研究人员开发了使用向量场进行约束生成模型的新方法。其中一种方法，逻辑引导向量场（LGVF），将符号逻辑集成到流匹配模型中，在生成过程中强制执行声明式约束，与标准方法相比，显著减少了约束违规。另一个框架，具有受控向量场的参数高效生成模型，通过学习到的标量控制来调制固定的向量场来构建富有表现力的流，提供了参数高效且可解释的替代方案。 AI
影响 约束生成模型和参数效率方面的这些进步可能导致更强大、更具可解释性的 AI 系统，能够遵守复杂的规则。
排序理由 该集群包含两篇 arXiv 论文，详细介绍了生成模型技术的新研究。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2602.02009v2 Announce Type: replace Abstract: Neuro-symbolic systems aim to combine the expressive structure of symbolic logic with the flexibility of neural learning; yet, generative models typically lack mechanisms to enforce declarative constraints at generation time. We…
arXiv:2605.28267v1 Announce Type: cross Abstract: We introduce a continuous-time generative modeling framework, motivated by the Chow-Rashevskii theorem, that builds expressive flows from a small set of fixed vector fields and learned scalar controls. Instead of learning an uncon…
We introduce a continuous-time generative modeling framework, motivated by the Chow-Rashevskii theorem, that builds expressive flows from a small set of fixed vector fields and learned scalar controls. Instead of learning an unconstrained high-dimensional vector field, our framew…