PaperlayerView exact revision on arXiv

Source-linked AI summary

Language of Thought Shapes Output Diversity in Large Language Models

Shaoyang Xu, Wenxuan Zhang

arXiv:2601.11227v2cs.CLcs.CY

TL;DR

LLMs need diverse outputs for pluralistic alignment and creativity, but existing work largely emphasizes English-only or multilingual inputs. This paper controls intermediate language of thought, evaluates English outputs under single- and mixed-language repeated sampling, and finds that non-English and heterogeneous thinking increase diversity, with broader cultural and value coverage, while noting limitations in alignment interventions and cultural-pluralism evaluation.

  • Problem

    Existing diversity research largely focuses on English-only or multilingual inputs, leaving language of thought as a potential structural source of output diversity underexplored.

  • Method

    The paper maps multilingual thinking-space geometry and compares Single-Language and Mixed-Language Sampling while keeping final outputs in English.

  • Results

    Non-English thinking consistently increases output diversity, languages farther from English yield larger gains, and mixed-language sampling adds compositional improvements.

  • Takeaways & Limitations

    Language of thought is a structural control axis for output diversity and can broaden cultural-knowledge and value coverage in pluralistic alignment.

  • Takeaways & Limitations

    Pluralistic-alignment evaluation uses output entropy as a proxy for cultural pluralism and abstracts away real-world, context-dependent cultural-value alignment.

Abstract

from arXiv · show

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.

1 Introduction

The paper investigates whether controlling an LLM’s language of thought can provide a structural source of output diversity. Across repeated-sampling strategies, non-English thinking, linguistic heterogeneity, and thinking-space distance from English are associated with broader English outputs and pluralistic coverage.

  • Motivation: Output diversity measures whether LLMs generate multiple distinct responses to open-ended questions without ground-truth answers.The paper links diversity to pluralistic alignment, creativity, exploration, and novel idea generation.
  • Motivation: Existing approaches increase diversity through temperature scaling, decoding methods, output aggregation, prompt variation, and diversity-oriented training objectives.The authors contrast these approaches with exploiting multilingual properties during model thinking.
  • Research Question: The study asks whether language of thought can serve as a controllable, structural source of output diversity in multilingual LLMs.The motivation draws on multilingual reasoning capabilities and the possibility that languages encode distinct conceptual and structural biases.
  • Approach: Different thinking languages occupy distinct regions in the model’s thinking space, motivating controlled repeated sampling with English final outputs.The study compares Single-Language Sampling within one thinking language with Mixed-Language Sampling across languages.
  • Findings: Switching from English to non-English thinking consistently increases output diversity, with larger gains for languages farther from English in thinking space.The authors report negligible output-quality degradation overall.
  • Findings: Mixed-Language Sampling adds compositional diversity gains, and linguistic heterogeneity expands the model’s diversity ceiling as sampling scales.Removing multiple languages causes greater diversity degradation than removing any single language.
  • Applications: In pluralistic alignment scenarios, Mixed-Language Sampling broadens coverage of cultural knowledge and values beyond several comparison strategies.Comparisons include English sampling, high-temperature decoding, explicit diversity requests, and multilingual prompting.
  • Conclusion: The findings establish language of thought as a novel and effective control axis for enhancing output diversity.

2 Related Work

Prior work improves LLM output diversity through decoding, aggregation, prompting, and training, while multilingual reasoning research studies cross-language thought and generalization. This paper distinguishes itself by using inherent multilingual structure as a diversity source.

  • Output Diversity of LLMs: Output-diversity research evaluates lexical, syntactic, and semantic variation using metrics such as Self-BLEU and Sentence-BERT.Recent benchmarks also assess novelty and creativity in complex generation settings.
  • Output Diversity of LLMs: Existing diversity-improvement methods aggregate outputs from multiple LLMs, vary prompts, or optimize diversity-driven RLHF and SFT objectives.
  • Positioning: Unlike prior approaches, this work explores the inherent multilingual properties of LLMs as a structural source of output diversity.
  • Multilingual Reasoning: Multilingual reasoning research examines explicit intermediate reasoning and cross-language generalization in LLMs with shared yet imbalanced thought spaces.

3 Language Geometry of Thinking Space

The paper constructs language-specific representations of multilingual thinking and visualizes their geometry across model layers. Different thinking languages occupy separable regions and vary systematically in distance from English.

  • Representation Construction: Given English inputs, the study controls intermediate thinking language and collects hidden representations to examine structural differences in thinking space.The model performs intermediate thinking before producing English outputs by default.
  • Representation Construction: Language-specific thinking representations are formed by averaging hidden states across thinking tokens and then across samples at each layer.Repeating this procedure across languages yields one representation per language and layer.
  • Visualization: The visualization normalizes representations, measures each language’s cosine distance from English, and applies PCA for a two-dimensional layout.PCA determines angular arrangement, while radial distance is fixed to cosine distance from English.
  • Experimental Setup: The analysis uses English and 14 officially supported non-English languages in Qwen3-8B across representative layers.Figure 1 visualizes the resulting multilingual thinking-space geometry.
  • Observations: Thinking representations induced by different languages show clear geometric separation consistently across model layers.The separation also appears in intermediate layers that may be considered relatively abstract and less language-specific.
  • Observations: Languages such as zh, fr, es, and de lie closer to English, whereas iw, bg, and tl are embedded farther away.Thus, languages occupy distinct thinking regions with varied distances to English.

4 Repeated Sampling under Multilingual Thinking

The study evaluates multilingual thinking through repeated sampling while constraining final responses to English for fair comparison. It introduces Single-Language Sampling and Mixed-Language Sampling to test within-region and cross-region diversity.

  • Controlled Output Setting: The controlled evaluation keeps final outputs in English while varying the language used for intermediate thinking.An English prefix after the thinking segment guides final-response language, and a sanity check assesses language control.
  • Single-Language Sampling: Single-Language Sampling repeatedly samples under one fixed thinking language and aggregates the resulting English outputs for diversity evaluation.
  • Mixed-Language Sampling: Mixed-Language Sampling samples repeatedly while controlling each intermediate-thinking process to a different language, then aggregates the English outputs.This strategy tests compositional effects from distinct language-induced thinking regions.
  • Evaluation: The resulting output sets are evaluated using the same diversity procedure, with Table 1 reporting Single-Language Sampling results across models and thinking languages.

5 How Does Language of Thought Shape Output Diversity?

Experiments across two benchmarks, multiple models, and 15 thinking languages show that multilingual thinking systematically increases output diversity. Mixed-Language Sampling adds compositional gains and expands diversity as sampling scales, while non-English thinking can involve a modest quality trade-off.

  • Experiment Settings: The study evaluates output diversity on NOVELTYBENCH and INFINITY-CHAT using Distinct Score, Similarity Score, and output-quality measures across multiple LLMs and 15 thinking languages.Distinct Score measures the ratio C/M, where C is the number of functional-equivalence classes among M outputs.
  • Single-Language Sampling: Switching from English to non-English thinking improves Distinct Score by 5.3–7.7 points and reduces Similarity Score by 1.04–2.56 points on NOVELTYBENCH.INFINITY-CHAT exhibits similar patterns.
  • Single-Language Sampling: Thinking languages vary substantially: iw on Qwen3-8B improves Distinct Score by 12.78 points over English, while languages farther from English tend to yield more diverse outputs.The reported Pearson correlation ranges from 0.72 to 0.88, and Spearman’s ρ ranges from 0.58 to 0.89 across models.
  • Output Diversity and Quality: Non-English thinking produces only a modest quality decrease overall, with no consistent pattern that the most diverse languages have the lowest quality.Some languages, including sv and oc, perform strongly on both diversity and quality.
  • Mixed-Language Sampling: Mixed-Language Sampling consistently outperforms English and average non-English Single-Language Sampling, often matching or exceeding the best single-language setting.It aggregates outputs from multiple distinct thinking regions without requiring prior knowledge of the best individual language.
  • Mixed-Language Sampling: Removing one language changes Distinct Score by 2.7% on average, but removing multiple languages causes rapidly growing, superlinear degradation.This pattern indicates complementary, compositional contributions rather than dependence on one dominant language.
  • Scaling and Sampling Parameters: As sampling number increases, Mixed-Language Sampling saturates more slowly than Single-Language Sampling, widening its advantage and expanding the model’s diversity ceiling.Increasing temperature also yields additional gains alongside language-of-thought changes, indicating a compositional effect between the two controls.

6 Application: Pluralistic Alignment

The paper tests Mixed-Language Sampling for pluralistic alignment using cultural-knowledge and cultural-values benchmarks. Across models and benchmarks, it achieves the strongest cultural pluralism performance and broader representation of knowledge and value orientations than comparison strategies.

  • Setup: The application targets cultural knowledge and cultural values using the BLEND and WVS multiple-choice datasets.Cultural pluralism is measured from repeated-sampling output distributions using entropy over countries or value-oriented responses.
  • Sampling Strategies: The comparison includes English Sampling, High Temperature, Request Diversity, Multilingual Prompting, and Mixed-Language Sampling, with M = 15 for every strategy.For Multilingual Prompting and Mixed-Language Sampling, each language is sampled once.
  • Results: Mixed-Language Sampling consistently achieves the highest cultural pluralism performance across benchmarks and models.The results indicate broader reflection of cultural knowledge and value orientations in LLM outputs.
  • Results: Increasing temperature, explicitly requesting diversity, or using multilingual inputs does not produce improvements comparable to Mixed-Language Sampling.This supports diversifying the language of thought as a practical strategy for pluralistic alignment.

7 Conclusion

Controlling the language of thought provides a structural source of LLM output diversity, with multilingual sampling improving diversity and broadening cultural coverage.

  • Switching from English to non-English thinking languages increases output diversity, with stronger gains for languages farther from English in thinking space.
  • Aggregating samples across multiple thinking languages yields additional diversity improvements through compositional effects.
  • Scaling sampling with linguistic heterogeneity expands the model’s diversity ceiling.
  • These findings broaden LLM coverage of cultural knowledge and values in pluralistic alignment.

8 Limitations

The paper identifies two limitations: unresolved effects of cross-lingual representation alignment and the use of output entropy as an abstraction of real-world cultural pluralism.

  • The study observes a positive distance–diversity correlation but does not test whether aligning non-English representations toward English reduces associated diversity.Answering this question would require controlled model interventions, left for future work.
  • Cultural-pluralism evaluation relies on output entropy, which abstracts real-world deployment scenarios.Real deployments may require alignment with multiple specific, context-dependent cultural values under explicit constraints.
  • The studied sampling strategies may require culturally contextualized language-of-thought routing for settings with explicit cultural constraints.The paper leaves this adaptation for future investigation.

A.1 Language Control Details

Language control uses translated prefixes to steer intermediate thinking into target languages while guiding final responses into English; language-identification checks largely confirm the intended separation.

  • Translated prefixes inserted after the <think> token guide intermediate thinking in one of 15 target languages.
  • An additional output-language control signal guides the model to think in a specified language while producing English responses.
  • Language identification predominantly detects target languages in thinking segments and English in output segments across NOVELTYBENCH and INFINITY-CHAT.Although identification may introduce noise, the results indicate that intended language-control signals are largely reflected in generated text.

A.2 Output Quality Evaluation Details

The paper evaluates output quality and cultural pluralism alongside diversity, using repeated-sampling protocols and entropy-based measures across multiple datasets and sampling baselines.

  • Additional Results: Single-Language Sampling generally improves diversity over English thinking, while several non-English languages maintain comparable output quality.The results show no clear systematic diversity–quality trade-off across languages.
  • Additional Results: Figure 6 compares Distinct Score with normalized thinking distance to English and reports Pearson’s r and Spearman’s ρ across languages.The figure uses Single-Language Sampling on INFINITY-CHAT; the accompanying results report strong positive correlations for most models.
  • Additional Results: Mixed-Language Sampling consistently outperforms English-only and non-English-average baselines on Similarity Score, sometimes matching or exceeding the best single-language setting.Its advantage does not require selecting one best-performing language.
  • Baselines: The Request Diversity baseline appends a request for a novel answer, while Multilingual Prompting translates each English question into 14 non-English languages.
Loading 2601.11227v2…