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

TL;DR

This paper introduces FaMSS, a method that enhances truthfulness in multilingual large language models by selecting optimal language subsets and employing translation instruction tuning, effectively reducing disparities across languages.

Contribution

The paper presents FaMSS, a novel approach for cross-lingual truthfulness transfer that balances multilingual representations and improves truthfulness in diverse languages.

Findings

Effective reduction of multilingual truthfulness gaps

Improved cross-lingual truthfulness transfer performance

Balanced multilingual representations achieved

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.