Preference Tuning For Toxicity Mitigation Generalizes Across Languages

TL;DR

This paper demonstrates that preference tuning with English data can effectively reduce toxicity in multilingual LLMs across various languages, showing strong cross-lingual generalization and explaining it through mechanistic interpretability.

Contribution

It reveals that DPO preference tuning trained on English data generalizes to multiple languages and explains this phenomenon via the dual multilinguality of MLP layers.

Findings

Toxicity drops from 46.8% to 3.9% across 17 languages after DPO training.

DPO generalizes to models like BLOOM, Llama3, and Aya-23.

Bilingual sentence retrieval predicts cross-lingual transferability.

Abstract

Detoxifying multilingual Large Language Models (LLMs) has become crucial due to their increasing global use. In this work, we explore zero-shot cross-lingual generalization of preference tuning in detoxifying LLMs. Unlike previous studies that show limited cross-lingual generalization for other safety tasks, we demonstrate that Direct Preference Optimization (DPO) training with only English data can significantly reduce toxicity in multilingual open-ended generations. For example, the probability of mGPT-1.3B generating toxic continuations drops from 46.8% to 3.9% across 17 different languages after training. Our results also extend to other multilingual LLMs, such as BLOOM, Llama3, and Aya-23. Using mechanistic interpretability tools like causal intervention and activation analysis, we identified the dual multilinguality property of MLP layers in LLMs, which explains the cross-lingual generalization of DPO. Finally, we show that bilingual sentence retrieval can predict the cross-lingual transferability of DPO preference tuning.