Calculating the lowest energy state of molecules is incredibly difficult because the number of possible electron arrangements grows exponentially, making it impossible to check them all. This arXiv preprint (https://arxiv.org/abs/2603.24728) introduces a Selected Configuration Interaction algorithm that uses auto-regressive neural networks to learn patterns from the ground state and intelligently pick the most promising configurations to explore, creating smaller but effective search spaces that reach accurate energies faster. The method was benchmarked on molecular systems, though the abstract provides no specific sample sizes or detailed limitations. As a preprint, the work has not yet been peer-reviewed. It aims to blend neural network learning with classical subspace techniques for use in both standard computers and hybrid quantum-classical systems.