LiveCLKTBench: Towards Reliable Evaluation of Cross-Lingual Knowledge Transfer in Multilingual LLMs

TL;DR

LiveCLKTBench is an automated benchmark that isolates and measures cross-lingual knowledge transfer in multilingual large language models, revealing factors influencing transfer effectiveness across languages.

Contribution

The paper introduces LiveCLKTBench, a novel pipeline for evaluating cross-lingual transfer, and provides new insights into how linguistic distance and model size affect transfer performance.

Findings

Cross-lingual transfer varies with linguistic distance.

Larger models improve transfer but with diminishing returns.

Transfer asymmetry depends on language pairs and domains.

Abstract

Evaluating cross-lingual knowledge transfer in large language models is challenging, as correct answers in a target language may arise either from genuine transfer or from prior exposure during pre-training. We present LiveCLKTBench, an automated generation pipeline specifically designed to isolate and measure cross-lingual knowledge transfer. Our pipeline identifies self-contained, time-sensitive knowledge entities from real-world domains, filters them based on temporal occurrence, and verifies them against the model's knowledge. The documents of these valid entities are then used to generate factual questions, which are translated into multiple languages to evaluate transferability across linguistic boundaries. Using LiveCLKTBench, we evaluate several LLMs across five languages and observe that cross-lingual transfer is strongly influenced by linguistic distance and often asymmetric across language directions. While larger models improve transfer, the gains diminish with scale and vary across domains. These findings provide new insights into multilingual transfer and demonstrate the value of LiveCLKTBench as a reliable benchmark for future research.