Beyond Pattern Recognition: A New Paradigm Uniting Symbolicism and Connectionism

This research challenges the conventional view of neural networks as mere statistical pattern recognizers, proposing and validating a new paradigm capable of awakening their endogenous capacity for exact reasoning. By employing programmatically generated ideal data and a parallel, non-autoregressive solving framework, this paradigm transforms standard neural networks from probabilistic mimics into deterministic rule executors. I powerfully demonstrate the universality of this paradigm through its success across a series of tasks, including symbolic rule execution, algorithmic fitting, visual reasoning, physics simulation, and interpretability. Furthermore, I systematically evaluate multiple representative architectures (MLP, CNN, RNN, UNET, Diffusion) on the task of multi-step evolution of a 1D cellular automaton, confirming that various non-Transformer models possess similar capabilities. This finding provides strong evidence that the ability for precise algorithmic execution is an inherent potential of connectionist systems, not the patent of a specific architecture. Through the Neural Sculpting Paradigm, this work, for the first time, systematically demonstrates that standard neural networks possess the intrinsic potential to execute arbitrary algorithms with precision, opening up a novel path toward building highly reliable and interpretable artificial general intelligence systems. Code is available at: https://github.com/ball-lightning6/neural-sculpting-paradigm