This paper introduces a Generation-Fact Graph (GFG) framework unifying training, learning, and inference dynamics in neural systems, developing a predictive model using three coordinates — target-boundary state, update geometry, and parameter state — that reaches 91.43% accuracy predicting training transitions. The authors validate the approach across nanoGPT, ResNet/CIFAR-100, and diffusion models, reframing how neural networks reorganize computational support during training and how that constrains inference behavior.
