Preprint reports neural model for larger quantum ground-state evaluations

Authors of a new arXiv preprint report Hamilton-Zero, a roughly 0.5-billion-parameter neural tensor-network model trained across hundreds of thousands of quadratic qubit Hamiltonians. They say they trained it on systems up to 64 qubits, fine-tuned it on held-out systems up to 1,024 qubits, and evaluated it on systems up to 8,100 qubits. The work represents quantum ground-state learning as optimization over functions on SU(2) rather than explicit Hilbert-space amplitudes.