Selected Languages are All You Need for Cross-lingual Truthfulness Transfer

TL;DR

The paper tackles the problem of truthfulness in multilingual LLMs by introducing FaMSS, a practical method for cross-lingual truthfulness transfer that relies on fact-aware multilingual data and a selective language strategy. It constructs MTruthfulQA to evaluate multilingual truthfulness and uses a language bias probe to identify a core subset of languages for efficient transfer via translation instruction tuning. Empirical results show FaMSS reduces multilingual representation disparity and improves cross-lingual truthfulness transfer across several base models and tasks, including gains on MTruthfulQA and Cross-MMLU, while demonstrating that selective language data outperforms indiscriminate multilingual mixing. The work highlights the value of targeted data selection and fact-aware content in multilingual alignment, though it also notes limitations in scope (e.g., context-free, English-centric setup) and outlines avenues for broader generalization and task types in future work.

Abstract

Truthfulness stands out as an essential challenge for Large Language Models (LLMs). Although many works have developed various ways for truthfulness enhancement, they seldom focus on truthfulness in multilingual scenarios. Meanwhile, contemporary multilingual aligning technologies struggle to balance numerous languages and often exhibit serious truthfulness gaps across different languages, especially those that differ greatly from English. In our work, we extend truthfulness evaluation to multilingual contexts and propose a practical method for cross-lingual truthfulness transfer called Fact-aware Multilingual Selective Synergy (FaMSS). FaMSS is able to select an optimal subset of all tested languages by language bias and transfer contributions, and then employ translation instruction tuning for cross-lingual truthfulness transfer. Experimental results demonstrate that our approach can effectively reduce the multilingual representation disparity and boost cross-lingual truthfulness transfer of LLMs.

Figures (9)

• Figure 1: FaMSS is able to align the multilingual capabilities of LLMs, making them more truthful in answering multilingual questions.
• Figure 2: Mean bias between languages. $dis_{x2x}$ represents the average distance value of all language pairs in one layer.
• Figure 3: Visualization results in the representation space of LLMs on FLORES-200 before and after utilizing FaMSS. Our method successfully enhances the multilingual representation alignment in the middle layers, where models encode more semantic-level information. In contrast, the top and bottom layers contain more syntactic information, resulting in little overlap in representations across different languages.
• Figure 4: Language bias movement when fine-tuned on language Y and probed on language X. All displayed values are calculated on the 14th layer of Gemma-7B. The value can also be formulated as $\mathcal{D}_s^Y[\text{"English"}][X]-\mathcal{D}_s[\text{"English"}][X]$.
• Figure 5: Performance on MTruthfulQA when fine-tuned with different language sets.