Mitigating Cross-Lingual Cultural Inconsistencies in LLMs via Consensus-Driven Preference Optimisation

2026-04-02

AI Summary

This work addresses the issue that multilingual large language models often generate culturally inconsistent outputs when prompted in different languages, even with a fixed user persona—such as assigning distinct national cultural identities to the same British character. To mitigate this, the authors propose C-3PO, a consensus-driven preference optimization framework that enforces cross-lingual cultural consistency, enabling models to faithfully preserve specified personas while remaining culturally neutral across languages. They introduce Singleton Fleiss’s κ_S, a novel metric to quantify cross-lingual cultural inconsistency, and integrate preference optimization, representation intervention, and intermediate-layer decoding during training. Experiments demonstrate that C-3PO improves κ_S by up to 0.13 over strong baselines, with particularly pronounced gains in low-resource languages such as Indonesian and Persian.

Abstract

Despite their impressive capabilities, multilingual large language models (MLLMs) frequently exhibit inconsistent behaviour when the prompt's language changes. While such adaptation is generally desirable, it becomes a critical failure when a user's identity is explicitly defined. For instance, given a fixed British persona and an ambiguous everyday knowledge query about literature, the prompt's language frequently overwrites the system persona -- yielding Shakespeare in English but Cervantes in Spanish. To robustly quantify this Cross-lingual Cultural Inconsistency, we introduce Singleton Fleiss's $\kappa_S$, a metric mathematically resilient to hallucinations. For mitigation, we propose Cross-lingual Cultural Consistent Preference Optimisation (C-3PO), a consensus-driven alignment framework. C-3PO achieves up to a 0.10-point absolute increase in $\kappa_S$ over unaligned models, outperforming strong prompting and representation steering baselines. Empirical evaluations show this inconsistency disproportionately affects lower-resource languages like Indonesian and Persian. A layer-wise interpretability analysis reveals the underlying mechanism: by early-decoding intermediate layer representations, we find that MLLMs implicitly personalise outputs towards the prompt language's stereotypical culture as forward-pass representations stabilise.

Problem

• Cross-lingual Cultural Inconsistency
• Multilingual Large Language Models
• User Identity Preservation
• Cultural Bias
• Language-dependent Behavior

Innovation

• Cross-lingual Cultural Inconsistency
• Preference Optimisation
• Consensus-driven Alignment
• Singleton Fleiss's κ_S
• Multilingual LLMs