Cognitive Priming Prompting Facilitates Knowledge Elicitation in Multilingual Large Language Models

Multilingual large language models (MLLMs) typically underperform when answering questions in non-native languages compared to their native language. Although existing translate-then-answer prompting methods partially alleviate this issue, their performance remains suboptimal compared to directly answering questions in the native language. Moreover, current studies lack a clear explanation for this gap. In this study, we attribute this issue to incomplete Cognitive Priming, a phenomenon observed from human cognition. And we point out that while existing methods achieve Language Priming (LP), they overlook Domain Priming (DP). To address this, we propose Cognitive Priming Prompting (CogPrim), which employs a Role-Enhanced Multi-MLLM Collaboration strategy to ensure both LP and DP, thereby improving knowledge elicitation for non-native QA tasks. Across five language question-answering benchmarks, CogPrim outperforms all state-of-the-art related methods. This approach contributes to advancing the understanding of human-like cognitive behaviors in MLLMs, fostering better MLLM service to a broader range of multilingual users.