A paper accepted at ICLR 2026 reveals that the 'creativity' of diffusion models is not a random fluke but a mathematical consequence of interpolation effects caused by score smoothing in neural networks. This study demystifies the black-box nature of generative AI by providing a theoretical framework for how models construct complex structures from noise.
Termsslug: score-function
Score Function
得分函数
The gradient of the log probability density function, commonly used in diffusion models to guide denoising processes.
Curated Intelligence
Google ResearchT1·Models
68 pts
Hugging Face BlogT1·Models
68 pts
Allen AI introduces DiScoFormer, a transformer-based model that simultaneously estimates data distribution density and score. It eliminates the need for per-problem retraining, significantly reducing computational costs in high-dimensional scenarios while improving cross-distribution generalization.