CrossCult-KIBench: A Benchmark for Cross-Cultural Knowledge Insertion in MLLMs

2026-05-07

AI Summary

This work addresses the challenge of culturally inappropriate responses in multimodal large language models (MLLMs), which predominantly rely on English-centric training data and thus struggle in cross-cultural contexts. The study introduces the novel task of “cross-cultural knowledge injection” and presents CrossCult-KIBench, the first fine-grained evaluation benchmark supporting English, Chinese, and Arabic, comprising 49 cultural visual scenarios and 9,800 image-text samples. It further proposes a new evaluation dimension that jointly considers target-culture appropriateness and preservation of non-target cultural behaviors. To tackle this task, the authors develop MCKI, an external memory retrieval method leveraging frozen MLLM representations and conditional prompting to inject culturally aligned image-text knowledge. Experimental results reveal that existing approaches fail to balance cultural adaptation with retention of original model behaviors, highlighting a key challenge in building culturally sensitive MLLMs.

Abstract

Multimodal Large Language Models (MLLMs), trained primarily on English-centric data, frequently generate culturally inappropriate or misaligned responses in cross-cultural settings. To mitigate this, we introduce the task of cross-cultural knowledge insertion, which focuses on adapting models to specific cultural contexts while preserving their original behavior in other cultures. To facilitate research in this area, we introduce CrossCult-KIBench, a comprehensive evaluation benchmark for assessing both the effectiveness of knowledge insertion and its unintended side effects on non-target cultures. The benchmark includes 9,800 image-grounded cases covering 49 culturally relevant visual scenarios across English, Chinese, and Arabic language-culture groups. It supports evaluation in both single-insert and sequential-insert settings. We also propose Memory-Conditioned Knowledge Insertion (MCKI) as a baseline method. MCKI retrieves relevant cultural knowledge from an external memory using frozen MLLM representations, prepending matched entries as conditional prompts when applicable. Extensive experiments on CrossCult-KIBench reveal that current approaches struggle to balance effective cultural adaptation with behavioral preservation, highlighting a key challenge in developing culturally-aware MLLMs. Our work thus underscores an important research direction for developing more culturally adaptive and responsible MLLMs.

Problem

• cross-cultural knowledge insertion
• Multimodal Large Language Models
• cultural adaptation
• behavioral preservation
• cross-cultural evaluation

Innovation

• cross-cultural knowledge insertion
• Multimodal Large Language Models
• CrossCult-KIBench
• Memory-Conditioned Knowledge Insertion
• cultural adaptation