Language of Thought Shapes Output Diversity in Large Language Models

TL;DR

This work investigates the language of thought as a structural control axis for augmenting output diversity in large language models (LLMs). By forcing intermediate thinking in various languages and evaluating English-final outputs, the authors demonstrate that different thinking languages occupy distinct regions in a thinking space and that thinking distances to English, $d_j(l,\text{en})$, correlate positively with diversity metrics. Two sampling strategies are proposed—Single-Language Sampling and Mixed-Language Sampling—and both show that non-English thinking expands diversity, with larger gains for languages farther from English and with compositional benefits when multiple languages are used. The approach yields practical benefits in pluralistic alignment, increasing coverage of cultural knowledge and value orientations, as evidenced on datasets like Blend and WVS, and under varying temperature and sampling counts. The work provides publicly available code and highlights a new paradigm for leveraging multilingual thinking to broaden the thinking space of LLMs.

Abstract

Output diversity is crucial for Large Language Models as it underpins pluralism and creativity. In this work, we reveal that controlling the language used during model thinking-the language of thought-provides a novel and structural source of output diversity. Our preliminary study shows that different thinking languages occupy distinct regions in a model's thinking space. Based on this observation, we study two repeated sampling strategies under multilingual thinking-Single-Language Sampling and Mixed-Language Sampling-and conduct diversity evaluation on outputs that are controlled to be in English, regardless of the thinking language used. Across extensive experiments, we demonstrate that switching the thinking language from English to non-English languages consistently increases output diversity, with a clear and consistent positive correlation such that languages farther from English in the thinking space yield larger gains. We further show that aggregating samples across multiple thinking languages yields additional improvements through compositional effects, and that scaling sampling with linguistic heterogeneity expands the model's diversity ceiling. Finally, we show that these findings translate into practical benefits in pluralistic alignment scenarios, leading to broader coverage of cultural knowledge and value orientations in LLM outputs. Our code is publicly available at https://github.com/iNLP-Lab/Multilingual-LoT-Diversity.

Figures (6)

• Figure 1: Language geometry of thinking space on Qwen3-8B, with different distance scales across layers for visualization purposes.
• Figure 2: Correlation between the Distinct Score and the thinking distance to English across languages. Pearson’s $r$ and Spearman’s $\rho$ are reported for each model. Distinct Scores are obtained under Single-Language Sampling on NoveltyBench. Thinking distances are normalized to the range $[0,1]$ for visualization.
• Figure 3: Relative deviation in Distinct Score under the removal of $k$ languages in Mixed-Language Sampling.
• Figure 4: Effects of sampling parameters on output diversity. (a) Distinct sample count as a function of the sampling number $M$ at a fixed temperature ($0.6$). (b) Distinct Score (%) under different temperatures with a fixed sampling number ($M = 15$).
• Figure 5: Prefix translations used for Thinking Language Control.