Language Models Entangle Language and Culture

TL;DR

The paper addresses the problem that LLMs deliver uneven answer quality and cultural content across languages, which can disadvantage users of low-resource languages. It proposes a multilingual evaluation pipeline using generative, open-ended questions derived from WildChat and an LLM-as-Judge, supplemented by a translated CulturalBench subset to probe cultural knowledge. Findings show consistent quality gaps across languages, with language also shaping the cultural context of responses; larger and more specialized models exhibit tighter cross-language consistency, but language-dependent cultural cues persist. The work highlights the need for richer multilingual training data and methodologies to reduce language-driven biases and to better align LLM outputs with diverse cultural contexts in real-world use.

Abstract

Users should not be systemically disadvantaged by the language they use for interacting with LLMs; i.e. users across languages should get responses of similar quality irrespective of language used. In this work, we create a set of real-world open-ended questions based on our analysis of the WildChat dataset and use it to evaluate whether responses vary by language, specifically, whether answer quality depends on the language used to query the model. We also investigate how language and culture are entangled in LLMs such that choice of language changes the cultural information and context used in the response by using LLM-as-a-Judge to identify the cultural context present in responses. To further investigate this, we evaluate LLMs on a translated subset of the CulturalBench benchmark across multiple languages. Our evaluations reveal that LLMs consistently provide lower quality answers to open-ended questions in low resource languages. We find that language significantly impacts the cultural context used by the model. This difference in context impacts the quality of the downstream answer.

Figures (7)

• Figure 1: We show an example of our evaluation methodology. (1) Each query is translated to multiple languages. (2) We provide the translated and original English query to a LLM and a response is generated for each language. (3) Responses are scored out of 5 using a LLM as a Judge. (We show the responses translated to English for the visualization. Responses are evaluated in the original language itself.)
• Figure 2: Comparison of answer quality across languages by model evaluated using LLM as a Judge. The results show all models provide worse responses in at least one language and all models show the best performance in English.
• Figure 3: Results show the proportion of responses classified as each culture by language. X-axis shows the language of the query and Y-axis shows the culture to which the response was classified using LLM as a Judge.
• Figure 4: Accuracy on translated subset of CulturalBench by language and country for Qwen3-14b
• Figure 5: Comparison of different LLM-as-a-Judge configurations, evaluated on the basis of alignment with ground truth scores in terms of pearson correlation and Cohen's Kappa score. We note that addition of examples (denoted by 'ex' in the graph) lead to higher alignment, with 8 examples resulting in the highest alignment.