研究人员推出了 FactorDiff，一个用于离散扩散模型的新颖框架，通过将样本分解为更小的因子来增强组合生成。这种方法允许每个因子动态地路由到最相关的专家，从而提高泛化和推理能力。FactorDiff 在 ARC-AGI 基准测试上得到了验证，在需要逻辑一致性和空间解耦的任务上表现优于全局加权方案。 AI
影响 引入了一种新方法来提高离散扩散模型的推理和泛化能力。
排序理由 该集群包含一篇详细介绍离散扩散模型新方法的学术论文。
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
arXiv:2607.11758v1 Announce Type: new Abstract: Discrete diffusion models offer a powerful framework for solving complex reasoning tasks, particularly through compositional generation, which combines multiple pre-trained experts to generalize beyond their individual training data…
Discrete diffusion models offer a powerful framework for solving complex reasoning tasks, particularly through compositional generation, which combines multiple pre-trained experts to generalize beyond their individual training data. Recent theoretical corrections introduce time-…