Amortized generative recovery advantage across problem families

Establish whether a configuration-importance model trained once across a family of fermionic Hamiltonian instances can complete the physical-sector tail of a new under-sampled instance without per-instance retraining and outperform a strong classical selector.

Background

The paper evaluates a generative tail-completer and an amortized configuration-importance model for recovering spectral-support configurations from under-sampled quantum data. A single-instance generative model does not show a statistically resolved improvement over a cheap classical CIPSI baseline.

An amortized model trained across a six-instance Hubbard family generalizes to held-out instances and matches the classical baseline, but does not yet demonstrate a statistically resolved advantage. The authors therefore leave unresolved whether amortized learning can deliver a genuine improvement, particularly in high-noise and high-correlation regimes.

References

Amortization without per-instance retraining is therefore feasible---a train-once model generalizes across the family---but shows \emph{no statistically resolved advantage} over the classical baseline yet, the status of the open question: the single-instance negative result sets the bar, the amortized result meets but does not clear it, and---consistent with both---any advantage, if it exists, would be confined to the high-noise, high-correlation regime, never on the easy instances where the classical prior already wins.

— Dynamical spectral functions from bitstring-sampled quantum subspaces: entanglement, not one-body magic, tracks the sampling cost  (2608.16436 - Vargas, 17 Aug 2026) in Section 6, Discussion, subsection “Machine learning that preserves, rather than removes, the quantum sampling role”