Accurate simulation of strongly correlated fermionic systems is limited by the combinatorial growth of configuration space. We propose a learning-guided reduced-subspace construction framework in which a reinforcement learning (RL) agent selects determinants under fixed budget constraints using only non-oracular, pre-solution structural features. The method is evaluated on exact-reference Fermi-Hubbard benchmarks and compared against random selection and a diagonal-energy (Diag-E) heuristic. Results show consistent improvement over random selection in overlap mass across all tested conditions, with gains ranging from ×1.7 to ×17.2. At small to moderate budgets, the RL approach reduces reduced-subspace energy error by 7–49%, with gains scaling with correlation strength. The Diag-E heuristic remains a strong competitor, particularly in energy, highlighting the importance of benchmarking against classical methods. The results demonstrate that non-oracular learning-guided determinant selection is physically meaningful and constitutes a viable optimization axis in hybrid quantum simulation workflows.
Nadal Vidal Jose Israel (Tue,) studied this question.