As malware evolves and adapts, traditional detection systems struggle to identify novel and unseen threats. This challenge highlights the critical need for building generalizable models that can effectively detect unknown malware types. In this paper, we propose meta-learning as a tool to explore the adaptability of malware detection systems. Our approach focuses on understanding how much model updating is required to extend detection capabilities to previously unseen malware samples. By leveraging meta-learning, we aim to identify the most useful data for building generalizable models, optimizing the trade-off between data efficiency and detection accuracy. Through this investigation, we seek to provide insights into creating more robust and adaptable malware detection systems capable of addressing the constantly evolving threat landscape. Our results suggest that, among three popular representations of malware data, the combination of static and dynamic analysis reports is the most helpful in building generalizable models.