研究人员正在探索图神经网络 (GNN) 的表达能力，以解决复杂的优化问题。一篇论文表明，虽然标准 GNN 在线性半定规划 (SDP) 方面存在困难，但更具表达力的架构可以模拟求解器更新并实现显著的加速。另一项研究调查了具有全局读出的 GNN，表明它们可以捕获某些一阶属性，并确定了它们的表达能力与分级模态逻辑一致的条件。第三篇论文介绍了一种用于验证量化 GNN 的逻辑语言，证明了这种验证是可判定的，但计算上是棘手的，尽管量化模型轻巧且准确。 AI
影响 GNN 表达能力和验证方面的进步可能导致更高效、更可靠的优化和复杂数据分析的 AI 系统。
排序理由 该集群包含多篇 arXiv 预印本，讨论图神经网络的理论方面和验证。
AI 生成摘要 · Google Gemini · 来自 6 个来源。 我们如何撰写摘要 →
AI-assisted development tools enable rapid prototyping of services but often lack awareness of architectural constraints, infrastructure dependencies, and organizational standards required in production environments. Consequently, generated artifacts may exhibit brittle behavior …
arXiv:2604.27786v1 Announce Type: new Abstract: Semidefinite programs (SDPs) are a powerful framework for convex optimization and for constructing strong relaxations of hard combinatorial problems. However, solving large SDPs can be computationally expensive, motivating the use o…
Semidefinite programs (SDPs) are a powerful framework for convex optimization and for constructing strong relaxations of hard combinatorial problems. However, solving large SDPs can be computationally expensive, motivating the use of machine learning models as fast computational …
Semidefinite programs (SDPs) are a powerful framework for convex optimization and for constructing strong relaxations of hard combinatorial problems. However, solving large SDPs can be computationally expensive, motivating the use of machine learning models as fast computational …
arXiv:2604.22870v1 Announce Type: new Abstract: We study the expressive power of message-passing aggregate-combine-readout graph neural networks (ACR-GNNs). Particularly, we focus on the first-order (FO) properties expressible by this formalism. While a tight logical characterisa…
arXiv:2510.08045v2 Announce Type: replace-cross Abstract: We introduce a logical language for reasoning about quantized aggregate-combine graph neural networks with global readout (ACR-GNNs). We provide a logical characterization and use it to prove that verification tasks for qu…