betaFeedback

WorksPhysics and Astronomy

Variational Relevance-Informed Sparsification of Term Measurements for Variational Quantum Algorithms

Abstract

Repeated measurements are a major bottleneck in variational quantum algorithms (VQAs). We uncover and characterize parameter-specific sparsity in the matrix of qubit-wise commuting (QWC) group contributions to parameter-shift gradients: most absolute contribution mass concentrates in a subset of groups whose relevance varies across parameters and optimization states. Building on this structure, we introduce a learned controller that predicts group relevance from Hamiltonian, circuit, light-cone, and current-state features, then selects measurements through a tunable coverage threshold. Exact contributions provide offline supervision without being required at deployment. Across 15 molecular VQE settings, our method saves 49.6% of nominal gradient-measurement shots with a mean absolute energy gap of 5.67 mHa to Full-QWC. We then ask whether large language models (LLMs) can solve the same selection problem without fine-tuning. The results reveal both promise and a clear limitation: LLMs can identify useful measurement groups, yet without guidance, local models do not consistently prioritize the physical information that matters most. Our learned controller uses these cues more effectively. This contrast points toward a broader direction for AI-assisted science: combining the flexibility of general-purpose models with explicit physical representations, task-specific objectives, and domain-grounded validation.

Introduction

The variational quantum eigensolver (VQE) is the flagship variational quantum algorithm (VQA) for estimating molecular ground-state energies ([33, 28]). Much of the research on VQAs has focused on trainability, particularly on mitigating barren plateaus (BPs) ([29]). Measurement overhead presents an equally fundamental challenge, yet has received comparatively less attention. Even with hardware-efficient circuits, each VQE iteration requires repeated estimation of Hamiltonian observables and gradients. For quantum chemistry, this measurement burden often dominates the practical cost ([27, 10]).

To identify where this measurement burden can be reduced, we first decompose the cost of a single gradient update. Because unknown quantum states cannot be cloned and quantum measurements are destructive ([40]), expectation values must be estimated through repeated circuit preparations. On this sampling-based interface, parameter-shift rules are widely used because they provide hardware-compatible analytic gradients ([34]). Let a qubit Hamiltonian contain GG QWC measurement groups and the ansatz PP trainable parameters. A parameter-shift gradient evaluates two shifted circuits per parameter and, under the standard Full-QWC estimator, measures all GG groups with SS shots, so one optimization step costs 2PGS2PGS circuit shots. Existing methods shrink one factor at a time: grouping compatible Pauli operators reduces GG ([9, 6]), adaptive shot allocation reduces SS ([25, 13, 44]), operator sampling subsamples the Hamiltonian ([4]), and parameter freezing reduces PP ([32]). None of these asks the finer question: when parameter jj is evaluated, which QWC groups are actually needed for its gradient?

The answer is strongly parameter dependent: every parameter sees the same Hamiltonian groups, yet the shifted expectation differences depend on gate location, ansatz light cone, and current circuit state, so the contribution matrix over parameter–group pairs is concentrated but not captured by any global coefficient ranking. Exploiting this structure faces a circularity: computing the matrix exactly at each iteration would require precisely the Full-QWC measurements we seek to avoid.

Our method resolves the circularity by learning to predict the latent matrix. Separate Transformer encoders contextualize parameter and QWC-group tokens; a pairwise scorer estimates a normalized relevance distribution for every parameter; a row-wise top-mass rule converts the prediction into a measurement mask. τ\tau thereby becomes a direct post-training accuracy-cost control; exact contributions are used only offline, never at deployment. Our contributions are threefold:

  • We uncover parameter-specific sparsity in the QWC contribution matrix of variational quantum gradients in Figure 1: most absolute contribution mass concentrates in a subset of measurement groups whose relevance varies across parameters and optimization states.

  • We develop a size-flexible learned controller that directly predicts parameter–group relevance from physical and current-state features, using exact contributions only as offline supervision. A single threshold converts these predictions into measurement subsets, enabling adjustable accuracy–cost trade-offs without retraining.

  • We investigate whether LLMs can perform this scientific selection task without fine-tuning. Frontier models achieve millihartree-scale gaps with retrieved training examples, while local-model audits reveal uneven utility of the supplied physical information. These findings demonstrate the potential of LLMs for scientific decision making and motivate domain-specific guidance as a promising route toward more reliable AI-assisted research.

Figure 1. Parameter-specific QWC sparsity. Cells show contribution-mass percentages in a 5×55 \times5 visualization of the exact parameter–group matrix; training and evaluation use the unaggregated matrix. The corresponding result under an IBM FakeGuadalupeV2 calibration-derived noise model ([19 b]) is reported in Appendix § 21, Figure 23.

Dominant groups vary by parameter, yet each row concentrates on a small fraction of the measurement set; this is the sparse structure our method exploits, and it persists under the calibrated-noise audit in Appendix § 21, indicating that it is not merely an artifact of idealized, noiseless execution.

Problem Formulation and Method

A molecular electronic-structure Hamiltonian mapped to qubits can be written as

H=h0I+∑ℓ=1LhℓPℓ,E(θ)=⟨ψ(θ)∣H∣ψ(θ)⟩,(1)H = h_0 I + \sum_{\ell=1}^{L} h_{\ell}P_{\ell}, \qquad E(\theta) = \langle\psi(\theta) \lvert H \rvert\psi(\theta) \rangle, \tag*{(1)}

where PℓP_{\ell} is a Pauli string and hℓh_{\ell} is fixed. Qubit-wise commuting strings share a tensor-product basis and form groups {Gk}k=1G\{\mathcal{G}_k\}_{k=1}^{G} ([6] (Verteletsky et al. 2020)).

For RY/RZ gates, the parameter-shift rule gives ([34, 39])

gj(θ)=∂E∂θj=12[E(θ+π2ej)−E(θ−π2ej)].(2)g_j(\theta) = \frac{\partial E}{\partial\theta_j} = \frac{1}{2}\left[E\left(\theta+ \frac{\pi}{2}e_j\right) - E\left(\theta- \frac{\pi}{2}e_j\right)\right]. \tag*{(2)}

Expanding each shifted energy over QWC groups exposes the signed contribution of group kk to jj:

Cj,k(θ)=12∑ℓ∈Gkhℓ[⟨Pℓ⟩θ+π2ej−⟨Pℓ⟩θ−π2ej],gj=∑k=1GCj,k.(3)C_{j,k}(\theta) = \frac{1}{2}\sum_{\ell\in\mathcal{G}_k} h_{\ell}\left[\langle P_{\ell}\rangle_{\theta+\frac{\pi}{2}e_j} - \langle P_{\ell}\rangle_{\theta-\frac{\pi}{2}e_j}\right], \qquad g_j = \sum_{k=1}^{G} C_{j,k}. \tag*{(3)}
Method overview in four stages

Figure 2. Method overview in four stages. (a) A Hamiltonian is partitioned into QWC groups and evaluated with a parameterized ansatz. (b) Full-QWC parameter shift measures every group for every parameter at cost 2PGS2PGS. (c) The learned controller predicts the parameter–group relevance matrix and retains the smallest per-parameter subset covering top-τ\tau mass. (d) Closed-loop VQE measures only those groups, forms sparse gradients, and updates all parameters at reduced cost.

Figure 2 summarizes our framework containing 4 stages. The coefficients and QWC partition stay fixed, but Cj,kC_{j,k} changes with the ansatz state. At an offline training snapshot, we evaluate Eq. (3) and normalize its absolute contribution,

wj,k(θ)=∣Cj,k(θ)∣∑r=1G∣Cj,r(θ)∣.(4)w_{j,k}(\theta) = \frac{\lvert C_{j,k}(\theta)\rvert}{\sum_{r=1}^{G}\lvert C_{j,r}(\theta)\rvert}. \tag*{(4)}

Exact CC supplies offline labels and simulator diagnostics; selectors never receive current oracle contributions. Separate Transformer encoders (Vaswani et al., [37]) contextualize parameter and group tokens (position, angle/time, coefficient, Pauli support, light cones). A pair scorer receives

zj,k=[pj,qk,pj⊙qk,∣pj−qk∣,xj,kpair],(5)z_{j,k} = [p_j, q_k, p_j \odot q_k, \lvert p_j-q_k\rvert, x_{j,k}^{\mathrm{pair}}], \tag*{(5)}

and predicts row-normalized relevance w^j,k\widehat{w}_{j,k}. Training combines distribution, top-set, ranking, budget, and entropy losses:

L=LCE+λmLmask+λrLrank+λbLbudget+λeLentropy.(6)\mathcal{L} = \mathcal{L}_{\mathrm{CE}} + \lambda_m\mathcal{L}_{\mathrm{mask}} + \lambda_r\mathcal{L}_{\mathrm{rank}} + \lambda_b\mathcal{L}_{\mathrm{budget}} + \lambda_e\mathcal{L}_{\mathrm{entropy}}. \tag*{(6)}

For each parameter, our method chooses the smallest set Sj(τ)S_j(\tau) satisfying ∑k∈Sj(τ)w^j,k≥τ\sum_{k\in S_j(\tau)}\widehat{w}_{j,k}\ge\tau.

Only groups in Sj(τ)S_j(\tau) are measured on the two shifted circuits; with SS shots per setting,

BFull=2SPG,BOurs=2S∑j=1P∣Sj(τ)∣,gain=1−BOursBFull.(7)B_{\mathrm{Full}} = 2SPG,\qquad B_{\mathrm{Ours}} = 2S\sum_{j=1}^{P}\lvert S_j(\tau)\rvert,\qquad\mathrm{gain}=1-\frac{B_{\mathrm{Ours}}}{B_{\mathrm{Full}}}. \tag*{(7)}

For fixed predictions, increasing τ\tau nests the selected masks and directly controls measurement coverage. In closed-loop optimization, each mask also shapes subsequent circuit states. The omitted signed contributions determine the truncation bias; Appendix M gives its bound, the complete losses, and implementation details. Appendix N.2 identifies the checkpoint used in each sweep. The formulation covers Pauli-observable VQAs; related work is in Appendix A.

Experimental Setup

PySCF/OpenFermion STO-3G Hamiltonians ([36, 30, 17]) for H2, HeH+, H3+, H4, and LiH map under Jordan–Wigner (Seeley et al., [35]) to 4–12 qubits and 5–179 QWC groups; the main ansatz is layered RY/RZ with linear entanglers (Kandala et al., [24]), and structural-OOD tests use named ansatze and spin Hamiltonians.

Training uses depths 1-3 and 8 seeds. Four held-out seeds run 28 Adam steps with exact masked expectations and nominal cost at 1024 shots/setting. We report ΔE=Emethod−EFull\Delta E = E_{\mathrm{method}} - E_{\mathrm{Full}} (Ha) and shot gain; no worse means ΔE≤1.6\Delta E \leq1.6 mHa. Metric and estimator definitions are in Appendix § 23.§ 4; 400-step shot-noise checks are in Appendix § 22.

Main Results

At τ=0.995\tau= 0.995, our method saves 49.6% of parameter-shift shots at a 5.67 mHa seed-level mean absolute gap, attains the smallest ∣ΔE∣|\Delta E| in eight of ten molecule–depth cells of Table 1, and satisfies the one-sided no-worse criterion in 8/10 cells, more than any alternative (Appendix § 12, Table 5). Tightening to τ=0.997\tau= 0.997 reduces the mean gap to 3.52 mHa while still saving 45.5%; the complete τ\tau frontier, including its behavior under depth shift, is in Appendix § 13 (Figure 8).

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+0.0011 ±\pm 0.0008+0.0015 ±\pm 0.0012+0.0005 ±\pm 0.0006−0.0006 ±\pm 0.0024+0.0023 ±\pm 0.0016
Depth 1ΔE\Delta E vs Full ↓\downarrowRandom+0.2359 ±\pm 0.0507+1.2220 ±\pm 0.2247+0.8916 ±\pm 0.1887+1.4117 ±\pm 0.2745+3.4982 ±\pm 0.0793
Depth 1ΔE\Delta E vs Full ↓\downarrowGlobal coeff.+0.0038 ±\pm 0.0053+0.0019 ±\pm 0.0041+0.0122 ±\pm 0.0093+0.0089 ±\pm 0.0074+0.0016 ±\pm 0.0026
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN†^{\dagger}+0.0128 ±\pm 0.0170+0.1299 ±\pm 0.0870+0.0772 ±\pm 0.0483−0.0642 ±\pm 0.0758−0.0648 ±\pm 0.1287
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN†^{\dagger}+Ours+0.0127 ±\pm 0.0160+0.1297 ±\pm 0.0845+0.0776 ±\pm 0.0483−0.0649 ±\pm 0.0752−0.0646 ±\pm 0.1286
Depth 1Shot gain ↑\uparrowOurs40.4 ±\pm 6.3%58.9 ±\pm 3.5%63.0 ±\pm 2.8%74.7 ±\pm 1.1%86.9 ±\pm 0.5%
Depth 1Shot gain ↑\uparrowRandom42.0 ±\pm 4.6%60.7 ±\pm 1.5%66.5 ±\pm 2.2%79.2 ±\pm 1.3%90.2 ±\pm 0.3%
Depth 1Shot gain ↑\uparrowGlobal coeff.41.7 ±\pm 6.5%60.0 ±\pm 3.3%62.9 ±\pm 2.9%75.0 ±\pm 0.3%87.0 ±\pm 0.6%
Depth 1Shot gain ↑\uparrowTITAN†^{\dagger}43.8 ±\pm 6.2%59.4 ±\pm 6.3%62.7 ±\pm 6.4%70.3 ±\pm 5.7%81.8 ±\pm 5.2%
Depth 1Shot gain ↑\uparrowTITAN†^{\dagger}+Ours66.8 ±\pm 3.7%91.7 ±\pm 2.0%95.0 ±\pm 2.7%97.9 ±\pm 0.9%99.9 ±\pm 0.1%
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs+0.0003 ±\pm 0.0007−0.0011 ±\pm 0.0026−0.0015 ±\pm 0.0041+0.0158 ±\pm 0.0443−0.0049 ±\pm 0.0097
Depth 2ΔE\Delta E vs Full ↓\downarrowRandom+0.1656 ±\pm 0.1105+0.4704 ±\pm 0.1737+0.6527 ±\pm 0.0767+0.9343 ±\pm 0.1414+4.0043 ±\pm 0.3648
Depth 2ΔE\Delta E vs Full ↓\downarrowGlobal coeff.+0.0015 ±\pm 0.0091−0.0004 ±\pm 0.0088+0.0233 ±\pm 0.0429+0.0379 ±\pm 0.0540−0.0049 ±\pm 0.0123
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN†^{\dagger}+0.0075 ±\pm 0.0093+0.0116 ±\pm 0.0467+0.0153 ±\pm 0.0618+0.0472 ±\pm 0.0558+0.2903 ±\pm 0.1308
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN†^{\dagger}+Ours+0.0072 ±\pm 0.0082+0.0129 ±\pm 0.0483+0.0140 ±\pm 0.0616+0.0493 ±\pm 0.0560+0.2920 ±\pm 0.1356
Depth 2Shot gain ↑\uparrowOurs25.6 ±\pm 2.9%31.6 ±\pm 2.6%44.9 ±\pm 1.3%51.8 ±\pm 0.5%82.6 ±\pm 0.4%
Depth 2Shot gain ↑\uparrowRandom26.3 ±\pm 2.5%33.2 ±\pm 1.5%47.7 ±\pm 2.1%53.7 ±\pm 0.4%83.4 ±\pm 1.3%
Depth 2Shot gain ↑\uparrowGlobal coeff.25.7 ±\pm 2.8%31.9 ±\pm 2.4%44.7 ±\pm 0.8%51.7 ±\pm 0.5%82.5 ±\pm 0.3%
Depth 2Shot gain ↑\uparrowTITAN†^{\dagger}23.8 ±\pm 4.7%39.0 ±\pm 14.5%41.9 ±\pm 9.4%46.8 ±\pm 9.9%82.7 ±\pm 5.3%
Depth 2Shot gain ↑\uparrowTITAN†^{\dagger}+Ours40.2 ±\pm 5.9%64.0 ±\pm 10.6%74.5 ±\pm 6.6%79.0 ±\pm 6.3%99.3 ±\pm 0.3%

Table 1. Cost-matched online VQE at τ=0.995\tau= 0.995. Exact-expectation results; shot gains are nominal. Entries are mean ±\pm standard deviation over four seeds; ΔE\Delta E is in Ha. Random and coefficient controls use comparable budgets; TITAN†^{\dagger} is cell-wise shot matched. Green/red shading indicates the one-sided 1.6 mHa criterion. Bold marks the lowest ∣ΔE∣|\Delta E| or largest gain; underlining marks the second-largest gain. Metrics, full controls, and noise extensions are in Appendices § 23.§ 4, § 12, and § 21.

Each control isolates one candidate explanation of the savings. Count-matched Random inherits our method’s per-parameter budgets and attains slightly larger shot gains in every cell, yet drifts by 0.17-4.00 Ha: the savings come from the budget size, while the accuracy comes from the learned ranking. Static coefficient ranking (natural training-free ordering) performs no worse in only 3/10 cells. Shot-matched TITAN†^{\dagger} freezing (Peng et al., [32]) drifts by tens to hundreds of mHa, so the gain is not explained by evaluating fewer parameters; conversely, TITAN†^{\dagger}+Ours achieves the largest gain, showing that relevance selection composes with, but does not repair, aggressive freezing.

The endpoint and trajectory views resolve complementary aspects of the trade-off. LiH depth 1 lies 0.7 mHa above the no-worse threshold, while H4 depth 2 (+15.8 ±\pm 44.3 mHa) exhibits strong seed dependence. Exact-expectation trajectories (Figure 7) reach Full-QWC parity at 74.4% (H2) and 55.1% (H3+) of its nominal budget. Separate 400-step runs with simulated shot noise (Appendix § 22) keep every molecule’s mean final gap within 0.18 mHa of Full-QWC while saving 24.7–92.1%.

Figure 7. Shots-to-chemical-parity dynamics at depth 2. Curves report the mean across four paired seeds for H2_2, H3+_3^+, an H4_4 chain, and LiH; shaded regions are 95% bootstrap confidence intervals. The horizontal axis is cumulative parameter-shift shots normalized by each seed’s Full-QWC budget. The vertical axis is the one-sided energy excess above the paired Full-QWC final energy, so lower energies incur no penalty; the dashed line marks 1.6 mHa chemical parity. Stars mark the first optimizer checkpoint where the four-seed mean reaches parity, while crosses mark methods that never reach it. Our method first reaches parity at 74.4% of the Full-QWC budget on H2_2 and 55.1% on H3+_3^+; it does not reach parity on H4_4 and finishes narrowly above it on LiH (1.77 mHa at 17.4% budget). Unlike Table 1, which compares final endpoints, this figure isolates convergence efficiency and shots-to-target.

The learned ranking is also not tied to one architecture: swapping the Transformer for message-passing GNN (Gilmer et al., [8]), MLP, or ResNet-18 (He et al., [16]) predictors (Appendix § 19) preserves the qualitative picture, while the Transformer has the lowest tail error on the measurement-intensive molecules and the strongest molecule-OOD transfer. Architecture details are in Appendix § 23.§ 3. The consistency across model families supports parameter-group relevance as a useful supervised learning target. The Transformer’s lower tail error further motivates its use as our default predictor, with robustness across molecular systems guiding the architecture choice.

Allocation and Structural Generalization

Separate transfer studies test four shift axes: Figure 3 shows our method beats count-matched Random on three of four unseen Hamiltonian families and on 44/45 transferable named-ansatz cells. TFIM at h=1.2h=1.2 and StronglyEntanglingLayers identify challenging structural shifts (Appendix H). Table 2 demonstrates composition with allocation policies: our method+variance-aware saves 56.8% and raises the no-worse rate from 72.5% to 82.5%. The common Adam protocol controls optimizer dynamics while comparing allocation and selection; exact policy settings are in Appendix N.1. Complete results are in Appendices F, E–H, D, and K; predictor diagnostics are in Appendix I.

Figure 3. Matched-budget structural OOD. Mean absolute gaps for unseen Hamiltonian families (a) and ansatz topologies (b); Random receives the same parameter-specific group counts as our method. Exact values and gains are in Appendix F.

Methodb=0.25b = 0.25 Mean ∣ΔE∣|\Delta E| ↓Shot gain ↑No-worse ↑b=0.50b = 0.50 Mean ∣ΔE∣|\Delta E| ↓Shot gain ↑No-worse ↑b=1.00b = 1.00 Mean ∣ΔE∣|\Delta E| ↓Shot gain ↑No-worse ↑
Full-QWC uniform0.000000.0%100.0%0.000000.0%100.0%0.000000.0%100.0%
iCANS ([25])0.17283−35.0%10.0%0.11977−49.8%37.5%0.09657−76.6%17.5%
gCANS ([13])0.2044992.9%7.5%0.1405992.2%25.0%0.1078991.8%27.5%
Rosalin ([4])0.06008−78.6%45.0%0.09137−90.9%72.5%0.06912−97.8%67.5%
Variance-aware ([44])0.066940.0%65.0%0.065210.0%80.0%0.062780.0%72.5%
Ours + uniform0.0547257.5%72.5%0.0606457.1%65.0%0.0637657.1%65.0%
Ours + variance-aware0.0721157.1%65.0%0.0687657.1%80.0%0.0603156.8%82.5%
Ours + gCANS0.1230097.3%17.5%0.1613497.2%27.5%0.0714897.2%30.0%

Table 2. Shot-allocation Pareto benchmark. Five molecules, depths 1–2, and four seeds under common Adam with simulated shot noise. ΔE\Delta E is in Ha; gain is relative to uniform Full-QWC at budget bb; no-worse means ΔE≤1.6\Delta E \le1.6 mHa. Bold/underline mark best/second-best non-Full values. Adapted-policy settings and detailed results are in Appendices N.1 and D.

LLM Controls and Mechanism Audit

Do the gains require task-specific learning, or can a capable general model be prompted into the same decisions? Figure 4 shows the shared interface: our method supplies each parameter’s cardinality kjk_j and structured physical features, a frozen zero-shot LLM selects the group identities, and the validated mask is reused for K=7K=7 steps. Matching refresh cadence controls invocation frequency. We evaluate four frozen local LLMs: Qwen3-8B ([41]), Llama-3.1-8B ([11]), DeepSeek-R1-Distill-Llama-8B ([14]), and GPT-OSS-20B ([1]). In Table 3, relative to per-step Ours-K=1K=1, the four-refresh Ours-K=7K=7 control increases the cross-cell mean absolute gap from 0.00567 to 0.00715 Ha while changing mean shot saving from 49.6% to 51.4%; stale masks do cost reliability (the no-worse count falls from 12/15 to 8/15 cells. Even so, Ours-K=7K=7 remains closer to Full-QWC than Qwen, Llama, and DeepSeek in all 15 cells, and closer than GPT-OSS in 13/15, at nearly identical aggregate savings. GPT-OSS is the strongest frozen LLM (0.0694 Ha mean absolute gap), followed by DeepSeek (0.172), Llama (0.421), and Qwen (0.786). Against the protocol-matched Random control, Qwen, Llama, DeepSeek, and GPT-OSS win 7/15, 11/15, 13/15, and 15/15 cells, respectively, while both proposed-method controls win 15/15: physics-conditioned prompting beats random choice, yet the aggregate LLM gap stays almost an order of magnitude above that of our refresh-matched method; the advantage comes from learned parameter-specific ranking, not from invocation frequency.

Clean frozen-LLM protocol diagram

Figure 4. Clean frozen-LLM protocol. Our learned controller supplies kjk_j and features; the LLM returns select/exclude QWC IDs; the mask is reused for K=7K=7.

SettingRanking methodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1Ours gain, K=1K=1 ↑40.4 ± 6.3%58.9 ± 3.5%63.0 ± 2.8%74.7 ± 1.1%86.9 ± 0.5%
Depth 1Ours gain, K=7K=7 ↑41.2 ± 5.3%60.7 ± 3.1%65.1 ± 2.6%77.3 ± 0.7%88.2 ± 0.4%
Depth 1Random (K=7K=7)+0.1847 ± 0.0868+1.2191 ± 0.2759+0.7247 ± 0.2856+0.9736 ± 0.3101+2.9030 ± 0.6058
Depth 1Qwen3-8B+0.1321 ± 0.0938 ✓+1.3928 ± 0.9006 ✗+0.9972 ± 0.1834 ✗+0.4765 ± 0.2133 ✓+2.0656 ± 0.7546 ✓
Depth 1Llama-3.1-8B+0.2840 ± 0.1350 ✗+1.2347 ± 0.3492 ✗+0.6195 ± 0.3195 ✓+0.4219 ± 0.3064 ✓+1.3266 ± 0.8740 ✓
Depth 1DeepSeek-R1-8B+0.2405 ± 0.1463 ✗+1.0993 ± 0.5373 ✓+0.2766 ± 0.1157 ✓+0.0522 ± 0.0653 ✓+0.0504 ± 0.0576 ✓
Depth 1GPT-OSS-20B+0.0471 ± 0.0463 ✓+0.2783 ± 0.2758 ✓+0.0568 ± 0.0867 ✓+0.0379 ± 0.0641 ✓+0.0018 ± 0.0032 ✓
Depth 1Ours (K=7K=7)+0.0013 ± 0.0003 ✓+0.0035 ± 0.0015 ✓+0.0004 ± 0.0006 ✓−0.0001 ± 0.0060 ✓+0.0030 ± 0.0016 ✓
Depth 1Ours (K=1K=1)+0.0011 ± 0.0008 ✓+0.0015 ± 0.0012 ✓+0.0005 ± 0.0006 ✓−0.0006 ± 0.0024 ✓+0.0023 ± 0.0016 ✓
Depth 2Ours gain, K=1K=1 ↑25.6 ± 2.9%31.6 ± 2.6%44.9 ± 1.3%51.8 ± 0.5%82.6 ± 0.4%
Depth 2Ours gain, K=7K=7 ↑27.3 ± 2.5%32.5 ± 2.4%47.9 ± 1.4%55.1 ± 0.5%83.7 ± 0.4%
Depth 2Random (K=7K=7)+0.0930 ± 0.0297+0.5668 ± 0.4463+0.3054 ± 0.1633+0.9327 ± 0.3273+2.6969 ± 0.2816
Depth 2Qwen3-8B+0.0992 ± 0.0689 ✗+0.8415 ± 0.1479 ✗+0.6751 ± 0.2741 ✗+0.1842 ± 0.0985 ✓+2.6722 ± 0.9356 ✓
Depth 2Llama-3.1-8B−0.0595 ± 0.1790 ✓+0.2644 ± 0.2592 ✓+0.6780 ± 0.0990 ✗−0.0543 ± 0.0502 ✓+0.4541 ± 0.2847 ✓
Depth 2DeepSeek-R1-8B+0.0051 ± 0.1945 ✓+0.0991 ± 0.0653 ✓+0.3239 ± 0.1195 ✗+0.0167 ± 0.0328 ✓+0.0332 ± 0.0342 ✓
Depth 2GPT-OSS-20B+0.0011 ± 0.0098 ✓+0.0049 ± 0.0044 ✓+0.0322 ± 0.0484 ✓+0.0167 ± 0.0328 ✓+0.0603 ± 0.1773 ✓
Depth 2Ours (K=7K=7)−0.0007 ± 0.0015 ✓+0.0076 ± 0.0176 ✓+0.0006 ± 0.0031 ✓+0.0162 ± 0.0438 ✓−0.0040 ± 0.0118 ✓
Depth 2Ours (K=1K=1)+0.0003 ± 0.0007 ✓−0.0011 ± 0.0026 ✓−0.0015 ± 0.0041 ✓+0.0158 ± 0.0443 ✓−0.0049 ± 0.0097 ✓
Depth 3Ours gain, K=1K=1 ↑18.1 ± 2.7%24.9 ± 1.3%29.9 ± 1.0%35.2 ± 0.9%75.4 ± 1.8%
Depth 3Ours gain, K=7K=7 ↑19.0 ± 2.4%26.1 ± 1.2%32.1 ± 1.2%38.6 ± 0.7%76.6 ± 1.3%
Depth 3Random (K=7K=7)+0.0730 ± 0.0859+0.2617 ± 0.2130+0.2750 ± 0.0765+0.2773 ± 0.2665+2.3928 ± 0.4840
Depth 3Qwen3-8B+0.0742 ± 0.0922 ✗+0.8192 ± 0.3035 ✗+0.3073 ± 0.1583 ✗−0.0046 ± 0.1338 ✓+0.9203 ± 0.3562 ✓
Depth 3Llama-3.1-8B+0.0631 ± 0.0983 ✓+0.3042 ± 0.1041 ✗+0.1583 ± 0.0876 ✓−0.0230 ± 0.0343 ✓−0.0493 ± 0.3344 ✓
Depth 3DeepSeek-R1-8B−0.0279 ± 0.0752 ✓+0.0302 ± 0.0339 ✓+0.0823 ± 0.0558 ✓−0.0093 ± 0.0205 ✓+0.0306 ± 0.0379 ✓
Depth 3GPT-OSS-20B+0.0436 ± 0.0577 ✓+0.1632 ± 0.1611 ✓+0.0091 ± 0.0196 ✓−0.0033 ± 0.0142 ✓−0.0392 ± 0.1522 ✓
Depth 3Ours (K=7K=7)+0.0010 ± 0.0027 ✓+0.0108 ± 0.0118 ✓+0.0029 ± 0.0095 ✓−0.0029 ± 0.0113 ✓−0.0046 ± 0.0040 ✓
Depth 3Ours (K=1K=1)+0.0005 ± 0.0015 ✓+0.0092 ± 0.0131 ✓+0.0005 ± 0.0077 ✓−0.0014 ± 0.0097 ✓−0.0017 ± 0.0036 ✓

Table 3. Table 3: Frozen local-LLM QWC ranking. Four zero-shot LLMs receive predicted counts and refresh every K=7K=7 steps. Entries are mean ± s.d. over four seeds (ΔE\Delta E: Ha; gain: %). Green/red shading indicates the 1.6 mHa no-worse split; ✓/✗ compares ∣ΔE∣|\Delta E| with Random. Bold/underline mark the smallest/second-smallest ∣ΔE∣|\Delta E| per cell. Protocols are in Appendices § 20 and § 24.§ 4.

The repair-free audit (Appendix § 20, Table 50) jointly changes the output contract and replaces coefficient completion with Full-QWC fallback after failed retries. Models retain 52.0–52.4% mean shot saving and beat Random in 13/15, 10/15, 11/15, and 7/15 cells (Qwen, Llama, DeepSeek, GPT-OSS), with mean absolute gaps of 0.4482–0.6986 Ha. Both proposed-method controls win 15/15. This second interface tests the same scientific selection task under a distinct validation policy. Protocol versions and direct-JSON inference settings are detailed in Appendix § 24.§ 4; Figure 12 retains the earlier schematic.

Controlled diagnostics for frozen local LLMs

Figure 12. Controlled diagnostics for frozen local LLMs. The schematic contrasts failure modes (red) and desired behavior (blue) for cardinality compliance, context awareness, parameter-specific QWC relevance ranking, and sparse-refresh consistency.

Table 50: Clean frozen local-LLM QWC-selection audit. This table duplicates the molecule–depth protocol and layout of Table 3, but replaces its repaired rank-only outputs with the stricter exact select/exclude contract. Each frozen LLM must emit unique QWC IDs at the requested per-parameter cardinality; invalid responses receive schema feedback for at most two retries, with no coefficient-based repair. Persistent failures use a logged Full-QWC safety fallback and pay the corresponding shots. Random uses the same per-parameter cardinality predicted by our method and K=7K=7 held-mask cadence, but samples group identities uniformly. All four models complete 60 online trajectories over seeds 97/101/107/109; their aggregate realized shot savings lie between 52.0% and 52.4%, versus 52.7% for Random. Energy entries are mean ±\pm standard deviation in Ha relative to Full-QWC, using exact-expectation gradients and nominal shot accounting (Appendix N.1). Random cells are intentionally unshaded. For every other method, green/red satisfies/violates the one-sided 1.6 mHa no-worse criterion (ΔE≤1.6\Delta E \le1.6 mHa; lower energy is accepted); deeper color indicates a lower energy/larger violation. The black check/cross separately indicates whether the absolute mean ΔE\Delta E is smaller/not smaller than Random. Bold/underline mark the smallest/second-smallest absolute mean ΔE\Delta E per cell. Figure 4 summarizes the audited protocol.

SettingSelection methodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1Ours gain, K=1K=1 ↑40.4 ± 6.3%58.9 ± 3.5%63.0 ± 2.8%74.7 ± 1.1%86.9 ± 0.5%
Depth 1Ours gain, K=7K=7 ↑41.2 ± 5.3%60.7 ± 3.1%65.1 ± 2.6%77.3 ± 0.7%88.2 ± 0.4%
Depth 1Random (K=7K=7)+0.1847 ± 0.0868+1.2191 ± 0.2759+0.7247 ± 0.2856+0.9736 ± 0.3101+2.9030 ± 0.6058
Depth 1Qwen3-8B+0.2012 ± 0.1085 ✗+0.9101 ± 0.4857 ✓+0.0200 ± 0.0203 ✓+0.9166 ± 0.5180 ✓+1.6190 ± 0.9510 ✓
Depth 1Llama-3.1-8B+0.3306 ± 0.0615 ✗+0.8140 ± 0.2673 ✓+0.5751 ± 0.2637 ✓+0.4804 ± 0.3100 ✓+0.6219 ± 0.5640 ✓
Depth 1DeepSeek-R1-8B+0.2245 ± 0.0696 ✗+1.0313 ± 0.3889 ✓+0.5259 ± 0.0514 ✓+0.6296 ± 0.1083 ✓+1.2793 ± 0.7197 ✓
Depth 1GPT-OSS-20B+0.3930 ± 0.1936 ✗+0.8238 ± 0.2303 ✓+0.7300 ± 0.4383 ✗+0.5671 ± 0.3164 ✓+1.5503 ± 0.7644 ✓
Depth 1Ours (K=7K=7)+0.0013 ± 0.0003 ✓+0.0035 ± 0.0015 ✓+0.0004 ± 0.0006 ✓+0.0001 ± 0.0060 ✓+0.0030 ± 0.0016 ✓
Depth 1Ours (K=1K=1)+0.0011 ± 0.0008 ✓+0.0015 ± 0.0012 ✓+0.0005 ± 0.0006 ✓−0.0006 ± 0.0024 ✓+0.0023 ± 0.0016 ✓
Depth 2Ours gain, K=1K=1 ↑25.6 ± 2.9%31.6 ± 2.6%44.9 ± 1.3%51.8 ± 0.5%82.6 ± 0.4%
Depth 2Ours gain, K=7K=7 ↑27.3 ± 2.5%32.5 ± 2.4%47.9 ± 1.4%55.1 ± 0.5%83.7 ± 0.4%
Depth 2Random (K=7K=7)+0.0330 ± 0.0297+0.5668 ± 0.4463+0.3054 ± 0.1633+0.9327 ± 0.3273+2.6969 ± 0.2816
Depth 2Qwen3-8B+0.0861 ± 0.0637 ✓+0.5614 ± 0.3196 ✓+0.0206 ± 0.0054 ✓+0.3825 ± 0.1254 ✓+2.7674 ± 0.5685 ✗
Depth 2Llama-3.1-8B+0.0184 ± 0.2098 ✓+0.8950 ± 0.1642 ✗+0.5715 ± 0.0934 ✓+0.2758 ± 0.1151 ✓+0.7222 ± 0.3456 ✓
Depth 2DeepSeek-R1-8B+0.1366 ± 0.1334 ✓+0.3298 ± 0.3566 ✓+0.5573 ± 0.1298 ✗+0.7876 ± 0.2497 ✓+2.4135 ± 0.3556 ✓
Depth 2GPT-OSS-20B+0.2148 ± 0.1220 ✓+0.6807 ± 0.7081 ✗+0.4691 ± 0.2024 ✓+0.5616 ± 0.2675 ✓+1.2894 ± 0.3290 ✓
Depth 2Ours (K=7K=7)−0.0007 ± 0.0015 ✓−0.0076 ± 0.0176 ✓+0.0006 ± 0.0031 ✓+0.0162 ± 0.0438 ✓−0.0040 ± 0.0118 ✓
Depth 2Ours (K=1K=1)+0.0003 ± 0.0007 ✓−0.0011 ± 0.0026 ✓−0.0015 ± 0.0044 ✓+0.0158 ± 0.0443 ✓−0.0049 ± 0.0097 ✓
Depth 3Ours gain, K=1K=1 ↑18.2 ± 2.7%24.9 ± 1.3%29.9 ± 1.0%35.2 ± 0.9%75.4 ± 1.1%
Depth 3Ours gain, K=7K=7 ↑19.0 ± 2.4%26.1 ± 1.2%32.1 ± 1.2%38.6 ± 0.7%76.6 ± 1.3%
Depth 3Random (K=7K=7)+0.0730 ± 0.0859+0.2617 ± 0.2130+0.2750 ± 0.0765+0.2773 ± 0.2665+2.3928 ± 0.4840
Depth 3Qwen3-8B−0.0435 ± 0.1539 ✓+0.1295 ± 0.1997 ✓+0.0721 ± 0.0496 ✓−0.0803 ± 0.1900 ✓+2.0654 ± 0.2623 ✓
Depth 3Llama-3.1-8B+0.0883 ± 0.2356 ✓+0.2766 ± 0.1085 ✓+0.2085 ± 0.1798 ✓+0.2084 ± 0.2366 ✓+0.2603 ± 0.2330 ✓
Depth 3DeepSeek-R1-8B+0.0220 ± 0.1312 ✓+0.2061 ± 0.0660 ✓+0.1788 ± 0.1528 ✓+0.3246 ± 0.1489 ✓+1.6082 ± 0.4131 ✓
Depth 3GPT-OSS-20B+0.1633 ± 0.1815 ✓+0.3438 ± 0.1193 ✓+0.1786 ± 0.0759 ✓+0.5089 ± 0.3596 ✓+0.5647 ± 0.4252 ✓
Depth 3Ours (K=7K=7)+0.0010 ± 0.0027 ✓+0.0108 ± 0.0118 ✓+0.0029 ± 0.0095 ✓+0.0029 ± 0.0113 ✓−0.0046 ± 0.0004 ✓
Depth 3Ours (K=1K=1)+0.0005 ± 0.0015 ✓−0.0092 ± 0.0131 ✓+0.0005 ± 0.0077 ✓−0.0014 ± 0.0097 ✓−0.0017 ± 0.0036 ✓

Table 50.

Data-exposed frontier models. We next test whether scale plus data access closes the gap: five GPT and four Claude variants receive two train-only same-molecule/depth/parameter neighbors and an eight-neighbor summary (retrieval details in Appendix § 24.§ 3). They choose both group IDs and counts, yielding model-specific accuracy–cost operating points. The held-out contribution matrix is never shown, tools are disabled, and an invalid response is charged as Full-QWC; all methods share the 28-step, K=7K=7 protocol and held-out seeds. Table 4 shows depth 1 with five GPT variants ((OpenAI, 2026)) and four Claude variants ([3]); Appendix § 20, Table 51–Table 52, give depths 2–3. Data-exposed frontier models do reach the mHa regime; the relevance structure is learnable in context, but the outcome is a cost–accuracy frontier rather than uniform dominance: Claude Fable 5 attains the best overall mean ∣ΔE∣|\Delta E| (0.00302 Ha) at 35.8% saving, Ours-K=7K=7 operates at a more aggressive 51.4% saving with 0.00715 Ha, and no model dominates every cell.

MethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Random (K=7K=7)+0.1847 ± 0.0868+1.2191 ± 0.2759+0.7247 ± 0.2856+0.9736 ± 0.3101+2.9030 ± 0.6058
GPT-5.4+0.0002 ± 0.0003 ✓−0.0007 ± 0.0017 ✓+0.0010 ± 0.0011 ✓+0.0021 ± 0.0015 ✓+0.0010 ± 0.0005 ✓
GPT-5.5+0.0001 ± 0.0007 ✓+0.0006 ± 0.0011 ✓+0.0008 ± 0.0009 ✓+0.0026 ± 0.0022 ✓+0.0017 ± 0.0012 ✓
GPT-5.6 Sol+0.0002 ± 0.0005 ✓−0.0007 ± 0.0023 ✓+0.0015 ± 0.0015 ✓+0.0002 ± 0.0023 ✓+0.0018 ± 0.0007 ✓
GPT-5.6 Terra−0.0000 ± 0.0007 ✓+0.0009 ± 0.0015 ✓+0.0041 ± 0.0050 ✓+0.0017 ± 0.0033 ✓+0.0016 ± 0.0005 ✓
GPT-5.6 Luna+0.0003 ± 0.0004 ✓+0.0001 ± 0.0010 ✓+0.0011 ± 0.0013 ✓+0.0035 ± 0.0024 ✓+0.0018 ± 0.0009 ✓
Claude Haiku 4.5−0.0001 ± 0.0004 ✓+0.0002 ± 0.0004 ✓+0.0006 ± 0.0010 ✓+0.0020 ± 0.0033 ✓+0.0011 ± 0.0007 ✓
Claude Sonnet 5+0.0000 ± 0.0004 ✓+0.0004 ± 0.0010 ✓+0.0003 ± 0.0010 ✓+0.0017 ± 0.0022 ✓+0.0016 ± 0.0005 ✓
Claude Opus 5+0.0002 ± 0.0004 ✓+0.0005 ± 0.0009 ✓+0.0004 ± 0.0009 ✓+0.0024 ± 0.0040 ✓+0.0015 ± 0.0008 ✓
Claude Fable 5+0.0003 ± 0.0007 ✓−0.0000 ± 0.0006 ✓+0.0004 ± 0.0010 ✓+0.0010 ± 0.0044 ✓+0.0011 ± 0.0005 ✓
Ours (K=7K=7)+0.0013 ± 0.0003 ✓+0.0035 ± 0.0015 ✓+0.0004 ± 0.0006 ✓+0.0001 ± 0.0060 ✓+0.0030 ± 0.0016 ✓
Random (K=7K=7)42.7 ± 3.8%61.3 ± 2.6%68.2 ± 2.5%80.2 ± 1.2%90.8 ± 0.1%
GPT-5.417.2 ± 1.0%46.7 ± 4.2%52.2 ± 3.7%64.6 ± 1.9%80.4 ± 0.3%
GPT-5.518.6 ± 1.6%45.4 ± 3.8%51.8 ± 3.3%70.4 ± 2.6%82.9 ± 2.2%
GPT-5.6 Sol18.1 ± 1.4%47.3 ± 4.5%53.3 ± 3.7%64.3 ± 1.2%80.7 ± 0.3%
GPT-5.6 Terra18.1 ± 1.0%46.7 ± 2.3%53.6 ± 2.5%64.1 ± 1.1%80.2 ± 0.4%
GPT-5.6 Luna19.4 ± 2.5%46.8 ± 4.1%52.3 ± 3.6%64.7 ± 1.0%80.8 ± 0.3%
Claude Haiku 4.518.1 ± 1.6%46.9 ± 3.3%52.0 ± 3.1%65.1 ± 1.8%81.3 ± 1.1%
Claude Sonnet 519.7 ± 1.4%43.8 ± 3.3%50.3 ± 3.1%69.6 ± 0.9%85.7 ± 0.6%
Claude Opus 517.2 ± 1.1%42.3 ± 3.1%44.7 ± 3.0%68.9 ± 2.2%81.9 ± 1.0%
Claude Fable 518.3 ± 1.5%44.3 ± 3.4%50.4 ± 3.0%67.2 ± 2.1%80.1 ± 0.5%
Ours (K=7K=7)41.2 ± 5.3%60.7 ± 3.1%65.1 ± 2.6%77.3 ± 0.7%88.2 ± 0.4%

Table 4. Data-exposed frontier models at depth 1. Train-only retrieval informs group IDs and counts under four-refresh K=7K=7. Entries are mean ± s.d. over four seeds. Green/red shading indicates the 1.6 mHa no-worse split; ✓/✗ compares ∣ΔE∣|\Delta E| with Random. Among non-Random methods, bold/underline mark best/second-best ∣ΔE∣|\Delta E| ↓ and gain ↑; gray marks largest gain. Protocols and full results: Appendices § 24.§ 3 and § 20.

MethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Energy fidelity ↓\downarrow: ΔE\Delta E vs. Full (Ha)
Random (K=7K=7)+0.0930 ±\pm 0.0297+0.5668 ±\pm 0.4463+0.3054 ±\pm 0.1633+0.9327 ±\pm 0.3273+2.6969 ±\pm 0.2816
GPT-5.4+0.0032 ±\pm 0.0064 ✓+0.0090 ±\pm 0.0164 ✓+0.0011 ±\pm 0.0056 ✓+0.0000 ±\pm 0.0098 ✓+0.0020 ±\pm 0.0041 ✓
GPT-5.5−0.0002 ±\pm 0.0021 ✓+0.0026 ±\pm 0.0057 ✓+0.0051 ±\pm 0.0077 ✓+0.0031 ±\pm 0.0178 ✓+0.0016 ±\pm 0.0042 ✓
GPT-5.6 Sol+0.0117 ±\pm 0.0135 ✓+0.0205 ±\pm 0.0326 ✓+0.0009 ±\pm 0.0074 ✓+0.0231 ±\pm 0.0383 ✓+0.0015 ±\pm 0.0046 ✓
GPT-5.6 Terra+0.0088 ±\pm 0.0150 ✓+0.0219 ±\pm 0.0367 ✓+0.0033 ±\pm 0.0062 ✓+0.0022 ±\pm 0.0063 ✓+0.0026 ±\pm 0.0055 ✓
GPT-5.6 Luna+0.0027 ±\pm 0.0057 ✓+0.0141 ±\pm 0.0188 ✓+0.0100 ±\pm 0.0099 ✓+0.0211 ±\pm 0.0385 ✓+0.0025 ±\pm 0.0057 ✓
Claude Haiku 4.5+0.0030 ±\pm 0.0056 ✓−0.0001 ±\pm 0.0013 ✓+0.0030 ±\pm 0.0070 ✓+0.0146 ±\pm 0.0514 ✓+0.0012 ±\pm 0.0028 ✓
Claude Sonnet 5+0.0003 ±\pm 0.0003 ✓−0.0001 ±\pm 0.0009 ✓+0.0016 ±\pm 0.0062 ✓−0.0039 ±\pm 0.0115 ✓+0.0043 ±\pm 0.0092 ✓
Claude Opus 5+0.0000 ±\pm 0.0005 ✓+0.0001 ±\pm 0.0003 ✓+0.0054 ±\pm 0.0073 ✓−0.0031 ±\pm 0.0186 ✓+0.0037 ±\pm 0.0082 ✓
Claude Fable 5+0.0086 ±\pm 0.0147 ✓+0.0001 ±\pm 0.0009 ✓+0.0017 ±\pm 0.0063 ✓+0.0077 ±\pm 0.0136 ✓+0.0015 ±\pm 0.0030 ✓
Ours (K=7K=7)−0.0007 ±\pm 0.0015 ✓+0.0076 ±\pm 0.0176 ✓+0.0006 ±\pm 0.0031 ✓+0.0162 ±\pm 0.0438 ✓−0.0040 ±\pm 0.0118 ✓
Measurement efficiency: shot gain (%; higher is better ↑\uparrow)
Random (K=7K=7)28.0 ±\pm 2.2%34.0 ±\pm 1.4%49.2 ±\pm 1.4%56.7 ±\pm 0.8%84.7 ±\pm 0.6%
GPT-5.413.0 ±\pm 0.3%21.4 ±\pm 1.3%29.6 ±\pm 1.0%39.2 ±\pm 1.2%72.0 ±\pm 1.8%
GPT-5.515.1 ±\pm 0.8%22.3 ±\pm 1.6%33.1 ±\pm 0.3%60.6 ±\pm 3.7%73.5 ±\pm 0.5%
GPT-5.6 Sol16.3 ±\pm 1.3%25.3 ±\pm 1.8%32.7 ±\pm 2.1%40.5 ±\pm 0.8%67.6 ±\pm 0.8%
GPT-5.6 Terra13.7 ±\pm 1.8%23.6 ±\pm 2.6%30.6 ±\pm 0.6%39.7 ±\pm 1.2%67.1 ±\pm 1.0%
GPT-5.6 Luna15.0 ±\pm 0.9%26.2 ±\pm 1.3%36.3 ±\pm 1.5%41.0 ±\pm 1.2%68.9 ±\pm 1.3%
Claude Haiku 4.513.0 ±\pm 1.6%21.6 ±\pm 1.0%32.2 ±\pm 1.2%41.6 ±\pm 1.9%68.2 ±\pm 1.0%
Claude Sonnet 511.6 ±\pm 0.6%21.3 ±\pm 1.4%29.5 ±\pm 0.9%57.6 ±\pm 4.4%77.4 ±\pm 4.4%
Claude Opus 512.4 ±\pm 1.0%20.7 ±\pm 1.0%28.9 ±\pm 0.7%54.0 ±\pm 3.6%76.0 ±\pm 2.1%
Claude Fable 514.6 ±\pm 2.2%21.1 ±\pm 1.5%29.4 ±\pm 0.9%29.0 ±\pm 6.8%66.9 ±\pm 1.0%
Ours (K=7K=7)27.3 ±\pm 2.5%32.5 ±\pm 2.4%47.9 ±\pm 1.4%55.1 ±\pm 0.5%83.7 ±\pm 0.4%

Table 51. Data-exposed frontier-model online VQE at depth 2. The train/test split, τ=0.995\tau= 0.995, four-refresh K=7K=7 protocol, and reporting conventions match Table 4.

MethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Random (K=7K=7)+0.0730±0.0859+0.0730 \pm0.0859+0.2617±0.2130+0.2617 \pm0.2130+0.2750±0.0765+0.2750 \pm0.0765+0.2773±0.2665+0.2773 \pm0.2665+2.3928±0.4840+2.3928 \pm0.4840
GPT-5.4+0.0023±0.0032+0.0023 \pm0.0032 ✓+0.0006±0.0016+0.0006 \pm0.0016 ✓−0.0018±0.0029-0.0018 \pm0.0029 ✓+0.0019±0.0031+0.0019 \pm0.0031 ✓+0.0018±0.0076+0.0018 \pm0.0076 ✓
GPT-5.5+0.0010±0.0036+0.0010 \pm0.0036 ✓−0.0003±0.0027-0.0003 \pm0.0027 ✓+0.0007±0.0036+0.0007 \pm0.0036 ✓+0.0134±0.0121+0.0134 \pm0.0121 ✓−0.0012±0.0122-0.0012 \pm0.0122 ✓
GPT-5.6 Sol+0.0014±0.0025+0.0014 \pm0.0025 ✓−0.0001±0.0052-0.0001 \pm0.0052 ✓+0.0017±0.0049+0.0017 \pm0.0049 ✓+0.0080±0.0056+0.0080 \pm0.0056 ✓+0.0012±0.0099+0.0012 \pm0.0099 ✓
GPT-5.6 Terra+0.0005±0.0018+0.0005 \pm0.0018 ✓+0.0004±0.0072+0.0004 \pm0.0072 ✓+0.0079±0.0100+0.0079 \pm0.0100 ✓+0.0057±0.0053+0.0057 \pm0.0053 ✓−0.0036±0.0046-0.0036 \pm0.0046 ✓
GPT-5.6 Luna−0.0010±0.0046-0.0010 \pm0.0046 ✓+0.0018±0.0044+0.0018 \pm0.0044 ✓+0.0028±0.0121+0.0028 \pm0.0121 ✓+0.0097±0.0088+0.0097 \pm0.0088 ✓+0.0037±0.0125+0.0037 \pm0.0125 ✓
Claude Haiku 4.5−0.0011±0.0050-0.0011 \pm0.0050 ✓+0.0012±0.0015+0.0012 \pm0.0015 ✓+0.0008±0.0039+0.0008 \pm0.0039 ✓+0.0631±0.1113+0.0631 \pm0.1113 ✓+0.0037±0.0101+0.0037 \pm0.0101 ✓
Claude Sonnet 5−0.0021±0.0020-0.0021 \pm0.0020 ✓+0.0001±0.0014+0.0001 \pm0.0014 ✓+0.0016±0.0011+0.0016 \pm0.0011 ✓+0.0131±0.0133+0.0131 \pm0.0133 ✓−0.0026±0.0122-0.0026 \pm0.0122 ✓
Claude Opus 5−0.0003±0.0009-0.0003 \pm0.0009 ✓−0.0005±0.0032-0.0005 \pm0.0032 ✓+0.0004±0.0035+0.0004 \pm0.0035 ✓+0.0095±0.0099+0.0095 \pm0.0099 ✓+0.0023±0.0109+0.0023 \pm0.0109 ✓
Claude Fable 5+0.0014±0.0022+0.0014 \pm0.0022 ✓+0.0001±0.0012+0.0001 \pm0.0012 ✓+0.0001±0.0009+0.0001 \pm0.0009 ✓+0.0019±0.0045+0.0019 \pm0.0045 ✓−0.0003±0.0039-0.0003 \pm0.0039 ✓
Ours (K=7K=7)+0.0010±0.0027+0.0010 \pm0.0027 ✓+0.0108±0.0118+0.0108 \pm0.0118 ✓+0.0029±0.0095+0.0029 \pm0.0095 ✓+0.0029±0.0113+0.0029 \pm0.0113 ✓−0.0046±0.0040-0.0046 \pm0.0040 ✓
Measurement efficiency: shot gain (%; higher is better ↑)
Random (K=7K=7)19.3±1.6%19.3 \pm1.6\%26.3±1.0%26.3 \pm1.0\%32.1±1.4%32.1 \pm1.4\%39.0±0.9%39.0 \pm0.9\%77.5±0.9%77.5 \pm0.9\%
GPT-5.47.7±1.5%7.7 \pm1.5\%13.9±1.3%13.9 \pm1.3\%19.6±0.8%19.6 \pm0.8\%27.2±2.3%27.2 \pm2.3\%58.6±1.9%58.6 \pm1.9\%
GPT-5.59.7±0.4%9.7 \pm0.4\%15.9±0.8%15.9 \pm0.8\%17.6±0.3%17.6 \pm0.3\%39.8±2.6%39.8 \pm2.6\%70.5±2.7%70.5 \pm2.7\%
GPT-5.6 Sol10.1±0.6%10.1 \pm0.6\%18.3±0.7%18.3 \pm0.7\%19.7±1.4%19.7 \pm1.4\%28.8±0.7%28.8 \pm0.7\%56.4±0.5%56.4 \pm0.5\%
GPT-5.6 Terra7.9±0.1%7.9 \pm0.1\%17.2±0.7%17.2 \pm0.7\%20.6±1.6%20.6 \pm1.6\%28.4±0.8%28.4 \pm0.8\%55.5±1.0%55.5 \pm1.0\%
GPT-5.6 Luna10.0±1.1%10.0 \pm1.1\%19.2±0.7%19.2 \pm0.7\%24.4±0.9%24.4 \pm0.9\%28.6±0.5%28.6 \pm0.5\%57.1±2.5%57.1 \pm2.5\%
Claude Haiku 4.57.6±1.3%7.6 \pm1.3\%14.1±1.1%14.1 \pm1.1\%19.1±0.9%19.1 \pm0.9\%29.0±0.9%29.0 \pm0.9\%56.6±0.8%56.6 \pm0.8\%
Claude Sonnet 57.5±0.4%7.5 \pm0.4\%12.5±0.2%12.5 \pm0.2\%18.0±0.9%18.0 \pm0.9\%28.1±7.3%28.1 \pm7.3\%70.6±1.4%70.6 \pm1.4\%
Claude Opus 57.3±0.8%7.3 \pm0.8\%14.7±1.0%14.7 \pm1.0\%21.2±0.9%21.2 \pm0.9\%36.9±3.5%36.9 \pm3.5\%71.3±1.2%71.3 \pm1.2\%
Claude Fable 59.8±0.5%9.8 \pm0.5\%13.5±0.8%13.5 \pm0.8\%17.2±0.6%17.2 \pm0.6\%19.1±1.2%19.1 \pm1.2\%55.8±1.7%55.8 \pm1.7\%
Ours (K=7K=7)19.0±2.4%19.0 \pm2.4\%26.1±1.2%26.1 \pm1.2\%32.1±1.2%32.1 \pm1.2\%38.6±0.7%38.6 \pm0.7\%76.6±1.3%76.6 \pm1.3\%

Table 52. Data-exposed frontier-model online VQE at depth 3. The train/test split, τ=0.995\tau= 0.995, four-refresh K=7K=7 protocol, and reporting conventions match Table 4.

Appendix Table 53 records 17.0–32.4M tokens per frontier model; our controller has 984,454 parameters and a 3.78-MB checkpoint. These are distinct resource measures; shot gain excludes offline training and classical costs (Appendix § 24.§ 1).

Prompt, context, and feature audit.

Which parts of the interface improve scientific selection? We freeze 12 Full-QWC snapshots and test the four local LLMs under 15 prompt, context, output, feature-deletion, and shuffle conditions (720 cases). The oracle matrix is audit-only, and cardinality is fixed. Figure 5 shows one observable Llama decision; the corresponding Qwen, DeepSeek, and GPT-OSS traces are in Appendix Figures 13, 14, and 15. These are observable outputs paired with external feature audits. The harness performs validation and VQE updates; frozen LLMs produce structured selection JSON through text inference with tool execution disabled.

Llama-3.1-8B observable decision trace

Figure 5. Llama-3.1-8B observable decision trace. On the shared H+^{+} snapshot, the explanation-prompt action captures 1.797% of the exact p14p_{14} contribution mass. Under the separate reference prompt, deleting coefficient fields raises local mass from 1.442% to 94.843%, identifying a harmful coefficient effect on that reference decision. Box (8) reports reference-prompt feature utility; prompt-specific actions and metrics are documented in Appendix § 24.§ 4.

The Qwen, DeepSeek, and GPT-OSS traces (Appendix Figures 13, 14, 15) capture 95.0%, 94.4%, and 93.0% mass, versus Llama’s 1.8%. Together, these cases reveal substantial model-specific difference within the same physical task. Appendix § 24.§ 4 links each action to its prompt and feature audit. Table 56 reports the 720-case study; Figure 16–Figure 20 resolve input-to-output attribution. At the same measurement budget, valid outputs can therefore differ sharply in their physical relevance.

Figure 16. Token-to-QWC-group attribution for one representative frozen-Qwen decision. The audited state is H3+H^{+}_{3} with STO-3G, depth 2, seed 97, iteration 0, and parameter p14p_{14}; our method supplies k14=6k_{14}=6, and Qwen emits [g0,g1,g2,g3,g4,g5g_{0}, g_{1}, g_{2}, g_{3}, g_{4}, g_{5}]. Panels (a,b) report the signed share of integrated-gradient attribution from each input group’s coefficient and parameter–group light-cone spans to each emitted ID: red raises and blue lowers the grammar-conditioned target score relative to the numeric-zero baseline. Panel (c) aggregates absolute attribution mass by feature family, while panel (d) divides that mass by the number of input tokens in each family. The observable action comes from the audit-panel Ollama qwen3:8b run; gradients are computed with the official frozen Qwen/Qwen3-8B checkpoint (Yang et al., [41]) in NF4 4-bit mode because the Ollama/GGUF runtime does not expose input gradients. This matched white-box surrogate is not bitwise identical to the generation runtime. The 32-point completeness audit has maximum residual 0.122; for the informative first decision g0g_{0}, this is 1.61% of its score change, but later columns have large relative errors. The plot attributes the teacher-forced output score; Appendix § 24.§ 5 reports the per-column numerical audit.

Figure 20. Model-native feature-to-QWC attribution for the trained controller. We freeze the saved Transformer checkpoint and use the same H3+H_{3}^{+}/STO-3G, depth-2, seed-97, iteration-0, parameter-p14p_{14} snapshot as in Figure 16–Figure 19. The six displayed columns are our method’s highest-scoring groups (g0,g14,g2,g7,g1,g4g_{0},g_{14},g_{2},g_{7},g_{1},g_{4}), fixing the LLM comparison cardinality; the native rounded budget head predicts five groups. Panels (a) and (b) trace coefficient and light-cone inputs from every QWC token to each scored group, (c) aggregates absolute attribution by feature family, and (d) normalizes that mass by the number of scalar features in each family. Attributions use 1,024-point Gauss–Legendre integrated gradients from the training-mean baseline in standardized feature space and include all parameter tokens, group tokens, and parameter–group pair features. The maximum completeness residual is 0.0126 log-probability units, with maximum relative residual 0.46%. This panel measures native feature-score dependence; the LLM panels measure token-score dependence under their respective baselines.

Figure 19. Token-to-QWC-group attribution for GPT-OSS-20B. The accepted reference-prompt action is [g0,g1,g2,g6,g7,g3][g_0, g_1, g_2, g_6, g_7, g_3] under the same snapshot and 32-point path. The action comes from Ollama gpt-oss:20b; gradients use the official frozen OpenAI GPT-OSS-20B native-MXFP4 checkpoint [1]. Flex Attention and one forward graph per output bound memory without changing the scalar scores or quadrature nodes. The runtime is not bitwise identical to Ollama, and the measured maximum completeness residual is 1.414 log-probability units; the full numerical audit appears in Appendix § 24.§ 5.

FactorInterventionΔ\Delta mass (pp) ↑\uparrowProtocol success ↑\uparrowMain observation
PromptMinimal instruction−1.0-1.0100.0%Shorter wording improves validity but not relevance quality.
PromptTwo demonstrations−4.8-4.897.9%Few-shot prompting does not provide a universal gain.
OutputDirect selected IDs+3.5+3.595.8%The strongest reliable output contract on average.
OutputFull ranking+7.9†+7.9^{\dagger}45.8%Long structured outputs are brittle and strongly survivor-biased.
Context8K tokens−1.2-1.260.4%Frequent context splitting and incomplete structured outputs.
Context16K tokens−2.4-2.481.2%Higher reliability than 8K, without a quality gain on valid cases.
Context32K reference0.097.9%Long context primarily prevents protocol failure.
PhysicsRemove coefficients−6.4-6.497.9%Coefficient information is useful overall, especially to Qwen and Llama.
PhysicsRemove light-cone features−4.9-4.993.8%Light-cone information is useful overall but strongly model-dependent.
PhysicsRemove θ\theta/iteration features+1.8+1.897.9%No stable aggregate benefit from current-state fields.
PhysicsShuffle coefficients−1.5-1.591.7%The largest shuffle effect is in Qwen rather than universal.
PhysicsShuffle light-cone features+2.3+2.393.8%Opposite model-level effects cancel in the aggregate.

Table 56.

How to read the trace. The upper half records the model-visible input, frozen inference, returned IDs, and schema validation under the explanation prompt. The lower half attaches exact contributions and probes feature utility under the reference prompt. These views answer complementary questions: what action the model returns, how much physical relevance that action retains, and how the reference selection responds to changes in its inputs.

Observed decision. Llama returns six unique, in-range IDs, {g7,…,g12}\{g_7,\ldots,g_{12}\}, with the requested cardinality and a valid schema. The action is directly executable, yet its captured mass is 1.797%. Because the measurement count is already fixed, the scientific distinction lies in the identities of the measured groups. The example makes the learned ranking task concrete: the controller must allocate the same number of measurements to the groups carrying the parameter’s actual contribution.

Physical information does not guarantee effective selection

  • Physical descriptors can help or mislead the same LLM: removing coefficients improves Llama’s local selection but reduces its average contribution coverage.

  • Removing light-cone or trajectory cues sharply degrades gradient alignment, showing that their importance extends beyond the contribution captured for a single parameter.

Task-specific relevance guides scientific decisions

  • Our controller learns which measurement groups matter for each parameter. Matched-budget comparisons support the value of this learned ranking for energy–cost performance.

  • These findings motivate expert-guided LLM workflows that prioritize task-relevant physical information and validate decisions through their scientific outcomes.

From one decision to closed-loop performance. Captured mass measures absolute contribution coverage for one parameter, while gradient cosine assesses alignment with the full gradient. Closed-loop evaluations assess the cumulative impact of repeated selections on final energy and measurement cost, extending the local trace analysis to complete optimization trajectories.

Across the feature audit. Over the paired audit, deleting coefficients or light-cone features reduces captured mass by 6.4 and 4.9 percentage points (Figure 22). Figure 21 relates these reference-prompt effects to separately reported tags. Positive and negative deletion effects distinguish beneficial and harmful feature utility, with model-specific responses to the same physical descriptors.

Figure 21. Self-reported reasons versus measured perturbation sensitivity. We audit three input families: coefficient denotes static Hamiltonian-coefficient summaries for each QWC group; light cone denotes parameter–group support and coefficient overlap; and trajectory denotes only the current optimization-state proxy, not a stored history. Specifically, for parameter jj it contains [sin⁡θj,cos⁡θj,tanh⁡θj,t/T][\sin\theta_j, \cos\theta_j, \tanh\theta_j, t/T], together with prompt-level iteration tt and total iterations TT. It does not include previous energies, gradients, Adam moments, measurement outcomes, or earlier masks. For each frozen LLM and feature family, the horizontal axis is the fraction of explained rows that name the corresponding reason code (the trajectory rate combines θ\theta and iteration tags). The vertical axis is the paired decrease in exact captured contribution mass when that entire feature family is removed from the same Full-QWC snapshot, without retraining or changing the requested cardinality. Positive values indicate beneficial input utility under the reference prompt; negative values indicate harmful utility. Near-zero effects may reflect redundancy. Reason tags come from a separate explanation prompt, giving a cross-prompt association between reported tags and measured feature utility. Black points are measured model–feature pairs and adjacent labels carry the model-family logo. The rank correlation is descriptive over only four models, with prompt provenance in Appendix § 24.§ 4.

Figure 22. Frozen LLMs use physical feature families differently. Each cell is the paired captured-mass drop, in percentage points, after removing one feature family from the 32K full-feature reference. Positive red values indicate useful dependence; negative blue values indicate that removal improves selection. The aggregate benefit is model dependent, and per-snapshot effects may reverse sign.

The learned controller uses optimization state. Our controller combines circuit/support information (63.24%) with a clear current-state contribution (12.16%) in Figure 6. Trajectory denotes [sin⁡θj\sin\theta_j, cos⁡θj\cos\theta_j, tanh⁡θj\tanh\theta_j, t/Tt/T]: the angle and normalized optimization step. Native-checkpoint attribution identifies their explicit use in QWC ranking, combining observable structure with the circuit’s current state. The fixed-snapshot audit in Appendix § 24.§ 6 corroborates consistently low trajectory attribution in all four frozen LLMs across quadrature and precision settings.

Figure 6. Feature-family attribution on a shared H3+_3^+ snapshot. Panels (a)–(d): frozen LLMs; (e): ours. Cells show within-model absolute IG mass. Despite receiving the full input feature set, frozen LLMs do not automatically make substantial use of trajectory information without targeted intervention: it ranks last throughout the numerical audit (Appendix § 24.§ 6). Our controller assigns 12.16% to current-state features. Scores and baselines: Appendix § 24.§ 5; full maps: Appendix Figures 16–20.

Context capacity and ranking quality play different roles. In Table 54, increasing capacity from 8K to 16K and 32K raises protocol success from 60.4% to 81.2% and 97.9%; valid-case captured mass is 0.562, 0.564, and 0.554. A minimal prompt reaches 100% success at 140.6k prompt tokens per case, versus 160.2k for the domain reference. Direct-ID output improves paired mass by 3.5 percentage points at 95.8% success (Table 56). Thus context capacity improves executable coverage, while physical ranking measures selection quality. The best tested GPT-OSS prompt yields ΔE=+0.5056±0.6562\Delta E = +0.5056 \pm0.6562 Ha at 52.4% saving online (Appendix Tables 55 and 57).

Ablation axisSettingCaptured mass ↑Grad. cosine ↑Success ↑Prompt kTok ↓Output kTok ↓Calls ↓
ReferenceDomain + compressed, 32K0.5540.62897.9%160.212.010.1
PromptMinimal + compressed, 32K0.5490.655100.0%140.68.79.0
PromptDomain + 2-shot, 32K0.5100.50997.9%189.113.410.8
OutputDomain + direct IDs, 32K0.5940.63395.8%169.813.211.5
OutputDomain + full ranking, 32K†0.6770.78945.8%114.616.512.5
ContextDomain + compressed, 8K0.5620.59260.4%56.75.98.0
ContextDomain + compressed, 16K0.5640.58381.2%113.510.29.5
ContextDomain + compressed, 32K0.5540.62897.9%160.212.010.1

Table 54. Prompt, output, and context ablations with token cost. Values are macro-averaged across Qwen3-8B, Llama-3.1-8B, DeepSeek-R1-8B, and GPT-OSS-20B on the 12 fixed audit snapshots per model. Token and call costs include retries, context splits, and invalid outputs, and are divided by all 48 attempted model–snapshot cases. Performance metrics are averaged over protocol-successful cases, whose identities can change across conditions; this is not an all-case matched accuracy comparison. Each ablation is compared with the first-row 32K reference; that reference is repeated in the context block for a direct 8K/16K/32K comparison. Within each numeric column, bold marks the best reported value and underlining marks the second-best distinct value, following the indicated arrow; tied values share the same formatting, and the repeated reference is counted once. These column-wise marks do not imply one overall-best protocol. The full-ranking row (†) succeeds in only 22/48 cases and is survivor-biased.

Long-horizon validation. Appendix § 22, Table 62, extends quantum optimization to 400 Adam steps with simulated shot noise and frozen predictor weights. All 15 molecule–seed pairs reach within 1.6 mHa of their attainable exact-Adam reference at a checkpoint. Each molecule’s mean final energy stays within 0.18 mHa of Full-QWC, with 24.7–92.1% shot savings. FCI anchors the absolute energy scale; the paired Full-QWC comparison isolates sparse measurement under the same ansatz. Savings persist across the long-horizon optimization.

MoleculeMethodExact floor (mHa)Final FCI gap (mHa)ΔE\Delta E to exact ref. (mHa)Shot savingParity
H2Full-QWC20.283 ± 0.30320.672 ± 0.106+0.389 ± 0.3500.0 ± 0.0%3/3
Random, same budget20.283 ± 0.30321.027 ± 0.222+0.744 ± 0.30526.3 ± 1.7%3/3
TITAN, matched cost20.283 ± 0.30320.238 ± 0.718−0.045 ± 0.41526.0 ± 1.9%3/3
Ours20.283 ± 0.30320.673 ± 0.072+0.390 ± 0.37024.7 ± 1.0%3/3
HeH+Full-QWC9.459 ± 0.0169.613 ± 0.159+0.154 ± 0.1580.0 ± 0.0%3/3
Random, same budget9.459 ± 0.0169.663 ± 0.057+0.204 ± 0.07229.4 ± 0.4%3/3
TITAN, matched cost9.459 ± 0.0169.446 ± 0.206−0.014 ± 0.19730.5 ± 0.2%3/3
Ours9.459 ± 0.0169.499 ± 0.111+0.040 ± 0.10329.8 ± 0.9%3/3
H3+Full-QWC11.939 ± 0.00012.415 ± 0.187+0.477 ± 0.1870.0 ± 0.0%3/3
Random, same budget11.939 ± 0.00013.127 ± 0.699+1.188 ± 0.69956.8 ± 0.8%3/3
TITAN, matched cost11.939 ± 0.00012.220 ± 0.019+0.281 ± 0.01958.5 ± 1.1%3/3
Ours11.939 ± 0.00012.274 ± 0.134+0.335 ± 0.13457.1 ± 1.3%3/3
H4 chainFull-QWC51.371 ± 0.00051.981 ± 0.034+0.610 ± 0.0340.0 ± 0.0%3/3
Random, same budget51.371 ± 0.00053.116 ± 0.897+1.745 ± 0.89776.7 ± 0.3%2/3
TITAN, matched cost51.371 ± 0.00052.075 ± 0.161+0.704 ± 0.16177.6 ± 0.1%3/3
Ours51.371 ± 0.00051.849 ± 0.100+0.479 ± 0.10076.9 ± 0.7%3/3
LiHFull-QWC18.415 ± 0.00019.921 ± 0.437+1.507 ± 0.4370.0 ± 0.0%3/3
Random, same budget18.415 ± 0.000728.571 ± 144.340+710.156 ± 144.34092.0 ± 0.1%0/3
TITAN, matched cost18.415 ± 0.00019.673 ± 0.646+1.259 ± 0.64692.3 ± 0.0%3/3
Ours18.415 ± 0.00020.098 ± 0.455+1.684 ± 0.45592.1 ± 0.2%3/3

Table 62. Table 62: Long-horizon shot-noise-simulated VQE with an FCI-anchored diagnostic. Gradients use Gaussian sampling from exact grouped-observable variances, not physical-QPU execution. All methods use the same depth-3 layered RY/RZ ansatz, three paired held-out seeds, 400 Adam steps, and 4096 shots per measured shifted QWC setting. Our method uses τ=0.997\tau= 0.997; Random receives its per-parameter group count, and TITAN [32] uses privileged exact-gradient ranking with the nearest Full-QWC parameter cost. The particle-number-penalized objective is H+(N−Ne)2H + (N - N_e)^2, while reported energies are those of HH. FCI and exact statevector energies are diagnostics only and are never supplied to the finite-shot optimizer or stopping rule. Exact floor is the best exact-gradient Adam gap to FCI. Parity counts seeds reaching within 1.6 mHa of their exact-Adam reference at any checkpoint. No method reaches FCI chemical accuracy because the depth-3 variational floor remains 9.459–51.371 mHa above FCI. Green/red final-error and parity entries satisfy/violate the attainable-reference criterion. Within each molecule, bold marks the best final FCI gap, absolute final error, or shot saving; underline marks the corresponding runner-up. Shot savings remain black.

Takeaway. Figure 6 suggests that the learned controller better exploits trajectory information, while the tested frozen LLMs struggle to prioritize these cues without explicit human guidance. This motivates expert-guided scientific workflows that help LLMs recognize and effectively use information critical to the task. This raises an interesting question: What happens if we explicitly guide LLMs to use trajectory information? We will explore this question across a broader range of quantum tasks in future work.

Conclusion

In this work, we uncover parameter-specific sparsity in QWC contributions to variational quantum gradients and develop a compact learned controller that exploits this structure for adjustable accuracy–cost trade-offs. Our evaluation of LLMs without fine-tuning reveals both capabilities and limitations: frontier models with retrieved training examples reach millihartree-scale gaps to Full-QWC, while local-model audits reveal uneven utility of the supplied physical information. These findings motivate AI-assisted scientific workflows in which human domain expertise guides problem representation, task objectives, and physical validation. By characterizing model capabilities and failure modes in a concrete scientific task, this work provides an important first step toward developing and evaluating such expert-guided workflows.

AI Use Statement

In this work, we used generative AI tools for research ideation and feedback on methodology and experimental design, assistance with implementation, literature discovery, and editing the manuscript for clarity. The LLMs evaluated as measurement-selection baselines and their protocols are described separately in the paper. The authors reviewed AI-assisted material, checked reported numerical claims against experimental artifacts, and take full responsibility for the final text, claims, and research artifacts.

Ethics Statement

This study evaluates measurement selection in simulated quantum optimization using computationally generated Hamiltonians and structured physical features. It does not involve human participants or personal data.

Reproducibility Statement

Appendices M–N specify implementations, metrics, and protocol provenance; preceding appendices retain all experimental results.

Extended Related Work

Measurement-frugal VQE. Commuting partitioning, unitary partitioning, and compatible-observable measurement reduce the number of bases used to estimate the Hamiltonian [43, 22, 42]; classical-shadow variants instead reuse randomized measurements across many observables [18, 15]. A complementary line allocates a finite shot budget according to gradient or observable variance [25, 13, 44], while operator sampling constructs stochastic gradient estimates from selected Hamiltonian components [4]. These approaches optimize the global grouping, the number of shots, or a sampled operator distribution. Our method operates after a QWC partition is fixed and asks which groups should be retained for each individual parameter.

Learned measurement adaptation. Learning has also been used to adapt generalized measurements and reduce observation cost [7, 26, 2]. Our output space is different: it is a variable-size P×GP \times G relevance matrix tied directly to parameter-shift gradient decomposition. To our knowledge, prior work does not jointly learn this parameter-by-QWC grid and charge cost at the physical measurement-setting level.

Ansatz adaptation and parameter freezing. Ansatz adaptation and parameter freezing reduce optimization cost along an orthogonal axis. ADAPT-VQE grows an operator pool selectively [12, 21], whereas TITAN predicts parameters that can be frozen along a trajectory [32]. Meta-VQE learns across related Hamiltonians to accelerate energy-profile estimation [5]. Our method neither changes the ansatz nor requires a parameter to be frozen: for every parameter that is evaluated, it sparsifies the QWC groups entering that gradient. We therefore include both TITAN and TITAN+Ours to separate and compose the two cost axes.

Additional In-Distribution Results

Complete Main-Panel Baselines

Table 5: Complete version of Table 1. This table preserves the default-threshold TITAN [32] and TITAN+Ours rows, the cost-matched † rows, and all one-sided no-worse indicators omitted from the main table for readability. All molecules, seeds, online-VQE settings, and highlighting rules are unchanged.

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+0.0011+0.0015+0.0005−0.0006+0.0023
Depth 1ΔE\Delta E vs Full ↓\downarrowRandom+0.2359+1.2220+0.8916+1.4117+3.4982
Depth 1ΔE\Delta E vs Full ↓\downarrowGlobal coeff.+0.0038+0.0019+0.0122+0.0089+0.0016
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN (Peng et al.)+0.0228+0.1040+0.0871+0.0220−0.0881
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN+Ours+0.0227+0.1040+0.0869+0.0218−0.0875
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN†+0.0128+0.1299+0.0772−0.0642−0.0648
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN†+Ours+0.0127+0.1297+0.0776−0.0649−0.0646
Depth 1Shot gain ↑\uparrowOurs40.4%58.9%63.0%74.7%86.9%
Depth 1Shot gain ↑\uparrowRandom42.0%60.7%66.5%79.2%90.2%
Depth 1Shot gain ↑\uparrowGlobal coeff.41.7%60.0%62.9%75.0%87.0%
Depth 1Shot gain ↑\uparrowTITAN (Peng et al.)61.8%56.2%58.2%62.7%71.9%
Depth 1Shot gain ↑\uparrowTITAN+Ours80.0%90.5%93.5%96.1%99.7%
Depth 1Shot gain ↑\uparrowTITAN†43.8%59.4%62.7%70.3%81.8%
Depth 1Shot gain ↑\uparrowTITAN†+Ours66.8%91.7%95.0%97.9%99.9%
Depth 1No-worse vs FullOurs✓✓✓✓✗
Depth 1No-worse vs FullRandom✗✗✗✗✗
Depth 1No-worse vs FullGlobal coeff.✗✗✗✗✗
Depth 1No-worse vs FullTITAN (Peng et al.)✗✗✗✗✓
Depth 1No-worse vs FullTITAN+Ours✗✗✗✗✓
Depth 1No-worse vs FullTITAN†✗✗✗✓✓
Depth 1No-worse vs FullTITAN†+Ours✗✗✗✓✓
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs+0.0003−0.0011−0.0015+0.0158−0.0049
Depth 2ΔE\Delta E vs Full ↓\downarrowRandom+0.1656+0.4704+0.6527+0.9343+4.0043
Depth 2ΔE\Delta E vs Full ↓\downarrowGlobal coeff.+0.0015−0.0004+0.0233+0.0379−0.0049
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN (Peng et al.)−0.0200−0.1194−0.0031+0.1521+0.0806
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN+Ours−0.0200−0.1200−0.0032+0.1538+0.0784
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN†+0.0075+0.0116+0.0153+0.0472+0.2903
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN†+Ours+0.0072+0.0129+0.0140+0.0493+0.2920
Depth 2Shot gain ↑\uparrowOurs25.6%31.6%44.9%51.8%82.6%
Depth 2Shot gain ↑\uparrowRandom26.3%33.2%47.7%53.7%83.4%
Depth 2Shot gain ↑\uparrowGlobal coeff.25.7%31.9%44.7%51.7%82.5%
Depth 2Shot gain ↑\uparrowTITAN (Peng et al.)54.0%54.4%50.0%59.1%77.1%
Depth 2Shot gain ↑\uparrowTITAN+Ours70.9%75.7%79.9%86.1%98.7%
Depth 2Shot gain ↑\uparrowTITAN†23.8%39.0%41.9%46.8%82.7%
Depth 2Shot gain ↑\uparrowTITAN†+Ours40.2%64.0%74.5%79.0%99.3%
Depth 2No-worse vs FullOurs✓✓✓✗✓
Depth 2No-worse vs FullRandom✗✗✗✗✗
Depth 2No-worse vs FullGlobal coeff.✓✓✗✗✓
Depth 2No-worse vs FullTITAN (Peng et al.)✓✓✓✗✗
Depth 2No-worse vs FullTITAN+Ours✓✓✓✗✗
Depth 2No-worse vs FullTITAN†✗✗✗✗✗
Depth 2No-worse vs FullTITAN†+Ours✗✗✗✗✗

Table 5. Complete version of Table 1. This table preserves the default-threshold TITAN [32] and TITAN+Ours rows, the cost-matched † rows, and all one-sided no-worse indicators omitted from the main table for readability. All molecules, seeds, online-VQE settings, and highlighting rules are unchanged.

In-Distribution Aggregate and Budget Details

SettingMean ∣ΔE∣\lvert\Delta E\rvert ↓No-worseMean shot gain ↑
Full-QWC reference0.00000✓0.0%
Ours depth 10.00150✓64.8%
Ours depth 20.00978✗47.3%
Ours depth 30.00571✓36.7%
Random depth 11.45188✗67.7%
Random depth 21.24548✗48.9%
Random depth 30.82189✗37.3%
Ours all0.00567✓49.6%
Random all1.17308✗51.3%

Table 6. Depth-wise aggregate at τ=0.995\tau= 0.995. Mean ∣ΔE∣\lvert\Delta E\rvert averages seed-level absolute gaps over four seeds and five molecules at each depth. No-worse marks whether the aggregate mean signed ΔE\Delta E is at most 1.6 mHa above Full.

MoleculeMethodFull QWC groups/param.Groups / param. ↓ D1Groups / param. ↓ D2Groups / param. ↓ D3Shot gain ↑ D1Shot gain ↑ D2Shot gain ↑ D3No-worse vs Full D1No-worse vs Full D2No-worse vs Full D3Best ΔE\Delta E ↓
H2 (4q)Ours52.983.724.0940.4%25.6%18.1%✓✓✓+0.0003
H2 (4q)Random52.903.684.1142.0%26.3%17.9%✗✗✓−0.0145
HeH+ (4q)Ours114.527.538.2658.9%31.6%24.9%✓✓✗−0.0011
HeH+ (4q)Random114.327.358.3060.7%33.2%24.6%✗✗✗+0.0513
H3+ (6q)Ours197.0310.4713.3363.0%44.9%29.9%✓✓✓+0.0005
H3+ (6q)Random196.379.9413.3166.5%47.7%29.9%✗✗✗+0.2864
H4 chain (8q)Ours7518.9636.1548.6074.7%51.8%35.2%✓✗✓−0.0006
H4 chain (8q)Random7515.6134.7147.4379.2%53.7%36.8%✗✗✗+0.2903
LiH (12q)Ours17923.4031.1644.0986.9%82.6%75.4%✗✓✓−0.0017
LiH (12q)Random17917.5729.7840.8290.2%83.4%77.2%✗✗✗+3.3113

Table 7. Detailed budget usage at τ=0.995\tau= 0.995. All molecular Hamiltonians use the STO-3G minimal basis; molecule labels report the resulting qubit counts. Groups/param. reports the mean selected QWC groups per parameter, shot gain is relative to full 2P2P GS, and No-worse vs Full uses the one-sided 1.6 mHa criterion (ΔE≤1.6\Delta E \le1.6 mHa; lower energy is accepted). Boldface marks the better Ours-vs-Random entry for each molecule-depth comparison.

Shots-to-Chemical-Parity Dynamics

Figure 7 complements the endpoint comparison with cumulative-shot convergence trajectories.

Post-hoc budget control. We vary only the cumulative top-mass threshold τ\tau after training; model weights, labels, and online oracle access remain unchanged. One selector trained on depths 1–4 is evaluated on unseen depths 5–7. Increasing τ\tau from 0.98 to 0.997 reduces the proposed method’s mean absolute energy error from 22.5 to 7.2 mHa while retaining 16.2% shot saving. At closely matched budgets, Random remains at 264–606 mHa. The result exposes a controllable accuracy–measurement frontier without retraining.

Figure 8. Depth-OOD threshold Pareto frontier. Bars show mean ∣ΔE∣|\Delta E| to Full QWC; labels above them count no-worse trajectories out of 60 under the one-sided 1.6 mHa criterion. The line shows the proposed method’s shot saving; Random uses an almost identical budget. Full values are in Appendix C, Table 8.

Threshold Control and Depth-OOD Pareto Details

Depth-OOD Pareto Visualization

Complete Numerical Threshold Frontier

ThresholdOursRandom same budget
Mean ∣ΔE∣|\Delta E| ↓Mean ΔE\Delta ENo-worse seedsShot gain ↑Groups/param.Mean ∣ΔE∣|\Delta E| ↓No-worse seedsShot gain ↑
τ\tau0.02246+0.006627/6036.6%24.430.606455/6037.0%
0.990.01791+0.002531/6026.8%30.470.468577/6027.4%
0.9950.00974−0.001032/6019.8%35.670.318678/6020.4%
0.9970.00716−0.001237/6016.2%38.900.2643112/6016.7%

Table 8. Complete numerical counterpart of Figure 8. A single model trained on depths 1–4 is reused across thresholds and evaluated on unseen depths 5–7. Only the inference top-mass threshold changes; energy gaps are in Ha and costs are nominal. This duplicates the depth-OOD aggregate in Table 10, not an independent replication. No-worse seeds have final energy at most 1.6 mHa above Full QWC; lower final energy is accepted. Further threshold, depth-wise, and molecule-wise results appear in Tables 10, 14, and 15.

Top-Mass Threshold Sweep Summary

The current same-family ansatz-OOD table uses τ=0.997\tau= 0.997. To make the budget dial explicit, Tables 9 and 10 summarize the existing threshold sweeps. In all cases, τ\tau is the cumulative top-mass threshold used to convert the predicted parameter-by-QWC relevance matrix into selected QWC groups:

∑k∈Sj(τ)w^j,k≥τ.(8)\sum_{k \in S_j(\tau)} \widehat{w}_{j,k} \ge\tau. \tag*{(8)}

For a fixed predicted row, larger τ\tau retains a nested set of groups. This does not guarantee monotone final energy or closed-loop average cost, and the historical points below are not all evaluations of one checkpoint.

ThresholdOursOursOursOursRandom same budgetRandom same budgetRandom same budget
τ\tauMean ∣ΔE∣|\Delta E| ↓\downarrowMean ΔE\Delta EShot gain ↑\uparrowGroups/param.Mean ∣ΔE∣|\Delta E| ↓\downarrowShot gain ↑\uparrowGroups/param.
0.950.00917+0.007362.3%12.191.3351263.4%11.24
0.980.00984+0.004052.8%17.591.0720253.8%16.63
0.990.00439+0.001145.5%21.540.9358946.5%20.32
0.9950.00567+0.001449.6%17.621.1730851.3%16.41
0.9970.00352+0.000045.5%19.841.1002647.4%18.59

Table 9. Historical held-out-seed operating points, not one checkpoint’s sweep. Depths 1–3 use exact-expectation VQE and nominal shot accounting. Earlier 0.95/0.98/0.99 points belong to separately trained models; those models used the auxiliary count head. Only 0.995/0.997 reuse the τtrain=0.99\tau_{\mathrm{train}} = 0.99 checkpoint with top-mass post-processing. Energy gaps are in Ha. Checkpoint and selection-mode provenance are in Appendix § 24.§ 2; detailed results follow.

ThresholdOursOursOursOursOursRandom same budgetRandom same budgetRandom same budget
τ\tauMean ∣ΔE∣|\Delta E| ↓\downarrowMean ΔE\Delta ENo-worse seedsShot gain ↑\uparrowGroups/param.Mean ∣ΔE∣|\Delta E| ↓\downarrowNo-worse seedsShot gain ↑\uparrow
0.980.02246+0.006627/6036.6%24.430.606455/6037.0%
0.990.01791+0.002531/6026.8%30.470.468577/6027.4%
0.9950.00974-0.001032/6019.8%35.670.318678/6020.4%
0.9970.00716-0.001237/6016.2%38.900.2643112/6016.7%

Table 10. Depth-OOD threshold sweep. A single depth-OOD model is trained on depths 1–4 and evaluated on unseen depths 5–7. Only the inference top-mass threshold changes; energy gaps are in Ha and shot costs are nominal. This differs from the historical table above (Appendix § 24.§ 2). No-worse seeds count trajectories whose final energy is not more than 1.6 mHa above Full-QWC; lower energy is accepted.

Stricter Point: τ=0.997\tau= 0.997

The τ=0.997\tau= 0.997 point uses the same saved relevance model and increases only the top-mass post-processing threshold. It is intended as a conservative operating point for cases where τ=0.995\tau= 0.995 is close but not always within the no-worse criterion.

SettingMetricMethodH2/STO-3GHeH+/STO-3GH3+/STO-3GH4 chain/STO-3GLiH/STO-3G
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+0.0005+0.0001−0.0002−0.0013+0.0020
Depth 1ΔE\Delta E vs Full ↓\downarrowRandom+0.2192+1.3164+0.7630+1.2211+3.5105
Depth 1Shot gain ↑\uparrowOurs37.2%55.1%59.5%71.7%84.9%
Depth 1Shot gain ↑\uparrowRandom41.3%56.6%62.7%75.5%88.9%
Depth 1No-worse vs FullOurs✓✓✓✓✗
Depth 1No-worse vs FullRandom✗✗✗✗✗
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs−0.0001−0.0024−0.0030+0.0017−0.0021
Depth 2ΔE\Delta E vs Full ↓\downarrowRandom+0.1092+0.3799+0.5018+0.7784+3.7420
Depth 2Shot gain ↑\uparrowOurs21.8%27.3%39.8%46.6%79.1%
Depth 2Shot gain ↑\uparrowRandom23.0%29.3%42.3%47.8%80.0%
Depth 2No-worse vs FullOurs✓✓✓✗✓
Depth 2No-worse vs FullRandom✗✗✗✗✗
Depth 3ΔE\Delta E vs Full ↓\downarrowOurs+0.0007+0.0074−0.0023+0.0004−0.0010
Depth 3ΔE\Delta E vs Full ↓\downarrowRandom+0.0092+0.1772+0.1448+0.1285+3.3875
Depth 3Shot gain ↑\uparrowOurs14.8%20.3%24.7%30.0%70.4%
Depth 3Shot gain ↑\uparrowRandom14.9%20.6%25.2%30.7%72.6%
Depth 3No-worse vs FullOurs✓✗✓✓✓
Depth 3No-worse vs FullRandom✗✗✗✗✗

Table 11. Matched-budget online VQE baseline comparison at τ=0.997\tau= 0.997. The same saved relevance model is reused; only the top-mass post-processing threshold is increased to τ=0.997\tau= 0.997. For each depth and metric, our method is paired with a Random same-budget baseline that keeps the same number of selected QWC groups per parameter but chooses groups uniformly at random. Boldface marks the better method for each molecule-specific comparison.

SettingMean ∣ΔE∣|\Delta E| ↓\downarrowNo-worseMean shot gain ↑\uparrow
Full-QWC reference0.00000✓0.0%
Ours depth 10.00112✓61.7%
Ours depth 20.00448✓42.9%
Ours depth 30.00497✓32.0%
Random depth 11.40602✗65.0%
Random depth 21.10226✗44.5%
Random depth 30.79248✗32.8%
Ours all0.00352✓45.5%
Random all1.10026✗47.4%

Table 12. Depth-wise aggregate at τ=0.997\tau= 0.997. Mean ∣ΔE∣|\Delta E| averages the five molecule columns at each ansatz depth. No-worse marks whether the aggregate mean signed ΔE\Delta E is at most 1.6 mHa above Full.

MoleculeMethodFull QWC groups/param.Groups / param. ↓ D1Groups / param. ↓ D2Groups / param. ↓ D3Shot gain ↑ D1Shot gain ↑ D2Shot gain ↑ D3No-worse vs Full D1No-worse vs Full D2No-worse vs Full D3Best ΔE\Delta E ↓
H2/STO-3GOurs53.143.914.2637.2%21.8%14.8%✓✓✓−0.0001
H2/STO-3GRandom52.943.854.2641.3%23.0%14.9%✗✗✗+0.0092
HeH+/STO-3GOurs114.948.008.7755.1%27.3%20.3%✓✓✗+0.0001
HeH+/STO-3GRandom114.777.788.7456.6%29.3%20.6%✗✗✗+0.1772
H3+/STO-3GOurs197.6911.4414.3159.5%39.8%24.7%✓✓✓−0.0002
H3+/STO-3GRandom197.0810.9714.2062.7%42.3%25.2%✗✗✗+0.1448
H4 chain/STO-3GOurs7521.2640.0352.5271.7%46.6%30.0%✓✗✓+0.0004
H4 chain/STO-3GRandom7518.4039.1451.9875.5%47.8%30.7%✗✗✗+0.1285
LiH/STO-3GOurs17926.9937.4652.9284.9%79.1%70.4%✗✓✓−0.0010
LiH/STO-3GRandom17919.8935.7549.0688.9%80.0%72.6%✗✗✗+3.3875

Table 13. Detailed budget usage at τ=0.997\tau= 0.997. Groups/param. reports the mean selected QWC groups per parameter, shot gain is relative to full 2PGS2PGS, and No-worse vs Full uses the one-sided 1.6 mHa criterion (ΔE≤1.6\Delta E \le1.6 mHa; lower energy is accepted). Boldface marks the better our method-vs-Random entry for each molecule-depth comparison.

Depth-OOD Pareto Sweep

This sweep reuses the depth-OOD relevance model and changes only the top-mass threshold. This isolates the post-hoc budget dial: lower thresholds save more QWC measurements, while higher thresholds recover more of the Full-QWC gradient signal.

ThresholdDepthOursRandom same budget
τ\tauddMean ΔE±\Delta E \pm s.d.CAShot gain ↑Group gain ↑Mean ΔE±\Delta E \pm s.d.CAShot gain ↑Group gain ↑
0.985+0.0219±0.0618+0.0219 \pm0.0618✗39.4%62.3%+0.7969±1.2005+0.7969 \pm1.2005✗39.9%64.4%
0.986−0.0054±0.0243-0.0054 \pm0.0243✓36.5%57.4%+0.5774±0.9583+0.5774 \pm0.9583✗36.9%59.3%
0.987+0.0035±0.0124+0.0035 \pm0.0124✗33.9%53.5%+0.4024±0.7036+0.4024 \pm0.7036✗34.1%55.0%
0.995+0.0048±0.0555+0.0048 \pm0.0555✗29.7%52.4%+0.7167±1.1610+0.7167 \pm1.1610✗30.4%54.8%
0.996+0.0026±0.0256+0.0026 \pm0.0256✗26.7%46.9%+0.4179±0.6588+0.4179 \pm0.6588✗27.2%49.0%
0.997+0.0002±0.0130+0.0002 \pm0.0130✓24.1%42.5%+0.2481±0.4441+0.2481 \pm0.4441✗24.6%44.1%
0.9955−0.0016±0.0276-0.0016 \pm0.0276✓22.6%43.8%+0.5389±0.9643+0.5389 \pm0.9643✗23.5%46.2%
0.9956−0.0014±0.0090-0.0014 \pm0.0090✓19.6%37.8%+0.2580±0.4218+0.2580 \pm0.4218✗20.1%39.2%
0.9957−0.0001±0.0107-0.0001 \pm0.0107✓17.3%33.2%+0.1470±0.2903+0.1470 \pm0.2903✗17.7%34.5%
0.9975−0.0025±0.0155-0.0025 \pm0.0155✓18.9%38.3%+0.4968±0.8809+0.4968 \pm0.8809✗19.5%39.9%
0.9976−0.0006±0.0105-0.0006 \pm0.0105✓16.0%32.1%+0.1665±0.3072+0.1665 \pm0.3072✗16.4%33.4%
0.9977−0.0005±0.0098-0.0005 \pm0.0098✓13.8%27.6%+0.0916±0.2070+0.0916 \pm0.2070✗14.1%28.7%

Table 14. Depth-wise view of the OOD Pareto sweep. A single Transformer relevance model is trained on ansatz depths 1–4 and evaluated without retraining on unseen depths 5–7. Each row pools the five PySCF/STO-3G molecular Hamiltonians (H2, HeH+, H3+, H4 chain, and LiH) and four held-out seeds at one unseen evaluation depth. Mean ΔE\Delta E is reported as mean ±\pm standard deviation over the 20 paired trajectories. CA marks whether the mean signed ΔE\Delta E is no more than 1.6 mHa above the Full-QWC reference; lower final energy is accepted. Group gain is the relative reduction in selected QWC groups per parameter versus Full-QWC.

MoleculeOursRandom same budget
Mean ∣ΔE∣\lvert\Delta E\rvert ↓Mean ΔE\Delta EShot gain ↑Mean ∣ΔE∣\lvert\Delta E\rvert ↓Mean ΔE\Delta EShot gain ↑
H2/STO-3G0.00140+0.0003+0.00034.8%0.01690+0.0066+0.00664.7%
HeH+/STO-3G0.00735−0.0024-0.00247.6%0.06000+0.0424+0.04247.5%
H3+/STO-3G0.00668−0.0036-0.00368.9%0.03097+0.0160+0.01608.9%
H4 chain/STO-3G0.01298−0.0011-0.001114.9%0.09763+0.0771+0.077115.0%
LiH/STO-3G0.00742+0.0009+0.000945.0%1.11605+1.1160+1.116047.1%

Table 15. Molecule-wise OOD summary at τ=0.997\tau= 0.997. This conservative point gives the best global mean ∣ΔE∣\lvert\Delta E\rvert in the sweep while still retaining nonzero shot savings, especially for richer Hamiltonians with many QWC groups.

Previous Threshold Sweep Tables

The main paper reports the stricter post-processing point τ=0.995\tau= 0.995. For completeness, we keep the earlier threshold-sweep tables below. These tables use the same held-out-seed online VQE protocol and matched-budget random baseline, but reflect earlier trained operating points τ∈{0.95,0.98,0.99}\tau\in\{0.95, 0.98, 0.99\}, including count-head selection rather than a single-checkpoint inference-threshold sweep. Appendix N.2 gives provenance. All energy gaps are in Ha and shot gains use nominal accounting.

Previous Threshold: τ=0.98\tau= 0.98

Table 16: Matched-budget online VQE baseline comparison (Main Results: threshold τ=0.98\tau= 0.98). The learned selector is trained on RY/RZ VQE trajectories from depths 1–3 using row-wise top-mass target threshold τ=0.98\tau= 0.98 and evaluated in closed-loop online VQE on held-out random seeds for the same five PySCF molecules and depths. For each depth and metric, our method is paired with a Random same-budget baseline that keeps the same predicted number of QWC groups per parameter but chooses groups uniformly at random. Boldface marks the better method for each molecule-specific comparison.

SettingMetricMethodH2/STO-3GHeH+/STO-3GH3+/STO-3GH4 chain/STO-3GLiH/STO-3G
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs−0.0000+0.0028+0.0040−0.0021+0.0012
Depth 1ΔE\Delta E vs Full ↓\downarrowRandom+0.2500+1.3729+0.7713+1.1147+3.2882
Depth 1Shot gain ↑\uparrowOurs41.2%59.2%66.2%73.7%86.4%
Depth 1Shot gain ↑\uparrowRandom43.7%62.9%69.0%77.0%89.8%
Depth 1No-worse vs FullOurs✓✗✗✓✓
Depth 1No-worse vs FullRandom✗✗✗✗✗
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs+0.0244+0.0103−0.0021+0.0214−0.0032
Depth 2ΔE\Delta E vs Full ↓\downarrowRandom+0.1560+0.4481+0.4674+0.8634+3.4465
Depth 2Shot gain ↑\uparrowOurs31.9%40.5%49.6%56.3%79.9%
Depth 2Shot gain ↑\uparrowRandom30.8%39.3%49.7%56.7%81.2%
Depth 2No-worse vs FullOurs✗✗✓✗✓
Depth 2No-worse vs FullRandom✗✗✗✗✗
Depth 3ΔE\Delta E vs Full ↓\downarrowOurs−0.0022+0.0121−0.0026−0.0020−0.0014
Depth 3ΔE\Delta E vs Full ↓\downarrowRandom+0.0586+0.1960+0.2986+0.2677+3.0320
Depth 3Shot gain ↑\uparrowOurs24.9%31.3%37.4%41.8%71.6%
Depth 3Shot gain ↑\uparrowRandom24.0%30.6%36.7%42.1%73.0%
Depth 3No-worse vs FullOurs✓✗✓✓✓
Depth 3No-worse vs FullRandom✗✗✗✗✗

Table 16.

Table 16 shows that at threshold τ=0.98\tau= 0.98, our method reaches mean absolute final-energy gap 0.00984 while saving 52.8% of full-gradient measurement shots. The matched-budget random baseline has mean absolute gap 1.07202 with 53.8% shot saving, so the selected QWC groups matter beyond the budget alone.

Table 17: Depth-wise aggregate (Main Results: threshold $\tau = 0.98). Same protocol as Table 16 with top-mass target threshold τ=0.98\tau= 0.98. Mean ∣ΔE∣|\Delta E| aggregates the five molecule columns at each ansatz depth; No-worse marks whether the aggregate mean signed ΔE\Delta E is at most 1.6 mHa above Full.

SettingMean ∣ΔE∣|\Delta E| ↓\downarrowNo-worseMean shot gain ↑\uparrow
Full-QWC reference0.00000✓0.0%
Ours depth 10.00217✓65.3%
Ours depth 20.01616✗51.6%
Ours depth 30.01119✓41.4%
Random depth 11.35940✗68.5%
Random depth 21.07628✗51.5%
Random depth 30.78037✗41.3%
Ours all0.00984✗52.8%
Random all1.07202✗53.8%

Table 17.

MoleculeMethodFull QWC groups/param.D1D2D3D1D2D3D1D2D3Best ΔE\Delta E
H2/STO-3GOurs52.943.413.7541.2%31.9%24.9%✓✗✓−0.0000
H2/STO-3GRandom52.823.463.8043.7%30.8%24.0%✗✗✗+0.0586
HeH+/STO-3GOurs114.496.557.5659.2%40.5%31.3%✗✗✗+0.0028
HeH+/STO-3GRandom114.086.687.6362.9%39.3%30.6%✗✗✗+0.1960
H3+/STO-3GOurs196.429.5811.8966.2%49.6%37.4%✗✓✓−0.0021
H3+/STO-3GRandom195.899.5512.0369.0%49.7%36.7%✗✗✗+0.2986
H4 chain/STO-3GOurs7519.7232.7743.6373.7%56.3%41.8%✓✗✓−0.0021
H4 chain/STO-3GRandom7517.2232.4643.4677.0%56.7%42.1%✗✗✗+0.2677
LiH/STO-3GOurs17924.3836.0250.7586.4%79.9%71.6%✓✓✓+0.0012
LiH/STO-3GRandom17918.3333.7348.3189.8%81.2%73.0%✗✗✗+3.0320

Table 18. Detailed budget usage (Main Results: threshold τ=0.98\tau= 0.98). Same table format with top-mass target threshold τ=0.98\tau= 0.98. Groups/param. reports the mean selected QWC groups per parameter, shot gain is relative to full 2PGS2PGS, and No-worse vs Full uses the one-sided 1.6 mHa criterion (ΔE≤1.6\Delta E \le1.6 mHa; lower energy is accepted). Boldface marks the better Ours-vs-Random entry for each molecule-depth comparison.

Previous Threshold: τ=0.95\tau= 0.95

Table 19 shows that at threshold τ=0.95\tau= 0.95, our method reaches mean absolute final-energy gap 0.00917 while saving 62.3% of full-gradient measurement shots. The matched-budget random baseline has mean absolute gap 1.33512 with 63.4% shot saving, so the selected QWC groups matter beyond the budget alone.

SettingMetricMethodH2/STO-3GHeH+/STO-3GH3+/STO-3GH4 chain/STO-3GLiH/STO-3G
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+0.0023+0.0023−0.0006-0.0006+0.0072+0.0072+0.0017+0.0017+0.0028+0.0028
Depth 1ΔE\Delta E vs Full ↓\downarrowRandom+0.2411+0.2411+1.4286+1.4286+1.0475+1.0475+1.6624+1.6624+3.4272+3.4272
Depth 1Shot gain ↑\uparrowOurs43.8%65.2%74.6%81.0%91.5%
Depth 1Shot gain ↑\uparrowRandom44.3%68.5%77.2%85.2%93.2%
Depth 1No-worse vs FullOurs×\times✓\checkmark×\times×\times×\times
Depth 1No-worse vs FullRandom×\times×\times×\times×\times×\times
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs+0.0364+0.0364+0.0158+0.0158+0.0243+0.0243+0.0077+0.0077−0.0017-0.0017
Depth 2ΔE\Delta E vs Full ↓\downarrowRandom+0.2499+0.2499+0.8874+0.8874+0.7010+0.7010+1.2117+1.2117+4.1073+4.1073
Depth 2Shot gain ↑\uparrowOurs40.2%54.1%62.0%66.4%88.3%
Depth 2Shot gain ↑\uparrowRandom39.6%51.2%60.8%68.8%89.1%
Depth 2No-worse vs FullOurs×\times×\times×\times×\times✓\checkmark
Depth 2No-worse vs FullRandom×\times×\times×\times×\times×\times
Depth 3ΔE\Delta E vs Full ↓\downarrowOurs−0.0029-0.0029+0.0234+0.0234−0.0047-0.0047+0.0022+0.0022−0.0039-0.0039
Depth 3ΔE\Delta E vs Full ↓\downarrowRandom+0.1469+0.1469+0.3339+0.3339+0.5480+0.5480+0.5354+0.5354+3.4985+3.4985
Depth 3Shot gain ↑\uparrowOurs33.6%43.8%50.5%55.8%84.1%
Depth 3Shot gain ↑\uparrowRandom34.2%44.1%51.9%58.1%85.5%
Depth 3No-worse vs FullOurs✓\checkmark×\times✓\checkmark×\times✓\checkmark
Depth 3No-worse vs FullRandom×\times×\times×\times×\times×\times

Table 19. Matched-budget online VQE baseline comparison (previous threshold τ=0.95\tau= 0.95). The learned selector is trained on RY/RZ VQE trajectories from depths 1–3 using row-wise top-mass target threshold τ=0.95\tau= 0.95 and evaluated in closed-loop online VQE on held-out random seeds for the same five PySCF molecules and depths. For each depth and metric, our method is paired with a Random same-budget baseline that keeps the same predicted number of QWC groups per parameter but chooses groups uniformly at random. Boldface marks the better method for each molecule-specific comparison.

SettingMean ∣ΔE∣|\Delta E| ↓\downarrowNo-worseMean shot gain ↑\uparrow
Full-QWC reference0.00000✓\checkmark0.0%
Ours depth 10.00290×\times71.2%
Ours depth 20.01719×\times62.2%
Ours depth 30.00742×\times53.6%
Random depth 11.56136×\times73.7%
Random depth 21.43145×\times61.9%
Random depth 31.01255×\times54.8%
Ours all0.00917×\times62.3%
Random all1.33512×\times63.4%

Table 20. Depth-wise aggregate (previous threshold τ=0.95\tau= 0.95). Same protocol as Table 19 with top-mass target threshold τ=0.95\tau= 0.95. Mean ∣ΔE∣|\Delta E| aggregates the five molecule columns at each ansatz depth; No-worse marks whether the aggregate mean signed ΔE\Delta E is at most 1.6 mHa above Full.

Table 21: Detailed budget usage (previous threshold τ=0.95\tau= 0.95). Same table format with top-mass target threshold τ=0.95\tau= 0.95. Groups/param. reports the mean selected QWC groups per parameter, shot gain is relative to full 2PGS2PGS, and No-worse vs Full uses the one-sided 1.6 mHa criterion (ΔE≤1.6\Delta E \leq1.6 mHa; lower energy is accepted). Boldface marks the better Ours-vs-Random entry for each molecule-depth comparison.

MoleculeMethodFull QWC groups/param.Groups / param. ↓\downarrowShot gain ↑\uparrowNo-worse vs FullBest ΔE↓\Delta E \downarrow
D1D2D3D1D2D3D1D2D3
H2/STO-3GOurs52.812.993.3243.8%40.2%33.6%✗✗✓+0.0023
H2/STO-3GRandom52.793.023.2944.3%39.6%34.2%✗✗✗+0.1469
HeH+/STO-3GOurs113.835.046.1865.2%54.1%43.8%✓✗✗−0.0006
HeH+/STO-3GRandom113.475.376.1568.5%51.2%44.1%✗✗✗+0.3339
H3+/STO-3GOurs194.827.239.4174.6%62.0%50.5%✗✗✓−0.0047
H3+/STO-3GRandom194.347.449.1377.2%60.8%51.9%✗✗✗+0.5480
H4 chain/STO-3GOurs7514.2825.1633.1481.0%66.4%55.8%✗✗✗+0.0017
H4 chain/STO-3GRandom7511.0823.3731.4585.2%68.8%58.1%✗✗✗+0.5354
LiH/STO-3GOurs17915.2120.9328.4791.5%88.3%84.1%✗✓✓−0.0017
LiH/STO-3GRandom17912.2119.5825.9593.2%89.1%85.5%✗✗✗+3.4272

Table 21. Detailed budget usage (previous threshold τ=0.95\tau= 0.95). Same table format with top-mass target threshold τ=0.95\tau= 0.95. Groups/param. reports the mean selected QWC groups per parameter, shot gain is relative to full 2PGS2PGS, and No-worse vs Full uses the one-sided 1.6 mHa criterion (ΔE≤1.6\Delta E \leq1.6 mHa; lower energy is accepted). Boldface marks the better Ours-vs-Random entry for each molecule-depth comparison.

Previous Threshold: τ=0.99\tau= 0.99

Table 22 shows that at threshold τ=0.99\tau= 0.99, our method reaches mean absolute final-energy gap 0.00439 while saving 45.5% of full-gradient measurement shots. The matched-budget random baseline has mean absolute gap 0.93589 with 46.5% shot saving, so the selected QWC groups matter beyond the budget alone.

SettingMetricMethodH2/STO-3GHeH+/STO-3GH3+/STO-3GH4 chain/STO-3GLiH/STO-3G
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+0.0013+0.0013+0.0008−0.0008+0.0013
Depth 1ΔE\Delta E vs Full ↓\downarrowRandom+0.2060+1.0934+0.5968+1.0145+3.0269
Depth 1Shot gain ↑\uparrowOurs36.6%51.3%58.2%67.4%81.7%
Depth 1Shot gain ↑\uparrowRandom35.5%54.3%60.1%72.7%86.9%
Depth 1No-worse vs FullOurs✓✓✓✓✓
Depth 1No-worse vs FullRandom✗✗✗✗✗
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs−0.0008+0.0037−0.0015+0.0012−0.0009
Depth 2ΔE\Delta E vs Full ↓\downarrowRandom+0.1683+0.3132+0.3132+0.5964+3.4087
Depth 2Shot gain ↑\uparrowOurs25.8%32.1%41.2%47.7%74.7%
Depth 2Shot gain ↑\uparrowRandom25.2%32.3%41.6%47.2%75.7%
Depth 2No-worse vs FullOurs✓✗✓✓✓
Depth 2No-worse vs FullRandom✗✗✗✗✗
Depth 3ΔE\Delta E vs Full ↓\downarrowOurs+0.0028+0.0087−0.0006−0.0009+0.0008
Depth 3ΔE\Delta E vs Full ↓\downarrowRandom+0.1022+0.2128+0.1860+0.1159+2.6667
Depth 3Shot gain ↑\uparrowOurs18.0%23.6%28.5%33.1%63.1%
Depth 3Shot gain ↑\uparrowRandom17.4%23.4%28.2%33.0%64.7%
Depth 3No-worse vs FullOurs✗✗✓✓✓
Depth 3No-worse vs FullRandom✗✗✗✗✗

Table 22. Matched-budget online VQE baseline comparison (conservative threshold τ=0.99\tau= 0.99). The learned selector is trained on RY/RZ VQE trajectories from depths 1–3 using row-wise top-mass target threshold τ=0.99\tau= 0.99 and evaluated in closed-loop online VQE on held-out random seeds for the same five PySCF molecules and depths. For each depth and metric, our method is paired with a Random same-budget baseline that keeps the same predicted number of QWC groups per parameter but chooses groups uniformly at random. Boldface marks the better method for each molecule-specific comparison.

SettingMean ∣ΔE∣|\Delta E| ↓\downarrowNo-worseMean shot gain ↑\uparrow
Full-QWC reference0.00000✓0.0%
Ours depth 10.00139✓59.0%
Ours depth 20.00560✓44.3%
Ours depth 30.00620✗33.2%
Random depth 11.18750✗61.9%
Random depth 20.95994✗44.4%
Random depth 30.66023✗33.3%
Ours all0.00439✓45.5%
Random all0.93589✗46.5%

Table 23. Depth-wise aggregate (conservative threshold τ=0.99\tau= 0.99). Same protocol as Table 22 with top-mass target threshold τ=0.99\tau= 0.99. Mean ∣ΔE∣|\Delta E| aggregates the five molecule columns at each ansatz depth; No-worse marks whether the aggregate mean signed ΔE\Delta E is at most 1.6 mHa above Full.

MoleculeMethodFull QWC groups/param.Groups / param. ↓ D1Groups / param. ↓ D2Groups / param. ↓ D3Shot gain ↑ D1Shot gain ↑ D2Shot gain ↑ D3No-worse vs Full D1No-worse vs Full D2No-worse vs Full D3Best ΔE\Delta E ↓
H2/STO-3GOurs53.173.714.1036.6%25.8%18.0%✓✓✗+0.0013
H2/STO-3GRandom53.223.744.1335.5%25.2%17.4%✗✗✗+0.1022
HeH+/STO-3GOurs115.357.478.4151.3%32.1%23.6%✓✗✗+0.0013
HeH+/STO-3GRandom115.037.458.4254.3%32.3%23.4%✗✗✗+0.2128
H3+/STO-3GOurs197.9411.1613.5958.2%41.2%28.5%✓✓✓+0.0008
H3+/STO-3GRandom197.5811.0913.6560.1%41.6%28.2%✗✗✗+0.1860
H4 chain/STO-3GOurs7524.4539.2250.1767.4%47.7%33.1%✓✓✓−0.0008
H4 chain/STO-3GRandom7520.4539.5950.2472.7%47.2%33.0%✗✗✗+0.1159
LiH/STO-3GOurs17932.8445.3466.0981.7%74.7%63.1%✓✓✓+0.0013
LiH/STO-3GRandom17923.4743.4663.2186.9%75.7%64.7%✗✗✗+2.6667

Table 24. Detailed budget usage (conservative threshold τ=0.99\tau= 0.99). Same table format with top-mass target threshold τ=0.99\tau= 0.99. Groups/param. reports the mean selected QWC groups per parameter, shot gain is relative to full 2PGS2PGS, and No-worse vs Full uses the one-sided 1.6 mHa criterion (ΔE≤1.6\Delta E \le1.6 mHa; lower energy is accepted). Boldface marks the better Ours-vs-Random entry for each molecule-depth comparison.

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1ΔE\Delta E vs Full ↓\downarrowFull-QWC uniform+0.0000+0.0000+0.0000+0.0000+0.0000
Depth 1ΔE\Delta E vs Full ↓\downarrowiCANS (Kubler et al.)+0.0038+0.0793+0.0321+0.0086+0.0177
Depth 1ΔE\Delta E vs Full ↓\downarrowgCANS (Gu et al.)+0.0205+0.0881+0.0703−0.0062+0.0315
Depth 1ΔE\Delta E vs Full ↓\downarrowRosalin (Arrasmith et al.)+0.0070−0.0092+0.0005+0.0060+0.0076
Depth 1ΔE\Delta E vs Full ↓\downarrowVariance-aware (Zhu et al.)−0.0028−0.0116+0.0020+0.0033+0.0136
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs uniform−0.0024−0.0221−0.0157−0.0082+0.0006
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+variance-aware (Zhu et al.)−0.0034−0.0232−0.0230−0.0137+0.0024
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+gCANS (Gu et al.)+0.0108+0.0335+0.0145+0.0099−0.0020
Depth 1Shot gain ↑\uparrowFull-QWC uniform0.0%0.0%0.0%0.0%0.0%
Depth 1Shot gain ↑\uparrowiCANS (Kubler et al.)−73.0%−64.3%−73.8%−89.1%−108.9%
Depth 1Shot gain ↑\uparrowgCANS (Gu et al.)92.7%93.8%93.3%93.0%90.2%
Depth 1Shot gain ↑\uparrowRosalin (Arrasmith et al.)−85.0%−62.8%−90.3%−114.7%−135.7%
Depth 1Shot gain ↑\uparrowVariance-aware (Zhu et al.)0.0%0.0%0.0%0.0%0.0%
Depth 1Shot gain ↑\uparrowOurs uniform41.2%59.6%64.2%76.2%87.9%
Depth 1Shot gain ↑\uparrowOurs+variance-aware (Zhu et al.)41.5%59.7%63.7%75.8%87.7%
Depth 1Shot gain ↑\uparrowOurs+gCANS (Gu et al.)96.3%97.6%97.8%98.5%99.3%
Depth 1No-worse vs FullFull-QWC uniform✓✓✓✓✓
Depth 1No-worse vs FulliCANS (Kubler et al.)✗✗✗✗✗
Depth 1No-worse vs FullgCANS (Gu et al.)✗✗✗✓✗
Depth 1No-worse vs FullRosalin (Arrasmith et al.)✗✓✓✗✗
Depth 1No-worse vs FullVariance-aware (Zhu et al.)✓✓✗✗✗
Depth 1No-worse vs FullOurs uniform✓✓✓✓✓
Depth 1No-worse vs FullOurs+variance-aware (Zhu et al.)✓✓✓✓✗
Depth 1No-worse vs FullOurs+gCANS (Gu et al.)✗✗✗✗✓
Depth 2ΔE\Delta E vs Full ↓\downarrowFull-QWC uniform+0.0000+0.0000+0.0000+0.0000+0.0000
Depth 2ΔE\Delta E vs Full ↓\downarrowiCANS (Kubler et al.)+0.1044+0.1569+0.0239+0.1216+0.2486
Depth 2ΔE\Delta E vs Full ↓\downarrowgCANS (Gu et al.)+0.0279−0.0474+0.0853+0.1612+0.3004
Depth 2ΔE\Delta E vs Full ↓\downarrowRosalin (Arrasmith et al.)+0.0570−0.1309−0.0221+0.0157−0.1710
Depth 2ΔE\Delta E vs Full ↓\downarrowVariance-aware (Zhu et al.)+0.0645+0.0997−0.0050+0.0044−0.1812
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs uniform+0.0954−0.0222−0.0962+0.0720−0.0427
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs+variance-aware (Zhu et al.)+0.0336+0.0237−0.0175−0.0000−0.1850
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs+gCANS (Gu et al.)−0.0319+0.0938+0.0138+0.1896+0.0891
Depth 2Shot gain ↑\uparrowFull-QWC uniform0.0%0.0%0.0%0.0%0.0%
Depth 2Shot gain ↑\uparrowiCANS (Kubler et al.)−67.5%−13.3%−60.1%−71.6%−144.6%
Depth 2Shot gain ↑\uparrowgCANS (Gu et al.)92.8%93.7%93.1%90.4%84.6%
Depth 2Shot gain ↑\uparrowRosalin (Arrasmith et al.)−68.1%−42.6%−86.8%−110.4%−181.2%
Depth 2Shot gain ↑\uparrowVariance-aware (Zhu et al.)0.0%0.0%0.0%0.0%0.0%
Depth 2Shot gain ↑\uparrowOurs uniform27.8%33.2%46.3%52.0%82.4%
Depth 2Shot gain ↑\uparrowOurs+variance-aware (Zhu et al.)27.5%31.9%45.6%52.1%82.8%
Depth 2Shot gain ↑\uparrowOurs+gCANS (Gu et al.)95.3%95.8%96.6%96.4%98.3%
Depth 2No-worse vs FullFull-QWC uniform✓✓✓✓✓
Depth 2No-worse vs FulliCANS (Kubler et al.)✗✓✗✗✗
Depth 2No-worse vs FullgCANS (Gu et al.)✗✓✗✗✗
Depth 2No-worse vs FullRosalin (Arrasmith et al.)✗✗✓✗✗
Depth 2No-worse vs FullVariance-aware (Zhu et al.)✗✗✗✗✓
Depth 2No-worse vs FullOurs uniform✗✗✗✗✓
Depth 2No-worse vs FullOurs+variance-aware (Zhu et al.)✗✗✓✓✓
Depth 2No-worse vs FullOurs+gCANS (Gu et al.)✓✗✗✗✓

Table 25. Molecule-wise shot-allocation benchmark at b=1.0b = 1.0. This expands Table 2 using the layout of Table 1: five STO-3G molecules, depths 1–2, four held-out seeds, 28 online iterations, and 1024 nominal shots per shifted selected parameter–group setting. ΔE\Delta E and shot gain are paired against Full-QWC uniform; adaptive allocators choose their realized shot counts. ✓/✗ applies the one-sided mean criterion ΔE≤1.6\Delta E \le1.6 mHa. Boldface marks the closest non-Full mean energy and bold-underlining marks the largest shot gain in each molecule–depth cell.

Shot-Allocation Details

Molecule-Wise In-Distribution Allocation

Depth-OOD Shot-Allocation Benchmark

Shot-allocation benchmark in the Table 29 depth-OOD protocol at b=1.0b = 1.0. The saved selector is trained on depths 1–4 and evaluated on unseen depths 5 and 6 over five PySCF/STO-3G molecules and four held-out seeds; depth 7 is reported in Appendix § 14.§ 3. All methods use 28 common Adam updates and the same moment-matched Gaussian finite-shot estimator with 1024 nominal shots per selected parameter–group pair and τ=0.997\tau= 0.997. iCANS (Kübler et al., [25]), gCANS (Gu et al., [13]), Rosalin (Arrasmith et al., [4]), and variance-aware allocation (Zhu et al., [44]) are controlled allocation policies rather than reproductions of their complete optimizer stacks. Green/red ΔE\Delta E and ✓/✗ use the one-sided 1.6 mHa criterion. Bold marks the closest non-Full mean ΔE\Delta E; bold-underlining marks the largest saving in each cell.

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 5ΔE\Delta E vs Full ↓Full-QWC uniform+0.0000+0.0000+0.0000+0.0000+0.0000
Depth 5ΔE\Delta E vs Full ↓iCANS+0.0380+0.0300+0.0334−0.0279−0.0371
Depth 5ΔE\Delta E vs Full ↓gCANS+0.0690+0.1606+0.0775+0.1232+0.0538
Depth 5ΔE\Delta E vs Full ↓Rosalin−0.0192−0.0242−0.0014−0.0212−0.2266
Depth 5ΔE\Delta E vs Full ↓Variance-aware−0.0232−0.0382−0.0447−0.1764−0.2449
Depth 5ΔE\Delta E vs Full ↓Ours uniform−0.0026+0.0484−0.0274−0.0407−0.0139
Depth 5ΔE\Delta E vs Full ↓Ours+variance-aware−0.0094−0.0339−0.0543−0.1588−0.2644
Depth 5ΔE\Delta E vs Full ↓Ours+gCANS+0.0158+0.0569−0.0597−0.1106−0.2085
Depth 5Shot gain ↑Full-QWC uniform0.0%0.0%0.0%0.0%0.0%
Depth 5Shot gain ↑iCANS−37.2%−8.0%−79.9%−116.5%−190.5%
Depth 5Shot gain ↑gCANS89.9%93.2%73.7%57.1%39.5%
Depth 5Shot gain ↑Rosalin−43.2%−21.3%−104.4%−162.1%−229.2%
Depth 5Shot gain ↑Variance-aware0.0%0.0%0.0%0.0%0.0%
Depth 5Shot gain ↑Ours uniform5.6%8.5%10.6%17.1%55.7%
Depth 5Shot gain ↑Ours+variance-aware5.6%8.5%10.6%17.0%53.7%
Depth 5Shot gain ↑Ours+gCANS90.2%93.6%78.7%61.6%68.8%
Depth 5No-worse vs FullFull-QWC uniform✓✓✓✓✓
Depth 5No-worse vs FulliCANS✗✗✗✓✓
Depth 5No-worse vs FullgCANS✗✗✗✗✗
Depth 5No-worse vs FullRosalin✓✓✓✓✓
Depth 5No-worse vs FullVariance-aware✓✓✓✓✓
Depth 5No-worse vs FullOurs uniform✓✗✓✓✓
Depth 5No-worse vs FullOurs+variance-aware✓✓✓✓✓
Depth 5No-worse vs FullOurs+gCANS✗✗✓✓✓
Depth 6ΔE\Delta E vs Full ↓Full-QWC uniform+0.0000+0.0000+0.0000+0.0000+0.0000
Depth 6ΔE\Delta E vs Full ↓iCANS+0.0659+0.0090+0.0690+0.1038+0.0150
Depth 6ΔE\Delta E vs Full ↓gCANS+0.0555+0.0643+0.1207+0.0554+0.2048
Depth 6ΔE\Delta E vs Full ↓Rosalin+0.0101−0.0008+0.0200−0.0183−0.3586
Depth 6ΔE\Delta E vs Full ↓Variance-aware−0.0021−0.0258−0.0009+0.0039−0.2956
Depth 6ΔE\Delta E vs Full ↓Ours uniform+0.0320−0.0262−0.0307+0.0214−0.0280
Depth 6ΔE\Delta E vs Full ↓Ours+variance-aware+0.0084−0.0126−0.0173−0.0265−0.3267
Depth 6ΔE\Delta E vs Full ↓Ours+gCANS+0.0385−0.0262−0.0305−0.0211−0.2811
Depth 6Shot gain ↑Full-QWC uniform0.0%0.0%0.0%0.0%0.0%
Depth 6Shot gain ↑iCANS−58.3%−7.4%−97.1%−133.9%−196.7%
Depth 6Shot gain ↑gCANS89.7%92.8%57.3%38.0%36.7%
Depth 6Shot gain ↑Rosalin−59.3%−29.8%−110.5%−181.7%−241.1%
Depth 6Shot gain ↑Variance-aware0.0%0.0%0.0%0.0%0.0%
Depth 6Shot gain ↑Ours uniform4.8%7.6%8.7%14.5%46.1%
Depth 6Shot gain ↑Ours+variance-aware4.8%7.6%8.7%14.5%45.1%
Depth 6Shot gain ↑Ours+gCANS89.9%93.1%61.6%41.4%36.4%
Depth 6No-worse vs FullFull-QWC uniform✓✓✓✓✓
Depth 6No-worse vs FulliCANS✗✗✗✗✗
Depth 6No-worse vs FullgCANS✗✗✗✗✗
Depth 6No-worse vs FullRosalin✗✓✗✗✓
Depth 6No-worse vs FullVariance-aware✗✓✓✗✓
Depth 6No-worse vs FullOurs uniform✗✓✗✗✗
Depth 6No-worse vs FullOurs+variance-aware✗✓✓✓✓
Depth 6No-worse vs FullOurs+gCANS✗✓✓✓✓

Table 26.

Across all 60 held-out seed-level cases, Ours uniform is the most faithful non-Full policy, with mean ∣ΔE∣=0.0568|\Delta E| = 0.0568 Ha while saving 16.6% of shots. Composing selection with gCANS increases saving to 63.3% and improves on gCANS alone in mean ∣ΔE∣|\Delta E| (0.0973 vs. 0.1014 Ha) and seed-level no-worse rate (60.0% vs. 10.0%). The variance-aware composition reaches the highest Ours-family no-worse rate (68.3%) but with a larger absolute gap. Thus Ours-uniform is the clean fidelity–saving point, whereas Ours+gCANS is the aggressive Pareto point.

Depth-7 Shot-Allocation Detail

Table 27: Depth-7 continuation of the depth-OOD shot-allocation benchmark. This table uses the same trained depth-1–4 selector, five PySCF/STO-3G molecules, four held-out seeds, finite-shot estimator, b=1.0b = 1.0, and τ=0.997\tau= 0.997 as Table 26.

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 7ΔE\Delta E vs Full ↓Full-QWC uniform+0.0000+0.0000+0.0000+0.0000+0.0000
Depth 7ΔE\Delta E vs Full ↓iCANS+0.1195+0.0256+0.0540+0.0295+0.0791
Depth 7ΔE\Delta E vs Full ↓gCANS+0.0151+0.1313+0.0640+0.0598+0.1809
Depth 7ΔE\Delta E vs Full ↓Rosalin+0.0517+0.0421−0.0208−0.0153−0.2483
Depth 7ΔE\Delta E vs Full ↓Variance-aware−0.0276−0.0240−0.0199−0.0429−0.2975
Depth 7ΔE\Delta E vs Full ↓Ours uniform−0.0109−0.0544+0.0231+0.0099+0.0250
Depth 7ΔE\Delta E vs Full ↓Ours+variance-aware−0.0125−0.0392−0.0045−0.0712−0.2989
Depth 7ΔE\Delta E vs Full ↓Ours+gCANS−0.0025+0.0291+0.0091−0.0303−0.3097
Depth 7Shot gain ↑Full-QWC uniform0.0%0.0%0.0%0.0%0.0%
Depth 7Shot gain ↑iCANS−48.9%−12.8%−89.5%−147.0%−200.9%
Depth 7Shot gain ↑gCANS87.2%92.0%65.0%5.3%−31.4%
Depth 7Shot gain ↑Rosalin−52.0%−31.4%−111.2%−191.9%−249.3%
Depth 7Shot gain ↑Variance-aware0.0%0.0%0.0%0.0%0.0%
Depth 7Shot gain ↑Ours uniform3.8%6.7%7.5%13.1%39.2%
Depth 7Shot gain ↑Ours+variance-aware3.6%6.6%7.4%13.1%38.0%
Depth 7Shot gain ↑Ours+gCANS86.5%91.6%64.0%−0.0%−8.2%
Depth 7No-worse vs FullFull-QWC uniform✓✓✓✓✓
Depth 7No-worse vs FulliCANS✗✗✗✗✗
Depth 7No-worse vs FullgCANS✗✗✗✗✗
Depth 7No-worse vs FullRosalin✗✗✓✓✓
Depth 7No-worse vs FullVariance-aware✓✓✓✓✓
Depth 7No-worse vs FullOurs uniform✓✓✗✗✗
Depth 7No-worse vs FullOurs+variance-aware✓✓✓✓✓
Depth 7No-worse vs FullOurs+gCANS✓✗✗✓✓

Table 27.

Testing HamiltonianDepthMetricFull-UiCANSgCANSRosalinVar.-awareOurs-UOurs+Var.Ours+gCANS
TFIM h=0.5h = 0.5Depth 1ΔE\Delta E vs Full ↓\downarrow+0.0000−0.0120−0.0000−0.0050+0.0041−0.0013−0.0012+0.0009
Depth 1Shot gain ↑\uparrow0.0%−126.6%72.1%−106.4%0.0%2.1%2.0%69.9%
Depth 1No-worse vs Full✓✓✓✓✗✓✓✓
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0245+0.0427+0.0028−0.0071−0.0002−0.0022+0.0445
Depth 4Shot gain ↑\uparrow0.0%−88.5%77.0%−95.0%0.0%2.2%1.9%77.4%
Depth 4No-worse vs Full✓✗✗✗✓✓✓✗
TFIM h=0.7h = 0.7Depth 1ΔE\Delta E vs Full ↓\downarrow+0.0000−0.0122+0.0081−0.0031+0.0093−0.0020+0.0014+0.0192
Depth 1Shot gain ↑\uparrow0.0%−116.5%81.4%−91.5%0.0%3.2%3.1%80.9%
Depth 1No-worse vs Full✓✓✗✓✗✓✓✗
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0182+0.0718+0.0087−0.0129−0.0161−0.0179+0.0606
Depth 4Shot gain ↑\uparrow0.0%−88.5%73.2%−81.2%0.0%2.1%1.9%75.6%
Depth 4No-worse vs Full✓✗✗✗✓✓✓✗
TFIM h=0.9h = 0.9Depth 1ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0041+0.0300−0.0037+0.0163+0.0811+0.2016+0.1387
Depth 1Shot gain ↑\uparrow0.0%−106.5%82.7%−80.4%0.0%25.0%25.0%88.7%
Depth 1No-worse vs Full✓✗✗✓✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0321+0.0813+0.0561−0.0146+0.0496+0.0307+0.0691
Depth 4Shot gain ↑\uparrow0.0%−88.3%72.3%−76.2%0.0%6.2%6.2%79.5%
Depth 4No-worse vs Full✓✗✗✗✓✗✗✗
Heis. XXZ+X fieldDepth 1ΔE\Delta E vs Full ↓\downarrow+0.0000+0.4624+0.7663+0.3730−0.1707−0.2163+0.1527+0.6480
Depth 1Shot gain ↑\uparrow0.0%−55.0%89.8%−62.1%0.0%16.7%16.7%92.8%
Depth 1No-worse vs Full✓✗✗✗✓✓✗✗
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.5043+1.1720−0.0388+0.0759−0.0737−0.0484+0.8442
Depth 4Shot gain ↑\uparrow0.0%−59.6%82.2%−61.5%0.0%4.2%4.2%82.2%
Depth 4No-worse vs Full✓✗✗✗✗✓✓✗
Heis. J1–J2Depth 1ΔE\Delta E vs Full ↓\downarrow+0.0000−0.2970+0.6094−0.0259−0.3164−0.1839−0.0655+0.3957
Depth 1Shot gain ↑\uparrow0.0%−49.9%91.4%−53.7%0.0%16.7%16.7%93.9%
Depth 1No-worse vs Full✓✗✗✗✓✓✓✗
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.1698+1.0128+0.3410+0.2490+0.1508+0.3493+0.8673
Depth 4Shot gain ↑\uparrow0.0%−52.0%80.7%−64.7%0.0%4.2%4.2%80.9%
Depth 4No-worse vs Full✓✗✗✗✓✗✗✗
Heis. XXY+Z fieldDepth 1ΔE\Delta E vs Full ↓\downarrow+0.0000+0.3060+0.2416+0.2757−0.2531−0.0757+0.0825+0.5081
Depth 1Shot gain ↑\uparrow0.0%−69.4%86.0%−63.3%0.0%16.7%16.7%89.5%
Depth 1No-worse vs Full✓✗✗✗✓✓✗✗
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.3343+0.4644+0.0554−0.3029−0.4193−0.1713+0.8496
Depth 4Shot gain ↑\uparrow0.0%−69.7%77.0%−59.8%0.0%4.2%4.2%74.5%
Depth 4No-worse vs Full✓✗✗✗✓✓✓✗
JW-like parity spinDepth 1ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0465+0.1333+0.0172−0.0230+0.0185−0.0198+0.0440
Depth 1Shot gain ↑\uparrow0.0%−94.4%92.7%−91.5%0.0%36.2%36.1%95.7%
Depth 1No-worse vs Full✓✗✗✗✓✓✓✗
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000−0.0141+0.1421−0.0724−0.0814−0.1151−0.0997+0.0190
Depth 4Shot gain ↑\uparrow0.0%−104.6%72.2%−118.7%0.0%11.5%11.6%75.2%
Depth 4No-worse vs Full✓✓✗✓✓✓✓✗

Table 28. Shot-allocation benchmark on the Table 39 Hamiltonian panel at b=1.0b = 1.0. We reuse the saved depth-1–4 selector and evaluate seven held-out 8-qubit Hamiltonians at depths 1 and 4 over seeds 97/101/107/109, with 28 Adam updates, τ=0.997\tau= 0.997, and 1024 nominal shots per shifted selected parameter–group setting. All methods are rerun against Full-QWC uniform with the same moment-matched Gaussian finite-shot estimator and common Adam optimizer; iCANS [25], gCANS [13], Rosalin [4], and variance-aware allocation [44] are therefore controlled allocation policies rather than reproductions of their complete optimizer stacks. Green/red ΔE\Delta E and ✓/✗ use the one-sided 1.6 mHa criterion. Bold marks the closest non-Full mean ΔE\Delta E; bold-underlining marks the largest shot gain in each cell.

Hamiltonian-Panel Allocation

Across all 56 paired seed-level cases, Ours-uniform saves 10.8% of shots with mean ∣ΔE∣=0.1889|\Delta E| = 0.1889 Ha, while Full-QWC variance-aware allocation has a smaller 0.1614 Ha gap but no shot saving. Composing our method with gCANS reaches an aggressive 82.6% saving and slightly improves on gCANS alone in both saving (80.8%) and mean absolute gap (0.3748 vs. 0.4001 Ha), but both pass only 14.3% of seed-level comparisons. In contrast, iCANS and Rosalin spend 83.5% and 79.3% more shots than Full on average at this operating point. The panel therefore supports composability, while also showing that allocation policies calibrated on easier molecular regimes do not automatically preserve energy under Hamiltonian-family shift.

Depth-OOD Transfer

Depth-OOD Transfer Baselines

This depth-OOD benchmark evaluates the conservative τ=0.997\tau= 0.997 selector on unseen RY/RZ depths 5–7; the table reports depths 5 and 6. A matched-budget control adds a cell-wise cost-matched TITAN comparison: TITAN{sup}† chooses the freezing threshold whose shot gain is closest to Ours, while TITAN{sup}†+Ours uses that same active-parameter mask and additionally sparsifies QWC groups.

Table 29 separates parameter-specific QWC selection from parameter freezing at both the original and cost-matched operating points. The dagger rows remove shot budget as an explanation for differences between our method and TITAN-only training; the composed row shows the additional saving available when the two orthogonal controls are combined.

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 5ΔE\Delta E vs Full ↓\downarrowOurs-0.0003+0.0053-0.0120+0.0012-0.0068
Depth 5ΔE\Delta E vs Full ↓\downarrowRandom+0.0179+0.0684+0.0503+0.1580+2.1893
Depth 5ΔE\Delta E vs Full ↓\downarrowTITAN (Peng et al.)+0.0239+0.0130+0.0146+0.1011+0.2254
Depth 5ΔE\Delta E vs Full ↓\downarrowTITAN+Ours+0.0251+0.0048+0.0111+0.1009+0.2250
Depth 5ΔE\Delta E vs Full ↓\downarrowTITAN{sup}†`-0.0028-0.0059-0.0044+0.0416+0.0914
Depth 5ΔE\Delta E vs Full ↓\downarrowTITAN{sup}†`+Ours-0.0035-0.0056-0.0112+0.0392+0.0928
Depth 5Shot gain ↑\uparrowOurs5.7%8.4%10.6%17.1%52.9%
Depth 5Shot gain ↑\uparrowRandom5.7%8.5%10.7%17.2%55.4%
Depth 5Shot gain ↑\uparrowTITAN (Peng et al.)32.0%32.7%38.7%42.3%73.4%
Depth 5Shot gain ↑\uparrowTITAN+Ours34.3%37.9%44.9%52.2%93.2%
Depth 5Shot gain ↑\uparrowTITAN{sup}†`6.9%12.7%13.1%14.4%52.4%
Depth 5Shot gain ↑\uparrowTITAN{sup}†`+Ours10.3%19.5%22.2%28.4%83.2%
Depth 5No-worse vs FullOurs✓✗✓✓✓
Depth 5No-worse vs FullRandom✗✗✗✗✗
Depth 5No-worse vs FullTITAN (Peng et al.)✗✗✗✗✗
Depth 5No-worse vs FullTITAN+Ours✗✗✗✗✗
Depth 5No-worse vs FullTITAN{sup}†`✓✓✓✗✗
Depth 5No-worse vs FullTITAN{sup}†`+Ours✓✓✓✗✗
Depth 6ΔE\Delta E vs Full ↓\downarrowOurs+0.0000-0.0036+0.0003-0.0081+0.0086
Depth 6ΔE\Delta E vs Full ↓\downarrowRandom+0.0046+0.0005+0.0048+0.0785+0.7440
Depth 6ΔE\Delta E vs Full ↓\downarrowTITAN (Peng et al.)+0.0139-0.0088-0.0426+0.0070+0.3083
Depth 6ΔE\Delta E vs Full ↓\downarrowTITAN+Ours+0.0133-0.0113-0.0429+0.0082+0.3092
Depth 6ΔE\Delta E vs Full ↓\downarrowTITAN{sup}†`-0.0070-0.0242+0.0130-0.0034+0.0825
Depth 6ΔE\Delta E vs Full ↓\downarrowTITAN{sup}†`+Ours-0.0069-0.0280+0.0162-0.0071+0.0798
Depth 6Shot gain ↑\uparrowOurs4.8%7.6%8.7%14.6%44.3%
Depth 6Shot gain ↑\uparrowRandom4.7%7.5%8.7%14.6%46.3%
Depth 6Shot gain ↑\uparrowTITAN (Peng et al.)32.8%30.5%28.1%39.4%71.2%
Depth 6Shot gain ↑\uparrowTITAN+Ours34.6%34.4%33.8%47.6%88.5%
Depth 6Shot gain ↑\uparrowTITAN{sup}†`7.0%9.6%7.5%12.4%37.0%
Depth 6Shot gain ↑\uparrowTITAN{sup}†`+Ours10.1%15.5%15.0%24.8%67.6%
Depth 6No-worse vs FullOurs✓✓✓✓✗
Depth 6No-worse vs FullRandom✗✓✗✗✗
Depth 6No-worse vs FullTITAN (Peng et al.)✗✓✓✗✗
Depth 6No-worse vs FullTITAN+Ours✗✓✓✗✗
Depth 6No-worse vs FullTITAN{sup}†`✓✓✗✓✗
Depth 6No-worse vs FullTITAN{sup}†`+Ours✓✓✗✓✗

Table 29. Table 29: Depth-OOD online VQE transfer baselines at τ=0.997\tau= 0.997. The relevance model is trained on depths 1–4; the table reports unseen depths 5 and 6 over five PySCF/STO-3G molecules and four held-out seeds. Random is group-budget matched to our method. TITAN{sup}† uses a cell-wise freezing threshold selected to match our method’s shot gain; TITAN{sup}†+Ours reuses that threshold and adds group selection. Green/red ΔE\Delta E indicates the one-sided 1.6 mHa no-worse criterion. Bold marks the closest mean ΔE\Delta E to Full; bold-underlined shot gain marks the largest saving.

Additional Depth-OOD Transfer Tables

The tables below report aggregate performance and budget usage for depth-OOD transfer at τ=0.995\tau= 0.995.

SettingMean ∣ΔE∣\lvert\Delta E\rvert ↓\downarrowNo-worseMean shot gain ↑\uparrow
Full-QWC reference0.00000✓0.0%
Ours depth 50.01540✓22.6%
Ours depth 60.00697✓19.6%
Ours depth 70.00684✓17.3%
Random depth 50.54065✗23.5%
Random depth 60.26326✗20.1%
Random depth 70.15211✗17.7%
Ours all0.00974✓19.8%
Random all0.31867✗20.4%

Table 30. Depth-OOD aggregate at τ=0.995\tau= 0.995. Mean ∣ΔE∣\lvert\Delta E\rvert averages the five molecule columns at each unseen depth. No-worse uses the one-sided criterion ΔE≤1.6\Delta E \le1.6 mHa against the Full-QWC reference, so lower final energy is accepted.

MoleculeMethodFull QWC groups/param.Groups / param. ↓\downarrow D5Groups / param. ↓\downarrow D6Groups / param. ↓\downarrow D7Shot gain ↑\uparrow D5Shot gain ↑\uparrow D6Shot gain ↑\uparrow D7No-worse vs Full D5No-worse vs Full D6No-worse vs Full D7Best ΔE\Delta E ↓\downarrow
H2/STO-3GOurs54.654.704.777.0%6.0%4.7%✓✗✗+0.0008
H2/STO-3GRandom54.654.704.777.0%5.9%4.7%✗✗✓+0.0015
HeH+/STO-3GOurs119.739.829.9511.5%10.7%9.6%✗✓✓+0.0042
HeH+/STO-3GRandom119.739.859.9511.5%10.5%9.6%✗✗✗+0.0207
H3+/STO-3GOurs1916.3416.7917.0514.0%11.6%10.2%✓✗✓+0.0015
H3+/STO-3GRandom1916.3616.7417.0613.9%11.9%10.2%✗✗✗+0.0481
H4 chain/STO-3GOurs7558.9960.9662.2521.3%18.7%17.0%✗✓✗−0.0055
H4 chain/STO-3GRandom7558.6961.1162.1321.8%18.5%17.2%✗✗✗+0.0151
LiH/STO-3GOurs17972.6587.4898.9159.4%51.1%44.7%✓✓✓−0.0001
LiH/STO-3GRandom17965.9283.2795.4263.2%53.5%46.7%✗✗✗+0.6498

Table 31. Depth-OOD budget usage at τ=0.995\tau= 0.995. Groups/param. reports the mean selected QWC groups per parameter on unseen depths 5–7. Shot gain is relative to full 2PGS2PGS parameter-shift QWC measurement.

Depth-OOD Baselines by Molecule and Depth

Table 32: Aggregate depth-OOD baselines at τ=0.997\tau= 0.997. This appendix table summarizes the same experiment as Table 29 across all 15 molecule-depth cells and 60 held-out trajectories. CA seeds count held-out trajectories with ∣ΔE∣≤1.6|\Delta E| \le1.6 mHa against Full-QWC. Cells won counts molecule-depth cells where the method has the smallest mean ∣ΔE∣|\Delta E| among all listed methods.

MethodMean ΔE\Delta EMean ∣ΔE∣|\Delta E| ↓CA seedsShot gain ↑Groups/param.Cells won
Ours-0.001180.0071618/6016.2%38.9012/15
Global coeff. same budget+0.002260.0113911/6016.2%38.883/15
Random same budget+0.251630.264312/6016.7%38.130/15
TITAN (Peng et al.)+0.056830.099853/6041.1%24.450/15
TITAN+Ours+0.056470.099814/6048.6%16.520/15

Table 32. Aggregate depth-OOD baselines at τ=0.997\tau= 0.997.

Table 33: Molecule-depth detail for depth-OOD baselines at τ=0.997\tau= 0.997. Each entry reports mean ∣ΔE∣|\Delta E| in Hartree followed by shot saving. Rows are grouped by molecular Hamiltonian and unseen evaluation depth. The model is trained on depths 1–4 and evaluated on depths 5–7 with four held-out seeds per molecule-depth cell. Green bold marks the lowest mean ∣ΔE∣|\Delta E| in each row.

MoleculeDepthOursGlobal coeff.TITAN (Peng et al.)TITAN+OursRandom
H2 (4q)50.00119 / 5.7%0.00786 / 5.7%0.03143 / 32.0%0.03200 / 34.3%0.01792 / 5.7%
H2 (4q)60.00101 / 4.8%0.00915 / 4.7%0.01577 / 32.8%0.01678 / 34.6%0.02628 / 4.7%
H2 (4q)70.00199 / 3.8%0.00352 / 3.8%0.01781 / 29.5%0.01761 / 31.3%0.00650 / 3.8%
HeH+ (4q)50.00572 / 8.4%0.01661 / 8.4%0.08943 / 32.7%0.08406 / 37.9%0.06841 / 8.5%
HeH+ (4q)60.00484 / 7.6%0.01717 / 7.6%0.11144 / 30.5%0.11416 / 34.4%0.03489 / 7.5%
HeH+ (4q)70.01149 / 6.7%0.00857 / 6.7%0.09757 / 31.3%0.09613 / 34.6%0.07671 / 6.6%
H3+ (6q)50.01204 / 10.6%0.01437 / 10.7%0.05297 / 38.7%0.05186 / 44.9%0.05029 / 10.7%
H3+ (6q)60.00422 / 8.7%0.01294 / 8.6%0.04326 / 28.1%0.04484 / 33.8%0.02628 / 8.7%
H3+ (6q)70.00378 / 7.4%0.00473 / 7.4%0.04070 / 31.6%0.04348 / 35.8%0.01635 / 7.3%
H4 chain (8q)50.02352 / 17.1%0.04048 / 17.0%0.12680 / 42.3%0.12740 / 52.2%0.15799 / 17.2%
H4 chain (8q)60.00929 / 14.6%0.01119 / 14.6%0.05705 / 39.4%0.05890 / 47.6%0.07854 / 14.6%
H4 chain (8q)70.00614 / 13.0%0.00704 / 13.1%0.04519 / 36.1%0.04317 / 43.9%0.05636 / 13.0%
LiH (12q)50.00877 / 52.9%0.00208 / 52.9%0.22542 / 73.4%0.22498 / 93.2%2.18927 / 55.4%
LiH (12q)60.01041 / 44.3%0.01414 / 44.4%0.34075 / 71.2%0.33967 / 88.5%0.74399 / 46.3%
LiH (12q)70.00307 / 37.9%0.00095 / 38.0%0.20216 / 66.6%0.20205 / 82.5%0.41488 / 39.7%

Table 33. Molecule-depth detail for depth-OOD baselines at τ=0.997\tau= 0.997.

Controlled Structural OOD Transfer Summary

Figure 3 presents the matched-budget energy comparison compactly; Table 34 gives the corresponding exact energy gaps, success counts, shot gains, and selected-group budgets.

Test HamiltonianMethodFull groupsMean ΔE\Delta EMean ∣ΔE∣|\Delta E| ↓No-worse vs FullShot gain ↑Groups/param.
TFIM h=0.7h = 0.7Ours2−1.55×10−15-1.55 \times10^{-15}3.66×10−153.66 \times10^{-15}16/166.13%1.877
TFIM h=0.7h = 0.7Random2−0.00253-0.002530.007106/166.20%1.876
TFIM h=1.2h = 1.2Ours20.211810.227770/1613.02%1.740
TFIM h=1.2h = 1.2Random20.030080.039881/1613.02%1.740
Heisenberg XXZ+fieldOurs3−0.00011-0.000110.0001116/168.69%2.739
Heisenberg XXZ+fieldRandom30.371280.430000/168.69%2.739
Heisenberg J1–J2Ours30.005630.0056714/168.72%2.738
Heisenberg J1–J2Random30.225920.307620/168.70%2.739
All targetsOurs–0.054330.0583946/649.14%2.274
All targetsRandom–0.156240.196157/649.15%2.273

Table 34. Controlled OOD transfer at τ=0.997\tau= 0.997. (a) Each target Hamiltonian is absent from training; no-worse counts final energies at most 1.6 mHa above Full-QWC, accepting lower energies. (b) Each named ansatz is unseen during training; strict CA counts seeds with ∣ΔE∣≤1.6|\Delta E| \le1.6 mHa, and our method wins counts molecule–depth cells with a smaller mean absolute gap than Random. In both panels, Random uses the same parameter-specific QWC-group count as our method, isolating relevance ordering at a matched measurement budget. Within each paired setting, bold marks the lower mean ∣ΔE∣|\Delta E| and larger success count; underline marks the runner-up. Budget columns are intentionally unranked. Complete TITAN controls are Tables 38 and 41 in Appendices G and H.

OOD ansatzMethodMean ΔE\Delta EMean ∣ΔE∣|\Delta E| ↓Strict CA seedsShot gain ↑Ours wins
EfficientSU2Ours+0.00300.011913/609.5%–
EfficientSU2Random+0.03440.06302/609.4%14/15
HEA-RYRZ-CircularOurs+0.00330.008812/6012.9%–
HEA-RYRZ-CircularRandom+0.06270.06902/6013.2%15/15
RealAmplitudesOurs−0.0006-0.00060.003029/6013.5%–
RealAmplitudesRandom+0.18670.21821/6013.8%15/15
StronglyEntanglingLayersOurs+0.03950.04530/6011.1%–
StronglyEntanglingLayersRandom+0.02710.03902/6011.2%4/15

Table 34. (b) Named-ansatz OOD

Molecule- and Hamiltonian-Family OOD Transfer

The compact comparison between our method and Random on Hamiltonian transfer appears in panel (a) of Table 34 in Appendix § 16; Table 38 below preserves the corresponding TITAN controls.

Test HamiltonianMethodFull groupsMean ΔE\Delta EMean ∣ΔE∣\lvert\Delta E\rvert ↓No-worse vs FullShot gain ↑Groups/param.
TFIM h=0.7h = 0.7Ours2−1.55×10−15-1.55 \times10^{-15}3.66×10−153.66 \times10^{-15}16/166.13%1.877
TFIM h=0.7h = 0.7Random2−0.00253-0.002530.007106/166.20%1.876
TFIM h=0.7h = 0.7TITAN ([32])2−0.00587-0.005870.019641/1656.11%0.878
TFIM h=0.7h = 0.7TITAN+Ours2−0.00587-0.005870.019641/1657.37%0.853
TFIM h=1.2h = 1.2Ours20.211810.227770/1613.02%1.740
TFIM h=1.2h = 1.2Random20.030080.039881/1613.02%1.740
TFIM h=1.2h = 1.2TITAN ([32])20.022930.058891/1649.22%1.016
TFIM h=1.2h = 1.2TITAN+Ours20.227500.242890/1653.81%0.924
Heisenberg XXZ+fieldOurs3−0.00011-0.000110.0001116/168.69%2.739
Heisenberg XXZ+fieldRandom30.371280.430000/168.69%2.739
Heisenberg XXZ+fieldTITAN ([32])30.500140.719730/1645.31%1.641
Heisenberg XXZ+fieldTITAN+Ours30.499670.719260/1646.79%1.596
Heisenberg J1–J2Ours30.005630.0056714/168.72%2.738
Heisenberg J1–J2Random30.225920.307620/168.70%2.739
Heisenberg J1–J2TITAN ([32])30.582580.671151/1644.13%1.676
Heisenberg J1–J2TITAN+Ours30.581650.670701/1645.64%1.631
All targetsOurs–0.054330.0583946/649.14%2.274
All targetsRandom–0.156240.196157/649.15%2.273
All targetsTITAN ([32])–0.274940.367353/6448.69%1.303
All targetsTITAN+Ours–0.325740.413122/6450.90%1.251

Table 38. Full Hamiltonian-family OOD baseline comparison at τ=0.997\tau= 0.997. The selector is trained on molecular Hamiltonians plus one Heisenberg Hamiltonian and evaluated on unseen TFIM/Heisenberg-family Hamiltonians at the same depths 1–4. This appendix table preserves the TITAN and TITAN+Ours controls omitted from the compact comparison in Table 34. No-worse vs Full reports seeds whose final energy is not more than 1.6 mHa above full-QWC training, accepting lower energies. Full groups is the number of QWC groups in the target Hamiltonian. Bold green marks the lowest mean ∣ΔE∣\lvert\Delta E\rvert within each test Hamiltonian; underlined green marks the largest shot gain.

Clean Molecule-OOD Hamiltonian Transfer

For each split, one molecule is held out and the relevance model is trained on the remaining four molecules at depths 1–4. The matched TITAN comparison uses the corresponding holdout-specific checkpoint, so neither our method nor either dagger composition has seen the testing molecule.

Table 35 directly compares group selection and parameter freezing under Hamiltonian shift. The dagger columns control the TITAN-only shot budget cell by cell, while the final column measures the additional savings from composing the matched active-parameter mask with learned QWC-group selection.

MoleculeDepthMetricOursRandomTITAN (Peng et al.)TITAN+OursTITAN†TITAN†+Ours
H2 (4q)Depth 1ΔE\Delta E vs Full ↓+0.0016+0.3323+0.0228+0.0258+0.0262+0.0272
Depth 1Shot gain ↑55.2%56.1%61.8%88.8%56.0%86.0%
Depth 1Chem. acc.✗✗✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓-0.0002+0.0249-0.0003-0.0029+0.0010+0.0007
Depth 4Shot gain ↑9.9%9.7%40.8%45.3%10.2%16.0%
Depth 4Chem. acc.✓✗✓✓✓✓
HeH+ (4q)Depth 1ΔE\Delta E vs Full ↓+0.0021+0.5282+0.1040+0.1063+0.0090+0.0114
Depth 1Shot gain ↑32.2%30.4%56.2%75.6%30.9%58.3%
Depth 1Chem. acc.✗✗✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓-0.0022+0.0452-0.0642-0.0650+0.0111+0.0112
Depth 4Shot gain ↑8.1%8.1%38.3%42.7%4.7%12.5%
Depth 4Chem. acc.✓✗✓✓✗✗
H3+ (6q)Depth 1ΔE\Delta E vs Full ↓+0.0051+0.9872+0.0871+0.0867+0.0871+0.0867
Depth 1Shot gain ↑56.2%60.6%58.2%94.3%58.2%94.3%
Depth 1Chem. acc.✗✗✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓-0.0052+0.0504+0.0116+0.0128+0.0049+0.0017
Depth 4Shot gain ↑10.0%10.0%45.9%52.0%6.2%16.0%
Depth 4Chem. acc.✓✗✗✗✗✗
H4 chain (8q)Depth 1ΔE\Delta E vs Full ↓+0.0038+1.5543+0.0220+0.0236+0.0220+0.0236
Depth 1Shot gain ↑63.7%68.3%62.7%97.4%62.7%97.4%
Depth 1Chem. acc.✗✗✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓+0.0652+0.3159+0.1904+0.1887+0.0770+0.0752
Depth 4Shot gain ↑28.5%29.4%44.7%63.2%28.6%50.4%
Depth 4Chem. acc.✗✗✗✗✗✗
LiH (12q)Depth 1ΔE\Delta E vs Full ↓+0.0022+0.4404-0.0881-0.0880+0.0387+0.0385
Depth 1Shot gain ↑40.4%37.9%71.9%88.5%47.1%71.9%
Depth 1Chem. acc.✗✗✓✓✗✗
Depth 4ΔE\Delta E vs Full ↓-0.0023+0.2301-0.0100-0.0092+0.0516+0.0510
Depth 4Shot gain ↑17.9%17.7%76.2%81.9%20.2%33.8%
Depth 4Chem. acc.✓✗✓✓✗✗

Table 35. Leave-one-molecule-out Hamiltonian transfer at τ=0.997\tau= 0.997. Representative depths 1 and 4 are shown; depths 2 and 3 appear in Appendix Table 37. TITAN† uses a cell-wise freezing threshold selected to match our method’s shot gain; TITAN†+Ours uses the same active-parameter mask and the matching holdout-specific selector. Green/red ΔE\Delta E and ✓/✗ follow the one-sided 1.6 mHa no-worse criterion. Bold marks the closest mean ΔE\Delta E to Full; bold-underlined shot gain marks the largest saving.

Leave-One-Molecule-Out Uncertainty and Intermediate Depths

Table 36: Uncertainty view for the leave-one-molecule-out comparison at τ=0.997\tau= 0.997. This appendix table reports mean ΔE±\Delta E \pm s.d. for the representative depths 1 and 4 shown in Table 35. Green/red coloring follows the mean ΔE\Delta E rule used in Table 35.

MoleculeDepthOursRandomTITAN (Peng et al.)TITAN+Ours
H2H_2 (4q)Depth 1+0.0016 ± 0.0035+0.3323 ± 0.0518+0.0228 ± 0.0259+0.0258 ± 0.0255
H2H_2 (4q)Depth 4−0.0002 ± 0.0099+0.0249 ± 0.0284−0.0003 ± 0.0206−0.0029 ± 0.0213
HeH+ (4q)Depth 1+0.0021 ± 0.0034+0.5282 ± 0.1378+0.1040 ± 0.0586+0.1063 ± 0.0578
HeH+ (4q)Depth 4−0.0022 ± 0.0065+0.0452 ± 0.0291−0.0642 ± 0.0614−0.0650 ± 0.0628
H3+H_3^+ (6q)Depth 1+0.0051 ± 0.0041+0.9872 ± 0.0979+0.0871 ± 0.0466+0.0867 ± 0.0453
H3+H_3^+ (6q)Depth 4−0.0052 ± 0.0072+0.0504 ± 0.0457+0.0116 ± 0.0819+0.0128 ± 0.0771
H4 chain (8q)Depth 1+0.0038 ± 0.0051+1.5543 ± 0.1952+0.0220 ± 0.0193+0.0236 ± 0.0171
H4 chain (8q)Depth 4+0.0652 ± 0.0666+0.3159 ± 0.1015+0.1904 ± 0.0841+0.1887 ± 0.0819
LiH (12q)Depth 1+0.0022 ± 0.0031+0.4404 ± 0.1245−0.0881 ± 0.0470−0.0880 ± 0.0466
LiH (12q)Depth 4−0.0023 ± 0.0049+0.2301 ± 0.1702−0.0100 ± 0.2342−0.0092 ± 0.2339
Table 36.

Table 67.

MoleculeDepthMetricOursRandomTITAN ((Peng et al.))TITAN+Ours
H2 (4q)Depth 2ΔE\Delta E vs Full ↓\downarrow+0.0046 ± 0.0063+0.1387 ± 0.1420−0.0200 ± 0.1784−0.0180 ± 0.1775
Depth 2Shot gain ↑\uparrow29.9%30.1%54.0%72.3%
Depth 2Chem. acc.✗✗✓✓
Depth 3ΔE\Delta E vs Full ↓\downarrow+0.0576 ± 0.1057+0.0814 ± 0.0453+0.0223 ± 0.0453+0.0205 ± 0.0409
Depth 3Shot gain ↑\uparrow19.3%21.3%39.7%52.0%
Depth 3Chem. acc.✗✗✗✗
HeH+ (4q)Depth 2ΔE\Delta E vs Full ↓\downarrow+0.0167 ± 0.0183+0.2844 ± 0.0547−0.1194 ± 0.1021−0.1191 ± 0.1008
Depth 2Shot gain ↑\uparrow22.2%23.7%54.4%70.5%
Depth 2Chem. acc.✗✗✓✓
Depth 3ΔE\Delta E vs Full ↓\downarrow+0.0133 ± 0.0145+0.0677 ± 0.1411+0.0112 ± 0.0424+0.0102 ± 0.0444
Depth 3Shot gain ↑\uparrow13.2%13.6%43.3%52.2%
Depth 3Chem. acc.✗✗✗✗
H3+ (6q)Depth 2ΔE\Delta E vs Full ↓\downarrow+0.0013 ± 0.0018+0.3811 ± 0.0782−0.0031 ± 0.0610−0.0036 ± 0.0629
Depth 2Shot gain ↑\uparrow27.6%27.3%50.0%71.0%
Depth 2Chem. acc.✓✗✓✓
Depth 3ΔE\Delta E vs Full ↓\downarrow+0.0017 ± 0.0076+0.1683 ± 0.1216+0.0338 ± 0.0777+0.0332 ± 0.0800
Depth 3Shot gain ↑\uparrow16.2%16.1%47.1%57.7%
Depth 3Chem. acc.✗✗✗✗
H4 chain (8q)Depth 2ΔE\Delta E vs Full ↓\downarrow+0.0065 ± 0.0112+1.1356 ± 0.1572+0.1521 ± 0.1283+0.1556 ± 0.1269
Depth 2Shot gain ↑\uparrow52.3%57.1%59.1%88.9%
Depth 2Chem. acc.✗✗✗✗
Depth 3ΔE\Delta E vs Full ↓\downarrow+0.0107 ± 0.0102+0.3099 ± 0.0388−0.0650 ± 0.1313−0.0615 ± 0.1294
Depth 3Shot gain ↑\uparrow39.3%40.6%48.3%72.9%
Depth 3Chem. acc.✗✗✓✓
LiH (12q)Depth 2ΔE\Delta E vs Full ↓\downarrow+0.0026 ± 0.0026+1.0068 ± 0.1733+0.0806 ± 0.0923+0.0803 ± 0.0926
Depth 2Shot gain ↑\uparrow32.1%30.6%77.1%87.1%
Depth 2Chem. acc.✗✗✗✗
Depth 3ΔE\Delta E vs Full ↓\downarrow+0.0038 ± 0.0133+0.5742 ± 0.2506+0.3972 ± 0.2246+0.3975 ± 0.2246
Depth 3Shot gain ↑\uparrow24.1%23.6%76.4%84.2%
Depth 3Chem. acc.✗✗✗✗

Table 37. Appendix depth-2/3 view for leave-one-molecule-out Hamiltonian transfer at τ=0.997\tau= 0.997. This table uses the same protocol and layout as Table 35: one molecule is held out for testing, the remaining four molecules train the selector, and depths are kept fixed between train and test. ΔE\Delta E entries report mean ± s.d. across held-out seeds. ✓ denotes the one-sided no-worse chemical-accuracy criterion ΔE≤1.6\Delta E \leq1.6 mHa relative to Full-QWC training; ✗ denotes a larger positive energy gap. In ΔE\Delta E rows, green means ΔE≤1.6\Delta E \leq1.6 mHa (including negative gaps) and red means a larger positive gap.

Full Hamiltonian-Family Baseline Comparison

This Hamiltonian-transfer experiment asks whether the learned relevance selector transfers beyond molecular electronic-structure Hamiltonians. We train one shared selector on five PySCF/OpenFermion molecular Hamiltonians (H2, HeH+, H3+, H4 chain, and LiH; STO-3G basis) plus one 8-qubit Heisenberg-chain Hamiltonian, using ansatz depths 1–4. We then evaluate closed-loop online VQE at the same depths on four unseen spin Hamiltonians: two transverse-field Ising models (TFIM) and two Heisenberg variants. The ansatz is the same layered RY/RZ circuit with a linear entangler, and the top-mass threshold is fixed at τ=0.997\tau= 0.997.

The result separates Hamiltonian-family transfer from ordinary molecular holdout. Our method remains the most accurate method on three of the four unseen spin Hamiltonians, including both Heisenberg variants and TFIM at h=0.7h = 0.7. In contrast, TITAN-style parameter freezing obtains much larger shot savings but usually changes the optimization trajectory enough to lose final-energy fidelity. TFIM at h=1.2h = 1.2 is the failure case: the learned molecular-plus-Heisenberg relevance ordering is not yet robust to this field regime. This suggests that Hamiltonian-family coverage, not just more molecular seeds, is the next dataset axis for a stronger general-purpose selector.

Hamiltonian-Panel Stress Test

This expanded panel systematically tests Hamiltonian-family OOD transfer. We train one shared selector on five molecular Hamiltonians plus two 8-qubit Heisenberg-chain Hamiltonians, using depths 1–4 and the same layered RY/RZ ansatz. We then evaluate the saved model on held-out transverse-field Ising models, unseen Heisenberg variants, and a larger-group JW-like parity spin Hamiltonian at depths 1–4 with τ=0.997\tau= 0.997. The baseline comparison adds cell-wise cost-matched TITAN controls so that parameter freezing and QWC-group selection are compared at similar measurement budgets.

Table 39 shows a useful but bounded generalization result. Our method is full-QWC-equivalent on the held-out Heisenberg variants and low-field TFIM, whereas shot-matched parameter freezing still incurs large positive drift on most Heisenberg cells. On low-field TFIM, our method saves only 2–3%, so the nearest TITAN-only operating point is no freezing; the composed dagger column then reduces exactly to our method group selection. The JW-like target remains a harder family shift, although matched TITAN substantially narrows the gap relative to aggressive freezing. These controls show that the proposed method’s stability is not explained by a more conservative parameter-freezing threshold.

Testing HamiltonianDepthMetricOursRandomTITAN (Peng et al.)TITAN+OursTITAN†TITAN†+Ours
TFIM h=0.5h = 0.5Depth 1ΔE\Delta E vs Full ↓+0.0000-0.0043-0.0139-0.0139+0.0000+0.0000
Depth 1Shot gain ↑2.0%2.0%56.8%57.5%0.0%2.0%
Depth 1Chem. acc.✓✓✓✓✓✓
Depth 4ΔE\Delta E vs Full ↓-0.0002-0.0008+0.0062+0.0063+0.0000-0.0002
Depth 4Shot gain ↑2.6%2.6%59.0%59.3%0.0%2.6%
Depth 4Chem. acc.✓✓✗✗✓✓
TFIM h=0.7h = 0.7Depth 1ΔE\Delta E vs Full ↓+0.0000-0.0138-0.0227-0.0227+0.0000+0.0000
Depth 1Shot gain ↑2.9%2.9%56.4%57.6%0.0%2.9%
Depth 1Chem. acc.✓✓✓✓✓✓
Depth 4ΔE\Delta E vs Full ↓+0.0000-0.0018+0.0107+0.0107+0.0000+0.0000
Depth 4Shot gain ↑2.4%2.4%53.0%53.5%0.0%2.4%
Depth 4Chem. acc.✓✓✗✗✓✓
TFIM h=0.9h = 0.9Depth 1ΔE\Delta E vs Full ↓+0.0771-0.0010-0.0018+0.0784-0.0002+0.0771
Depth 1Shot gain ↑25.0%25.0%54.4%59.8%27.1%37.6%
Depth 1Chem. acc.✗✓✓✗✓✗
Depth 4ΔE\Delta E vs Full ↓+0.0302+0.0064+0.0028+0.0351-0.0012+0.0306
Depth 4Shot gain ↑6.2%6.2%47.1%49.0%6.0%11.1%
Depth 4Chem. acc.✗✗✗✗✓✗
Heis. XXZ+X fieldDepth 1ΔE\Delta E vs Full ↓+0.0000+0.8893+1.2891+1.2891+0.2391+0.2391
Depth 1Shot gain ↑16.7%16.7%52.2%53.5%19.4%29.6%
Depth 1Chem. acc.✓✗✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓+0.0000+0.3996+0.5969+0.5969-0.1146-0.1146
Depth 4Shot gain ↑4.2%4.2%36.3%38.1%3.7%7.7%
Depth 4Chem. acc.✓✗✗✗✓✓
Heis. J1–J2Depth 1ΔE\Delta E vs Full ↓+0.0000+0.6602+0.9835+0.9835+0.0863+0.0863
Depth 1Shot gain ↑16.7%16.7%50.6%52.2%12.7%25.1%
Depth 1Chem. acc.✓✗✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓+0.0000-0.0010+0.5678+0.5678+0.1550+0.1550
Depth 4Shot gain ↑4.2%4.2%34.2%36.0%3.5%7.4%
Depth 4Chem. acc.✓✓✗✗✗✗
Heis. XXY+Z fieldDepth 1ΔE\Delta E vs Full ↓+0.0000+0.9213+0.8904+0.8904+0.1382+0.1382
Depth 1Shot gain ↑16.7%16.7%49.6%51.9%18.9%29.4%
Depth 1Chem. acc.✓✗✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓+0.0000+0.0742-0.0137-0.0137+0.0219+0.0219
Depth 4Shot gain ↑4.2%4.2%31.1%33.3%3.6%7.5%
Depth 4Chem. acc.✓✗✓✓✗✗
JW-like parity spinDepth 1ΔE\Delta E vs Full ↓+0.0312+0.6318+0.1336+0.1387+0.0049+0.0350
Depth 1Shot gain ↑35.7%35.6%55.5%76.9%31.4%59.5%
Depth 1Chem. acc.✗✗✗✗✓✗
Depth 4ΔE\Delta E vs Full ↓-0.0046+0.0076+0.0822+0.0738+0.0012-0.0198
Depth 4Shot gain ↑11.6%11.5%45.7%52.1%9.8%19.6%
Depth 4Chem. acc.✓✗✗✗✓✓

Table 39. Hamiltonian-panel transfer at τ=0.997\tau= 0.997. Each testing Hamiltonian is held out from training; representative depths 1 and 4 are shown. TITAN† uses the cell-wise freezing threshold closest to our method’s shot gain, and TITAN†+Ours reuses that active-parameter mask before group selection. Green/red ΔE\Delta E and ✓/✗ follow the one-sided 1.6 mHa no-worse criterion. Bold marks the closest mean ΔE\Delta E to Full; bold-underlined shot gain marks the largest saving.

Joint Molecule+Depth OOD Stress Test

This joint-OOD experiment is preserved as a strict stress test rather than the clean molecule-OOD experiment. The model is trained on H2H_2, HeH+\mathrm{HeH}^{+}, H3+\mathrm{H_3}^{+}, and LiH at RY/RZ depths 1–4, then evaluated without retraining on H4H_4 chain at unseen depths 5–7. Thus both the molecular Hamiltonian and the ansatz depth are held out. We do not claim this is the first cross-molecule VQE transfer experiment broadly: prior work such as Meta-VQE [5] studies learning over parameterized Hamiltonians. Here the role of this appendix result is narrower: it probes whether a learned parameter-by-QWC measurement policy survives simultaneous Hamiltonian and depth shift.

Table 40 shows that joint molecule+depth OOD transfer is substantially harder than depth-OOD within the same molecule panel. Our method is much more stable than matched-budget random selection and wins the hardest H4H_4 depth-5 cell, but the static Global coefficient baseline is competitive on depths 6–7 and obtains the lowest aggregate mean absolute gap. We therefore treat this joint shift as a stress test and diagnostic rather than the clean molecule-OOD result: the current learner captures useful transferable structure, but stronger Hamiltonian/QWC group features are needed for robust cross-Hamiltonian generalization.

MethodDepth 5Depth 6Depth 7Mean ∣ΔE∣\lvert\Delta E\rvert ↓No-worse seedsShot gain ↑
Ours+0.0342 / 23.5%+0.0030 / 20.3%+0.0060 / 18.8%0.02516/1220.9%
Random+0.1682 / 23.6%+0.1171 / 20.4%+0.0322 / 18.9%0.10641/1221.0%
Global coeff.+0.0380 / 23.8%−0.0020 / 20.2%+0.0049 / 18.9%0.02015/1220.9%
TITAN (Peng et al.)+0.1011 / 42.3%+0.0070 / 39.4%−0.0289 / 36.1%0.07635/1239.3%
TITAN+Ours+0.1040 / 56.8%+0.0071 / 52.7%−0.0300 / 48.3%0.07915/1252.6%

Table 40. Joint molecule+depth OOD transfer to H4H_4 at τ=0.997\tau= 0.997. The relevance model is trained on H2H_2, HeH+\mathrm{HeH}^{+}, H3+\mathrm{H_3}^{+}, and LiH with RY/RZ depths 1–4 and evaluated on held-out H4H_4 chain at unseen depths 5–7 with seeds 97/101/107/109. Each depth cell reports mean ΔE\Delta E versus full-QWC followed by mean shot saving. Mean ∣ΔE∣\lvert\Delta E\rvert averages the three H4H_4 depths. No-worse counts held-out seeds with ΔE≤1.6\Delta E \le1.6 mHa, so lower final energy is accepted. Global coefficient uses the same per-parameter group count as Ours but ranks QWC groups by Hamiltonian coefficient magnitude.

Ansatz-Family OOD Transfer

The compact comparison between our method and Random on named-ansatz transfer appears in panel (b) of Table 34 in Appendix F; Table 41 below preserves the corresponding cost-matched TITAN controls.

Table 41: Full zero-shot baseline comparison on named ansatz families. The saved depth-OOD selector is trained on HEA-RYRZ-linear depths 1–4 and evaluated without retraining on depths 5–7, five STO-3G molecules, and four held-out seeds per molecule-depth cell at τ=0.997\tau= 0.997. This appendix table preserves the TITAN{sup}† and TITAN{sup}†+Ours controls omitted from the compact comparison in Table 34. Random uses the same per-parameter group budget as our method. TITAN{sup}† directly chooses the active-parameter count in each molecule-depth cell so its shot gain matches our method; TITAN{sup}†+Ours applies group sparsification inside that active set. Strict CA counts seeds with ∣ΔE∣≤1.6|\Delta E| \le1.6 mHa. The final column reports cells, out of 15, where our method has a smaller mean absolute gap than the method in that row.

OOD ansatzMethodMean ΔE\Delta EMean ∣ΔE∣|\Delta E| ↓Strict CA seedsShot gain ↑Ours wins
EfficientSU2Ours+0.00300.011913/609.5%–
EfficientSU2Random+0.03440.06302/609.4%14/15
EfficientSU2TITAN{sup}†−0.00480.011814/609.5%7/15
EfficientSU2TITAN{sup}†+Ours−0.00100.01756/6017.4%9/15
HEA-RYRZ-CircularOurs+0.00330.008812/6012.9%–
HEA-RYRZ-CircularRandom+0.06270.06902/6013.2%15/15
HEA-RYRZ-CircularTITAN{sup}†−0.00120.019714/6013.0%9/15
HEA-RYRZ-CircularTITAN{sup}†+Ours+0.00110.02098/6023.0%11/15
RealAmplitudesOurs−0.00060.003029/6013.5%–
RealAmplitudesRandom+0.18670.21821/6013.8%15/15
RealAmplitudesTITAN{sup}†+0.01510.04298/6013.5%15/15
RealAmplitudesTITAN{sup}†+Ours+0.01550.04355/6023.4%15/15
StronglyEntanglingLayersOurs+0.03950.04530/6011.1%–
StronglyEntanglingLayersRandom+0.02710.03902/6011.2%4/15
StronglyEntanglingLayersTITAN{sup}†−0.00320.008034/6011.1%0/15
StronglyEntanglingLayersTITAN{sup}†+Ours−0.03430.04070/6020.0%5/15

Table 41.

Full Named-Ansatz Zero-Shot Baselines

The preceding depth-OOD experiments keep the ansatz family fixed and increase depth. We next test a harder transfer axis: the saved depth-OOD model is trained only on the HEA-RYRZ-linear ansatz at depths 1–4, then evaluated zero-shot on named ansatz families at depths 5–7. The molecule panel, held-out seeds, top-mass threshold, and online VQE protocol are unchanged: H2, HeH+, H3+, H4 chain, and LiH in STO-3G; four held-out seeds per molecule-depth cell; τ=0.997\tau= 0.997; and 1024 shots per parameter-shift measurement setting.

Table 41 shows that the learned relevance ordering transfers well to EfficientSU2, HEA-RYRZ-Circular, and RealAmplitudes, where our method beats Random on 44 of 45 cells. Cost matching rules out a universal advantage from parameter freezing: TITAN{sup}† is comparable on EfficientSU2, but our method has smaller gaps on 9/15 HEA-RYRZ-Circular cells and all 15 RealAmplitudes cells. The ordering reverses on StronglyEntanglingLayers, where TITAN{sup}† reaches a 0.0080 Ha mean absolute gap while our method fails. Composing the controls adds 8.9–10.0 percentage points of shot saving beyond the matched TITAN arm, but generally increases energy error. Thus parameter freezing and group selection are complementary cost axes, while their reliability remains ansatz dependent.

Predictor Architecture Under Named-Ansatz OOD

OOD ansatzPredictorMean ΔE\Delta EMean ∣ΔE∣\lvert\Delta E\rvert ↓Strict CA seedsShot gain ↑Transformer wins
EfficientSU2Transformer+0.00300.011913/609.5%–
EfficientSU2MLP+0.00050.00889/608.7%6/15
EfficientSU2ResNet-18−0.00030.006623/6011.0%‾\underline{11.0\%}3/15
EfficientSU2GNN+0.00090.008922/605.1%2/15
HEA-RYRZ-CircularTransformer+0.00330.008812/6012.9%‾\underline{12.9\%}–
HEA-RYRZ-CircularMLP+0.00380.010215/6012.0%8/15
HEA-RYRZ-CircularResNet-18+0.00140.004831/6012.7%4/15
HEA-RYRZ-CircularGNN+0.00380.009416/6010.7%5/15
RealAmplitudesTransformer−0.00060.003029/6013.5%–
RealAmplitudesMLP−0.00330.010617/6017.6%‾\underline{17.6\%}14/15
RealAmplitudesResNet-18−0.00200.008023/6015.5%9/15
RealAmplitudesGNN+0.00310.01499/6017.0%14/15
StronglyEntanglingLayersTransformer+0.03950.04530/6011.1%–
StronglyEntanglingLayersMLP+0.01480.02435/6011.0%1/15
StronglyEntanglingLayersResNet-18−0.00020.007327/6011.3%‾\underline{11.3\%}0/15
StronglyEntanglingLayersGNN+0.00180.011610/609.8%0/15

Table 42. Predictor comparison under named-ansatz OOD at τ=0.997\tau= 0.997. Transformer denotes the depth-OOD predictor used by our method in Table 34. MLP, ResNet-18, and GNN use the same labels, token features, train/evaluation splits, and online VQE protocol; only the relevance predictor changes. The final column counts cells, out of 15, where Transformer has a smaller mean absolute gap than the method in that row. Boldface marks the smallest mean absolute gap and largest strict-CA count per ansatz; bold-underlined values mark the largest shot gain.

ResNet-18 is the strongest ansatz-OOD predictor on three of four targets: it reduces the mean absolute gap to 0.0066 Ha on EfficientSU2, 0.0048 Ha on HEA-RYRZ-Circular, and 0.0073 Ha on StronglyEntanglingLayers while retaining 11.0–12.7% shot savings. The Transformer remains best on RealAmplitudes at 0.0030 Ha. This result refines the earlier architecture conclusion: token cross-attention is strongest for molecule OOD, whereas convolutional grid inductive bias is unexpectedly robust to unseen ansatz topology. No single predictor dominates every transfer axis.

Multi-Ansatz Training Improves Transfer

Single-ansatz training isolates a hard transfer question: a selector trained only on HEA-RYRZ-linear circuits can transfer to some named ansatz families, but fails on StronglyEntanglingLayers. Multi-ansatz training keeps the architecture, molecule panel, online VQE protocol, threshold, held-out seeds, and depth-OOD setting fixed, but changes the training distribution. The learner is now trained on 11,520 snapshots from four ansatz families—HEA-RYRZ-linear, HEA-RYRZ-circular, EfficientSU2, and RealAmplitudes—across depths 1–4, then evaluated on depths 5–7.

Table 43: Multi-ansatz training improves named-ansatz transfer. The single-ansatz baseline trains the relevance selector only on HEA-RYRZ-linear circuits. The multi-ansatz variant trains the same model architecture on a mixed ansatz panel: HEA-RYRZ-linear, HEA-RYRZ-circular, EfficientSU2, and RealAmplitudes. Both settings evaluate online VQE on the same five STO-3G molecular Hamiltonians, depths 5–7, four held-out seeds per molecule-depth cell, τ=0.997\tau= 0.997, and 1024 shots per parameter-shift measurement setting. “Abs-gap gain” is the reduction in mean ∣ΔE∣|\Delta E| from single-ansatz to multi-ansatz training; positive values mean multi-ansatz training improved the final-energy match to Full-QWC.

Target ansatzTraining regimeMean ΔE\Delta EMean ∣ΔE∣|\Delta E| ↓Abs-gap gain ↑Strict CA seedsShot gain ↑Cells won vs Random
HEA-RYRZ-CircularSingle ansatz+0.00330.0088–12/6012.9%15/15
HEA-RYRZ-CircularOurs multi-ansatz+0.00150.0032+0.005734/6014.0%14/15
EfficientSU2Single ansatz+0.00300.0119–13/609.5%14/15
EfficientSU2Ours multi-ansatz+0.00020.0069+0.005024/6015.1%15/15
RealAmplitudesSingle ansatz−0.00060.0030–29/6013.5%15/15
RealAmplitudesOurs multi-ansatz−0.00080.0034-0.000422/6020.8%15/15
StronglyEntanglingLayersSingle ansatz+0.03950.0453–0/6011.1%4/15
StronglyEntanglingLayersOurs multi-ansatz+0.03600.0476-0.00232/6013.4%3/15

Table 43. Multi-ansatz training improves named-ansatz transfer. The single-ansatz baseline trains the relevance selector only on HEA-RYRZ-linear circuits. The multi-ansatz variant trains the same model architecture on a mixed ansatz panel: HEA-RYRZ-linear, HEA-RYRZ-circular, EfficientSU2, and RealAmplitudes. Both settings evaluate online VQE on the same five STO-3G molecular Hamiltonians, depths 5–7, four held-out seeds per molecule-depth cell, τ=0.997\tau= 0.997, and 1024 shots per parameter-shift measurement setting. “Abs-gap gain” is the reduction in mean ∣ΔE∣|\Delta E| from single-ansatz to multi-ansatz training; positive values mean multi-ansatz training improved the final-energy match to Full-QWC.

The multi-ansatz training distribution substantially improves transfer to HEA-RYRZ-Circular and EfficientSU2: mean absolute gap drops from 0.0088 to 0.0032 Ha for HEA-RYRZ-Circular and from 0.0119 to 0.0069 Ha for EfficientSU2, while strict chemical-accuracy seed counts increase from 12/60 to 34/60 and from 13/60 to 24/60, respectively. RealAmplitudes remains a strong case but shifts toward a more measurement-frugal operating point, increasing shot saving from 13.5% to 20.8% at a small cost in mean absolute gap. StronglyEntanglingLayers remains the hard OOD failure case. This suggests that our method learns transferable structure-conditioned relevance rules when the training panel covers comparable entangler families, but substantially different entangling layers require either explicit gate/topology embeddings or inclusion in the training distribution.

Controlled Same-Family Ansatz OOD Transfer With Baselines

This comparison extends the controlled same-family ansatz-transfer setting with TITAN-style parameter-freezing baselines under the identical online VQE protocol. For each ansatz family, the learned selector is trained on two topology variants and evaluated on held-out same-family topology variants. The test ansatz is unseen during training, but the transfer is controlled: same rotation family, same molecular panel, train depths 1–4, unseen evaluation depths 5–7, four held-out seeds per molecule-depth cell, τ=0.997\tau= 0.997, and 1024 shots per parameter-shift measurement setting.

Table 44 shows that the controlled same-family transfer result is not explained by measurement budget alone. Matched-budget Random keeps similar shot savings but is substantially less accurate. TITAN-style freezing can reduce cost by dropping parameters, and TITAN+Ours can compound this saving, but these rows often trade away final-energy fidelity. The cleanest read is that our method targets a different axis from parameter freezing: for parameters that are still measured, it learns which QWC groups are worth measuring.

Family / train topologiesHeld-out targetMethodMean ΔE\Delta EMean ∣ΔE∣|\Delta E| ↓Shot gain ↑CA excess (mHa) ↓Cells won
HEA-RYRZ / linear + reverse-linearCircularOurs+0.00170.005512.40%0.0914/15
HEA-RYRZ / linear + reverse-linearCircularRandom+0.06140.071612.70%59.80–
HEA-RYRZ / linear + reverse-linearCircularTITAN (Peng et al. [32])+0.00740.045833.00%5.776/15
HEA-RYRZ / linear + reverse-linearCircularTITAN+Ours+0.00740.045140.10%5.826/15
HEA-RYRZ / linear + reverse-linearFullOurs−0.00020.00858.50%0.0014/15
HEA-RYRZ / linear + reverse-linearFullRandom+0.02800.04608.50%26.45–
HEA-RYRZ / linear + reverse-linearFullTITAN (Peng et al. [32])+0.04840.100740.80%46.822/15
HEA-RYRZ / linear + reverse-linearFullTITAN+Ours+0.04810.101044.60%46.542/15
EfficientSU2 / linear + reverse-linearCircularOurs+0.00190.005612.40%0.2715/15
EfficientSU2 / linear + reverse-linearCircularRandom+0.06120.071612.60%59.59–
EfficientSU2 / linear + reverse-linearCircularTITAN (Peng et al. [32])+0.00740.045833.00%5.776/15
EfficientSU2 / linear + reverse-linearCircularTITAN+Ours+0.00730.045040.10%5.696/15
EfficientSU2 / linear + reverse-linearFullOurs−0.00020.00898.60%0.0012/15
EfficientSU2 / linear + reverse-linearFullRandom+0.03770.06488.50%36.11–
EfficientSU2 / linear + reverse-linearFullTITAN (Peng et al. [32])+0.04840.100740.80%46.823/15
EfficientSU2 / linear + reverse-linearFullTITAN+Ours+0.04810.100944.60%46.513/15
RealAmplitudes / linear + reverse-linearCircularOurs+0.00520.017719.60%3.6313/15
RealAmplitudes / linear + reverse-linearCircularRandom+0.06840.085819.90%66.75–
RealAmplitudes / linear + reverse-linearCircularTITAN (Peng et al. [32])−0.02450.051218.00%0.0011/15
RealAmplitudes / linear + reverse-linearCircularTITAN+Ours−0.02400.052733.40%0.0010/15
RealAmplitudes / linear + reverse-linearFullOurs+0.00120.023513.90%0.0014/15
RealAmplitudes / linear + reverse-linearFullRandom+0.06480.088014.00%63.18–
RealAmplitudes / linear + reverse-linearFullTITAN (Peng et al. [32])−0.00140.088522.80%0.009/15
RealAmplitudes / linear + reverse-linearFullTITAN+Ours−0.00140.089933.90%0.009/15
StronglyEntangling / range-1 + range-2Layer-dependentOurs+0.00120.007712.40%0.0013/15
StronglyEntangling / range-1 + range-2Layer-dependentRandom+0.03430.044712.60%32.69–
StronglyEntangling / range-1 + range-2Layer-dependentTITAN (Peng et al. [32])+0.01040.045439.60%8.804/15
StronglyEntangling / range-1 + range-2Layer-dependentTITAN+Ours+0.01060.045645.50%9.015/15
StronglyEntangling / range-1 + range-2FullOurs+0.00230.02068.90%0.6613/15
StronglyEntangling / range-1 + range-2FullRandom+0.05960.09689.10%57.97–
StronglyEntangling / range-1 + range-2FullTITAN (Peng et al. [32])+0.04930.107147.30%47.684/15
StronglyEntangling / range-1 + range-2FullTITAN+Ours+0.05020.106650.70%48.594/15

Table 44. Controlled same-family ansatz OOD transfer with matched baselines at τ=0.997\tau= 0.997. Each block trains our method on two topology variants from one ansatz family and evaluates on held-out same-family topologies that do not appear in the training labels. TITAN (Peng et al. [32]) applies active-parameter freezing and full-QWC measurements for active parameters; TITAN+Ours composes the same active-parameter mask with our method group selection. CA excess reports 1000max⁡(ΔE−0.0016,0)1000 \max(\Delta E - 0.0016, 0) in mHa, so lower energies are accepted and smaller is better. “Cells won” counts molecule-depth cells, out of 15, where the method has smaller mean ∣ΔE∣|\Delta E| than matched-budget Random.

Cell-Wise Breakdown of Table 44

Table 45 expands the “Cells won” column from Table 44. Each entry reports the molecule-depth cell improvement in mean absolute final-energy gap, measured in mHa:

1000×(∣ΔE∣Random−∣ΔE∣Ours).(9)1000 \times\left(\left|\Delta E\right|_{\mathrm{Random}}-\left|\Delta E\right|_{\mathrm{Ours}}\right). \tag*{(9)}

Positive green values mean Ours is closer to Full-QWC for that exact molecule-depth cell; negative red values indicate a random same-budget win.

Held-out targetH2_2 D5H2_2 D6H2_2 D7HeH+^{+} D5HeH+^{+} D6HeH+^{+} D7H3+_3^{+} D5H3+_3^{+} D6H3+_3^{+} D7H4_4 D5H4_4 D6H4_4 D7LiH D5LiH D6LiH D7Wins
HEA-RYRZ / Circular+11.4+4.8-0.4+30.9+5.0+3.0+18.0+16.0+8.7+22.7+47.7+36.8+320.2+198.4+268.814/15
HEA-RYRZ / Full+18.1+22.6-0.0+37.0+10.2+10.8+46.5+32.6+0.7+22.9+8.3+51.8+42.9+109.2+148.214/15
EfficientSU2 / Circular+3.5+5.3+0.3+28.0+5.7+0.9+14.5+13.3+16.6+25.2+53.8+111.6+280.9+215.2+214.515/15
EfficientSU2 / Full+18.1+14.8-1.8+31.5-4.5+22.9+38.6+13.8-0.2+125.8+20.9+51.5+109.9+206.7+190.712/15
RealAmplitudes / Circular+19.9+7.8-2.1+94.3+0.5+42.1+24.5+18.2-8.1+80.6+33.9+63.8+188.2+190.3+267.013/15
RealAmplitudes / Full+34.1-1.6+0.0+33.1+27.2+99.1+48.6+121.0+39.5+121.7+18.1+5.6+148.9+75.9+195.214/15
StronglyEntangling / Layer-dependent-3.6+12.9+2.7+15.4+18.1+4.9+13.7-1.7+27.0+39.2+2.6+29.1+183.9+84.9+125.513/15
StronglyEntangling / Full+10.7+9.0+21.3+83.2-14.9+11.3+4.9+14.1+20.1+62.3-23.3+5.3+473.9+249.6+215.913/15

Table 45. Cell-wise breakdown of the same-family ansatz-OOD wins. Entries are Random-minus-Ours mean ∣ΔE∣\left|\Delta E\right| in mHa for each molecule and depth. Green positive entries are cells won by Ours; red negative entries are cells won by the matched-budget random baseline.

Molecule-Depth Breakdown of Table 44

Figure 9 expands the “Cells won” column from Table 44 into individual molecule-depth cells. Each entry reports

Molecule-depth breakdown of same-family ansatz transfer

Figure 9. Molecule-depth breakdown for Table 44. Each panel corresponds to one controlled same-family ansatz-OOD held-out target. Rows are molecules, columns are unseen evaluation depths 5–7, and entries are Random-minus-Ours mean ∣ΔE∣|\Delta E| in mHa. Green positive entries are cells where our method is closer to Full-QWC than matched-budget Random. The color scale is clipped for readability, but numeric labels show the unclipped values rounded to one decimal place.

1000×(∣ΔE∣Random−∣ΔE∣Ours),(10)1000 \times\left(|\Delta E|_{\mathrm{Random}} - |\Delta E|_{\mathrm{Ours}}\right), \tag*{(10)}

measured in mHa. Positive green values mean our method is closer to Full-QWC for that exact molecule-depth cell; negative red values mean the matched-budget random selector is closer.

Predictor Architecture Tables

Figure 10 gives the aggregate architecture comparison referenced in the main text. Figure 11 provides its cell-level counterpart. On this held-out snapshot, the Transformer captures 96.8% of the exact relevance mass while selecting 76.8% of parameter–group cells. ResNet-18 captures slightly more mass (98.6%) by selecting 89.3% of cells, whereas the GNN is sparser but misses more oracle-relevant groups. Similar scalar budgets can therefore arise from materially different parameter-specific masks.

Figure 10. Predictor robustness as QWC measurement complexity grows. We use leave-one-molecule-out transfer at τ=0.997\tau= 0.997 and order the five held-out molecular Hamiltonians by their Full-QWC group counts G∈{5,11,19,75,179}G \in\{5, 11, 19, 75, 179\}. Panels (a,b) report the mean and 90th-percentile absolute final-energy gap to Full QWC; panel (c) shows the error-quantile curve over the three measurement-intensive targets with G≥19G \ge19; and panel (d) reports mean shot gain. Lines are computed from four depths and four paired seeds, and bands are hierarchical-bootstrap 95% confidence intervals. The Transformer has the lowest P90 error at all three measurement-intensive group counts while retaining a measurement budget comparable to the other learned predictors. The dashed line marks 1.6 mHa. All features, labels, and online-VQE settings are fixed while only the predictor architecture changes. A cell-level mask diagnostic for the same predictors appears in Appendix § 19, Figure 11; underlying molecule-OOD values and matched held-out-seed/Hamiltonian controls appear in Table 47, Table 46, and Table 48.

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs+0.0011+0.0015+0.0005−0.0006+0.0023
Depth 1ΔE\Delta E vs Full ↓\downarrowMLP+0.0008+0.0028+0.0015−0.0021+0.0025
Depth 1ΔE\Delta E vs Full ↓\downarrowResNet-18+0.0000−0.0010+0.0029+0.0024+0.0016
Depth 1ΔE\Delta E vs Full ↓\downarrowGNN+0.0012+0.0022+0.0004−0.0021+0.0013
Depth 1Shot gain ↑\uparrowOurs40.4%58.9%63.0%74.7%86.9%
Depth 1Shot gain ↑\uparrowMLP47.3%62.5%68.2%75.9%87.6%
Depth 1Shot gain ↑\uparrowResNet-180.0%7.5%38.1%46.1%77.9%
Depth 1Shot gain ↑\uparrowGNN42.7%57.7%60.3%66.6%82.5%
Depth 1No-worse vs FullOurs✓✓✓✓✗
Depth 1No-worse vs FullMLP✓✗✓✓✗
Depth 1No-worse vs FullResNet-18✓✓✗✗✓
Depth 1No-worse vs FullGNN✓✗✓✓✓
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs+0.0003−0.0011−0.0015+0.0158−0.0049
Depth 2ΔE\Delta E vs Full ↓\downarrowMLP+0.0127−0.0014−0.0029+0.0006−0.0037
Depth 2ΔE\Delta E vs Full ↓\downarrowResNet-18+0.0000+0.0011+0.0007+0.0221+0.0065
Depth 2ΔE\Delta E vs Full ↓\downarrowGNN+0.0010−0.0134+0.0031+0.0254−0.0026
Depth 2Shot gain ↑\uparrowOurs25.6%31.6%44.9%51.8%82.6%
Depth 2Shot gain ↑\uparrowMLP28.3%33.4%46.1%48.4%78.0%
Depth 2Shot gain ↑\uparrowResNet-180.0%2.2%24.2%35.2%73.3%
Depth 2Shot gain ↑\uparrowGNN21.6%28.3%41.7%45.9%77.6%
Depth 2No-worse vs FullOurs✓✓✓✗✓
Depth 2No-worse vs FullMLP✗✓✓✓✓
Depth 2No-worse vs FullResNet-18✓✓✓✗✗
Depth 2No-worse vs FullGNN✓✓✗✗✓
Depth 3ΔE\Delta E vs Full ↓\downarrowOurs+0.0005+0.0092+0.0005−0.0014−0.0017
Depth 3ΔE\Delta E vs Full ↓\downarrowMLP+0.0035−0.0045−0.0008−0.0000−0.0039
Depth 3ΔE\Delta E vs Full ↓\downarrowResNet-18+0.0000+0.0044−0.0004−0.0064+0.0053
Depth 3ΔE\Delta E vs Full ↓\downarrowGNN+0.0039−0.0017+0.0020−0.0025+0.0001
Depth 3Shot gain ↑\uparrowOurs18.1%24.9%29.9%35.2%75.4%
Depth 3Shot gain ↑\uparrowMLP20.8%23.7%28.9%32.7%71.5%
Depth 3Shot gain ↑\uparrowResNet-180.0%2.0%19.9%28.4%73.0%
Depth 3Shot gain ↑\uparrowGNN16.8%22.1%28.4%31.7%66.9%
Depth 3No-worse vs FullOurs✓✗✓✓✓
Depth 3No-worse vs FullMLP✗✓✓✓✓
Depth 3No-worse vs FullResNet-18✓✗✓✓✗
Depth 3No-worse vs FullGNN✗✓✗✓✓

Table 46. Predictor architecture comparison at τ=0.995\tau= 0.995. This table copies the Table 1 held-out-seed online VQE setting and changes only the learned relevance predictor. Our method uses the Transformer predictor throughout the paper; MLP, ResNet-18, and GNN use the same labels, features, threshold, molecules, depths, and online protocol. Boldface marks the closest final energy; shot-gain values are green/red according to the no-worse criterion, and bold-underlined values mark the largest shot gain per cell.

Table 47: Predictor comparison under leave-one-molecule-out transfer at τ=0.997\tau= 0.997. The predictor is the only changed component: all methods use identical features, labels, molecule splits, depths, seeds, and closed-loop online VQE. Representative depths 1 and 4 are shown. Green/red follows the one-sided 1.6 mHa no-worse criterion; bold marks the closest mean energy to Full, and bold-underlined shot gain marks the largest saving per cell.

MoleculeDepthMetricOursMLPResNet-18GNN
H2 (4q)Depth 1ΔE\Delta E vs Full ↓+0.0016+0.0030+0.0000+0.0018
Depth 1Shot gain ↑55.2%56.7%0.0%46.9%
Depth 1Chem. acc.✗✗✓✗
Depth 4ΔE\Delta E vs Full ↓-0.0002-0.0020+0.0000-0.0000
Depth 4Shot gain ↑9.9%13.7%0.0%12.2%
Depth 4Chem. acc.✓✓✓✓
HeH+ (4q)Depth 1ΔE\Delta E vs Full ↓+0.0021-0.0038+0.0139-0.0032
Depth 1Shot gain ↑32.2%36.2%27.6%33.6%
Depth 1Chem. acc.✗✓✗✓
Depth 4ΔE\Delta E vs Full ↓-0.0022-0.0025-0.0071-0.0018
Depth 4Shot gain ↑8.1%6.0%9.8%6.0%
Depth 4Chem. acc.✓✓✓✓
H3+ (6q)Depth 1ΔE\Delta E vs Full ↓+0.0051-0.0004+0.0078+0.0034
Depth 1Shot gain ↑56.2%55.2%45.3%52.3%
Depth 1Chem. acc.✗✓✗✗
Depth 4ΔE\Delta E vs Full ↓-0.0052-0.0057-0.0320+0.0029
Depth 4Shot gain ↑10.0%8.9%16.1%12.3%
Depth 4Chem. acc.✓✓✓✗
H4 chain (8q)Depth 1ΔE\Delta E vs Full ↓+0.0038+0.0052+0.0030+0.0072
Depth 1Shot gain ↑63.7%67.1%59.8%66.3%
Depth 1Chem. acc.✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓+0.0652+0.0195+0.0335-0.0043
Depth 4Shot gain ↑28.5%35.7%46.5%28.9%
Depth 4Chem. acc.✗✗✗✓
LiH (12q)Depth 1ΔE\Delta E vs Full ↓+0.0022+0.8939-0.0314+0.0101
Depth 1Shot gain ↑40.4%47.1%41.4%51.9%
Depth 1Chem. acc.✗✗✓✗
Depth 4ΔE\Delta E vs Full ↓-0.0023+0.0037+0.1054+0.0073
Depth 4Shot gain ↑17.9%19.4%13.5%19.5%
Depth 4Chem. acc.✓✗✗✗

Table 47.

Testing HamiltonianDepthMetricOursMLPResNet-18GNN
TFIM h=0.5h = 0.5Depth 1ΔE\Delta E vs Full ↓\downarrow-0.0000-0.0000+0.0000+0.0000
Depth 1Shot gain ↑\uparrow2.0%5.7%0.0%2.1%
Depth 1Chem. acc.✓✓✓✓
Depth 4ΔE\Delta E vs Full ↓\downarrow-0.0002+0.0000+0.0000+0.0000
Depth 4Shot gain ↑\uparrow2.6%4.5%0.0%1.3%
Depth 4Chem. acc.✓✓✓✓
TFIM h=0.7h = 0.7Depth 1ΔE\Delta E vs Full ↓\downarrow-0.0000-0.0000+0.0000+0.0000
Depth 1Shot gain ↑\uparrow2.9%4.9%0.0%1.7%
Depth 1Chem. acc.✓✓✓✓
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000-0.0000+0.0000-0.0014
Depth 4Shot gain ↑\uparrow2.4%4.3%0.0%1.2%
Depth 4Chem. acc.✓✓✓✓
TFIM h=0.9h = 0.9Depth 1ΔE\Delta E vs Full ↓\downarrow+0.0771+0.0771+0.0000+0.0496
Depth 1Shot gain ↑\uparrow25.0%25.0%0.0%14.1%
Depth 1Chem. acc.✗✗✓✗
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0302+0.0302+0.0000+0.0255
Depth 4Shot gain ↑\uparrow6.2%6.2%0.0%3.8%
Depth 4Chem. acc.✗✗✓✗
Heis. XXZ+X fieldDepth 1ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0000+0.0000+0.0000
Depth 1Shot gain ↑\uparrow16.7%16.7%0.0%16.4%
Depth 1Chem. acc.✓✓✓✓
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0000+0.0000+0.0000
Depth 4Shot gain ↑\uparrow4.2%4.2%0.0%4.2%
Depth 4Chem. acc.✓✓✓✓
Heis. J1–J2Depth 1ΔE\Delta E vs Full ↓\downarrow-0.0000+0.0000+0.0000+0.0001
Depth 1Shot gain ↑\uparrow16.7%16.8%0.0%15.9%
Depth 1Chem. acc.✓✓✓✓
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0000+0.0000+0.0001
Depth 4Shot gain ↑\uparrow4.2%4.2%0.0%4.2%
Depth 4Chem. acc.✓✓✓✓
Heis. XXY+Z fieldDepth 1ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0001+0.0000+0.0000
Depth 1Shot gain ↑\uparrow16.7%16.7%0.0%15.8%
Depth 1Chem. acc.✓✓✓✓
Depth 4ΔE\Delta E vs Full ↓\downarrow+0.0000+0.0000+0.0000+0.0000
Depth 4Shot gain ↑\uparrow4.2%4.2%0.0%4.2%
Depth 4Chem. acc.✓✓✓✓
JW-like parity spinDepth 1ΔE\Delta E vs Full ↓\downarrow+0.0312+0.0290+0.0452+0.0337
Depth 1Shot gain ↑\uparrow35.7%23.0%65.8%31.9%
Depth 1Chem. acc.✗✗✗✗
Depth 4ΔE\Delta E vs Full ↓\downarrow-0.0046-0.0014+0.1674+0.0049
Depth 4Shot gain ↑\uparrow11.6%8.8%28.2%15.1%
Depth 4Chem. acc.✓✓✗✗

Table 48. Predictor comparison under Hamiltonian-family transfer at τ=0.997\tau= 0.997. One shared predictor per architecture is trained on the same five molecular and two Heisenberg Hamiltonians, then evaluated on seven held-out Hamiltonians with identical features, depths, seeds, and online VQE. Representative depths 1 and 4 are shown. A zero-gap ResNet-18 cell with zero shot gain denotes Full-QWC fallback rather than successful sparsification.

Figure 11. Oracle-relative parameter–QWC masks for four relevance predictors. All predictors are evaluated on the same held-out H3+H_{3}^{+} molecule-OOD snapshot (depth 2, seed 97, optimization iteration 14) at τ=0.997\tau= 0.997. Rows index VQE parameters and columns index QWC groups sorted by oracle relevance. Dark-blue cells are selected by both the predictor and oracle; gray cells are omitted by both; light-blue cells are predictor-only extra measurements; and red cells are oracle-selected groups missed by the predictor. “Captured” is the mean exact relevance mass covered by a predicted row mask, “Selected” is its mean selected-group fraction, and “Density” is captured mass divided by selected fraction. The snapshot is a qualitative structural diagnostic; aggregate performance is reported in main-text Figure 10 and Table 47 and Table 48.

Held-Out-Seed Predictor Comparison

Leave-One-Molecule-Out Predictor Comparison

Hamiltonian-Family Predictor Comparison

Frozen-LLM and Frontier-Model Details

Controlled intervention. The local-LLM experiment does not replace our full controller and does not fine-tune an LLM. Instead, it isolates one decision: ranking QWC groups for each parameter. At a refresh iteration tt, our saved controller first predicts a binary mask and therefore a per-parameter cardinality

kj,t=∑g=1GMj,gOurs(t).(11)k_{j,t} = \sum_{g=1}^{G} M^{\mathrm{Ours}}_{j,g}(t). \tag*{(11)}

The LLM receives kj,tk_{j,t} and the same structured parameter, group, and parameter–group features, but it never receives the oracle contribution matrix. It returns a ranking rj,tLLMr^{\mathrm{LLM}}_{j,t}, from which we form

Mj,gLLM(t)=I[g∈Top⁡kj,t(rj,tLLM)].(12)M^{\mathrm{LLM}}_{j,g}(t) = \mathbb{I}\left[g \in\operatorname{Top}_{k_{j,t}}\left(r^{\mathrm{LLM}}_{j,t}\right)\right]. \tag*{(12)}

Thus the comparison holds the count controller fixed at each query and tests whether a frozen language model can replace the learned group-ranking function.

Exact system prompt for the repaired rank-only protocol. All four models receive the following system message without demonstrations:

You are a measurement-selection controller for variational
quantum eigensolvers.
For each requested circuit parameter, choose exactly k
distinct QWC group IDs and order them from
most to least relevant to that parameter’s energy gradient.
Use coefficient, circuit-location,
theta, iteration, and parameter–group light-cone features
jointly. A globally large coefficient
does not imply equal relevance to every parameter. Every
returned ID must be an integer present in
the candidate groups. Return only the JSON object required
by the supplied schema.

Compact request and response. The request contains only feature columns that are nonzero in the current instance. The ellipses below suppress repeated rows, not additional natural-language instructions.

{
  "task": "rank_parameter_specific_qwc_groups",
  "context": {
    "ansatz": "ryrz", "depth": 2,
    "parameters": 48, "qwc_groups": 19,
    "iteration": 7, "total_iterations": 28
  },
  "columns": {
    "parameter_x": [...],
    "group_rows": ["group_id", ...],
    "pair_nonzero_rows": ["group_id", ...]
  },
  "groups": [[0, ...], [1, ...], ...],
  "parameters_to_rank": [
    {"id": 3, "k": 5, "x": [...],
     "pair_nonzero": [[0, ...], ...]}
  ]
}

The response is constrained by Ollama’s JSON schema and contains one row for every requested parameter:

{"selections": [
{"parameter_id": 3,
 "ranked_group_ids": [7, 2, 5, 1, 8]}

Validation and repair in the rank-only protocol. Later clean/explanation output contracts are different; Appendix § 24.§ 4 documents the clean and mechanism-audit protocols and their prompt-specific trace records. Boolean values, repeated IDs, noninteger values, and group IDs outside [0,G−1][0, G - 1] are discarded. If fewer than kj,tk_{j,t} valid distinct IDs remain, the missing positions are filled deterministically using descending global group-coefficient magnitude. We log the requested and repaired group counts, invalid batches, context splits, prompt/output tokens, and inference latency. This repair prevents malformed JSON from changing the requested cardinality while preserving a measurable repair-rate diagnostic.

Sparse-refresh execution. The LLM is queried only at t∈{0,7,14,21}t \in\{0, 7, 14, 21\} during each 28-step online trajectory. The resulting mask is reused for the following seven optimizer steps. Because the count targets predicted by our method can evolve between refreshes while the LLM mask is held fixed, an LLM’s realized cumulative shot gain need not equal the continuously updated reference gain from our method. We therefore treat the reference-gain rows in Table 3 as count-controller context rather than as measured LLM cost.

ComponentSetting
ModelsQwen3-8B, Llama-3.1-8B, DeepSeek-R1-8B, GPT-OSS-20B
AdaptationZero-shot; no demonstrations; no fine-tuning
DecodingTemperature 0, seed 123, think=false
Context/outputnum_ctx = 32768, num_predict = 1536
Parameter batchingUp to 16 parameters; adaptive halving on context overflow
Refresh scheduleIterations {0,7,14,21}\{0, 7, 14, 21\} of 28
VQE setting1024 shots per shift, Adam, layered RY/RZ ansatz
Evaluation panelFive STO-3G molecules, depths 1–3, seeds 97/101/107/109
Count controllerSaved Ours model at τ=0.995\tau= 0.995

Table 49. Frozen local-LLM inference settings used in the sparse-refresh comparison.

Controlled rank-only intervention. We retain our method only as the cardinality controller: at each query it supplies kjk_j, while a frozen local LLM must infer which QWC groups to select from structured parameter, group, and light-cone features. The oracle contribution matrix is never included in the prompt.

No model adaptation. Qwen3-8B, Llama-3.1-8B, DeepSeek-R1-8B, and GPT-OSS-20B are evaluated zero-shot with deterministic decoding and no fine-tuning. Schema validation and coefficient-based repair prevent malformed output from changing the requested cardinality.

Sparse refresh. The LLM is queried at t∈{0,7,14,21}t \in\{0, 7, 14, 21\} and each returned mask is reused for seven online-VQE steps. The complete prompt and settings are in Appendix J.

Clean Frozen-LLM Audit

Frontier Models at Deeper Circuits

MethodNon-cached inputCached inputOutputTotal60-cell totalPer cell
GPT-5.412.713.640.7817.1344.430.74
GPT-5.512.584.420.4517.4578.671.31
GPT-5.6 Sol16.501.460.5718.52100.351.67
GPT-5.6 Terra13.463.620.6117.7043.760.73
GPT-5.6 Luna14.741.740.52<u>17.00</u><u>18.01</u><u>0.30</u>
Claude Haiku 4.524.581.086.7532.4071.291.19
Claude Sonnet 525.104.811.4331.35183.343.06
Claude Opus 524.902.750.5528.20159.012.65
Claude Fable 524.241.331.1526.72327.155.45
Ours0.000.000.000.000.000.00

Table 53. Online controller token traffic and API-equivalent cost. The accounting covers the same 60 held-out VQE cells used in Table 4–Table 52: 28 optimizer steps per cell and four selector refreshes (240 total). Token categories come from run logs. OpenAI dollar entries are historical API-equivalent estimates retained from the original version; Claude entries are CLI-reported API totals. Dollar values describe historical API-equivalent accounting; token traffic is directly recorded in the experiment logs (Appendix § 24.§ 1). Our method uses local checkpoint inference and no LLM tokens or API credits. Lower is better. Bold marks minima; underline marks the lowest nonzero entry in the total-token and cost columns.

Accounting note. OpenAI input totals include cached tokens, so the non-cached column subtracts the logged cached subset before pricing; Claude cache-read tokens are logged separately. The OpenAI reconstruction uses OpenAI API pricing, while Claude uses each CLI response’s provider-reported total_cost_usd (see Anthropic pricing). These are normalized API-equivalent costs, not necessarily the subscription invoice. The one-time training cost of our method is excluded from this online-inference accounting; its 984,454-parameter checkpoint is 3.78 MiB and averages 7.29 ms per refresh on the local CPU.

Current takeaway

  • Stronger frontier LLMs better integrate long structured contexts and exploit retrieved task supervision.

  • No model consistently wins across molecules, depths, and fidelity–measurement trade-offs: current in-context policies are capable but not yet stable.

  • The progression from clean prompts to physics context and data exposure shows substantial potential to identify task-relevant physical structure.

Future implication: Physics-aware agent

  • Scale alone is insufficient; a physics-aware agent should combine retrieval, executable physics tools, memory, and verification.

  • A promising hybrid uses our method as an efficient prior and invokes LLM reasoning only for OOD or low-confidence states.

  • Crossed clean/data-exposed evaluations remain necessary before attributing the gain to frontier-model scale alone.

Observable Decision Traces

The main text retains the representative Llama trace. The complementary model traces below use the same fixed-state audit and expose only observable outputs, schema checks, selected-group evidence, and counterfactual feature tests.

Qwen3-8B observable decision trace

Figure 13. Qwen3-8B observable decision trace. On the fixed H3+H^+_3, depth-2, seed-97 snapshot, Qwen selects {g0,…,g5}\{g_0,\ldots,g_5\} and captures 94.963% of the exact p14p_{14} contribution mass. Removing coefficient or light-cone features reduces that mass to 0.636% and 3.714%, respectively, whereas removing θ\theta/time features leaves the reference mask unchanged. Box (4) uses the explanation prompt; Box (7) uses the separate reference prompt. Thus deletion measures reference-prompt selection quality. Raw reason tags are lightcone, theta, and circuit_location, as shown in Box (3). Box (8) reports reference-prompt utility. Exact contributions are audit-only (Appendix § 24.§ 4).

How to read the trace. Boxes (1)–(5) show the complete model-visible interaction: a fixed VQE snapshot with 24 circuit parameters and 19 QWC groups is serialized into numerical descriptors, the frozen LLM returns reason tags and six group IDs, and a deterministic validator checks only schema, range, and cardinality. Boxes (6)–(8) are post-hoc scientific audits. They attach the exact contribution matrix only after the decision and delete feature families under the reference prompt. Prompt identity distinguishes the explanation action from its companion feature audit.

Observed decision. Qwen’s executable mask {g0,…,g5}\{g_0,\ldots,g_5\} captures 94.963% of the local contribution mass without repair or fallback. The trace therefore separates two evaluation levels: executable output is assessed by the validator, and physical relevance is assessed by exact captured mass.

Reference-prompt feature utility

  • Removing coefficients lowers captured mass from 94.963% to 0.636%, a 94.327 percentage-point loss.

  • Removing light-cone features lowers it to 3.714%, a 91.248 percentage-point loss.

State sensitivity at this snapshot

  • Removing θ\theta/time leaves the selected p14p_{14} mask and its 94.963% local mass unchanged; mean mass across valid parameters changes from 0.854 to 0.837.

  • The unchanged local mask coexists with strong coefficient and light-cone effects. The explanation action and reference-prompt sensitivities provide complementary observations.

Scientific interpretation. Qwen’s reference selection responds strongly to physically meaningful coefficient and light-cone inputs, and its selected groups retain most of the exact row contribution. The aggregate online experiments extend this local view across molecules and optimization trajectories.

DeepSeek-R1-8B observable decision trace

Figure 14. DeepSeek-R1-8B observable decision trace. DeepSeek selects {g0,g2,g3,g4,g6,g7}\{g_0,g_2,g_3,g_4,g_6,g_7\} and captures 94.411% of the exact p14p_{14} mass; its full-feature reference mask captures 94.269%. Removing coefficient or light-cone features reduces local mass to 1.313% and 1.797% under the separate reference prompt. Removing θ\theta/time also reduces mass to 1.313%, identifying current-state sensitivity in the reference selection. Appendix N.4 records the explanation action and each reference-prompt intervention.

How to read the trace. The action panel contains only the IDs and observable reason tags emitted by the frozen model. Exact Cj,kC_{j,k} values in the evidence panel are never prompt features; they are attached afterward to score the action. The counterfactual panel tests reference-prompt feature utility alongside the separately logged explanation-producing action.

Observed decision. DeepSeek chooses {g0,g2,g3,g4,g6,g7}\{g_0,g_2,g_3,g_4,g_6,g_7\} and captures 94.411% of local mass, close to the 94.269% full-feature audit reference. Thus the model produces both a valid action and a high-quality local ranking on this snapshot.

Large reference-prompt quality effects

  • Removing coefficients reduces local mass to 1.313% (a 92.956-point loss); removing light-cone features reduces it to 1.797% (a 92.473-point loss).

  • Both input families support the high-quality local reference selection by retaining the dominant group g0g_0.

Broader feature sensitivity

  • Removing θ\theta/time under the reference prompt also reduces local mass to 1.313%. All three deletions omit the dominant group g0g_0.

  • The reference ranking responds to current-state information as well as coefficients and light cones; the reported tags summarize the separate explanation action.

Scientific interpretation. DeepSeek produces a high-quality action on this cell, while the companion audit reveals sensitivity to several physical feature families. Together, the action score and intervention profile characterize the model’s local selection behavior under the two documented prompt settings.

GPT-OSS-20B observable decision trace

Figure 15. GPT-OSS-20B observable decision trace. GPT-OSS reports coefficient evidence and selects {g18,g0,g17,g16,g15,g14}\{g_{18}, g_{0}, g_{17}, g_{16}, g_{15}, g_{14}\}, capturing 93.042% of the exact p14p_{14} mass; its reference mask captures 95.977%. Removing coefficient, light-cone, or θ\theta/time features reduces local mass to 5.090%, 3.192%, and 5.090%, respectively, under the separate reference prompt. This ranking shows multi-feature quality sensitivity alongside the explanation action’s coefficient tag (Appendix § 24.§ 4).

How to read the trace. This example pairs a high-mass action with reference-prompt interventions. The frozen LLM generates selection JSON from serialized physical features; the external Python harness performs validation and all VQE simulation. The lower audit scores reference-prompt masks using exact contributions and feature-family deletion, with LLM tool execution disabled.

Observed decision. GPT-OSS selects {g18,g0,g17,g16,g15,g14}\{g_{18},g_{0},g_{17},g_{16},g_{15},g_{14}\} and captures 93.042% of local mass. This is a strong mask, alongside the 95.977% full-feature reference. Both actions concentrate measurements on groups carrying most of the exact parameter-row contribution.

Reference-prompt feature utility

  • Removing coefficients lowers local captured mass to 5.090%, a 90.888 percentage-point loss.

  • Coefficients improve the reference-prompt selection. The separately prompted explanation action also reports a coefficient tag.

Multiple sensitive input families

  • Removing light-cone or θ\theta/time features lowers mass to 3.192% and 5.090%, respectively.

  • The reference ranking draws useful information from several physical families, complementing the coefficient tag returned under the explanation prompt.

Scientific interpretation. The deletion pattern is consistent with joint integration of several serialized physical inputs. The external harness evaluates this behavior through captured mass and feature sensitivity, with model-generated output restricted to selection JSON. The successful action and its companion audit provide two concrete views of GPT-OSS’s physical selection capability.

Prompt and Context Cost

Key finding. The saved estimates emphasize circuit/support, coefficient, and light-cone spans on this snapshot. The first decision has a 1.61% relative completeness residual; later continuation scores have larger relative residuals. The displayed profile describes this fixed-prefix attribution calculation.

Token- and Feature-Level Attribution

Figure 17. Token-to-QWC-group attribution for Llama-3.1-8B. The state and panel definitions match Figure 16, while the accepted reference-prompt action is [g4,g5,g6,g7,g8,g9][g_4,g_5,g_6,g_7,g_8,g_9]. The observable action comes from Ollama llama3.1:8b; gradients use a public NF4 conversion of the frozen Meta-Llama-3.1-8B-Instruct checkpoint (Grattafiori et al., [11]), which is not bitwise identical to the GGUF generation runtime. The 32-point maximum completeness residual is 1.149 log-probability units. Appendix § 24.§ 5 reports absolute and relative numerical residuals alongside the estimates.

Key finding. The saved Llama map reports a 0.73% trajectory share with the numerical residuals tabulated in Appendix § 24.§ 5. Its first ID g4g_4 has only 1.06% conditional probability (rank 15), whereas g5g_5–g9g_9 become rank-one continuations, exposing a strong autoregressive list-completion pattern in the white-box runtime.

Figure 18. Token-to-QWC-group attribution for DeepSeek-R1-Distill-Llama-8B. The accepted reference-prompt action is [g2,g5,g9,g6,g0,g7][g_2, g_5, g_9, g_6, g_0, g_7] under the same H3+\mathrm{H}_3^+ snapshot and attribution protocol. The action comes from Ollama deepseek-r1:8b; white-box gradients use the official frozen DeepSeek-R1-Distill-Llama-8B checkpoint ([14]) with runtime NF4 quantization. This surrogate is not bitwise identical to the GGUF runtime. The 32-point maximum completeness residual is 0.437 log-probability units; absolute and relative numerical residuals are reported in Appendix N.5.

Key finding. The saved DeepSeek map emphasizes static structure, coefficients, and light-cone spans. The accepted IDs receive only 0.14–30.85% conditional probability and ranks 2–15 in the white-box runtime. The panel records input sensitivity of these fixed-prefix scores, with runtime settings and integration residuals specified by the accompanying audit.

Key finding. GPT-OSS’s saved light-cone estimate is 21.36%. Four emitted IDs are high-confidence while g6g_6 and g3g_3 remain below 0.4%, indicating a heterogeneous fixed-prefix score profile in the white-box runtime.

Key finding. Our method combines strong circuit/support dependence (63.24%) with explicit use of trajectory features (12.16%). Its top-six ranking contains all five groups in the exact τ=0.995\tau= 0.995 top-mass mask. The native-checkpoint attribution connects current optimization state to the scores used for parameter-specific measurement selection.

Best-Prompt Online VQE and Explanation Faithfulness

Table 56: What drives frozen-LLM QWC selection? The audit evaluates four frozen local LLMs on the same 12 Full-QWC snapshots (H3+, H4 chain, and LiH; depths 2 and 3; seed 97; iterations 0 and 14), with no fine-tuning and identical per-parameter cardinalities. The reference uses a domain prompt, compressed output, 32K context, and all non-oracle coefficient, light-cone, parameter, and trajectory features. Δ\Delta captured mass is the paired, model-macro-averaged change from this reference in percentage points; positive is better. Protocol success is the fraction of the 48 model–snapshot cases yielding a valid exact-cardinality mask. The apparent ranking-output gain is conditional on only 22 successful cases and should not be interpreted as a reliable improvement.

Complete Mechanism-Audit Summary

Table 56 preserves the aggregate intervention summary behind the main-text attribution and cost views.

Complete Best-Prompt Online-VQE Controls

Table 57 retains the Random baseline and the refresh-matched and per-step our method controls omitted from the compact clean-versus-best comparison in the main text.

SettingSelection methodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1GPT clean gain, K=7K=7 ↑39.4 ± 1.8%62.6 ± 1.3%67.0 ± 1.5%79.6 ± 1.8%89.9 ± 0.8%
Depth 1GPT best gain, K=7K=7 ↑42.5 ± 4.6%61.7 ± 3.4%66.7 ± 1.5%79.3 ± 1.8%90.5 ± 0.4%
Depth 1Ours gain, K=7K=7 ↑41.2 ± 5.3%60.7 ± 3.1%65.1 ± 2.6%77.3 ± 0.7%88.2 ± 0.4%
Depth 1Ours gain, K=1K=1 ↑40.4 ± 6.3%58.9 ± 3.5%63.0 ± 2.8%74.7 ± 1.1%86.9 ± 0.5%
Depth 1Random (K=7K=7)+0.1847 ± 0.0868+1.2191 ± 0.2759+0.7247 ± 0.2856+0.9736 ± 0.3101+2.9030 ± 0.6058
Depth 1GPT-OSS (clean)+0.3930 ± 0.1936 ✗+0.8238 ± 0.2303 ✓+0.7300 ± 0.4383 ✗+0.5671 ± 0.3164 ✓+1.5503 ± 0.7644 ✓
Depth 1GPT-OSS (best)+0.3392 ± 0.1856 ✗+0.9079 ± 0.4907 ✓+0.3882 ± 0.2525 ✓+0.6337 ± 0.2763 ✓+1.9964 ± 0.9332 ✓
Depth 1Ours (K=7K=7)+0.0013 ± 0.0003 ✓+0.0035 ± 0.0015 ✓+0.0004 ± 0.0006 ✓+0.0001 ± 0.0060 ✓+0.0030 ± 0.0016 ✓
Depth 1Ours (K=1K=1)+0.0011 ± 0.0008 ✓+0.0015 ± 0.0012 ✓+0.0005 ± 0.0006 ✓−0.0006 ± 0.0024 ✓+0.0023 ± 0.0016 ✓
Depth 2GPT clean gain, K=7K=7 ↑29.5 ± 3.5%34.3 ± 1.7%49.8 ± 1.9%55.5 ± 1.6%84.1 ± 0.5%
Depth 2GPT best gain, K=7K=7 ↑29.7 ± 2.2%33.0 ± 2.3%49.8 ± 0.8%55.4 ± 0.4%84.3 ± 0.5%
Depth 2Ours gain, K=7K=7 ↑27.3 ± 2.5%32.5 ± 2.4%47.9 ± 1.4%55.1 ± 0.5%83.7 ± 0.4%
Depth 2Ours gain, K=1K=1 ↑25.6 ± 2.9%31.6 ± 2.6%44.9 ± 1.3%51.8 ± 0.5%82.6 ± 0.4%
Depth 2Random (K=7K=7)+0.0930 ± 0.0297+0.5668 ± 0.4463+0.3054 ± 0.1633+0.9327 ± 0.3273+2.6969 ± 0.2816
Depth 2GPT-OSS (clean)+0.2148 ± 0.1220 ✗+0.6807 ± 0.7081 ✗+0.4691 ± 0.2024 ✗+0.5616 ± 0.2675 ✓+1.2894 ± 0.3290 ✓
Depth 2GPT-OSS (best)−0.0510 ± 0.0402 ✓+0.1302 ± 0.0822 ✗+0.4776 ± 0.3130 ✗+0.4682 ± 0.1297 ✓+1.5294 ± 0.3237 ✓
Depth 2Ours (K=7K=7)−0.0007 ± 0.0015 ✓+0.0012 ± 0.0176 ✓+0.0006 ± 0.0016 ✓+0.0162 ± 0.0438 ✓−0.0040 ± 0.0118 ✓
Depth 2Ours (K=1K=1)+0.0003 ± 0.0007 ✓−0.0011 ± 0.0026 ✓−0.0015 ± 0.0041 ✓+0.0158 ± 0.0434 ✓−0.0049 ± 0.0118 ✓
Depth 3GPT clean gain, K=7K=7 ↑19.8 ± 1.9%25.1 ± 1.8%32.1 ± 1.4%39.0 ± 0.5%76.8 ± 1.5%
Depth 3GPT best gain, K=7K=7 ↑19.5 ± 1.6%26.3 ± 1.2%32.1 ± 1.0%38.7 ± 1.1%76.6 ± 1.3%
Depth 3Ours gain, K=7K=7 ↑19.0 ± 2.4%26.1 ± 1.2%32.1 ± 1.2%38.6 ± 0.7%76.6 ± 1.3%
Depth 3Ours gain, K=1K=1 ↑18.1 ± 2.7%24.9 ± 1.3%29.9 ± 1.0%35.2 ± 0.9%75.4 ± 1.8%
Depth 3Random (K=7K=7)+0.0730 ± 0.0859+0.2617 ± 0.2130+0.2750 ± 0.0765+0.2773 ± 0.2665+2.3928 ± 0.4840
Depth 3GPT-OSS (clean)+0.1633 ± 0.1815 ✗+0.3438 ± 0.1193 ✗+0.1786 ± 0.0759 ✓+0.5089 ± 0.3596 ✗+0.5647 ± 0.4252 ✓
Depth 3GPT-OSS (best)+0.0478 ± 0.0637 ✓+0.0162 ± 0.1316 ✓+0.1222 ± 0.1216 ✓−0.0449 ± 0.2110 ✓+0.5331 ± 0.3351 ✓
Depth 3Ours (K=7K=7)+0.0010 ± 0.0027 ✓+0.0108 ± 0.0118 ✓+0.0029 ± 0.0095 ✓−0.0029 ± 0.0113 ✓−0.0046 ± 0.0040 ✓
Depth 3Ours (K=1K=1)+0.0005 ± 0.0015 ✓+0.0092 ± 0.0131 ✓+0.0005 ± 0.0077 ✓−0.0014 ± 0.0097 ✓−0.0017 ± 0.0036 ✓

Table 57. Complete best-prompt closed-loop online-VQE comparison. This appendix table preserves the full matched-control version of Table 55. GPT-OSS (clean) is the clean exact select/exclude baseline, while GPT-OSS (best) uses the audit-selected domain prompt, direct-ID output, 32K context, and all physical features. Both models remain frozen, receive only the count kjk_j requested by the proposed method, and refresh group identities every K=7K=7 steps. Random uses the same cardinality and refresh cadence; our method is reported at both K=7K=7 and per-step K=1K=1. Gain rows report realized mean ± standard deviation over seeds 97/101/107/109; all safety fallbacks pay Full-QWC shots. Energy entries are mean ± standard deviation relative to Full-QWC. Green/red satisfies/violates the one-sided 1.6 mHa no-worse criterion; the black check/cross compares absolute mean gap with Random. Bold/underline mark the smallest/second-smallest absolute mean gap among the five methods in each molecule–depth cell.

Calibration-Derived Noise Diagnostics

Across 120 surrogate snapshots, clean and noisy contribution grids have 50.6% and 51.5% mean sparsity, respectively. Their masks have 0.970 mean Jaccard overlap, their relevance rankings have 0.996 mean Spearman correlation, and a clean top-mass mask retains 99.7% of noisy contribution mass. The smaller gate-level Aer check is deliberately more conservative: finite-shot transpiled circuits reduce mask overlap and rank correlation, but 15–26% of parameter–group cells remain unselected at τ=0.995\tau= 0.995. Neither experiment is a physical-QPU run.

Table 58: Noisy online VQE comparison at τ=0.995\tau= 0.995. This repeats the Table 1 protocol under a finite-shot Pauli-channel surrogate derived from the fixed IBM FakeGuadalupeV2 calibration snapshot [23, 19 b]. Final energies are evaluated noiselessly and paired against noisy Full-QWC training; † rows retain the noiseless cell-wise TITAN thresholds. The same noise study additionally tests contribution-mask stability (Figure 23 and Table 59), a gate-level Aer implementation (Table 60), and Hamiltonian-family transfer (Table 61).

SettingMetricMethodH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1ΔE\Delta E vs Full ↓\downarrowOurs−0.0076−0.0219−0.0252+0.0001−0.0304
Depth 1ΔE\Delta E vs Full ↓\downarrowRandom+0.2170+1.2823+0.9091+1.3470+3.3622
Depth 1ΔE\Delta E vs Full ↓\downarrowGlobal coeff.−0.0050+0.0155−0.0022−0.0020+0.0041
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN (Peng et al.)+0.0150+0.0723+0.0413+0.0116−0.1104
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN+Ours+0.0138+0.0746+0.0422+0.0138−0.1065
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN†+0.0053+0.0985+0.0293−0.0851−0.0705
Depth 1ΔE\Delta E vs Full ↓\downarrowTITAN†+Ours+0.0011+0.0988+0.0291−0.0864−0.0693
Depth 1Shot gain ↑\uparrowOurs43.2%59.9%64.0%75.9%88.1%
Depth 1Shot gain ↑\uparrowRandom42.6%60.5%67.9%79.1%90.3%
Depth 1Shot gain ↑\uparrowGlobal coeff.41.5%61.1%64.8%76.1%88.3%
Depth 1Shot gain ↑\uparrowTITAN (Peng et al.)65.0%56.2%58.6%63.1%73.7%
Depth 1Shot gain ↑\uparrowTITAN+Ours81.7%90.5%93.6%96.1%99.7%
Depth 1Shot gain ↑\uparrowTITAN†43.6%59.4%64.3%71.6%82.3%
Depth 1Shot gain ↑\uparrowTITAN†+Ours65.9%91.7%95.5%98.0%99.9%
Depth 1No-worse vs FullOurs✓✓✓✓✓
Depth 1No-worse vs FullRandom✗✗✗✗✗
Depth 1No-worse vs FullGlobal coeff.✓✗✓✓✗
Depth 1No-worse vs FullTITAN (Peng et al.)✗✗✗✗✓
Depth 1No-worse vs FullTITAN+Ours✗✗✗✗✓
Depth 1No-worse vs FullTITAN†✗✗✗✓✓
Depth 1No-worse vs FullTITAN†+Ours✓✗✗✓✓
Depth 2ΔE\Delta E vs Full ↓\downarrowOurs−0.0632−0.1081−0.0765−0.0583−0.1241
Depth 2ΔE\Delta E vs Full ↓\downarrowRandom+0.2196+0.5977+0.5653+0.8670+3.6871
Depth 2ΔE\Delta E vs Full ↓\downarrowGlobal coeff.−0.0965−0.1624−0.0270−0.0491+0.0889
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN (Peng et al.)−0.0650−0.2074−0.0688+0.0172−0.1387
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN+Ours−0.0603−0.2072−0.0736+0.0162−0.1369
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN†−0.0819−0.0621−0.0283−0.0596+0.0998
Depth 2ΔE\Delta E vs Full ↓\downarrowTITAN†+Ours+0.0120−0.0710−0.0400−0.0239+0.1058
Depth 2Shot gain ↑\uparrowOurs28.8%32.8%46.2%52.6%82.6%
Depth 2Shot gain ↑\uparrowRandom29.9%34.7%47.5%53.4%83.9%
Depth 2Shot gain ↑\uparrowGlobal coeff.27.9%33.2%46.7%52.0%83.0%
Depth 2Shot gain ↑\uparrowTITAN (Peng et al.)53.2%54.3%51.2%61.7%78.7%
Depth 2Shot gain ↑\uparrowTITAN+Ours70.5%75.5%80.9%87.6%98.8%
Depth 2Shot gain ↑\uparrowTITAN†25.3%40.2%42.0%48.2%84.3%
Depth 2Shot gain ↑\uparrowTITAN†+Ours43.1%64.2%74.8%80.9%99.4%
Depth 2No-worse vs FullOurs✓✓✓✓✓
Depth 2No-worse vs FullRandom✗✗✗✗✗
Depth 2No-worse vs FullGlobal coeff.✓✓✓✓✗
Depth 2No-worse vs FullTITAN (Peng et al.)✓✓✓✗✓
Depth 2No-worse vs FullTITAN+Ours✓✓✓✗✓
Depth 2No-worse vs FullTITAN†✓✓✓✓✗
Depth 2No-worse vs FullTITAN†+Ours✗✓✓✓✗

Table 58.

Figure 23. The same contribution matrices under IBM calibration-derived noise. Molecules, circuits, initialization, parameter shifts, block aggregation, and color normalization are identical to Figure 1. We extract median gate and readout errors, gate durations, and coherence metadata from the fixed Qiskit FakeGuadalupeV2 system snapshot [23, 19]. Following standard backend-derived noise modeling [20], our scalable Pauli-channel surrogate combines gate depolarization, T2T_{2}-limited decay, and readout attenuation. This is a calibrated noise simulation, not execution on a physical IBM QPU.

SystemDepthClean sparsityNoisy sparsityMask JaccardRank ρ\rhoCross-mass
H2 (4q)147.3%47.5%0.9960.98699.9%
H2 (4q)230.3%30.7%0.9940.99899.9%
HeH+ (4q)149.9%50.7%0.9810.99699.7%
HeH+ (4q)225.5%26.0%0.9870.99899.7%
H3+ (6q)154.2%54.6%0.9860.99699.7%
H3+ (6q)235.4%36.1%0.9810.99999.7%
H4 chain (8q)168.4%69.2%0.9670.99699.6%
H4 chain (8q)243.1%44.5%0.9690.99999.6%
LiH (12q)182.3%83.3%0.9380.99599.7%
LiH (12q)269.5%72.4%0.8990.99899.7%

Table 59. Relevance sparsity under calibration-derived noise. Metrics average four seeds at iterations 0, 14, and 27. Cross-mass is the noisy contribution mass retained by the clean top-mass mask.

SystemParams.QWC groupsCircuitsClean selectedNoisy selectedNoisy sparsityJaccard / Rank ρ\rho
H2 (4q)8580213415.0%0.571 / 0.635
HeH+ (4q)811176396526.1%0.529 / 0.441
H3+ (6q)12194568817224.6%0.477 / 0.335

Table 60. Gate-level Aer spot-check under the FakeGuadalupeV2 snapshot ([19 b]). Each depth-1 contribution grid uses seed 97 and 4096 shots for every shifted QWC-group circuit. The selected-cell counts use τ=0.995\tau= 0.995. This is a calibrated Aer simulation, not a physical-QPU run.

Table 61: Calibration-derived noisy Hamiltonian-panel transfer at τ=0.997\tau= 0.997. This table repeats Table 39 under a fixed IBM FakeGuadalupeV2-derived Pauli-channel surrogate [23, 19 b]. ΔE\Delta E is paired against noisy Full-QWC at the same seed; final energies are evaluated noiselessly. TITAN† reuses the clean, cell-wise cost-matched thresholds without noise-specific tuning. Green/red ΔE\Delta E and ✓/✗ follow the one-sided 1.6 mHa no-worse criterion; bold marks the closest mean ΔE\Delta E to noisy Full, and bold-underlined shot gain marks the largest saving. This is calibrated simulation, not physical-QPU execution.

Testing HamiltonianDepthMetricOursRandomTITAN (Peng et al.)TITAN+OursTITAN†TITAN†+Ours
TFIM h=0.5h = 0.5Depth 1ΔE\Delta E vs noisy Full ↓-0.0030-0.0112-0.0146-0.0113-0.0002-0.0004
Depth 1Shot gain ↑2.5%2.3%55.3%55.6%0.0%2.1%
Depth 1Chem. acc.✓✓✓✓✓✓
Depth 4ΔE\Delta E vs noisy Full ↓+0.0203+0.0104+0.0033+0.0040-0.0032+0.0110
Depth 4Shot gain ↑2.1%1.9%55.2%57.1%0.0%1.9%
Depth 4Chem. acc.✗✗✗✗✓✗
TFIM h=0.7h = 0.7Depth 1ΔE\Delta E vs noisy Full ↓+0.0010-0.0061-0.0157-0.0133+0.0075+0.0036
Depth 1Shot gain ↑3.2%3.0%53.3%54.3%0.0%2.8%
Depth 1Chem. acc.✓✓✓✓✗✗
Depth 4ΔE\Delta E vs noisy Full ↓+0.0063+0.0123+0.0013+0.0028-0.0077-0.0021
Depth 4Shot gain ↑1.9%2.0%47.6%49.8%0.0%1.8%
Depth 4Chem. acc.✗✗✓✗✓✓
TFIM h=0.9h = 0.9Depth 1ΔE\Delta E vs noisy Full ↓+0.1398+0.0032-0.0214+0.0998+0.0048+0.1120
Depth 1Shot gain ↑25.0%25.0%53.4%58.8%24.8%38.5%
Depth 1Chem. acc.✗✗✓✗✗✗
Depth 4ΔE\Delta E vs noisy Full ↓+0.0909+0.0181-0.0138+0.0887-0.0154+0.0357
Depth 4Shot gain ↑6.2%6.2%42.9%46.2%4.5%9.8%
Depth 4Chem. acc.✗✗✓✗✓✗
Heis. XXZ+X fieldDepth 1ΔE\Delta E vs noisy Full ↓-0.1190+0.3294+0.6037+0.6111+0.0398-0.0175
Depth 1Shot gain ↑16.7%16.7%52.2%53.7%18.9%27.6%
Depth 1Chem. acc.✓✗✗✗✗✓
Depth 4ΔE\Delta E vs noisy Full ↓+0.3771+0.0446+0.2029+0.1974-0.1167+0.7431
Depth 4Shot gain ↑4.2%4.2%36.4%37.8%4.3%7.5%
Depth 4Chem. acc.✗✗✗✗✓✗
Heis. J1–J2Depth 1ΔE\Delta E vs noisy Full ↓-0.0907+0.4036+0.5352+0.5145-0.0779-0.2654
Depth 1Shot gain ↑16.7%16.7%50.5%52.2%11.3%22.8%
Depth 1Chem. acc.✓✗✗✗✓✓
Depth 4ΔE\Delta E vs noisy Full ↓+0.6422-0.0145+0.4860+0.2510+0.4632+0.6444
Depth 4Shot gain ↑4.2%4.2%34.2%36.1%4.1%7.3%
Depth 4Chem. acc.✗✓✗✗✗✗
Heis. XXY+Z fieldDepth 1ΔE\Delta E vs noisy Full ↓+0.0889+0.3424+0.3749+0.2889-0.1381-0.2234
Depth 1Shot gain ↑16.7%16.7%49.3%52.1%14.1%25.9%
Depth 1Chem. acc.✗✗✗✗✓✓
Depth 4ΔE\Delta E vs noisy Full ↓-0.0345-0.1825-0.2512-0.0173-0.0943+0.1709
Depth 4Shot gain ↑4.2%4.2%31.8%32.7%3.3%7.0%
Depth 4Chem. acc.✓✓✓✓✓✗
JW-like parity spinDepth 1ΔE\Delta E vs noisy Full ↓+0.0102+0.7538+0.1011+0.0976-0.0114+0.0106
Depth 1Shot gain ↑36.4%36.1%56.2%77.1%29.6%59.4%
Depth 1Chem. acc.✗✗✗✗✓✗
Depth 4ΔE\Delta E vs noisy Full ↓-0.0024+0.0546+0.0515+0.0092-0.0346-0.0106
Depth 4Shot gain ↑11.6%11.6%47.1%51.0%10.7%19.3%
Depth 4Chem. acc.✓✗✗✗✓✓

Table 61.

Long-Horizon Convergence and FCI-Anchored Validation

Table 62 reports the complete long-horizon benchmark referenced in § 6. The exact-gradient Adam trajectories define attainable, seed-specific references for the fixed depth-3 ansatz; FCI is used only after optimization to quantify the remaining variational gap.

Implementation and Estimator Details

This appendix distinguishes the implemented learning objective, deployment inputs, and estimator used by each experiment. Classical predictor weights are fixed throughout quantum-parameter optimization. Updating circuit angles θ\theta with Adam is not fine-tuning the classical predictor or LLM.

Architecture and Feature Construction

The main predictor has two independent Transformer encoders, one for parameter tokens and one for QWC-group tokens. Each encoder has two pre-normalized layers, width 128, four attention heads, GELU activation, feed-forward width 512, and dropout 0.14. There is no cross-attention block: interaction occurs in the pairwise scorer described in § 2. The scorer concatenates parameter and group embeddings, their elementwise product and absolute difference, and two pair features. Its LayerNorm–MLP has hidden widths 256 and 128, GELU nonlinearities, dropout, and a scalar output. A separate parameter-only LayerNorm–MLP with hidden width 128 predicts a sigmoid selected-group fraction. Variable-length examples are padded within a batch; padding masks exclude invalid groups and parameters from attention, normalization of row probabilities, and losses. The main saved state contains 984,454 trainable scalar parameters.

Actual inputs, not the entire historical feature interface. The saved interface has 35 parameter, 30 group, and two pair channels. Parameter descriptors encode normalized parameter, layer, and qubit indices, sinusoidal index features, rotation channel, and last-layer status. Group descriptors encode group index, coefficient ℓ1\ell_1 magnitude and its log/rank transforms, group/term counts, and Pauli-support fractions. Structural parameter/group summaries aggregate light-cone reachability. Pair channels are the fraction of group terms overlapping the parameter’s light cone and the coefficient-weighted fraction of those terms. Current-state channels are exactly [sin⁡θj,cos⁡θj,tanh⁡θj,t/T][\sin\theta_j, \cos\theta_j, \tanh\theta_j, t/T]; they are not a history of energies, gradients, previous masks, or optimizer moments. Historical cache and probe slots remain in the common interface but are not populated with oracle contributions in the direct relevance learner: token construction passes an empty cache and uses fixed action descriptors. Thus a feature name containing “history” does not imply access to past exact gradients. For each token type, training-set channel means and standard deviations standardize inputs; scales below 10−610^{-6} are replaced by one. Held-out examples use these saved statistics without refitting.

Exact Training Objective

Let R\mathcal{R} be valid rows with ∑k∣Cj,k∣>10−10\sum_k \lvert C_{j,k}\rvert> 10^{-10}, aj,k=sj,k/Tsa_{j,k}=s_{j,k}/T_s, and qj=softmax⁡(aj)q_j=\operatorname{softmax}(a_j) over valid groups. Here jj indexes a snapshot and parameter, so reductions below cover all valid rows in a batch. Let mj,km_{j,k} be the binary oracle top-mass mask at the training threshold τtrain\tau_{\mathrm{train}}, let bj=Gj−1∑kmj,kb_j=G_j^{-1}\sum_k m_{j,k}, and let uju_j be the auxiliary budget-head logit. The five terms in Eq. (6) are implemented as

LCE=−1∣R∣∑j∈R∑k=1Gjwj,klog⁡qj,k,(13)\mathcal{L}_{\mathrm{CE}}=-\frac{1}{\lvert\mathcal{R}\rvert}\sum_{j\in\mathcal{R}}\sum_{k=1}^{G_j}w_{j,k}\log q_{j,k}, \tag*{(13)}
Lmask=∑j∈R∑k=1GjBCEWithLogits⁡(aj,k,mj,k)∑j∈RGj,(14)\mathcal{L}_{\mathrm{mask}}=\frac{\sum_{j\in\mathcal{R}}\sum_{k=1}^{G_j}\operatorname{BCEWithLogits}(a_{j,k},m_{j,k})}{\sum_{j\in\mathcal{R}}G_j}, \tag*{(14)}
Lrank=∑j∈R∑k,l(wj,k−wj,l)+softplus⁡(−(aj,k−aj,l)/Tr)max⁡{1,∑j∈R∑k,l(wj,k−wj,l)+},(15)\mathcal{L}_{\mathrm{rank}}=\frac{\sum_{j\in\mathcal{R}}\sum_{k,l}(w_{j,k}-w_{j,l})_{+}\operatorname{softplus}\left(-(a_{j,k}-a_{j,l})/T_r\right)}{\max\left\{1,\sum_{j\in\mathcal{R}}\sum_{k,l}(w_{j,k}-w_{j,l})_{+}\right\}}, \tag*{(15)}
Lbudget=1∣R∣∑j∈RSmoothL1⁡(σ(uj),bj),(16)\mathcal{L}_{\mathrm{budget}}=\frac{1}{\lvert\mathcal{R}\rvert}\sum_{j\in\mathcal{R}}\operatorname{SmoothL1}(\sigma(u_j),b_j), \tag*{(16)}
Lentropy=−1∣R∣∑j∈R∑kqj,klog⁡qj,k.(17)\mathcal{L}_{\mathrm{entropy}}=-\frac{1}{\lvert\mathcal{R}\rvert}\sum_{j\in\mathcal{R}}\sum_k q_{j,k}\log q_{j,k}. \tag*{(17)}

The entropy term has a positive coefficient and therefore penalizes diffuse predictions. BCE supervises the logits independently; it is not BCE applied to row-softmax probabilities. The budget head is an auxiliary training loss for the main top-mass results, not their deployment rule. Historical count-head results instead round its predicted fraction times GjG_j and keep that many top-ranked groups. These two selection modes must not be conflated.

Alternative Predictor Implementations

All four predictors receive the same standardized parameter, group, and pair features and the same training relevance labels in the main architecture comparison. The MLP, GNN, and ResNet-18 use separately trained checkpoints. Their saved settings agree with Table 63: width setting 128, dropout 0.14, 260 epochs, batch size eight, AdamW learning rate 3.5×10−43.5 \times10^{-4}, weight decay 2.5×10−42.5 \times10^{-4}, and initialization seed 123. The initialization and architecture differ, while the objective, normalization procedure, data split, and online protocol are fixed. Other transfer panels use their separately trained checkpoints.

ComponentSaved configuration / implementation
Training panelFive molecular Hamiltonians; layered RY/RZ; depths 1–3
Training seeds11, 23, 37, 41, 53, 67, 79, 83
Held-out seeds97, 101, 107, 109
Label generation18 exact-gradient Adam snapshots per cell; 2,160 snapshots
Angle initializationIndependent N(0,0.352)\mathcal{N}(0, 0.352) angles
Quantum optimizerAdam; learning rate 0.035; no gradient-norm clipping in the main 28-step runner
Classical training260 epochs; batch size 8; AdamW; learning rate 3.5×10−43.5 \times10^{-4}; weight decay 2.5×10−42.5 \times10^{-4}
Classical gradient clipNorm 1.0; initialization seed 123
Loss coefficients(λm,λr,λb,λe)=(0.15,0.35,0.60,0.010)(\lambda_m, \lambda_r, \lambda_b, \lambda_e) = (0.15, 0.35, 0.60, 0.010)
TemperaturesTraining Ts=0.75T_s = 0.75; ranking Tr=0.75T_r = 0.75; inference 0.55
ThresholdsMain training mask 0.99; main inference top-mass 0.995 or 0.997
Checkpoint selectionLowest epoch-average training loss; no held-out evaluation labels used for selection in this runner
Depth-OOD checkpointSeparately trained on depths 1–4; 2,880 snapshots; training mask 0.995; evaluated at depths 5–7

Table 63. Verified settings of the main molecular checkpoint and the separate depth-OOD checkpoint. Training and inference thresholds are different controls.

Pairwise MLP. Independent parameter and group encoders each apply Linear–LayerNorm–GELU–Dropout–Linear–GELU with hidden width 128. The pair scorer concatenates the two embeddings, their product, their absolute difference, and the two pair features, then uses the same 514→256→128→1514 \to256 \to128 \to1 head as the Transformer. The budget head uses LayerNorm and a 128→128→1128 \to128 \to1 MLP. This model has 224,902 parameters and no token-to-token attention.

Bipartite GNN. Parameter and group inputs are projected to width 128. A 2→128→1282 \to128 \to128 GELU MLP followed by a sigmoid produces a vector gate for every parameter-group edge. Two message-passing rounds alternately update parameter nodes from the mean gated valid-group embeddings, then group nodes from the updated valid-parameter embeddings. Each residual update applies LayerNorm and a 256→128→128256 \to128 \to128 GELU/dropout MLP to the node and its aggregated message. The pair and budget heads match the MLP above. Padding is excluded from message sums and denominators. This model has 407,942 parameters.

ResNet-18-style dense predictor. Linear projections map parameter, group, and pair inputs to 64 channels, which are added to form a P×G×64P \times G \times64 grid. A stride-one 3×33 \times3 convolution, one-group GroupNorm, and ReLU replace the image-classification stem. Four stages have two BasicBlocks each, channel widths 64, 128, 256, 512, and first-block strides 1, 2, 2, 2. Blocks contain two 3×33 \times3 convolutions, GroupNorm, ReLU, Dropout2d, and a projected skip when dimensions change. The final feature map is bilinearly interpolated to the input grid size with align_corners=False; GroupNorm–ReLU–1×11 \times1 convolution produces one score per pair. A separate parameter-only 64→64→164 \to64 \to1 budget head is used. There is no ImageNet pretraining or fixed-resolution image resize. The model has 11,214,338 parameters.

Variable shapes and masking. Training pads each batch to its maximum parameter and group counts after feature standardization. Invalid output logits are masked and invalid rows are omitted from the loss. The Transformer masks padded attention keys; the GNN masks messages. ResNet convolutions and GroupNorm operate on the padded grid before output masking, so this implementation should not be assumed invariant to padding. Prediction uses one example at a time. These are architecture substitutions in the implemented common protocol, not independently optimized best configurations for each architecture.

Classical scaling. Let dd be width and LeL_e the number of layers per encoder. With a fixed feed-forward multiplier, dense Transformer encoding requires

O(Le[(P2+G2)d+(P+G)d2])O\left(L_e[(P^2+G^2)d+(P+G)d^2]\right)

operations; the all-pairs MLP adds O(PGd2)O(PGd^2). Materialized attention and pair activations require O(Le(P2+G2)+PGd+Le(P+G)d)O(L_e(P^2+G^2)+PGd+L_e(P+G)d) memory for fixed head count, apart from weights and optimizer state. Sorting all rows for top-mass selection costs O(PGlog⁡G)O(PG\log G). These bounds describe inference after features and groups are constructed; they exclude offline simulation, grouping, and training. Efficient attention implementations can reduce materialized memory without removing the dense interaction cost.

Metric Definitions and Aggregation

For a molecule/ansatz/depth cell cc and paired seed ss, define

ΔEc,s=Ec,smethod(T)−Ec,sFull(T),ΔE‾c=1nc∑sΔEc,s.\Delta E_{c,s}=E^{\mathrm{method}}_{c,s}(T)-E^{\mathrm{Full}}_{c,s}(T),\qquad\overline{\Delta E}_c=\frac{1}{n_c}\sum_s\Delta E_{c,s}.

Energy entries are in Hartree (Ha), with 11 mHa =10−3=10^{-3} Ha. Energy ±\pm values in the molecular, local-LLM, and frontier tables use population standard deviation over the recorded seeds (denominator ncn_c). Shot-gain uncertainties in Table 1 use sample standard deviation (denominator nc−1n_c-1); its reused K=1K=1 rows retain that convention. The sparse-refresh and best-prompt shot-gain entries use population standard deviation. These values describe variation over the four paired seeds, not standard errors or confidence intervals.

Absolute gaps and table rankings. The main 5.67-mHa headline averages seed-level absolute gaps, 1∣C∣∑c1nc∑s∣ΔEc,s∣\frac{1}{|C|}\sum_c\frac{1}{n_c}\sum_s|\Delta E_{c,s}|. It is not the absolute value of the signed mean and need not equal the average of ∣ΔE‾c∣|\overline{\Delta E}_c|. The same seed-level convention is used for the frontier method summaries quoted in the main text. Bold/underline rankings and Random check/cross symbols in detailed tables compare unrounded cell means as specified by each caption. Coloring also uses unrounded means, so values printed identically can have different ranks or threshold outcomes. These are descriptive comparisons, not significance tests. Best ΔE\Delta E in historical budget tables selects a depth-wise reported endpoint; it is not an average over depths.

Passing criteria and reference energies. No worse is one-sided: ΔEc,s≤0.0016\Delta E_{c,s}\leq0.0016 Ha, so lower energy is accepted even outside the symmetric tolerance. Cell-level coloring applies this threshold to ΔE‾c\overline{\Delta E}_c; seed counts apply it to individual trajectories. FCI error uses the separate exact ground-state reference. Legacy raw fields named ca_vs_full use the symmetric test ∣ΔE∣≤0.0016|\Delta E|\leq0.0016 Ha; they must not be substituted for the one-sided no-worse counts. The single-no-horizon benchmark instead records whether a trajectory enters the 1.6-mHa neighborhood of its paired attainable exact-Adam reference at any checkpoint. Its final reference error, best observed reference attainment, and gap to FCI are separate quantities. The best exact-Adam energy found defines the empirical attainable reference for the chosen ansatz, initialization, and optimization schedule.

Measurement cost. For constant shots SS per shifted selected group, B=2S∑t∑j∣Sj,t∣B=2S\sum_t\sum_j|S_{j,t}| and gain =100(1−Bmethod/BFull)%=100(1-B_{\mathrm{method}}/B_{\mathrm{Full}})\%. With adaptive allocation nj,k,tn_{j,k,t}, use B=2∑t,j,knj,k,tB=2\sum_{t,j,k}n_{j,k,t}. Reported aggregate gains average the per-trajectory ratios. A negative gain means more shots than the paired reference. Groups/parameter averages selected pairs over iterations and parameters; it is neither a count of distinct bases across the whole run nor the fraction of optimizer steps skipped. Refresh reuse holds the selection mask, not previously measured expectation values. Nominal counts for exact-expectation runs and sampled-gradient protocols are distinguished in Appendix N.1.

Feature-audit quantities. Captured mass for row jj is mj=∑kMj,k∣Cj,k∣/∑k∣Cj,k∣m_j = \sum_k M_{j,k}|C_{j,k}|/\sum_k |C_{j,k}| on nonzero rows; mean mass averages valid rows. The gradient cosine compares ∑kMj,kCj,k\sum_k M_{j,k}C_{j,k} with the full gradient over all parameters. These quantities need not improve together because signed contributions can cancel. Jaccard compares the selected group sets against the full-feature reference action. A deletion effect in percentage points is 100(mreference−mdeletion)100(m^{\mathrm{reference}}-m^{\mathrm{deletion}}) at matched snapshot and protocol; it is not a percentage relative improvement. Audit quality and token costs are conditional on the validity rules of the corresponding table, so context-window comparisons also report the number of valid cases.

Deterministic Truncation Bias

Condition on the current angles and predicted subset SjS_j. If measurements of retained groups are unbiased, the sparse estimator obeys

E[g^j∣θ,Sj]−gj=−∑k∉SjCj,k,∣E[g^j]−gj∣≤∑k∉Sj∣Cj,k∣.(18)\mathbb{E}[\hat{g}_j \mid\theta, S_j] - g_j = -\sum_{k \notin S_j} C_{j,k}, \qquad\left|\mathbb{E}[\hat{g}_j] - g_j\right| \le\sum_{k \notin S_j} |C_{j,k}|. \tag*{(18)}

If the subset captures at least τ\tau of the exact absolute mass, the last quantity is at most (1−τ)∑k∣Cj,k∣(1-\tau)\sum_k |C_{j,k}|. This bound is conditional on exact retained mass; deployment applies the threshold to predicted mass. The implemented estimator uses deterministic truncation, whose bias is the omitted signed contribution above. The experiments evaluate the resulting closed-loop energy–cost trade-off. Offline exact-statevector label construction scales exponentially in qubit count, setting the computational challenge for larger training systems.

Protocol Provenance and Numerical Diagnostics

Estimator Fidelity and Resource Accounting

The main 28-step molecular, transfer, and local/frontier-LLM ranking runners form gradients as

∑kMj,kCj,k\sum_k M_{j,k} C_{j,k}

using exact simulated expectations. Their shot counts are nominal: each retained parameter–group pair is charged two shifted settings times the specified shots per setting. The shot-count argument does not inject sampling noise in these runners. Randomness there comes from angle initialization and, for Random, sampled group identities, not from quantum shot sampling. Their energy differences therefore measure closed-loop truncation effects, not finite-shot robustness. Exact contributions may be calculated inside the simulator, but are not provided to our learned selector or to an LLM prompt. TITAN-style freezing controls additionally use privileged exact-gradient importance in simulation; their ranking information is stronger than a deployable zero-cost controller.

The separate shot-allocation panel samples gradient noise using exact grouped observable moments under the specified allocation, and the long-horizon benchmark uses a Gaussian finite-shot approximation with variance determined by those moments. These are simulated noisy estimators, not hardware measurements or exact sampling of every bitstring. The long-horizon benchmark uses 4096 shots per shifted selected group, depth 3, three paired seeds, 400 steps, and objective H+(N−Ne)2H+(N-N_{\mathrm{e}})^2, reporting energy of HH. It is a distinct protocol, not a continuation of each 28-step run. It trains a separate depth-3 controller on 80-label-snapshot trajectories per cell, with training seeds 11/23/37 and validation seed 41; evaluation uses 97/101/107. Its model stays frozen during the 400-step quantum optimization. Its exact-Adam attainable references remain 9.459–51.371 mHa above FCI; Reference parity and absolute FCI error therefore quantify distinct aspects of the optimization outcome.

The calibration-derived Pauli-channel surrogate, finite-shot Aer grid spot check, and shot-noise VQE are also distinct experiments. The Aer check uses depth 1, seed 97, and 4096 shots per shifted group. Its 15–26% omitted cells describe grid sparsity, not realized learned-controller savings in a complete noisy online VQE. All reported noise studies use simulation. Online shot gain counts gradient measurement settings; it excludes offline label generation, classical training/inference, retrieval, compilation, energy diagnostics, and LLM token cost. Parameter counts, file sizes, tokens, and wall times are separate resource axes, not directly interchangeable costs. Historical dollar estimates use the original run-time accounting assumptions. Token traffic is the directly logged resource measure. The current scale comprises small molecular systems, four-seed short-horizon panels, and compact spin-model QWC partitions; larger systems require broader data generation and greater simulation resources.

Allocation controls and comparison scope. iCANS, gCANS, and Rosalin labels denote adapted allocation/operator-sampling policies inside a common Adam harness, with the caps, warm-up, and running moment estimates specified by the corresponding experiment. They do not use each paper’s native optimizer or stopping rule; Adam, the horizon, and the experimental budget are shared controls. The panels evaluate allocation and operator-sampling policies, including their composition with learned QWC selection. The empirical baseline scope is this fixed-partition selection/allocation setting; related work separately discusses learned POVM and shadow-based measurement design.

Recorded allocation settings. The main allocation panel uses five molecules, depths 1–2, seeds 97/101/107/109, 28 Adam steps at learning rate 0.035, and no gradient-norm clipping. Nominal shots are 1024 per shifted group, with ϵ∈{0.25,0.50,1.00}\epsilon\in\{0.25, 0.50, 1.00\}. The learned mask uses the main saved checkpoint, inference temperature 0.55, and top-mass threshold 0.995. The allocation state starts with zero gradient and coefficient-based variance prior max⁡(∥hgk∥1,10−4)2\max(\lVert h_{g_k}\rVert_1,10^{-4})^2. After observations, gradient and group-variance forecasts use exponential averaging with decay 0.9. Variance observations are simulated from the exact grouped moments using a chi-square sample with n−1n-1 degrees of freedom; the gradients use a Gaussian moment approximation.

The adapted CANS policies use L=10L=10, η=0.035\eta=0.035, and a=2ηL/(2−ηL)a=2\eta L/(2-\eta L). Their unrounded per-pair requests are

sjiCANS=a∑kv^j,kg^j2+10−8,sjgCANS=aσj∑iσi∑ig^i2+10−8,σj2=∑k∈Sjv^j,k.s_j^{\mathrm{iCANS}}=a\frac{\sum_k\hat v_{j,k}}{\hat g_j^2+10^{-8}},\qquad s_j^{\mathrm{gCANS}}=a\frac{\sigma_j\sum_i\sigma_i}{\sum_i\hat g_i^2+10^{-8}},\qquad\sigma_j^2=\sum_{k\in S_j}\hat v_{j,k}.

Requests are multiplied by bb, rounded upward, and clipped to [⌈64b⌉,⌈4096b⌉][\lceil64b\rceil, \lceil4096b\rceil]. The first uninitialized step uses the lower bound. These are shots per selected pair, so the row cost also scales with its number of retained groups.

Variance-aware allocation keeps each row’s nominal total and distributes it in proportion to predicted standard deviations, with a minimum of two shots per selected group and stable largest-remainder integer rounding. The combined learned-mask/gCANS control additionally uses this within-mask reallocation. Rosalin uses the adapted iCANS row request times the full group count, multinomial group sampling proportional to coefficient ℓ1\ell_1 magnitude, and inverse-probability correction. All methods then take the same Adam update.

Small estimated gradients can drive the CANS request toward its cap; therefore shared nominal scale does not enforce equal realized shot cost, and negative gains are possible. This explains how overspending arises in this implementation without establishing why a native algorithm would perform poorly. The stored run records one configuration for these constants, not a dedicated hyperparameter search for each native optimizer. The comparison supports composition under the stated harness.

Threshold and Checkpoint Provenance

Table 9 is a historical aggregate, not a single-checkpoint threshold sweep. Its 0.98 and 0.99 rows use separately trained checkpoints with different oracle training-mask thresholds and the historical count-head selection mode. Its 0.95 row is an earlier trained operating point. Only the 0.995 and 0.997 rows use the same saved τtrain=0.99\tau_{\mathrm{train}} = 0.99 checkpoint with top-mass post-processing. The main molecular baseline panel also loads that checkpoint. The Transformer checkpoint used for LLM cardinalities and native attribution has identical model-state and normalization values; only its serialized artifact/configuration hash differs. In contrast, Table 10 reuses one separately trained depth-OOD checkpoint over the displayed inference thresholds. Thus the nonmonotone group counts between historical 0.99 and 0.995 rows are not a violation of fixed-prediction top-mass nesting. Even for a fixed checkpoint, different masks change subsequent angles, so closed-loop average group counts and final energies need not be monotone in the threshold. A checkpoint transferred within a named study should not be described as one universal checkpoint across separately trained depth-, molecule-, and Hamiltonian-family panels. “Best ΔE\Delta E” columns are selected descriptive endpoints, not estimates of average reliability; depth-wise and seed-wise results remain the relevant evidence for that question.

Train-Only Retrieval for Frontier Models

The retrieval bank is restricted to the same molecule, ansatz, depth, and parameter ID, using training seeds only. For a parameter row, distances are squared Euclidean distances between [sin⁡θj,cos⁡θj,tanh⁡θj,t/T][\sin\theta_j,\cos\theta_j,\tanh\theta_j,t/T], standardized by the training bank’s per-channel standard deviation (values below 10−410^{-4} are replaced by one). Invalid zero-mass rows are excluded and ties use a stable sort. The prompt receives two explicit nearest-neighbor records and a summary over the eight-neighbor pool, including average training relevance and selection frequencies, with a limited list of salient group signals. It is not an unbounded dump of the full training dataset. Retrieval uses training relevance labels, never the current held-out contribution matrix. Both use exact offline relevance supervision, but the frontier bank uses 20 snapshots per training cell (2,400 total), versus 18 (2,160 total) for the main learned checkpoint. Data access is therefore not matched in snapshot count, and in-context examples are not equivalent to gradient-based training. Unlike the local fixed-cardinality tests, frontier models choose both identities and counts: the request explicitly removes each parameter’s kk field and accepts either a select or exclude list with at least one measured group. This differs in both data access and count policy from the local tests, so the comparison evaluates the combined model and data interface. The requested 0.995 mass is an objective informed by retrieved training labels; the validator checks output structure, range, and cardinality. The four-refresh schedule is shared with our K=7K = 7 control, but budgets are not matched; comparisons are accuracy-cost operating points. This evaluates data-exposed frozen selection within an external harness, not a fine-tuned LLM or autonomous code-executing agent.

Decision Traces, Prompt Identity, and Reason Tags

The action and reason tags in trace Box (4) come from the audit panel’s explain_tags prompt. The full-feature reference and deletions in Box (7) come from baseline, no_coeff, no_lightcone, and no_theta_time. All four decision traces use H+H^{+}, depth 2, seed 97, iteration 0, parameter 14, and cardinality six. The decision-trace state is initialized with θj∼N(0,0.352)\theta_j \sim\mathcal{N}(0, 0.35^2), giving θ14=+0.1067889170\theta_{14} = +0.1067889170 rad. The 24 circuit parameters and 19 QWC groups describe problem dimensions, not tokenizer counts. Local contribution shares, mean captured mass, and gradient cosine are evaluated on this same state. The integrated-gradient studies teacher-force the reference action using their separately serialized attribution inputs (Appendix N.5). The prompt-specific groups are:

ModelExplanation-prompt actionReference action / IG target
Qwen3-8B[0, 1, 2, 3, 4, 5][0, 1, 2, 3, 4, 5]
Llama-3.1-8B[7, 8, 9, 10, 11, 12][4, 5, 6, 7, 8, 9]
DeepSeek-R1-8B[0, 2, 3, 4, 6, 7][2, 5, 9, 6, 0, 7]
GPT-OSS-20B[18, 0, 17, 16, 15, 14][0, 1, 2, 6, 7, 3]

Table 68.

The raw Qwen explanation tags are lightcone, theta, and circuit_location; Llama emits coefficient, theta, iteration; DeepSeek emits coefficient, lightcone; GPT-OSS emits coefficient. Box (3) reproduces these logged tags. Each action’s exact contribution shares are computed after inference, outside the model, solely for evaluation. An intervention difference must be subtracted from the reference-prompt score, not from a different explanation-prompt score. Llama’s explanation action captures 1.797%; its reference action captures 1.442%, and deleting coefficients raises that local reference mass to 94.843% (+93.401 percentage points). Over valid parameters, mean mass instead falls from 0.483 to 0.395, distinguishing the local intervention effect from aggregate selection quality.

Dependence, usefulness, and faithfulness are distinct. A changed mask establishes input sensitivity in the reference protocol. A positive captured-mass drop means the deleted feature was useful there; a negative drop means it was harmful there, not that it was unused. Unchanged outputs can reflect redundancy or feature interactions. Reason tags and deletion effects are linked by snapshot while retaining their distinct prompt identities: explain_tags produces the reported tags, and reference-prompt interventions measure feature utility. The verdicts in the drawings are interpreted through this distinction. These traces record observable interaction and externally measured sensitivity. Local LLMs run frozen text inference with code-execution tools disabled; think=false specifies the direct-output configuration used in this comparison.

Protocol-specific output contracts. The rank-only/repaired local protocol in Appendix J emits a ranking. The clean protocol instead uses selections.p14 = {mode,group_ids}, emitting either the selected IDs or their shorter excluded complement at exact cardinality. The audit’s optional tags use the controlled vocabulary coefficient, lightcone, theta, circuit_location, iteration, and group_support, with one to three unique tags per row. The clean validator checks integer type (excluding booleans), range, uniqueness, mode, and length; schema feedback permits retries, but no coefficient completion repairs IDs. Persistent failure invokes logged Full-QWC safety fallback and charges all groups. The audit additionally varies output contracts and optionally returns reason_codes; the best-prompt evaluation tests its selected prompt online. The example {parameter_id,mode,group_ids,reason_codes} in a trace is a normalized logged record, not the complete literal clean-protocol wire schema. The best prompt was selected on the fixed audit panel, which includes seed 97, also present in the online evaluation. It is the best tested configuration on that tuning panel; the shared seed links prompt selection to the corresponding online evaluation cells. The pair_nonzero_rows key names the column dictionary, whereas pair_nonzero holds each parameter’s rows; the different names are intentional. These are frozen LLM inference protocols within our external Python harness, not learned tool-use agents.

Clean system instruction and explanation extension. The clean instruction, also used by the domain-prompt audit baseline, is:

You are a measurement-selection controller for variational
quantum eigensolvers.  For every requested circuit
parameter, follow its fixed output mode and emit exactly the
requested number of distinct QWC group IDs.  In select mode,
emit the groups most relevant to that parameter’s energy
gradient, ordered from most to least relevant.  In exclude
mode, emit the groups least relevant to that gradient,
ordered from least to most relevant; every unlisted
candidate will be measured.  Use Hamiltonian coefficient,
circuit location, theta, iteration, and parameter–group
light-cone features jointly.  A globally large coefficient
does not imply equal relevance to every parameter.  Return
only the JSON object required by the schema.  Never omit a
requested parameter and never invent an ID.  The exact
cardinality is mandatory.

For the explanation condition the appended instruction is:

For each parameter, also emit one to three reason_codes
chosen only from [’coefficient’, ’lightcone’, ’theta’,
’circuit_location’, ’iteration’, ’group_support’].  These are
concise evidence tags, not chain-of-thought.

The domain reference uses a 32,768-token context window and the compressed select/exclude contract. A row with requested count kj≤G−kjk_j \le G-k_j uses select mode and emits kjk_j IDs; otherwise it emits the G−kjG-k_j excluded IDs. For the shared trace with G=19G=19 and k14=6k_{14}=6, the explanation response has the following wire structure (shown with the accepted Qwen action):

{"selections": {
  "p14": {
    "mode": "select",
    "group_ids": [0, 1, 2, 3, 4, 5],
    "reason_codes": ["lightcone", "theta", "circuit_location"]
  }
}}

Without explanation tags the same reference contract omits reason_codes. The request also specifies each row’s selected_count, output_mode, and emit_count. The returned object is validated against those values, range and uniqueness constraints, and the complete set of requested parameter keys.

What the context ablation varies. The context ablation changes the configured context-window capacity to 8,192, 16,384, or 32,768 tokens. This is a capacity setting, not a promise that every request fills that many tokens. Its separate prompt variants use a minimal instruction or add two training demonstrations; output variants emit a direct selected list or a complete ranking. Deletion and shuffling conditions hold the Full-QWC snapshot and requested cardinalities fixed. Explanation tags are an additional condition. Reported differences compare configurations under this harness; context capacity, protocol success, and valid-case mask quality describe distinct parts of controller performance.

Observable computation boundary. The models serialize a selection as text/JSON; Python constructs features, retrieves training records where enabled, validates outputs, applies masks, and performs the VQE update. Code-execution and external tool calls are not available to the evaluated local models, and frontier tool use is disabled in the saved CLI protocol. These controls specify the observable interaction and the division of work between model and harness. The frontier GPT runs record low reasoning effort; the local direct-output tests record think=false. Conclusions concern those settings.

Integrated-Gradient Numerical Reliability

The four LLM attribution prompts use the serialized H3+\mathrm{H}_{3}^{+}, depth-2, seed-97, iteration-0 state with initialization scale 0.12 (θ14≃+0.0366\theta_{14} \simeq+0.0366 rad) and the fixed reference group lists above. The decision traces use their recorded scale-0.35 state; each analysis evaluates its own state and score while sharing the task and target IDs. For the LLM figures, we fix the observed action and teacher-force the JSON prefix for each emitted ID. The scalar score is the first-token log probability conditioned on grammar-valid ID prefixes, not the probability of the entire freely generated mask. Only marked numeric feature-token embeddings are interpolated. The input is tokenized once: the baseline replaces these embeddings with repeated embeddings of the token for numeric zero, preserving sequence length, positions, attention mask, syntax, instructions, and the fixed output prefix. It does not retokenize a string of shorter numerical values. This is a synthetic embedding-space baseline, not a physically valid alternative VQE state. Attribution sums are aggregated by labeled spans. Absolute-mass normalization summarizes the distribution of attribution magnitude across input families; its denominator is total absolute attribution, whereas energy accuracy is evaluated separately.

The Ollama/GGUF generation runtime is not differentiable in this setup; gradients come from matched Hugging Face model families under the quantization specified by each figure. Quantization, kernels, and routed experts can change scores and continuations. Our controller instead uses its native checkpoint, standardized scalar features, a training-mean baseline, and 1024 quadrature nodes. Therefore cross-model percentages are not calibrated on a common functional or numerical scale; each panel describes its own target score and baseline.

For a scored group, let D=f(x)−f(x0)D = f(x) - f(x_{0}) and R=D−∑iIGiR = D - \sum_i \mathrm{IG}_i. We report both absolute and relative error: small ∣R∣|R| alone is insufficient when ∣D∣|D| is near zero. The table below is computed from the saved metadata, not from a new experiment. Qwen’s first score has a 1.61% residual relative

ModelNodesMax ∣R∣|R|Max ∣R∣/∣D∣|R|/|D|∣R∣>∣D∣|R| > |D|
Qwen3-8B320.12248.564/6
Llama-3.1-8B321.149193.333/6
DeepSeek-R1-8B320.43734.632/6
GPT-OSS-20B321.41401.402/6
Our controller10240.01260.00460/6

Table 64. Completeness diagnostics for the six displayed group scores. The last column counts scores with ∣R∣>∣D∣|R| > |D|; these estimates fail even a 100% relative completeness check. Residuals accompany the finite-quadrature estimates for each model.

to its score change; several later near-saturated continuation scores have substantially larger relative errors. The Llama, DeepSeek, and GPT-OSS entries likewise show substantial finite-quadrature residuals. Their plotted family shares are finite-quadrature estimates under the specified scoring functions. The native controller uses 1024 nodes and has maximum relative residual 0.46%; its 12.16% current-state attribution identifies a resolved dependence of its group scores on optimization-state inputs. The LLM panels use different scoring functions and embedding baselines, so the family percentages describe within-model attribution profiles. Closed-loop energy and shot cost provide the common cross-model performance measures.

Fixed-Snapshot Attribution Stability Audit

We test whether the feature-family ordering in Figure 6 persists across numerical integration settings. This completed follow-up evaluates the same four serialized H3+_3^+ attribution cases described in Appendix § 24.§ 5: depth 2, seed 97, iteration 0, parameter 14, and initialization scale 0.12. Each model retains its archived prompt, token positions, reference group IDs, teacher-forced JSON prefix, score function, and numeric-zero embedding baseline. These are numerical repeats of fixed cases, not additional physical snapshots or online VQE runs.

Controlled refinement. We use 32, 64, 128, and 256 Gauss–Legendre nodes with float64 accumulation. All four models are evaluated under their original attribution runtime. For Qwen, Llama, and DeepSeek, a separate FP32-compute diagnostic preserves the NF4 weight bytes while promoting floating-point layers and quantized matrix-multiplication compute to FP32. This is a separate arithmetic regime, not an unquantized model or a bitwise-identical scoring implementation. GPT-OSS retains its native MXFP4 implementation and hard expert routing. The audit thus contains 28 model–precision–quadrature configurations, each scoring six fixed group IDs. No model is trained, no new action is generated, and the original plotted assets are retained.

Aggregation and ranking. For output group gg, let IgsI_{gs} be the signed IG summed over the tokens of numeric input span ss. Its feature-family share is

AgF=100∑s∈F∣Igs∣∑s∣Igs∣,A‾F=16∑gAgF.(19)A_{gF} = 100 \frac{\sum_{s \in F} \lvert I_{gs} \rvert}{\sum_s \lvert I_{gs} \rvert}, \qquad\overline{A}_F = \frac{1}{6}\sum_g A_{gF}. \tag*{(19)}

The family ranking uses A‾F\overline{A}_F. Percentages therefore summarize absolute span attribution, not selection probability, causal effect size, or energy contribution. Structure includes circuit location and Pauli support; budget/action covers controller-contract fields. Trajectory denotes the serialized current-angle and normalized-step features, rather than a complete history of measured gradients.

Consistent low trajectory attribution. Trajectory ranks fifth among the five families for every scored group in every configuration: 168/168 output–configuration combinations. The model-specific orderings and trajectory-share ranges are reported in Table 65. Under the original runtime, increasing from 128 to 256 nodes changes these means by only 0.024, 0.165, 0.007, and 0.037 percentage points. These counts describe numerical robustness within the fixed cases, rather than independent-sample statistical evidence.

ModelStable mean rankingTrajectory at 256 nodes (%) — OriginalTrajectory at 256 nodes (%) — FP32All-setting range (%)
Qwen3-8BS>C>L>B>TS > C > L > B > T0.771.040.59–1.06
Llama-3.1-8BS>C>L>B>TS > C > L > B > T0.950.760.67–0.95
DeepSeek-R1-8BS>C>L>B>TS > C > L > B > T0.991.010.96–1.01
GPT-OSS-20BS>L>C>B>TS > L > C > B > T2.49—2.49–2.71

Table 65. Feature-family ranking and trajectory-share stability in the fixed-snapshot audit. SS: circuit/support; CC: coefficient; LL: light cone; BB: budget/action; TT: trajectory. Each model’s mean ranking is identical across its tested settings. Shares are averaged over six scored group IDs; the range spans all tested node counts and arithmetic regimes, not a confidence interval. FP32 is a compute diagnostic with unchanged NF4 weights.

Implication for frozen selection. Despite receiving the full input feature set, frozen LLMs do not automatically make substantial use of trajectory information without targeted intervention in these audited cases: static physical inputs dominate their attribution profiles. This pattern persists across the tested quadrature and precision settings. It complements the feature-deletion and online evaluations in Table 56 and Table 55, which assess changes in decisions and closed-loop performance. The consistently low within-model trajectory ranking, rather than an exact cross-model percentage ratio, is the central result of this refinement.

SettingProtocolH2 (4q)HeH+ (4q)H3+ (6q)H4 chain (8q)LiH (12q)
Depth 1GPT clean gain, K=7K=7 ↑39.4±1.8%39.4 \pm1.8\%62.6±1.3%62.6 \pm1.3\%67.0±1.5%67.0 \pm1.5\%79.6±1.8%79.6 \pm1.8\%89.9±0.8%89.9 \pm0.8\%
Depth 1GPT best gain, K=7K=7 ↑42.5±4.6%42.5 \pm4.6\%61.7±3.4%61.7 \pm3.4\%66.7±1.5%66.7 \pm1.5\%79.3±1.8%79.3 \pm1.8\%90.5±0.4%90.5 \pm0.4\%
Depth 1GPT-OSS (clean), ΔE\Delta E+0.3930±0.1936+0.3930 \pm0.1936+0.8238±0.2303+0.8238 \pm0.2303+0.7300±0.4383+0.7300 \pm0.4383+0.5671±0.3164+0.5671 \pm0.3164+1.5503±0.7644+1.5503 \pm0.7644
Depth 1GPT-OSS (best), ΔE\Delta E+0.3392±0.1856+0.3392 \pm0.1856+0.9079±0.4907+0.9079 \pm0.4907+0.3882±0.2525+0.3882 \pm0.2525+0.6337±0.2763+0.6337 \pm0.2763+1.9964±0.9332+1.9964 \pm0.9332
Depth 2GPT clean gain, K=7K=7 ↑29.5±3.5%29.5 \pm3.5\%34.3±1.7%34.3 \pm1.7\%49.8±1.9%49.8 \pm1.9\%55.5±1.6%55.5 \pm1.6\%84.1±0.5%84.1 \pm0.5\%
Depth 2GPT best gain, K=7K=7 ↑29.7±2.2%29.7 \pm2.2\%33.0±2.3%33.0 \pm2.3\%49.8±0.8%49.8 \pm0.8\%55.4±0.4%55.4 \pm0.4\%84.3±0.5%84.3 \pm0.5\%
Depth 2GPT-OSS (clean), ΔE\Delta E+0.2148±0.1220+0.2148 \pm0.1220+0.6807±0.7081+0.6807 \pm0.7081+0.4691±0.2024+0.4691 \pm0.2024+0.5616±0.2675+0.5616 \pm0.2675+1.2894±0.3290+1.2894 \pm0.3290
Depth 2GPT-OSS (best), ΔE\Delta E−0.0510±0.0402-0.0510 \pm0.0402+0.1302±0.0822+0.1302 \pm0.0822+0.4776±0.3130+0.4776 \pm0.3130+0.4682±0.1297+0.4682 \pm0.1297+1.5294±0.3237+1.5294 \pm0.3237
Depth 3GPT clean gain, K=7K=7 ↑19.8±1.9%19.8 \pm1.9\%25.1±1.8%25.1 \pm1.8\%32.1±1.4%32.1 \pm1.4\%39.0±0.5%39.0 \pm0.5\%76.8±1.5%76.8 \pm1.5\%
Depth 3GPT best gain, K=7K=7 ↑19.5±1.6%19.5 \pm1.6\%26.3±1.2%26.3 \pm1.2\%32.1±1.0%32.1 \pm1.0\%38.7±1.1%38.7 \pm1.1\%76.6±1.3%76.6 \pm1.3\%
Depth 3GPT-OSS (clean), ΔE\Delta E+0.1633±0.1815+0.1633 \pm0.1815+0.3438±0.1193+0.3438 \pm0.1193+0.1786±0.0759+0.1786 \pm0.0759+0.5089±0.3596+0.5089 \pm0.3596+0.5647±0.4252+0.5647 \pm0.4252
Depth 3GPT-OSS (best), ΔE\Delta E+0.0478±0.0637+0.0478 \pm0.0637+0.0162±0.1316+0.0162 \pm0.1316+0.1222±0.1216+0.1222 \pm0.1216+0.0449±0.2110+0.0449 \pm0.2110+0.5331±0.3351+0.5331 \pm0.3351

Table 55. Does the selected prompt improve clean frozen-LLM selection online? GPT-OSS (clean) is the clean exact select/exclude baseline; GPT-OSS (best) uses the selected domain prompt, direct-ID output, 32K context, and all physical features. Both remain frozen, receive the same count kjk_j requested by our method, and refresh group identities every K=7K=7 steps. Gain and energy entries are mean ±\pm standard deviation over seeds 97/101/107/109, with ΔE\Delta E in Ha relative to Full-QWC; gradients use exact expectations and shot costs are nominal (Appendix N.1). An energy cell for the best prompt is green when its absolute mean gap is smaller than that of the clean baseline and red otherwise. The complete Random and our method controls are reported in Appendix Table 57.

Numerical checks and their relation to ranking stability. The audit predeclares a per-output completeness tolerance ∣R∣≤0.001+0.05∣D∣|R| \le0.001 + 0.05|D|, with DD and RR defined in Appendix § 24.§ 5. Between successive node counts, it also checks maximum family-cell change ≤1\le1 percentage point, maximum trajectory-cell change ≤0.5\le0.5 percentage points, and minimum per-output family-rank Spearman correlation ≥0.9\ge0.9. Refinement can stop after at least 128 nodes only when two consecutive levels meet all conditions; otherwise it continues through 256 nodes. All seven model–runtime pairs reach 256 nodes in this audit.

ModelComputeMax ∣R∣|R|Complete scoresΔF\Delta_FΔT\Delta_Tρmin⁡\rho_{\min}
Qwen3-8BOriginal1.82690/63.2960.2050.90
Qwen3-8BFP323.44252/62.5510.5280.90
Llama-3.1-8BOriginal0.96300/65.3810.5681.00
Llama-3.1-8BFP320.36022/62.1440.1341.00
DeepSeek-R1-8BOriginal0.21881/61.2760.0470.90
DeepSeek-R1-8BFP320.02224/60.8120.0131.00
GPT-OSS-20BOriginal1.41160/61.8380.3521.00

Table 66. Final numerical diagnostics, kept separate from the stable mean rankings in Table 65. Residual and completeness pass counts use 256 nodes. Changes and correlations compare 128 with 256 nodes: ΔF\Delta_F is the largest absolute family-cell change, ΔT\Delta_T the largest trajectory-cell change (both in percentage points), and ρmin⁡\rho_{\min} the minimum per-output rank correlation. Pass counts apply the combined absolute/relative completeness tolerance, not a relative-error-only threshold.

DeepSeek with FP32 compute meets the family-stability conditions at 256 nodes, with a maximum family-cell change of 0.812 percentage points and identical per-output family ranks. Its completeness tolerance is met by four of six scores. Across all seven model–runtime pairs, the number of complete scores ranges from zero to four, so the full conjunction of numerical criteria is not met. The other pairs retain larger family-cell changes even when their mean ordering and low trajectory shares persist. Thus the audit reports ranking robustness and completeness as separate properties; the percentages remain finite-quadrature estimates. In particular, normalizing absolute attribution to 100% is not a correction for residual error. No post-hoc rescaling is applied.

References

  1. [1]Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925, 2025.arxiv.org/abs/2508.10925
  2. [2]Christopher J Anders, Kim A Nicoli, Binting Wu, Naima Elesogui, Samuele Pedrielli, Lena Funcke, Karl Jansen, Stefan Kühn, and Shinichi Nakajima. Adaptive observation cost control for variational quantum eigensolvers. arXiv preprint arXiv:2502.01704, 2025.arxiv.org/abs/2502.01704
  3. [3]Anthropic. Anthropic’s transparency hub, 2026. URL https://www.anthropic.com/transparency.
  4. [4]Andrew Arrasmith, Lukasz Cincio, Rolando D Somma, and Patrick J Coles. Operator sampling for shot-frugal optimization in variational algorithms. arXiv preprint arXiv:2004.06252, 2020.
  5. [5]Alba Cervera-Lierta, Jakob S Kottmann, and Alán Aspuru-Guzik. Meta-variational quantum eigen-solver: Learning energy profiles of parameterized hamiltonians for quantum simulation. PRX Quantum, 2(2):020329, 2021.
  6. [6]Ophelia Crawford, Barnaby Van Straaten, Daochen Wang, Thomas Parks, Earl Campbell, and Stephen Brierley. Efficient quantum measurement of pauli operators in the presence of finite sampling error. Quantum, 5:385, 2021.DOI
  7. [7]Guillermo García-Pérez, Matteo AC Rossi, Boris Sokolov, Francesco Tacchino, Panagiotis Kl Barkoutsos, Guglielmo Mazzola, Ivano Tavernelli, and Sabrina Maniscalco. Learning to measure: Adaptive informationally complete generalized measurements for quantum algorithms. Prx quantum, 2(4):040342, 2021.
  8. [8]Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pp. 1263–1272. Pmlr, 2017.arxiv.org/abs/1704.01212
  9. [9]Pranav Gokhale, Olivia Angiuli, Yongshan Ding, Kaiwen Gui, Teague Tomesh, Martin Suchara, Margaret Martonosi, and Frederic T Chong. Minimizing state preparations in variational quantum eigensolver by partitioning into commuting families. arXiv preprint arXiv:1907.13623, 2019.arxiv.org/abs/1907.13623
  10. [10]Jérôme F Gonthier, Maxwell D Radin, Corneliu Buda, Eric J Doscocil, Elena M Abuan, and Jhonathan Romero. Measurements as a roadblock to near-term practical quantum advantage in chemistry: Resource analysis. Physical Review Research, 4(3):033154, 2022.DOI
  11. [11]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhari, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.arxiv.org/abs/2407.21783
  12. [12]Harper R Grimsley, Sophia E Economou, Edwin Barnes, and Nicholas J Mayhall. An adaptive variational algorithm for exact molecular simulations on a quantum computer. Nature communications, 10(1):3007, 2019.DOI
  13. [13]Andi Gu, Angus Lowe, Pavel A Dub, Patrick J Coles, and Andrew Arrasmith. Adaptive shot allocation for fast convergence in variational quantum algorithms. arXiv preprint arXiv:2108.10434, 2021.DOI
  14. [14]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-rl: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025.
  15. [15]Charles Hadfield, Sergey Bravyi, Rudy Raymond, and Antonio Mezzacapo. Measurements of quantum Hamiltonians with locally-biased classical shadows. Communications in Mathematical Physics, 391(3):951–967, 2022.arxiv.org/abs/2006.15788
  16. [16]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.arxiv.org/abs/1512.03385
  17. [17]Warren J Hehre, Robert F Stewart, and John A Pople. Self-consistent molecular-orbital methods. i. use of gaussian expansions of slater-type atomic orbitals. The Journal of Chemical Physics, 51 (6):2657–2664, 1969.DOI
  18. [18]Hsin-Yuan Huang, Richard Kueng, and John Preskill. Predicting many properties of a quantum system from very few measurements. Nature Physics, 16(10):1050–1057, 2020.DOI
  19. [19]IBM Quantum. Fake provider: qiskit_ibm_runtime.fake_provider, 2026a. URL https://qiskit.github.io/qiskit-ibm-runtime/apidocs/fake_provider.html. Accessed: 2026-07-17.
  20. [20]IBM Quantum. Build noise models, 2026b. URL https://quantum.cloud.ibm.com/docs/en/guides/build-noise-models. Accessed: 2026-07-17.
  21. [21]Azhar Ikhtiarudin, Gagus Ketut Sunnarディanto, Fadjар Fathurrahman, Mohammad Kemal Agusta, and Hermawan Kresno Dipopjono. Shot-efficient adapt-vqe via reused pauli measurements and variance-based shot allocation. arXiv preprint arXiv:2507.16879, 2025.DOI
  22. [22]Artur F Izmaylov, Tzu-Ching Yen, Robert A Lang, and Vladyslav Vereteltskyi. Unitary partitioning approach to the measurement problem in the variational quantum eigensolver method. Journal of chemical theory and computation, 16(1):190–195, 2020.arxiv.org/abs/1907.09040
  23. [23]Ali Javadi-Abhari, Matthew Treinish, Kevin Krulich, Christopher J Wood, Jake Lishman, Julien Gacon, Simon Martiel, Paul D Nation, Lev S Bishop, Andrew W Cross, et al. Quantum computing with qiskit. arXiv preprint arXiv:2405.08810, 2024.
  24. [24]Abhinav Kandala, Antonio Mezzacapo, Kristan Temme, Maika Takita, Markus Brink, Jerry M Chow, and Jay M Gambetta. Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets. nature, 549(7671):242–246, 2017.DOI
  25. [25]Jonas M Kübler, Andrew Arrasmith, Lukasz Cinicio, and Patrick J Coles. An adaptive optimizer for measurement-frugal variational algorithms. Quantum, 4:263, 2020.DOI
  26. [26]Senwei Liang, Linghua Zhu, Xiaolin Liu, Chao Yang, and Xiaosong Li. Artificial-intelligence-driven shot reduction in quantum measurement. Chemical Physics Reviews, 5(4), 2024.arxiv.org/abs/2405.02493
  27. [27]Sam McArdle, Suguru Endo, Alán Aspuru-Guzik, Simon C Benjamin, and Xiao Yuan. Quantum computational chemistry. Reviews of Modern Physics, 92(1):015003, 2020.DOI
  28. [28]Jarrod R McClean, Jonathan Romero, Ryan Babbush, and Alán Aspuru-Guzik. The theory of variational hybrid quantum-classical algorithms. New Journal of Physics, 18(2):023023, 2016.DOI
  29. [29]Jarrod R McClean, Sergio Boixo, Vadim N Smelyanskyi, Ryan Babbush, and Hartmut Neven. Barren plateaus in quantum neural network training landscapes. Nature communications, 9(1):4812, 2018.DOI
  30. [30]Jarrod R McClean, Nicholas C Rubin, Kevin J Sung, Ian D Kivlichan, Xavier Bonet-Monroig, Yudong Cao, Chengyu Dai, E Schuyler Fried, Craig Gidney, Brendan Gimby, et al. Openfermion: the electronic structure package for quantum computers. Quantum Science & Technology, 5(3):034014, 2020.arxiv.org/abs/1710.07629
  31. [32]Yifeng Peng, Xinyi Li, Samuel Yen-Chi Chen, Kaining Zhang, Zhiding Liang, Ying Wang, and Yuxuan Du. Titan: A trajectory-informed technique for adaptive parameter freezing in large-scale vqe. Advances in Neural Information Processing Systems, 38:47294–47315, 2026.arxiv.org/abs/2509.15193
  32. [33]Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man-Hong Yung, Xiao-Qi Zhou, Peter J Love, Alán Aspuru-Guzik, and Jeremy L O’Brien. A variational eigenvalue solver on a photonic quantum processor. Nature communications, 5(1):4213, 2014.DOI
  33. [34]Maria Schuld, Ville Bergholm, Christian Gogolin, Josh Isaac, and Nathan Killoran. Evaluating analytic gradients on quantum hardware. Physical Review A, 99(3):032331, 2019.DOI
  34. [35]Jacob T Seeley, Martin J Richard, and Peter J Love. The bravyi-kitaev transformation for quantum computation of electronic structure. The Journal of chemical physics, 137(22), 2012.DOI
  35. [36]Qiming Sun, Timothy C Berkelbach, Nick S Blunt, George H Booth, Sheng Guo, Zhendong Li, Junzi Liu, James D McClain, Elvira R Sayfutyarova, Sandeep Sharma, et al. PsycF: the python-based simulations of chemistry framework. Wiley Interdisciplinary Reviews: Computational Molecular Science, 8(1):e1340, 2018.DOI
  36. [37]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.arxiv.org/abs/1706.03762
  37. [39]David Wierichs, Josh Isaac, Cody Wang, and Cedric Yen-Yu Lin. General parameter-shift rules for quantum gradients. Quantum, 6:677, 2022.DOI
  38. [40]William K Wootters and Wojciech H Zurek. A single quantum cannot be cloned. Nature, 299(5886):802–803, 1982.DOI
  39. [41]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chenggen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025.arxiv.org/abs/2505.09388
  40. [42]Tzu-Ching Yen, Vladyslav Verteletskyi, and Artur F Izmaylov. Measuring all compatible operators in one series of single-qubit measurements using unitary transformations. Journal of chemical theory and computation, 16(4):2400–2409, 2020.DOI
  41. [43]Andrew Zhao, Andrew Tranter, William M Kirby, Shu Fay Ung, Akimasa Miyake, and Peter J Love. Measurement reduction in variational quantum algorithms. Physical Review A, 101(6):062322, 2020.DOI
  42. [44]Linghua Zhu, Senwei Liang, Chao Yang, and Xiaosong Li. Optimizing shot assignment in variational quantum eigensolver measurement. Journal of Chemical Theory and Computation, 20(6):2390–2403, 2024.arxiv.org/abs/2307.06504

Paper details

vlnc:W_8cyp98pu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

generative AI tools

View statement
Looking for related papers…
Contents