发表机构
Toyota Research Institute; Technical University of Denmark; CAPeX Pioneer Center for Accelerating P2X Materials Discovery; University of Toronto; Vector Institute for Artificial Intelligence; NVIDIA; Massachusetts Institute of Technology; Acceleration Consortium（丰田研究所; 丹麦技术大学; CAPeX P2X材料发现加速先锋中心; 多伦多大学; 向量人工智能研究所; 英伟达; 麻省理工学院; 加速联盟）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究通过明确分子身份并将其通过生成过程拉回，揭示了三种分子生成架构的内部储备呈分段常数区域，其组织受多种因素影响，需先表征内部组织才能将生成空间视为可化学导航。
AI 中文摘要
针对物质的生成模型通常被评估为输出表示的采样器，其潜在空间常被用作化学空间导航的代理，但人们对这些模型如何在这些表示内部安排离散的化学身份知之甚少。我们通过明确分子身份并将其通过生成过程拉回，来研究这种安排。通过这些拉回操作，我们探查生成相同对象的区域，揭示了训练好的模型的内部储备：一个固定的分区，决定模型可以生成哪些对象（新的或非新的）。在三种分子生成架构中，我们发现该储备被组织成分段常数区域，由反复出现的从粗到细的边界分隔。其组织方式取决于所探查的表示、身份约定、解码器随机性以及用于比较坐标的度量。训练期间，局部化学组织趋于稳定，而每个邻域内表示的不同分子身份的数量仍在变化。因此，在将生成空间视为可化学导航之前，必须对内部组织进行表征，而不是假设。
英文摘要
Generative models for matter are often evaluated as samplers over output representations, and their latent spaces are commonly used as proxies for navigating chemical space. Much less is known about how these models internally arrange discrete chemical identities within those representations. We study this arrangement by making molecular identity explicit and pulling it back through the generative process. Through these pullbacks we probe the regions that generate the same object, exposing the trained model's internal repertoire: a fixed partition that determines which objects (novel or not) the model can produce. Across three molecular generative architectures, we find that this repertoire is arranged into piecewise-constant regions separated by recurring coarse-to-fine boundaries. Its organization depends on the representation probed, the identity convention, decoder stochasticity, and the metric used to compare coordinates. During training, local chemical organization stabilizes while the number of distinct molecular identities represented within each neighborhood continues to change. Internal organization must therefore be characterized, rather than assumed, before a generative space can be treated as chemically navigable.
Comments22 pages (14 main text + 8 supporting information)
