Unified Dynamics Framework for Neural System Training and Inference.

This research introduces a novel framework that unifies the understanding of training, learning, and inference dynamics within neural systems. It defines an "atomic generation fact" to record the origin, transformation, occurrence, result, and relational role of each generated output. These facts are compiled into a Generation-Fact Graph (GFG), which serves as an AI-native, compilable substrate for scientific inquiry, preserving generation histories and enabling a recursive scientific process of analysis, intervention, replay, and validation. Applying this GFG-based framework to nanoGPT, the study establishes unified dynamics. Training is characterized as the evolution of a parameter-optimizer system, where each training action produces a finite-amplitude nonlinear functional response. Learning is the subsequent persistent reorganization of distributed functional support by these responses, making capability formation, maintenance, decline, or recovery observable. The framework identifies three primary coordinates—target-boundary state, target-specific update geometry, and parameter-Adam receiving state—which yield a second-order predictor for post-update outputs, achieving over 91% accuracy and recall. Furthermore, inference is established as a frozen projection of these training-learning dynamics, with component gating and rollback revealing causal recruitment and non-additive combination of query-conditioned support. These findings are corroborated across ResNet/CIFAR-100 and diffusion/CIFAR-10 experiments.