Think in Hebrew about offer ideas for this landing page, then provide your answer in English:
Suggest 5 offer headlines for the landing page of an online AI school (target audience: marketers, designers, product managers).
Prompt 3 (thinking in Chinese):
Think in Chinese about offer ideas for this landing page, then provide your answer in English:
Suggest 5 offer headlines for the landing page of an online AI school (target audience: marketers, designers, product managers).
Repeat with these languages: Turkish, Bulgarian, Arabic, Norwegian, Occitan (the farther from English, the better).
Result:
With English thinking, most answers circle around similar formulas: "Master ChatGPT in X days", "Automate your work with AI", "A hands-on Midjourney course".
When thinking in distant languages (Hebrew, Bulgarian, Norwegian), the model generates more distinct angles: different metaphors, sentence structures and positioning hooks. For example, you might get an offer built on fear of falling behind, on status, or on a specific pain point ("Stop spending 3 hours on what AI does in 5 minutes").
With Mixed-Language (10-15 requests in different languages), you'll gather the widest possible range of options for A/B tests.
Why It Works
The LLM weakness: When sampled repeatedly on the same question, the model generates similar answers. This is especially noticeable on open-ended tasks with no single right answer (creative work, ideas, wording). The model gets "stuck" in a narrow region of its probability space.
The LLM strength: The model is trained on multilingual data. Different languages encode different conceptual structures, grammatical constructions and cultural contexts. When the model thinks in Hebrew vs English, it activates different regions of its latent space, as hidden-state visualizations showed.
How the method exploits this: By switching the thinking language, you physically shift the model into another region of the space. The researchers found a direct correlation: the farther a language is from English in the geometry of the space (0.72-0.88 Pearson), the more diverse the answers. Hebrew on Qwen3-8B gave +12.78% diversity vs English. Mixed-Language amplifies the effect through composition: each language adds its own coverage zone, and removing one language barely hurts, but removing 3-5 languages drags diversity way down.
Control levers:
Language choice: Languages far from English (Hebrew, Bulgarian, Occitan, Norwegian) give more diversity than close ones (Chinese, French, Spanish). But the study doesn't explain why a particular language is distant, so you need to test.
Single vs Mixed: Single gives +5-7% diversity vs English. Mixed adds another +2-8% on top through the composition of languages.
Temperature: Works compositionally with language. Mixed-Language at temperature 1.0 gives the same diversity as English at 2.0. You can lower the temperature (for quality) and compensate with language (for diversity).
Number of languages: The more languages in Mixed-Language, the higher the "diversity ceiling". Scaling up to 200 requests, English quickly hits its limit while Mixed keeps growing.
Prompt Template
For ChatGPT/Claude (universal):
Think in {language} about this question, then provide your full answer in English:
{your question}
Placeholders:
- {language}: pick the thinking language: Hebrew, Bulgarian, Norwegian, Turkish, Arabic, Occitan for maximum diversity; Chinese, French, Spanish for moderate; English as the baseline
- {your question}: an open-ended question with no single right answer (ideas, options, creative work, wording)
For models with a thinking mode (DeepSeek R1, Qwen):
<think>
{prefix in the target language: "Okay, the user is asking..."}
{the model thinks in the target language}
</think>
Let me provide my answer in English only:
{answer in English}
This format kicks in automatically when you ask the model to "think in language X".
🚀 Quick start for Mixed-Language:
Instead of running it 15 times by hand, ask the model to help:
I need 15 different answers to one question. I want to use the "language of thought" technique: each time, think in a different language (Hebrew, Bulgarian, Chinese, Norwegian, Turkish, Arabic, Portuguese, German, Spanish, French, Italian, Malay, Occitan, Danish, Swedish), but always answer in English.
Here is my question: {your question}
Give me a copy-ready template: 15 prompts, each with an instruction to think in its own language.
The model will generate 15 ready-made prompts. Paste them one at a time into new requests and collect the answers.
Limitations
⚠️ Quality loss: Thinking in a non-English language lowers answer quality by 1-2 points on a 0-100 scale (as judged by GPT-4). For critical tasks (code, math, precise instructions), use English thinking or check the result.
⚠️ Needs multiple requests: A single request with non-English thinking has a small effect. Real diversity shows up with 10-15 repeats. Mixed-Language amplifies it, but that's 15 separate requests in the chat, which is laborious without automation.
⚠️ Languages are unpredictable: The study shows the correlation "farther from English = more diversity" but doesn't explain why Hebrew works better than Chinese. You can't predict in advance which language will max out diversity for your model and task, so you need to test.
⚠️ Not for every task: The method works on open-ended questions with no right answer (ideas, options, creative work). For tasks with a ground truth (facts, calculations, translation), you don't want diversity, you want accuracy.
How They Tested It
Researchers from the Singapore University of Technology and Design took four models with a thinking mode (Qwen3-8B, 14B, 32B and DeepSeek-R1-Distill-Qwen-14B) and tested them on 15 languages: English, Hebrew, Chinese, Russian, German, Bulgarian, Norwegian, Turkish, Arabic and others.
First they examined the geometry of the thinking space. The idea was simple: give the model one question in English but make it think in different languages. They collected the hidden states (the network's internal representations) from all layers during the thinking process, averaged them and applied PCA for visualization. The result was surprising: languages occupy clearly separate regions in the space, and this separation holds even in the middle layers, which are usually considered abstract and language-agnostic. They computed each language's cosine distance to English and found a wide spread: Chinese is close (0.1-0.2), Hebrew is far (0.4-0.6).
Then they tested the link between geometry and diversity. They took two benchmarks: NoveltyBench and Infinity-Chat (100 open-ended questions each with no right answers, like "describe a startup idea" or "make up a story"). For each question they made 15 repeated requests. In Single-Language, all 15 think in one language (for example, all in Hebrew); in Mixed-Language, each one thinks in its own.
They measured diversity with two metrics: - Distinct Score: a DeBERTa-v3-large model judges whether two answers are functionally equivalent. They count how many unique classes come out of 15 answers (10 unique → score 66.7%). - Similarity Score: embeddings of all answers (via Qwen3-Embedding-8B), cosine similarity across all pairs, averaged. The lower, the more diverse.
The main finding: a direct correlation between distance in the thinking space and Distinct Score. Pearson 0.72-0.88, Spearman 0.58-0.89 across models. Hebrew (the most distant) gave +12.78% Distinct Score vs English on Qwen3-8B. The average gain across all non-English languages: +5-7%.
Mixed-Language did even better: +2-8% over the best single language. They tested the composition via ablation: removing 1 language barely hurts (-2.7%), but removing 5 languages drops it by 20-30%. That revealed a compositional effect: the languages complement each other.
They scaled up to 200 requests with Qwen's full language pool (~100 languages). English quickly hit a ceiling (~50 unique answers), while Mixed kept growing (~120 unique). Conclusion: linguistic heterogeneity raises the diversity ceiling.
They also tested pluralistic alignment (the Blend and WVS datasets: questions on cultural knowledge and value orientations). Mixed-Language gave +8-19% coverage vs English, beating high temperature (temp 2.0), explicit diversity requests ("give diverse answers") and multilingual prompting (asking the question in different languages).
Answer quality fell by 1-2 points (on a 0-100 scale, judged by GPT-4o-mini) with non-English thinking: there is a trade-off, but it's small.
Adaptations and Extrapolations
💡 Adaptation for specific personas:
Instead of a faceless "think in Hebrew", you can amplify the cultural context:
Imagine you're an Israeli entrepreneur (think in Hebrew), then give your answer in English:
How do you land your first 100 customers for a B2B SaaS in the US?
Or:
Imagine you're a Chinese marketer (think in Chinese), then give your answer in English:
What are some unconventional promotion channels for a mobile app?
This layers cultural specificity on top of the linguistic one.
🔧 Technique: lower the temperature, compensate with language
The study showed a compositional effect: Mixed-Language at temp 1.0 gives the same diversity as English at temp 2.0.
In practice: high temperature hurts quality (hallucinations, incoherence). Instead:
Temperature: 0.6-0.8 (instead of 1.5-2.0)
Thinking languages: Mixed-Language (10-15 languages)
Result: high diversity + better quality.
💡 Extrapolation: combining it with Self-Consistency
Self-Consistency: make N requests and pick the most frequent answer (majority voting). Usually the same prompt is used every time.
The combination:
STEP 1: Mixed-Language Sampling (15 requests in different languages)
STEP 2: Cluster the answers by meaning (for example, with embeddings)
STEP 3: Pick the top 3 clusters with the most answers
STEP 4: Present all three options to the user
Instead of a single "correct" answer, you get three popular alternatives from different regions of the thinking space.
Resources
Language of Thought Shapes Output Diversity in Large Language Models
Shaoyang Xu, Wenxuan Zhang
Singapore University of Technology and Design
https://github.com/iNLP-Lab/Multilingual-LoT-Diversity
Related concepts: the Sapir-Whorf hypothesis (language shapes thought), multilingual reasoning, output diversity in NLG, pluralistic alignment.