兰德：2025年对严格的GPAI模型评估的初步建议报告（英文版）

The science of GPAI evaluations is nascent and evolving,9 and consensus is lacking on how to robustly measure model capabilities and impacts amidst rapid technological advancement. With this in mind, we offer these suggestions as preliminary considerations within an ongoing conversation. The intended audience of this document is technical in nature and includes providers of general-purpose AI models presenting systemic risk (GPAISR),10 for whom the EU AI Act11 lays out specific evaluation req