Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

A Hardware-Trained Generative Quantum Eigensolver for Sample-Based Quantum Diagonalization of Organotin Hydroxides up to 70 Qubits

Kunal Kumar, Kimberlee Keithley, Toshiya Iwamori, Shu Kanno, Ikko Hamamura, Takao Kobayashi, Akira Koseki, Qi Gao, Naoki Yamamoto, Hiroshi Horii, Yuri Kobayashi, Borja Peropadre, Kouhei Nakaji, Toru Imai, Takayuki Osogami

Published Nov 6, 2026• 0 citations
Publisher

Abstract

This dataset accompanies the manuscript "A Hardware-Trained Generative Quantum Eigensolver for Sample-Based Quantum Diagonalization of Organotin Hydroxides up to 70 Qubits". We trained amortized generative quantum eigensolver (GQE) policies—transformer models conditioned on molecular descriptors—to design the state-preparation circuits for sample-based quantum diagonalization (SQD). Training ran first on a noiseless simulator and then on the IBM Quantum processor ibm_kobe at 46 and 70 qubits. This record contains the zero-shot evaluation of the trained policies on the eleven active spaces of the organotin Sn(OH)3-Me family and its mono- and di-hydrate adducts (32 to 70 qubits), together with the training behavior of the policies and the data needed to verify every result of the evaluation. Two training configurations. The policies differ only in the reward structure of the hardware training stage: a Hartree–Fock penalty (which prevents the policy from producing only the reference determinant—the main configuration of the manuscript) and a reward whose weights were tuned jointly on a noisy simulator. Each configuration consists of four policies (two conditioning schemes, each trained for 46 and 70 qubits), evaluated across eleven active spaces, yielding 44 runs per configuration. What a run is. For each run, the policy generated ten gate sequences for an unseen molecule. The ten circuits were transpiled and measured on ibm_kobe with 100,000 shots each, without error mitigation. Each circuit was ranked by a one-shot diagonalization of its valid shots, and the best, second-best, merged, and two further selections were recovered with a 20,000-dimensional subspace across 25 iterations. Energies are compared against density-matrix renormalization group (DMRG) reference energies. What each report contains. Each report includes interactive HTML summaries of the training metrics, zero-shot evaluation benchmarks, and individual per-run details. Related material. Companion code repository: https://github.com/kunalroy332/sqd-gqe. The molecular active-space integrals and DMRG energies were provided by Mitsubishi Chemical Corporation. Deposit contents File / Directory Content report_hf_penalty.html Report for the Hartree–Fock-penalty configuration: training, evaluation of 44 trained runs, and per-run details. report_tuned_reward.html Report for the tuned-reward configuration: training, evaluation of 44 trained runs, cost analysis, and per-run details.

DOI
10.5281/zenodo.23995591

Related papers

A Hardware-Trained Generative Quantum Eigensolver for Sample-Based Quantum Diagonalization of Organotin Hydroxides up to 70 Qubits

Kunal Kumar, Kimberlee Keithley, Toshiya Iwamori +12 more

Quantum hardware can provide feedback for training circuit generators beyond the reach of exact classical simulation. We train generative quantum eigensolver (GQE) policies, transformers conditioned on molecular descriptors, to generate the state-preparation circuits for sample-based quantum diagonalization (SQD). Training runs first on a noiseless simulator and then with the IBM Heron processor ibm_kobe inside the loop, at 46 and 70 qubits ($1.2 \times 10^8$ shots for the four main training runs). A classical CCSD(T) reference energy of the training molecules enters only an admission gate and two auxiliary reward terms. During hardware training, a reward penalty prevents collapse to circuits that sample only the Hartree-Fock determinant. Policies trained on transition-metal complexes are applied without retraining to Sn(OH)$_3$-Me and its hydrates, minimal models of organotin photoresists for extreme-ultraviolet lithography, from 32 to 70 qubits. After one round of hardware sampling, followed by classical ranking and recovery of the cached counts with the molecular Hamiltonian and no further QPU calls, 27 of 28 runs at 46-70 qubits are within 15 mHa of the density-matrix renormalization group (DMRG) reference (median 4.5 mHa), and in all eight runs with an untrained control the trained policy is more accurate. At 56-70 qubits the selected circuits use 34-72 times fewer two-qubit gates than the local unitary cluster Jastrow (LUCJ) circuit transpiled for the same device. These results demonstrate that circuit generators can be trained with quantum hardware at 70 qubits.

arXiv:2611.13755•Nov 10, 2026•0 citations

The energy is the last thing to fail: properties, spin and controls for sample-based quantum diagonalization, from N₂ to an iron–sulfur cluster

Alain Chancé

The energy is the last thing to fail. Sample-based quantum diagonalization (SQD), the method behind the largest chemistry calculations on quantum processors, grades its results by the energy, the quantity least sensitive to errors in the state. This work asks what else the state gets right, on archived hardware data from two sources: the author's published N₂ run on IBM hardware, and IBM's archive for the 77-qubit iron–sulfur demonstration of 2025, the largest SQD calculation whose raw samples are public. Three density properties are followed alongside the energy against converged references: two stop improving while the energy still does; the third, set by the diffuse tail, oscillates while the energy descends smoothly. Four controls show whose work the energy is: random bitstrings seeded with five valid strings lag the hardware data by one iteration and match them at equal subspace size, a coupled-cluster prior matches or beats them there, and the device beats only the published control, uniform in-sector strings, whose loop has nothing to repair. The ordering repeats on the iron–sulfur cluster, where the spin is read alongside the energy and a penalty confirms the classical seed's lead as a singlet energy. Exact samples of the one-layer LUCJ circuit stall on a few hundred configurations: the noise feeds the repair. The recommendation is cheap: report a density property and the spin alongside the energy, at fixed subspace size, against a converged reference. Looking forward, exact samples of a time-evolved state broaden the support ten- to twentyfold over the LUCJ circuit, take the static loop to half a millihartree where LUCJ stalls, and, repaired from a noise channel, lead the hardware shots at equal subspace size at equilibrium and at the stretched bond; generative circuit compression would bring such a sampler inside the Nighthawk gate budget. v2 (2026-09-11): Table 10 gains the one-layer LUCJ circuit with optimized parameters, all-to-all and hardware pattern, as the rung between the coupled-cluster-parametrized circuit and the time-evolved state; §3.6.3 and §3.7.6 state that the configuration prior is coupled cluster's output, not the circuit's input. Changes in v3 (2026-09-13): the optimized-circuit rows of Table 10 restated from four converged optimizations (hardware pattern reproducible to chemical accuracy; three all-to-all optima, a perturbed start returning to one of them); a new §3.6.4 measuring the classical solver's envelope to fifty million determinants on a laptop (Table 12). Changes in v4: Table 10 gains the full-CI weight missing from each sample's string product space, 1 − W(U×U), and two full-CI yardstick rows; §3.6.3 reads it: the missing weight orders the static-loop error where configuration and string counts do not, and a millihartree needs less than about 3 × 10⁻⁴ missing. Changes in v5: a new §3.6.5, "Which root is the state", summarizes in one paragraph a measurement reported in full in a companion paper in preparation: at the stretched geometries the lowest root of a subspace closed under spin inversion is a spin mixture, which a DMRG-X-type overlap rule cannot repair; the abstract and §4 follow it. Table 10's hardware pattern read from the CCSD amplitudes is given as a range over ten builds, because round-off-level differences between coupled-cluster runs change that circuit; three optimizations "from the same starting state". Changes in v6: the companion paper’s reference updated; §3.6.5, the abstract and §4 reworded to match its results (a pure root is rare, not absent; the property statement scoped to the run of record); §3.6.1 cites van Rossum et al. and Reinholdt et al. for the mechanism and the concentration it measures; §3.7.3 cites Mostame et al.

•Sep 10, 2026•1 citation

SQD-Enabled Circuit Compression for Resource-Efficient Quantum Chemistry

Kangyu Zheng, Yidong Zhou, Jinglei Cheng +3 more

Sample-based Quantum Diagonalization (SQD) recovers ground-state energies by classically diagonalizing a Hamiltonian in the subspace spanned by quantum samples, requiring only bitstrings with sufficient ground-state overlap rather than an accurate variational energy. We reveal and exploit this underexplored robustness property: how much non-Clifford and variational expressivity can be removed from the sampling circuit before SQD accuracy degrades? We answer through two complementary compression techniques: gradient-based operator pruning, which discards low-impact excitation operators, and Clifford rounding, which snaps remaining parameters to the nearest Clifford angle. Both of these techniques can be applied to a VQE ansatz on a qubit-reduced Hamiltonian. A systematic ablation study across 21 molecules shows that median SQD error stays within chemical accuracy even at 50\% compression on both axes, while simulation speedup reaches $33\times$. Hardware validation on 6 molecules on IBM quantum hardware confirms up to $2.8\times$ transpiled-depth reduction with zero loss in SQD accuracy. Our implementation can be found at: https://github.com/zkysfls/cs-vqe-sqd

quant-pharXiv:2607.15076•Jul 17, 2026•0 citations

DF-SQD: Deterministic Fields for Sampling-Based Quantum Diagonalization

Kushagra Agarwal, Anupama Ray

Sampling-based quantum diagonalization method exploits Quantum-centric supercomputing platforms to sample bitstrings for Hamiltonian projection on a quantum computer, and then classically diagonalize the Hamiltonian to estimate the eigenvalues and eigenvectors. In current quantum devices an algorithm is useful when shallow quantum circuits with error mitigation support can discover better results while having either a proof of convergence or some method to explain trust in experiment. In this paper, we introduce DF-SQD, a hybrid algorithm that derives deterministic auxiliary-field circuits from selected double-factorization leaves of the two-electron tensor. The circuits propose occupation-number configurations, while selected configuration interaction evaluates the original active-space Hamiltonian and can recentre subsequent proposal rounds. On N2 (32 qubits; 6-31G basis) and a 40-qubit [Fe2S2(SCH3)4]2- active-space Hamiltonian, we show that DF-SQD improves the energy obtained from sampled determinant spaces while using shallow number-preserving circuits in both simulator and hardware runs. For N2, DF-SQD is 45x more accurate with a 11.23\% smaller subspace, and due to its ability to sample better bitstrings at lesser shots it is 2.93x faster than SQD in quantum devices. For the iron-sulfur cluster, DF-SQD generated a subspace dimension of 221M with 400K shots, while SQD needed 1.5M shots to generate a 238M subspace, thus we have better subspace recovery evident from the hardware at 3.75x reduced shots. At a matched 50M subspace dimension, DF-SQD is 1.32x more accurate (achieves a 24.5\% relative error reduction over standard SQD). So overall, our method is able to discover better results with shallower circuits, is sample efficient, uses configuration recovery (so has targeted error mitigation) and we have empirical convergence observation.

quant-pharXiv:2609.01264•Sep 2, 2026•0 citations

A Discrete Recovery Failure Mode in Sample-Based Quantum Diagonalization Carries Chemically Predictive Information

Takeshi Okuda

A previous report (Okuda, Zenodo DOI: 10.5281/zenodo.22469203) identified a discrete failure mode in sample-based quantum diagonalization (SQD): at low shot counts, classical configuration recovery can lock onto a reproducible but incorrect "wrong attractor," and injecting a single correct configuration into the raw sample can trigger a discrete escape from it. That finding was established on a shallow circuit (N2, CAS(6,6), 166 two-qubit gates) where the wrong-attractor structure itself was directly observable. Here we address two questions left open by that work. First, was the original demonstration merely a sample-size artifact rather than evidence that the specific injected configuration mattered? We answer this with a controlled random-injection experiment (Fe2S2, CAS(12,12), 2053 two-qubit gates, 76-80 independent trials per condition): recovery probability rises from 14.5% with no injection to 100.0% with one pre-specified escape configuration, while a random single-configuration injection of the same size produces only 15.0% recovery, and the escape itself is discrete while the residual energy error continues to improve smoothly with further injected configurations. Second, at circuit depths beyond roughly 1000 two-qubit gates, the wrong-attractor structure itself often becomes too noise-dominated to observe directly, so a practical recovery procedure cannot simply look for and inject "the" known wrong-attractor escape configuration. We present such a procedure: a small set of chemically motivated candidate configurations is generated and injected, and the best candidate is selected using one of three answer-free criteria (variational energy, escape rate from a known baseline attractor, or seed-to-seed energy variance), chosen by a cheap upfront classification of the target system. Across 23 independent hardware-data trials spanning two IBM Quantum backends and 21 distinct molecular systems and active-space combinations, this decision rule selected the correct candidate in all 23 trials, including one case improving accuracy by approximately 1000x. We report this as validation of the decision framework across a broad range of hardware and chemical conditions, not as evidence of answer-free prediction for arbitrary, genuinely unknown multi-reference chemistry, and report the framework's limitations candidly. Open-source implementation: https://github.com/okudat9/escape-config-predictor

•Sep 6, 2026•0 citations

Machine learning for sample-based quantum diagonalization: a review of generative configuration recovery and the classical-simulability frontier

Nicolás Bonilla Vargas

Sample-based quantum diagonalization (SQD), equivalently quantum-selected configuration interaction (QSCI), has become a centre of gravity of pre-fault-tolerant quantum chemistry: a processor samples electronic configurations and the Hamiltonian is diagonalized classically in the resulting subspace. Accuracy is governed entirely by which configurations enter it -- a machine-learning selection problem, made acute by a coupon-collector bottleneck. We review the generative and learned selectors by what each generates and the signal it exploits, and identify one gap: no reward-proportional generative-flow-network proposer has been built for tail discovery. On the field's central question -- whether the quantum sampler beats classical selected CI -- the negative verdict is not ours to claim: priority belongs to Reinholdt et al. [JCTC 21, 6811 (2025)], and polynomial-time classical estimation of the flagship circuits has reinforced it. We state that verdict at the precision a falsifiable claim requires -- it concerns reproducible, same-active-space comparisons on molecular electronic structure -- and weigh the claims outside those qualifiers. We show that alpha-string weights are not invariant under rotations inside degenerate orbital shells, so determinant counts are undefined until the orbital gauge is declared. We distil a ten-element benchmarking standard and apply it to our own deposit, which returned defects that changed numbers printed here and retracted one from v1. FCI-exact experiments confirm one prediction and refute another: the single generative advantage we find keeps no consistent sign along the dissociation coordinate at device-calibrated noise and reverses under a symmetric readout model. It does beat a noise-matched classical recovery loop on N2 by a margin five seeds cannot resolve, and loses by over a factor of two to a classical selector that needs no sampler.

quant-pharXiv:2608.05314•Aug 7, 2026•3 citations
This dataset accompanies the manuscript "A Hardware-Trained Generative Quantum Eigensolver for Sample-Based Quantum Diagonalization of Organotin Hydroxides up to 70 Qubits". We trained amortized generative quantum eigensolver (GQE) policies—transformer models conditioned on molecular descriptors—to design the state-preparation circuits for sample-based quantum diagonalization (SQD). Training ran first on a noiseless simulator and then on the IBM Quantum processor ibm_kobe at 46 and 70 qubits. This record contains the zero-shot evaluation of the trained policies on the eleven active spaces of the organotin Sn(OH)3-Me family and its mono- and di-hydrate adducts (32 to 70 qubits), together with the training behavior of the policies and the data needed to verify every result of the evaluation. Two training configurations. The policies differ only in the reward structure of the hardware training stage: a Hartree–Fock penalty (which prevents the policy from producing only the reference determinant—the main configuration of the manuscript) and a reward whose weights were tuned jointly on a noisy simulator. Each configuration consists of four policies (two conditioning schemes, each trained for 46 and 70 qubits), evaluated across eleven active spaces, yielding 44 runs per configuration. What a run is. For each run, the policy generated ten gate sequences for an unseen molecule. The ten circuits were transpiled and measured on ibm_kobe with 100,000 shots each, without error mitigation. Each circuit was ranked by a one-shot diagonalization of its valid shots, and the best, second-best, merged, and two further selections were recovered with a 20,000-dimensional subspace across 25 iterations. Energies are compared against density-matrix renormalization group (DMRG) reference energies. What each report contains. Each report includes interactive HTML summaries of the training metrics, zero-shot evaluation benchmarks, and individual per-run details. Related material. Companion code repository: https://github.com/kunalroy332/sqd-gqe. The molecular active-space integrals and DMRG energies were provided by Mitsubishi Chemical Corporation. Deposit contents File / Directory Content report_hf_penalty.html Report for the Hartree–Fock-penalty configuration: training, evaluation of 44 trained runs, and per-run details. report_tuned_reward.html Report for the tuned-reward configuration: training, evaluation of 44 trained runs, cost analysis, and per-run details.
Quantum hardware can provide feedback for training circuit generators beyond the reach of exact classical simulation. We train generative quantum eigensolver (GQE) policies, transformers conditioned on molecular descriptors, to generate the state-preparation circuits for sample-based quantum diagonalization (SQD). Training runs first on a noiseless simulator and then with the IBM Heron processor ibm_kobe inside the loop, at 46 and 70 qubits (1.2×1081.2 \times 10^81.2×108 shots for the four main training runs). A classical CCSD(T) reference energy of the training molecules enters only an admission gate and two auxiliary reward terms. During hardware training, a reward penalty prevents collapse to circuits that sample only the Hartree-Fock determinant. Policies trained on transition-metal complexes are applied without retraining to Sn(OH)3_33​-Me and its hydrates, minimal models of organotin photoresists for extreme-ultraviolet lithography, from 32 to 70 qubits. After one round of hardware sampling, followed by classical ranking and recovery of the cached counts with the molecular Hamiltonian and no further QPU calls, 27 of 28 runs at 46-70 qubits are within 15 mHa of the density-matrix renormalization group (DMRG) reference (median 4.5 mHa), and in all eight runs with an untrained control the trained policy is more accurate. At 56-70 qubits the selected circuits use 34-72 times fewer two-qubit gates than the local unitary cluster Jastrow (LUCJ) circuit transpiled for the same device. These results demonstrate that circuit generators can be trained with quantum hardware at 70 qubits.
The energy is the last thing to fail. Sample-based quantum diagonalization (SQD), the method behind the largest chemistry calculations on quantum processors, grades its results by the energy, the quantity least sensitive to errors in the state. This work asks what else the state gets right, on archived hardware data from two sources: the author's published N₂ run on IBM hardware, and IBM's archive for the 77-qubit iron–sulfur demonstration of 2025, the largest SQD calculation whose raw samples are public. Three density properties are followed alongside the energy against converged references: two stop improving while the energy still does; the third, set by the diffuse tail, oscillates while the energy descends smoothly. Four controls show whose work the energy is: random bitstrings seeded with five valid strings lag the hardware data by one iteration and match them at equal subspace size, a coupled-cluster prior matches or beats them there, and the device beats only the published control, uniform in-sector strings, whose loop has nothing to repair. The ordering repeats on the iron–sulfur cluster, where the spin is read alongside the energy and a penalty confirms the classical seed's lead as a singlet energy. Exact samples of the one-layer LUCJ circuit stall on a few hundred configurations: the noise feeds the repair. The recommendation is cheap: report a density property and the spin alongside the energy, at fixed subspace size, against a converged reference. Looking forward, exact samples of a time-evolved state broaden the support ten- to twentyfold over the LUCJ circuit, take the static loop to half a millihartree where LUCJ stalls, and, repaired from a noise channel, lead the hardware shots at equal subspace size at equilibrium and at the stretched bond; generative circuit compression would bring such a sampler inside the Nighthawk gate budget. v2 (2026-09-11): Table 10 gains the one-layer LUCJ circuit with optimized parameters, all-to-all and hardware pattern, as the rung between the coupled-cluster-parametrized circuit and the time-evolved state; §3.6.3 and §3.7.6 state that the configuration prior is coupled cluster's output, not the circuit's input. Changes in v3 (2026-09-13): the optimized-circuit rows of Table 10 restated from four converged optimizations (hardware pattern reproducible to chemical accuracy; three all-to-all optima, a perturbed start returning to one of them); a new §3.6.4 measuring the classical solver's envelope to fifty million determinants on a laptop (Table 12). Changes in v4: Table 10 gains the full-CI weight missing from each sample's string product space, 1 − W(U×U), and two full-CI yardstick rows; §3.6.3 reads it: the missing weight orders the static-loop error where configuration and string counts do not, and a millihartree needs less than about 3 × 10⁻⁴ missing. Changes in v5: a new §3.6.5, "Which root is the state", summarizes in one paragraph a measurement reported in full in a companion paper in preparation: at the stretched geometries the lowest root of a subspace closed under spin inversion is a spin mixture, which a DMRG-X-type overlap rule cannot repair; the abstract and §4 follow it. Table 10's hardware pattern read from the CCSD amplitudes is given as a range over ten builds, because round-off-level differences between coupled-cluster runs change that circuit; three optimizations "from the same starting state". Changes in v6: the companion paper’s reference updated; §3.6.5, the abstract and §4 reworded to match its results (a pure root is rare, not absent; the property statement scoped to the run of record); §3.6.1 cites van Rossum et al. and Reinholdt et al. for the mechanism and the concentration it measures; §3.7.3 cites Mostame et al.
Sample-based Quantum Diagonalization (SQD) recovers ground-state energies by classically diagonalizing a Hamiltonian in the subspace spanned by quantum samples, requiring only bitstrings with sufficient ground-state overlap rather than an accurate variational energy. We reveal and exploit this underexplored robustness property: how much non-Clifford and variational expressivity can be removed from the sampling circuit before SQD accuracy degrades? We answer through two complementary compression techniques: gradient-based operator pruning, which discards low-impact excitation operators, and Clifford rounding, which snaps remaining parameters to the nearest Clifford angle. Both of these techniques can be applied to a VQE ansatz on a qubit-reduced Hamiltonian. A systematic ablation study across 21 molecules shows that median SQD error stays within chemical accuracy even at 50\% compression on both axes, while simulation speedup reaches 33×33\times33×. Hardware validation on 6 molecules on IBM quantum hardware confirms up to 2.8×2.8\times2.8× transpiled-depth reduction with zero loss in SQD accuracy. Our implementation can be found at: https://github.com/zkysfls/cs-vqe-sqd
Sampling-based quantum diagonalization method exploits Quantum-centric supercomputing platforms to sample bitstrings for Hamiltonian projection on a quantum computer, and then classically diagonalize the Hamiltonian to estimate the eigenvalues and eigenvectors. In current quantum devices an algorithm is useful when shallow quantum circuits with error mitigation support can discover better results while having either a proof of convergence or some method to explain trust in experiment. In this paper, we introduce DF-SQD, a hybrid algorithm that derives deterministic auxiliary-field circuits from selected double-factorization leaves of the two-electron tensor. The circuits propose occupation-number configurations, while selected configuration interaction evaluates the original active-space Hamiltonian and can recentre subsequent proposal rounds. On N2 (32 qubits; 6-31G basis) and a 40-qubit [Fe2S2(SCH3)4]2- active-space Hamiltonian, we show that DF-SQD improves the energy obtained from sampled determinant spaces while using shallow number-preserving circuits in both simulator and hardware runs. For N2, DF-SQD is 45x more accurate with a 11.23\% smaller subspace, and due to its ability to sample better bitstrings at lesser shots it is 2.93x faster than SQD in quantum devices. For the iron-sulfur cluster, DF-SQD generated a subspace dimension of 221M with 400K shots, while SQD needed 1.5M shots to generate a 238M subspace, thus we have better subspace recovery evident from the hardware at 3.75x reduced shots. At a matched 50M subspace dimension, DF-SQD is 1.32x more accurate (achieves a 24.5\% relative error reduction over standard SQD). So overall, our method is able to discover better results with shallower circuits, is sample efficient, uses configuration recovery (so has targeted error mitigation) and we have empirical convergence observation.
A previous report (Okuda, Zenodo DOI: 10.5281/zenodo.22469203) identified a discrete failure mode in sample-based quantum diagonalization (SQD): at low shot counts, classical configuration recovery can lock onto a reproducible but incorrect "wrong attractor," and injecting a single correct configuration into the raw sample can trigger a discrete escape from it. That finding was established on a shallow circuit (N2, CAS(6,6), 166 two-qubit gates) where the wrong-attractor structure itself was directly observable. Here we address two questions left open by that work. First, was the original demonstration merely a sample-size artifact rather than evidence that the specific injected configuration mattered? We answer this with a controlled random-injection experiment (Fe2S2, CAS(12,12), 2053 two-qubit gates, 76-80 independent trials per condition): recovery probability rises from 14.5% with no injection to 100.0% with one pre-specified escape configuration, while a random single-configuration injection of the same size produces only 15.0% recovery, and the escape itself is discrete while the residual energy error continues to improve smoothly with further injected configurations. Second, at circuit depths beyond roughly 1000 two-qubit gates, the wrong-attractor structure itself often becomes too noise-dominated to observe directly, so a practical recovery procedure cannot simply look for and inject "the" known wrong-attractor escape configuration. We present such a procedure: a small set of chemically motivated candidate configurations is generated and injected, and the best candidate is selected using one of three answer-free criteria (variational energy, escape rate from a known baseline attractor, or seed-to-seed energy variance), chosen by a cheap upfront classification of the target system. Across 23 independent hardware-data trials spanning two IBM Quantum backends and 21 distinct molecular systems and active-space combinations, this decision rule selected the correct candidate in all 23 trials, including one case improving accuracy by approximately 1000x. We report this as validation of the decision framework across a broad range of hardware and chemical conditions, not as evidence of answer-free prediction for arbitrary, genuinely unknown multi-reference chemistry, and report the framework's limitations candidly. Open-source implementation: https://github.com/okudat9/escape-config-predictor
Sample-based quantum diagonalization (SQD), equivalently quantum-selected configuration interaction (QSCI), has become a centre of gravity of pre-fault-tolerant quantum chemistry: a processor samples electronic configurations and the Hamiltonian is diagonalized classically in the resulting subspace. Accuracy is governed entirely by which configurations enter it -- a machine-learning selection problem, made acute by a coupon-collector bottleneck. We review the generative and learned selectors by what each generates and the signal it exploits, and identify one gap: no reward-proportional generative-flow-network proposer has been built for tail discovery. On the field's central question -- whether the quantum sampler beats classical selected CI -- the negative verdict is not ours to claim: priority belongs to Reinholdt et al. [JCTC 21, 6811 (2025)], and polynomial-time classical estimation of the flagship circuits has reinforced it. We state that verdict at the precision a falsifiable claim requires -- it concerns reproducible, same-active-space comparisons on molecular electronic structure -- and weigh the claims outside those qualifiers. We show that alpha-string weights are not invariant under rotations inside degenerate orbital shells, so determinant counts are undefined until the orbital gauge is declared. We distil a ten-element benchmarking standard and apply it to our own deposit, which returned defects that changed numbers printed here and retracted one from v1. FCI-exact experiments confirm one prediction and refute another: the single generative advantage we find keeps no consistent sign along the dissociation coordinate at device-calibrated noise and reverses under a symmetric readout model. It does beat a noise-matched classical recovery loop on N2 by a margin five seeds cannot resolve, and loses by over a factor of two to a classical selector that needs no sampler.