语言模型训练中学习率的混合量子选择:终止准则与交互引导的调度预言机

arXiv cs.LG 论文

摘要

本文提出将量子最小值查找(结合 BBHT 搜索的 Dürr–Høyer 算法)与交互引导的调度预言机相结合,联合选择用于语言模型微调的各层组学习率调度方案,并展示了一个具有 15 亿参数的模型的闭环训练,其基准测试结果与采用统一学习率的基线相当。

arXiv:2611.21701v1 Announce Type: new Abstract: Learning-rate choices strongly affect language-model fine-tuning, but the best joint profile across layer groups is generally unknown. This paper studies quantum minimum finding for layer-group learning-rate selection within a hybrid classical-quantum training framework. It combines Durr-Hoyer minimum finding with BBHT search and compares fixed-budget and confidence-based stopping policies using a lookup oracle built from measured profile scores at fixed training checkpoints. The paper then introduces an Interaction-Guided Schedule Oracle that scores and marks joint schedules in superposition without a precomputed score table. Classical one-step loss measurements supply the main, pairwise and three-way interaction effects between layer groups. The circuit adds the effects for each schedule, compares the resulting score with the current baseline score, and marks improving schedules for amplitude amplification. Measured candidates are checked classically before acceptance. Uniform-rate training supplies the initial baseline; using schedules from classical adaptive methods as baselines for joint refinement is left for future work. In closed-loop fine-tuning of a 1.5-billion-parameter language model, each selected schedule drives the next 200 classical training steps. Benchmark accuracies lie within the range observed across five uniform-rate seeds. The results give a working example of how quantum superposition can support adaptive learning-rate selection during training; the remaining quantum limitations are discussed in the paper.
查看原文
查看缓存全文

缓存时间: 2026/11/19 03:34

# Hybrid Quantum Selection of Learning Ratesin Language-Model Training:Termination Criteria and anInteraction-Guided Schedule Oracle
Source: [https://arxiv.org/html/2611.21701](https://arxiv.org/html/2611.21701)
###### Abstract

Learning\-rate choices strongly affect language\-model fine\-tuning, and prior work shows that layer\-wise learning rates can improve training; however, the best layer\-group learning\-rate profile is generally unknown\. This paper studies quantum minimum finding as a layer\-group learning\-rate selector for a future coherent quantum training architecture\. The search combines Dürr–Høyer minimum finding with BBHT search and compares two classes of stopping policies: fixed\-budget and confidence\-based termination\. The stopping policies are first compared on a lookup oracle built from measured profile scores at fixed training checkpoints, which reproduces the required search behavior\. The paper then introduces an Interaction\-Guided Schedule Oracle, which needs no score table and works in two steps inside one simulated circuit\. First, it scores all joint schedules at once in superposition: classical one\-step loss measurements give the main effect of each layer group and the pairwise and three\-way interaction effects between groups, and the circuit adds up the effects that apply to each schedule\. Second, it compares every score with that of a baseline schedule and marks the schedules that beat it, so that amplitude amplification returns an improving candidate; after each round of Grover iterations, the measured candidate is confirmed classically to score better than the baseline\. The baseline in this paper is uniform\-rate training; a schedule proposed by any classical adaptive method could take its place, with the quantum search looking for joint refinements beyond it, a direction left for future research\. In closed\-loop fine\-tuning of a 1\.5\-billion\-parameter language model, each selected schedule drives the next 200 training steps, and benchmark accuracies lie within the range observed across five uniform\-rate seeds\. The results give a working example of how quantum superposition can be brought into adaptive learning\-rate selection during training; the limitations that remain on the quantum side are discussed in the paper\.

## 1Introduction

Layer\-wise learning rates can improve language\-model training\. Discriminative fine\-tuning assigns different rates across layers\([Howard & Ruder, 2018](https://arxiv.org/html/2611.21701#bib.bib19)\), layer\-wise adaptive rate scaling does the same for deep networks\([You et al\., 2017](https://arxiv.org/html/2611.21701#bib.bib31)\), and recent work derives layer\-dependent rates from model statistics or training dynamics\([He et al\., 2026](https://arxiv.org/html/2611.21701#bib.bib17);[Pang et al\., 2026](https://arxiv.org/html/2611.21701#bib.bib24)\); separate rates can also improve the two factors of a low\-rank adapter\([Hayou et al\., 2024](https://arxiv.org/html/2611.21701#bib.bib16)\)\. These methods determine rates individually or by a prescribed rule, rather than jointly selecting one profile from all admissible layer\-group combinations\. The number of joint profiles grows quickly with the number of groups and candidate rates: six groups with eight admissible rates already define86=262,1448^\{6\}=262\{,\}144possible profiles\. This combinatorial growth motivates the quantum\-search formulation developed here\.

This paper studies quantum minimum finding for joint learning\-rate selection acrossGGlayer groups, where each group chooses fromKKadmissible rates and one complete profile specifies the rates applied to the model\. The quantum search has one task: identify a low\-loss learning\-rate profile from these combinations, after which the selected rates are used for the next block of classical training\. In the first part of the paper, this process is simulated at fixed training checkpoints: candidate profiles are evaluated by shorth=25h=25\-step replays whose held\-out losses fill a score table and provide the search scores\. The checkpoints are spaced every200200training steps in the experimental design; in a practical system, when to invoke the selector could instead be determined by the training schedule or other monitored signals\.

Quantum search enters only at the profile\-selection step\. Grover’s search can identify marked candidates in an unstructured space ofNNpossibilities usingO⁡\(N\)O\(\\sqrt\{N\}\)oracle queries\([Grover, 1996](https://arxiv.org/html/2611.21701#bib.bib14)\), and[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)extend this idea to minimum finding by repeatedly lowering a threshold\. In this paper the search space containsN=KGN=K^\{G\}learning\-rate profiles, each represented by one basis\-state index of a quantum search register that holds all of them in superposition\. For example,N=8N=8candidate profiles require three search qubits, represented by\|000⟩,…,\|111⟩\|000\\rangle,\\ldots,\|111\\rangle\. The quantum search therefore acts on profile indices and uses the profile score only to determine which candidates improve on the current threshold\.

In the score\-table experiments, each search round prepares the register, marks the profiles below the current loss threshold, amplifies them by Grover diffusion and measures; because the measurement is probabilistic, a round may return no improving candidate even when one exists\. The rounds are simulated without noise, and a small hardware experiment verifies the marking behavior\. The marking is carried out by a lookup oracle built from measured profile scores\. Run on candidate spaces ofN=81N=81,128128and256256, these simulations provide the data from which the performance of the termination criteria is determined\.

Quantum minimum finding proceeds through both successful and unsuccessful search rounds\. A successful round finds a better candidate and lowers the current threshold, whereas an unsuccessful round leaves the threshold unchanged\. Because the true minimum is unknown, repeated failures do not by themselves reveal when no better profile remains\. The paper therefore compares termination criteria that decide when the search should stop: each rests on its own assumption and stops the search once the observed rounds meet its condition\. For each criterion, the work spent finding better profiles is counted, and so is the extra effort spent after the last improvement until the search stops\.

The second part of the paper replaces the lookup oracle\. The Interaction\-Guided Schedule Oracle \(Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4)\) borrows a notion familiar in machine learning, main effects and interaction effects\([Tsang et al\., 2018](https://arxiv.org/html/2611.21701#bib.bib28)\), and brings it into superposition\. Classical one\-step loss measurements give the main effect of each layer group and the interaction effects between groups; the oracle then works in two steps inside one simulated circuit\. First, the circuit computes a score for every schedule, one complete set of group rates, in superposition\. Second, it marks the schedules whose scores exceed the current threshold, initially the score of the uniform\-rate baseline\. Grover diffusion then amplifies the marked schedules before measurement\. The measured candidate’s score is checked classically: an improvement replaces the current best schedule and raises the threshold; otherwise, the threshold stays unchanged\. Search rounds continue, with the rounds and Grover iterations counted, until a termination criterion stops the search\. The best schedule found drives the next block of training \(Section[3\.2](https://arxiv.org/html/2611.21701#S3.SS2)\)\.

Machine learning already adapts learning rates during training without exhaustive search: population\-based training discovers a schedule of hyperparameter settings rather than one fixed set for the whole course of training\([Jaderberg et al\., 2017](https://arxiv.org/html/2611.21701#bib.bib21)\), and hypergradient descent updates the learning rate during optimization from the gradient of the update rule with respect to that rate\([Baydin et al\., 2018](https://arxiv.org/html/2611.21701#bib.bib3)\)\. The oracle offers a further refinement step: starting from a baseline schedule, it searches for joint changes across layer groups that achieve a higher interaction score\. This paper demonstrates that process with a uniform\-rate baseline\. More generally, a schedule proposed by a classical adaptive method could serve as the baseline: new one\-step measurements around that schedule would supply the main and interaction effects, and its score would set the initial threshold\. Much as a support vector machine can use different kernels within the same learning framework\([Cortes & Vapnik, 1995](https://arxiv.org/html/2611.21701#bib.bib8)\), the oracle could use effects measured around different baseline schedules while retaining the same quantum scoring and marking structure\. Extending this process to classical adaptive baselines is a direction for future research\. The contributions are a comparison of termination criteria for quantum minimum finding on measured score tables, and a demonstration, with uniform\-rate training as the baseline, in which the Interaction\-Guided Schedule Oracle searches for improved combinations of group rates and the selected rates are applied to successive blocks of training\.

## 2Methodology

### 2\.1From classical training to quantum selection

For scale, the original Transformer was trained for100,000100\{,\}000steps at about0\.40\.4s per step in its base configuration and for300,000300\{,\}000steps at1\.01\.0s per step in its large configuration\([Vaswani et al\., 2017](https://arxiv.org/html/2611.21701#bib.bib29)\)\. LoRA fine\-tuning keeps the same step\-based structure while reducing the number of trainable parameters\([Hu et al\., 2022](https://arxiv.org/html/2611.21701#bib.bib20)\)\. Forntrn\_\{\\mathrm\{tr\}\}optimizer steps with mean timettrt\_\{\\mathrm\{tr\}\}, the classical training time is

TC=ntr​ttr\.T\_\{C\}=n\_\{\\mathrm\{tr\}\}\\,t\_\{\\mathrm\{tr\}\}\.\(1\)
Nonuniform learning rates have been used in several forms: discriminative fine\-tuning assigns layer\-specific rates\([Howard & Ruder, 2018](https://arxiv.org/html/2611.21701#bib.bib19)\); layer\-wise adaptive scaling treats layers separately\([You et al\., 2017](https://arxiv.org/html/2611.21701#bib.bib31)\); model\-derived rates use layer statistics or training dynamics\([He et al\., 2026](https://arxiv.org/html/2611.21701#bib.bib17);[Pang et al\., 2026](https://arxiv.org/html/2611.21701#bib.bib24)\); and LoRA\-specific rates assign different rates to the two low\-rank factors\([Hayou et al\., 2024](https://arxiv.org/html/2611.21701#bib.bib16)\)\. Motivated by these results, this paper groups the trainable layers intoGGgroups and searches only the rateηg\\eta\_\{g\}assigned to each groupgg:

wg←wg−ηgdg,g=1,…,G,w\_\{g\}\\leftarrow w\_\{g\}\-\\eta\_\{g\}\\,d\_\{g\},\\qquad g=1,\\dots,G,\(2\)wherewgw\_\{g\}denotes the trainable parameters of groupgg, anddgd\_\{g\}is the update direction supplied by back\-propagation and the optimizer\.

From Eq\. \([2](https://arxiv.org/html/2611.21701#S2.E2)\), each group chooses one learning rate\. Letℰ\\mathcal\{E\}containKKadmissible rates\. A learning\-rate profilezzselects one member ofℰ\\mathcal\{E\}for each of theGGgroups, so the complete profile space containsN=KGN=K^\{G\}combinations:

ℰ=\{η\(1\),…,η\(K\)\},z=\(η1,…,ηG\)∈ℰG,N=\|ℰG\|=KG\.\\mathcal\{E\}=\\\{\\eta^\{\(1\)\},\\dots,\\eta^\{\(K\)\}\\\},\\qquad z=\(\\eta\_\{1\},\\dots,\\eta\_\{G\}\)\\in\\mathcal\{E\}^\{G\},\\qquad N=\|\\mathcal\{E\}^\{G\}\|=K^\{G\}\.\(3\)Each of theseNNprofiles is one candidate learning\-rate setting for the model\.

The quantum search register encodes the learning\-rate profiles of Eq\. \([3](https://arxiv.org/html/2611.21701#S2.E3)\)\. Let𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}denote the table of profiles that a search covers, one profilezzper row andNNrows in total\. A register ofqsrchq\_\{\\mathrm\{srch\}\}qubits has2qsrch2^\{q\_\{\\mathrm\{srch\}\}\}computational basis states\([Grover, 1996](https://arxiv.org/html/2611.21701#bib.bib14)\); therefore a𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}ofNNrows requiresqsrch=⌈log2⁡N⌉q\_\{\\mathrm\{srch\}\}=\\lceil\\log\_\{2\}N\\rceilsearch qubits\. Each profilezzis represented by its binary row index as the basis state\|z⟩\|z\\rangleof the register, and the search begins from the uniform superposition over theNNencoded profiles:

qsrch=⌈log2⁡N⌉,z⟷\|z⟩,\|ψ0⟩=1N​∑z\|z⟩\.q\_\{\\mathrm\{srch\}\}=\\lceil\\log\_\{2\}N\\rceil,\\qquad z\\;\\longleftrightarrow\\;\|z\\rangle,\\qquad\|\\psi\_\{0\}\\rangle=\\frac\{1\}\{\\sqrt\{N\}\}\\sum\_\{z\}\|z\\rangle\.\(4\)For the complete profile spaceℰG\\mathcal\{E\}^\{G\},𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}containsKGK^\{G\}profiles\. When only a subset is searched,NNis the number of rows included in𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}, and Eq\. \([4](https://arxiv.org/html/2611.21701#S2.E4)\) fixes the width of the search register, so the number of search qubits grows with the logarithm ofNN\(Figure[A2](https://arxiv.org/html/2611.21701#A5.F2)shows the mapping atN=8N=8\)\.

To find a useful profilezzduring training, the search over𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}follows the quantum minimum finding of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9), which applies Grover’s search\([Grover, 1996](https://arxiv.org/html/2611.21701#bib.bib14)\)repeatedly while lowering a threshold\. Each row carries one scoref⁡\(z\)f\(z\), the quantity to be minimized\. Letyybe the current threshold profile andf⁡\(y\)f\(y\)its score,MMthe number of profiles in theNN\-row𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}withf⁡\(z\)<f⁡\(y\)f\(z\)<f\(y\), andjjthe number of Grover iterations applied in one search round\. For example, at the scale of the held\-out losses in the simulations of Section[3](https://arxiv.org/html/2611.21701#S3), if the current threshold hasf⁡\(y\)=2\.70f\(y\)=2\.70and exactly three profiles in𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}have lower scores, thenM=3M=3; the search does not know this value in advance\. A round prepares\|ψ0⟩\|\\psi\_\{0\}\\rangleof Eq\. \([4](https://arxiv.org/html/2611.21701#S2.E4)\), appliesjjGrover iterations, measures once, and tests whether the measured profile lies below the current threshold\. A successful round returns some profilezzwithf⁡\(z\)<f⁡\(y\)f\(z\)<f\(y\), which becomes the new threshold profile\. For fixedNN,MM, andjj, the probability that a round returns a profile below the threshold is

pj​\(M\)=sin2⁡\(\(2​j\+1\)​θ\),θ=arcsin⁡M/N\.p\_\{j\}\(M\)=\\sin^\{2\}\\\!\\big\(\(2j\+1\)\\theta\\big\),\\qquad\\theta=\\arcsin\\sqrt\{M/N\}\.\(5\)
BecauseMMis unknown, the Grover iteration count that Eq\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\) would favor is unknown too, and too many iterations can rotate the state past the marked subspace\. The exponential search of[Boyer et al\. \(1998\)](https://arxiv.org/html/2611.21701#bib.bib4)therefore does not fix a single Grover iteration countjjin advance\. Instead, it starts with a small range of possible iteration counts and widens that range after unsuccessful rounds, so the search can reach deeper without assumingMMbeforehand\. Figure[1](https://arxiv.org/html/2611.21701#S2.F1)shows one round of this search, and a run is a sequence of such rounds, with failures that widen the range and successes that lower the threshold\. LetmBBHTm\_\{\\mathrm\{BBHT\}\}denote the upper scale of this range; in each roundjjis drawn uniformly from\{0,…,⌈mBBHT⌉−1\}\\\{0,\\dots,\\lceil m\_\{\\mathrm\{BBHT\}\}\\rceil\-1\\\}\. After a failed round the range expands, and after a successful threshold update it resets:

j∼𝒰⁡\{0,…,⌈mBBHT⌉−1\},failure:mBBHT←min\(λmBBHT,N\),λ=65,success:mBBHT←1,y←z\.\\begin\{array\}\[\]\{l@\{\\quad\}l\}\\lx@intercol\\hfil j\\sim\\mathcal\{U\}\\\{0,\\dots,\\lceil m\_\{\\mathrm\{BBHT\}\}\\rceil\-1\\\},\\hfil\\lx@intercol\\\\\[3\.0pt\] \\text\{failure:\}&m\_\{\\mathrm\{BBHT\}\}\\leftarrow\\min\(\\lambda m\_\{\\mathrm\{BBHT\}\},\\sqrt\{N\}\),\\qquad\\lambda=\\tfrac\{6\}\{5\},\\\\ \\text\{success:\}&m\_\{\\mathrm\{BBHT\}\}\\leftarrow 1,\\qquad y\\leftarrow z\.\\end\{array\}\(6\)For anyλ∈\(1,4/3\)\\lambda\\in\(1,4/3\),[Boyer et al\. \(1998\)](https://arxiv.org/html/2611.21701#bib.bib4)show that a search with unknownMMrequiresO⁡\(N/M\)O\(\\sqrt\{N/M\}\)Grover iterations in expectation\. Across the successive threshold updates,[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)bound the expected total number of Grover iterationsQQneeded to reach the minimum by

𝔼⁡\[Q\]≤454​N\.\\mathbb\{E\}\[Q\]\\leq\\tfrac\{45\}\{4\}\\sqrt\{N\}\.\(7\)This bound characterizes the expected search effort, but does not by itself specify a practical termination criterion; the criteria used in this paper are introduced in Section[2\.3](https://arxiv.org/html/2611.21701#S2.SS3)\.

Figure 1:Quantum threshold minimum finding: one search round and when to stop\. Each round prepares the search register, chooses an iteration countjj, appliesjjoracle–diffuser iterations, measures one row, and checks whether its score improves the current thresholdf⁡\(y\)f\(y\); a success moves the threshold and resets the rangemBBHTm\_\{\\mathrm\{BBHT\}\}from whichjjis drawn, a failure keeps it and enlarges that range\. At the illustrative sizeN=16N=16withM=2M=2rows below the threshold, a round withj=1j=1measures a marked row with probability78\.1%78\.1\\%by Eq\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\)\.
### 2\.2The lookup oracle and the short\-horizon score

Grover adaptive search uses a threshold oracle to distinguish improving candidates from the remaining search states\([Gilliam et al\., 2021](https://arxiv.org/html/2611.21701#bib.bib12)\)\. For the learning\-rate search, this paper uses the held\-out evaluation loss as that criterion: each profilezzreceives a scoref⁡\(z\)f\(z\), obtained by training from the current checkpoint under that profile\. Given the current threshold profileyy, profiles satisfyingf⁡\(z\)<f⁡\(y\)f\(z\)<f\(y\)are the improving candidates and are phase\-marked by the oracle for subsequent Grover amplification\.

Evaluating this score coherently would require the transformer\-training dynamics and the held\-out loss evaluation to be represented inside the quantum computation, which is beyond the implementation studied here\. The experiments of Section[3\.1](https://arxiv.org/html/2611.21701#S3.SS1)and Appendix[D](https://arxiv.org/html/2611.21701#A4)therefore compute the candidate scores classically and store them in a second table,𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}, with one score corresponding to each row of𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}\. Thus,𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}specifies the profiles being searched, whereas𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}supplies the scores used by the simulated oracle\.

Using𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}, the simulated lookup oracleOyO\_\{y\}applies the required threshold marking

Oy​\|z⟩=\{−\|z⟩,f⁡\(z\)<f⁡\(y\),\|z⟩,f⁡\(z\)≥f⁡\(y\)\.O\_\{y\}\\,\|z\\rangle=\\begin\{cases\}\-\\,\|z\\rangle,&f\(z\)<f\(y\),\\\\ \\phantom\{\-\}\\,\|z\\rangle,&f\(z\)\\geq f\(y\)\.\\end\{cases\}The lookup therefore preserves the marking action required by the search while replacing coherent score evaluation with access to the classically computed scores\. Figure[A3](https://arxiv.org/html/2611.21701#A5.F3)contrasts these two paths: the intended coherent oracle evaluatesf⁡\(z\)f\(z\)on the device, whereas the experiments obtain the same threshold decision from𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}\.

#### Oracle convention\.

The lookup experiments use Dürr–Høyer minimum finding\([Dürr & Høyer, 1996](https://arxiv.org/html/2611.21701#bib.bib9)\)with the search dynamics of Eqs\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\)–\([6](https://arxiv.org/html/2611.21701#S2.E6)\) and Figure[1](https://arxiv.org/html/2611.21701#S2.F1), with candidate losses supplied by𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}\. The quantum search acts on profile indices, while classical replay constructsf⁡\(z\)f\(z\); because the score depends on replay length,𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}is defined at a replay horizonhh\.

#### The replay horizon\.

At checkpointtt,𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}is constructed by replaying each candidate profilezzforhhoptimizer steps from the same saved model and optimizer state, using the same sequence of training batches\. LetLt\+h​\(z\)L\_\{t\+h\}\(z\)denote the resulting held\-out evaluation loss\. The checkpoint\-local score at horizonhhis

ft\(h\)​\(z\):=Lt\+h​\(z\)\.f\_\{t\}^\{\(h\)\}\(z\):=L\_\{t\+h\}\(z\)\.The score has the form of a bilevel hyperparameter objective: an outer loss evaluated on the model that a fixed number of inner optimization steps produces\([Franceschi et al\., 2018](https://arxiv.org/html/2611.21701#bib.bib10)\)\. The inner dynamics are thehhreplayed optimizer steps from the checkpoint state, and the outer loss is the held\-out evaluation loss\.

The horizon was fixed by measuring how well shorter replays preserve the candidate order of a200200\-step replay \(Appendix[A](https://arxiv.org/html/2611.21701#A1)\): at checkpoints200200and600600, measured before any𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}of this paper was built,h=25h=25is the shortest horizon with the highest rank agreement, whereas the one\-step ranking is unstable across checkpoints, in line with the bias of short\-horizon objectives toward small step sizes\([Wu et al\., 2018](https://arxiv.org/html/2611.21701#bib.bib30)\)and the long\-run loss of one\-step gains\([Roulet et al\., 2024](https://arxiv.org/html/2611.21701#bib.bib25)\)\. This paper then applies the held\-out loss afterh=25h=25replayed steps as the score,

f⁡\(z\)=ft\(25\)​\(z\)=Lt\+25​\(z\)\.f\(z\)=f\_\{t\}^\{\(25\)\}\(z\)=L\_\{t\+25\}\(z\)\.\(8\)

#### Selection target\.

The target of a selection at checkpointttis the minimizer of that checkpoint’s own𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}, computed using only measurements up tot\+25t\+25:

zt∗=arg⁡minz∈𝑇𝑎𝑏𝑙𝑒prof​ft\(25\)​\(z\)\.z^\{\*\}\_\{t\}=\\arg\\min\_\{z\\in\\mathit\{Table\}\_\{\\mathrm\{prof\}\}\}f\_\{t\}^\{\(25\)\}\(z\)\.\(9\)A search that returns the profilez^t\\hat\{z\}\_\{t\}instead ofzt∗z^\{\*\}\_\{t\}has the regretf⁡\(z^t\)−f⁡\(zt∗\)≥0f\(\\hat\{z\}\_\{t\}\)\-f\(z^\{\*\}\_\{t\}\)\\geq 0, zero when the minimizer is returned\.

### 2\.3Termination criteria

BecauseMM, the number of profiles better than the current threshold, is unknown, a failed round does not reveal whether a better profile remains\.[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)therefore end the procedure by a global time\-out, and[Boyer et al\. \(1998\)](https://arxiv.org/html/2611.21701#bib.bib4)handle the caseM=0M=0by anO⁡\(N\)O\(\\sqrt\{N\}\)time\-out\. Four termination criteria are compared in this paper, in two families: fixed\-budget criteria, and confidence\-based criteria that read the failed rounds as evidence\. Section[3\.1](https://arxiv.org/html/2611.21701#S3.SS1)compares them on the measured score tables to choose the criterion of the closed\-loop searches\.

#### Fixed\-budget termination\.

During a selection, each round adds itsjjGrover iterations to the running total, and each success opens a new threshold visit; both counts are recorded\. Criterion A is the global time\-out of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9), a limit on the effort of the whole run\. LetQQbe the Grover iterations spent so far andVVthe threshold visits so far, the initial one included, withlg\\lgthe base\-2 logarithm:

A:stop whenQ\+\(VlgN\)∗⁣∗≥22\.5N\+1\.4lg2N,returny\.\\text\{A:\}\\quad\\text\{stop when \}Q\+\\big\(V\\lg N\\big\)^\{\*\*\}\\ \\geq\\ 22\.5\\sqrt\{N\}\+1\.4\\lg^\{2\}N,\\ \\text\{ return \}y\.\(10\)Criterion B instead limits one threshold visit: when the Grover iterations of the current visit reach the budget that the proof of Theorem 3 in[Boyer et al\. \(1998\)](https://arxiv.org/html/2611.21701#bib.bib4)assigns to one exponential search, without a better profile found, the run stops:

B:stop when∑visitj≥12λλ−1N=3Natλ=65,returny\.\\text\{B:\}\\quad\\text\{stop when \}\\textstyle\\sum\_\{\\mathrm\{visit\}\}j\\ \\geq\\ \\tfrac\{1\}\{2\}\\tfrac\{\\lambda\}\{\\lambda\-1\}\\sqrt\{N\}\\;=\\;3\\sqrt\{N\}\\ \\text\{ at \}\\lambda=\\tfrac\{6\}\{5\},\\ \\text\{ return \}y\.\(11\)

#### Derivation note\.

A double asterisk marks a paper\-specific extension of a cited bound; the corresponding derivations are given in Appendix[B](https://arxiv.org/html/2611.21701#A2)\.

#### Confidence\-based termination\.

The second family treats failed rounds as evidence and stops when they bound \(criterion C\) or estimate \(criterion D\) the chance that a better profile remains below a confidence levelδ\\delta\. Both are computed classically from the observed rounds\.

Criterion C counts the consecutive failed rounds of one visit once the rangemBBHTm\_\{\\mathrm\{BBHT\}\}has saturated, where[Boyer et al\. \(1998, Lemma 2\)](https://arxiv.org/html/2611.21701#bib.bib4)guarantee a success probability of at least14\\tfrac\{1\}\{4\}wheneverM≥1M\\geq 1; if a better profile remains,kksuch failures occur with probability at most\(3/4\)k\(3/4\)^\{k\}, and the count restarts at every threshold update:

C:\(3/4\)k≤δ⟹kstop\(δ\)=⌈lnδ/ln\(3/4\)⌉∗⁣∗=25atδ=10−3\.\\text\{C:\}\\quad\(3/4\)^\{k\}\\leq\\delta\\ \\Longrightarrow\\ k\_\{\\mathrm\{stop\}\}\(\\delta\)=\\big\\lceil\\ln\\delta/\\ln\(3/4\)\\big\\rceil^\{\*\*\}\\;=\\;25\\ \\text\{ at \}\\delta=10^\{\-3\}\.\(12\)
Criterion D makes the confidence state\-dependent\. A success updates the possible rank of the threshold, so the next failure is interpreted from that new state rather than counted equally with earlier failures\. The resulting distributionπ⁡\(r\)\\pi\(r\)over the rankrrof the current threshold within𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}, uniform at the start,π0​\(r\)=1/N\\pi\_\{0\}\(r\)=1/N, is updated after every round of a visit from the distributionπu\\pi\_\{u\}held at its start, and the search stops when the probability that a better profile remains falls belowδ\\delta,∑r≥2π⁡\(r\)≤δ\\sum\_\{r\\geq 2\}\\pi\(r\)\\leq\\delta:

D∗⁣∗:failure with​j:π⁡\(r\)→π⁡\(r\)​\[1−pj​\(r−1\)\]/cπ,success:πu\+1​\(r\)=1cπ​∑s\>rπu​\(s\)s−1\.\\text\{D\}^\{\*\*\}\\\!:\\quad\\begin\{array\}\[\]\{ll\}\\text\{failure with \}j:&\\pi\(r\)\\to\\pi\(r\)\\,\[1\-p\_\{j\}\(r\-1\)\]\\,/\\,c\_\{\\pi\},\\\\\[4\.0pt\] \\text\{success:\}&\\pi\_\{u\+1\}\(r\)=\\dfrac\{1\}\{c\_\{\\pi\}\}\\displaystyle\\sum\_\{s\>r\}\\frac\{\\pi\_\{u\}\(s\)\}\{s\-1\}\.\\end\{array\}\(13\)cπc\_\{\\pi\}normalizesπ\\pito total mass one after each update, and the next visit starts fromπu\+1\\pi\_\{u\+1\}\.

#### Rounds and queries\.

The four termination criteria differ mainly in how long the search continues after the improving part of the run\. The total Grover\-iteration count of a selection under criterionx∈\{A,B,C,D\}x\\in\\\{\\mathrm\{A\},\\mathrm\{B\},\\mathrm\{C\},\\mathrm\{D\}\\\}therefore splits into the iterations that improve the threshold and those spent in the terminal visit before stopping:

R\(x\)=\#​rounds=\#​preparations=\#​checks,Q\(x\)=∑j=Qimp\(x\)\+Qstop\(x\)\.R^\{\(x\)\}=\\\#\\text\{rounds\}=\\\#\\text\{preparations\}=\\\#\\text\{checks\},\\qquad Q^\{\(x\)\}=\\textstyle\\sum j=Q\_\{\\mathrm\{imp\}\}^\{\(x\)\}\+Q\_\{\\mathrm\{stop\}\}^\{\(x\)\}\.\(14\)Here the improvement effortQimp\(x\)Q\_\{\\mathrm\{imp\}\}^\{\(x\)\}counts the iterations of visits that find a better threshold and the stopping effortQstop\(x\)Q\_\{\\mathrm\{stop\}\}^\{\(x\)\}those of the final visit before criterionxxstops; each round has one state preparation, one measurement and one classical check\. For criteria A–D, rounds and Grover iterations summarize search effort, while timing estimates and detailed cost accounting are given in Appendix[C](https://arxiv.org/html/2611.21701#A3)\.

### 2\.4Interaction\-Guided Schedule Oracle

After the termination criterion is chosen from the lookup\-oracle data, the Interaction\-Guided Schedule Oracle is applied to the same pipeline without a precomputed score table and works in two steps inside one circuit: it first scores all joint schedules at once in superposition, adding up the main and interaction effects of the layer groups that apply to each schedule, and then compares every score with that of the current threshold schedule and marks the schedules that beat it; Grover amplification increases the chance that the measured schedule is an improving candidate\. Figure[2](https://arxiv.org/html/2611.21701#S2.F2)connects the classical loss measurements with the quantum schedule selection and the subsequent training interval\.

Figure 2:Hybrid loop with the Interaction\-Guided Schedule Oracle: classical measurements construct the interaction score, quantum search selects a schedule, and training applies it\.#### Checkpoint evaluation and the interaction score\.

Following the intuition of main effects and interaction effects, a schedule is assumed to change the loss through the effect of each group alone and through interactions between groups\. Pair and triple interactions are kept in this first model; further terms can be added\. In units of the uniform rate, a schedulez=\(η1,…,ηG\)z=\(\\eta\_\{1\},\\dots,\\eta\_\{G\}\)of Eq\. \([3](https://arxiv.org/html/2611.21701#S2.E3)\) takesηg∈\{0\.5,0\.7,1\.0,1\.4\}\\eta\_\{g\}\\in\\\{0\.5,0\.7,1\.0,1\.4\\\}, and the uniform\-rate baselinezrefz\_\{\\mathrm\{ref\}\}has everyηg=1\.0\\eta\_\{g\}=1\.0\. For comparison, the checkpoint evaluation moves the rates of one, two or three groups away from the baseline and takes one trial optimizer step; the gain is the drop in held\-out loss against the same step at the baseline\. The main effectYt​\(g,η\)Y\_\{t\}\(g;\\eta\), the single\-group gain, is the gain of moving groupggalone to the rateη\\eta, measured at0\.50\.5,0\.70\.7and1\.41\.4\. Pairs and triples are measured only at0\.50\.5and1\.41\.4, a choice that reduces the measurement and computation requirements; their interactionsνt,g​g′\\nu\_\{t,gg^\{\\prime\}\}andνt,g​g′​g′′\\nu\_\{t,gg^\{\\prime\}g^\{\\prime\\prime\}\}are derived from the joint gains by removing the lower\-order effects\. An interaction term with a group at0\.70\.7shares the term derived at0\.50\.5, writtenη¯g=0\.5\\bar\{\\eta\}\_\{g\}=0\.5whenηg=0\.7\\eta\_\{g\}=0\.7andη¯g=ηg\\bar\{\\eta\}\_\{g\}=\\eta\_\{g\}otherwise\. The interaction score of a schedule is

Y^t​\(z\)=∑g=1GYt​\(g,ηg\)⏟single\-group gains\+∑g<g′νt,g​g′​\(η¯g,η¯g′\)⏟pair interactions\+∑g<g′<g′′νt,g​g′​g′′​\(η¯g,η¯g′,η¯g′′\)⏟triple interactions\.\\widehat\{Y\}\_\{t\}\(z\)=\\underbrace\{\\sum\_\{g=1\}^\{G\}Y\_\{t\}\(g;\\eta\_\{g\}\)\}\_\{\\text\{single\-group gains\}\}\+\\underbrace\{\\sum\_\{g<g^\{\\prime\}\}\\nu\_\{t,gg^\{\\prime\}\}\(\\bar\{\\eta\}\_\{g\},\\bar\{\\eta\}\_\{g^\{\\prime\}\}\)\}\_\{\\text\{pair interactions\}\}\+\\underbrace\{\\sum\_\{g<g^\{\\prime\}<g^\{\\prime\\prime\}\}\\nu\_\{t,gg^\{\\prime\}g^\{\\prime\\prime\}\}\(\\bar\{\\eta\}\_\{g\},\\bar\{\\eta\}\_\{g^\{\\prime\}\},\\bar\{\\eta\}\_\{g^\{\\prime\\prime\}\}\)\}\_\{\\text\{triple interactions\}\}\.\(15\)Compiled into the oracle circuit as integer constants, the terms of Eq\. \([15](https://arxiv.org/html/2611.21701#S2.E15)\) let the oracle score all4,0964\{,\}096schedules in superposition\. The interaction scoreY^t​\(z\)\\widehat\{Y\}\_\{t\}\(z\)plays a role analogous to the kernel of a support vector machine: its terms can change while the quantum scoring and marking structure is retained\. Appendix[J](https://arxiv.org/html/2611.21701#A10)gives the evaluation protocol, interaction derivation and integer score\.

#### Quantum search over candidate schedules\.

Each group’s rate takes two qubits, soqsrch=12q\_\{\\mathrm\{srch\}\}=12search qubits hold allN=KG=4,096N=K^\{G\}=4\{,\}096schedules ofK=4K=4rates overG=6G=6groups\. WithΦt​\(z\)\\Phi\_\{t\}\(z\)the integer form ofY^t​\(z\)\\widehat\{Y\}\_\{t\}\(z\)andyythe current threshold schedule, the oracle marks every schedule that scores above the threshold:

Oy\|z⟩=\(−1\)𝟏\[Φt\(z\)\>Φt\(y\)\]\|z⟩\.O\_\{y\}\\,\|z\\rangle=\(\-1\)^\{\\mathbf\{1\}\[\\Phi\_\{t\}\(z\)\>\\Phi\_\{t\}\(y\)\]\}\\,\|z\\rangle\.\(16\)The search starts with the uniform\-rate baselinezrefz\_\{\\mathrm\{ref\}\}, whose score is zero, and retains each improving schedule it samples as the new threshold\. When termination criterion D is satisfied, it returns the best schedule retained,z^t\\hat\{z\}\_\{t\}, which sets the group rates of the next200200training steps, and the next checkpoint repeats the checkpoint evaluation\.

### 2\.5Simulation, hardware, and full\-run validation

The tables ofN=81N=81,128128, and256256profiles, the noise\-free simulation of the search and the trial design are given in Appendix[E](https://arxiv.org/html/2611.21701#A5), and a1616\-profile instance of the checkpoint\-200200table is run on an IBM Heron r2 device to check the action of the oracle and the diffuser \(validation and clocks in Appendix[D](https://arxiv.org/html/2611.21701#A4)\)\. A search returns its threshold row as the selected profile of its checkpoint, the estimate ofzt∗z^\{\*\}\_\{t\}; over thenseln\_\{\\mathrm\{sel\}\}selection instantst1<⋯<tnselt\_\{1\}<\\dots<t\_\{n\_\{\\mathrm\{sel\}\}\}the minimizers form the seriesZ∗=\(zt1∗,…,ztnsel∗\)Z^\{\*\}=\(z^\{\*\}\_\{t\_\{1\}\},\\dots,z^\{\*\}\_\{t\_\{n\_\{\\mathrm\{sel\}\}\}\}\)and the selected profilesZ^\\hat\{Z\}\. Whether a selected series helps the fine\-tuning is checked by training under it: a full runFR⁡\(Z∗,seed\)\\mathrm\{FR\}\(Z^\{\*\},\\text\{seed\}\)trains thentrn\_\{\\mathrm\{tr\}\}steps under the uniform\-rate baselinezrefz\_\{\\mathrm\{ref\}\}until the first selection instant and under each minimizer until the next,

FR⁡\(Z∗,seed\):z⁡\(t\)=\{zref,t<t1,zti∗,ti≤t<ti\+1,i=1,…,nsel,tnsel\+1=ntr,\\mathrm\{FR\}\(Z^\{\*\},\\text\{seed\}\):\\qquad z\(t\)=\\begin\{cases\}z\_\{\\mathrm\{ref\}\},&t<t\_\{1\},\\\\ z^\{\*\}\_\{t\_\{i\}\},&t\_\{i\}\\leq t<t\_\{i\+1\},\\quad i=1,\\dots,n\_\{\\mathrm\{sel\}\},\\end\{cases\}\\qquad t\_\{n\_\{\\mathrm\{sel\}\}\+1\}=n\_\{\\mathrm\{tr\}\},\(17\)and the trained model is compared with the run underzrefz\_\{\\mathrm\{ref\}\}by its held\-out loss; the executed runs are recorded in Appendix[F](https://arxiv.org/html/2611.21701#A6)\. A closed\-loop run of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4)is Eq\. \([17](https://arxiv.org/html/2611.21701#S2.E17)\) withz^ti\\hat\{z\}\_\{t\_\{i\}\}in place ofzti∗z^\{\*\}\_\{t\_\{i\}\}, each search simulated noise\-free\.

Section[3\.3](https://arxiv.org/html/2611.21701#S3.SS3)repeats the closed loop with every search sampled on an IBM superconducting device, using the full circuit of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4); Appendix[K](https://arxiv.org/html/2611.21701#A11)gives the gate limits of the device and the bound they set on the iteration count\.

## 3Results

### 3\.1Selecting a termination criterion from measured score tables

Hardware validation of the lookup oracle and diffuser on a1616\-profile instance is reported in Appendix[D](https://arxiv.org/html/2611.21701#A4)\. The measuredh=25h=25score tables are used to compare the stopping effort and selection quality of criteria A–D and guide the choice of criterion for the subsequent closed\-loop searches\.

Table[1](https://arxiv.org/html/2611.21701#S3.T1)compares the four termination criteria on theN=81N=81tables, pooled over the five checkpoints\. All four spend the same improvement effort and reach comparable selection quality, with a mean regret below the replay drift of order10−410^\{\-4\}\(Appendix[E](https://arxiv.org/html/2611.21701#A5)\)\. Most Grover iterations go into stopping:76%76\\%\(D\) to96%96\\%\(A\) of them come after the last improvement, and criterion A uses about8\.78\.7times the stopping effortQstop\(x\)Q\_\{\\mathrm\{stop\}\}^\{\(x\)\}of criterion D\. Criteria B and D are therefore the practical choices\. B caps the iterations of each visit with a fixed budget that needs onlyNN; D reads the failed rounds of the visit through the classical confidence update and needs the least stopping effort\. This paper selects criterion D for the closed\-loop searches; detailed data are in Appendices[H](https://arxiv.org/html/2611.21701#A8)and[C](https://arxiv.org/html/2611.21701#A3)\.

Table 1:Search effort and selection quality under four termination criteria \(N=81N=81\)\.Criteria B and D stop after fewer search rounds while maintaining comparable selection quality\.5,0005\{,\}000simulated trials per criterion,δ=10−3\\delta=10^\{\-3\}for C and D; full table and setting in Appendix[G](https://arxiv.org/html/2611.21701#A7)\.

### 3\.2Closed\-loop training with a simulated Interaction\-Guided Schedule Oracle

The simulated Interaction\-Guided Schedule Oracle drove two closed\-loop fine\-tunes by selecting joint group\-rate schedules at each of five checkpoints\. Each selected schedule was applied to the next training interval, and both resulting models achieved benchmark accuracies within the range observed across five uniform\-rate seeds\. The two runs supply different interaction coefficients to the same Schedule Oracle structure: one takes the checkpoint evaluation on a single training batch, the other averages it over five batches\. The single\-batch and the averaged evaluation select different schedules \(Appendix[J](https://arxiv.org/html/2611.21701#A10)\); comparing the performance of the two is a direction for future research\. Table[2](https://arxiv.org/html/2611.21701#S3.T2)compares the runs with the uniform\-rate run of the same seed and start, the range over five uniform\-rate seeds and the base model\. On every benchmark the two oracle runs differ by at most0\.50\.5percentage points, the largest difference from the uniform\-rate run is0\.70\.7points \(single\-batch, MMLU\), and the single\-batch SciQ score lies at the lower end of the five\-seed range\. Comparable observed scores from one seed per run do not establish equivalence, better training or a quantum speedup: the searches are noise\-free statevector simulations, and no error bar exists for the loss of a single run\.

Table 2:Held\-out loss and benchmark accuracy of the closed\-loop fine\-tunes, seed22\.All runs start from the same checkpoint at step200200\(held\-out loss2\.6782\.678\)\. Held\-out loss on6464evaluation sequences; zero\-shot accuracy in percent, standard errors about1\.01\.0,1\.51\.5,1\.01\.0and1\.41\.4points\.

### 3\.3Closed\-loop training with the Interaction\-Guided Schedule Oracle on IBM hardware

The loop of Section[3\.2](https://arxiv.org/html/2611.21701#S3.SS2)was repeated with every search sampled on IBM hardware\. The five searches of Table[3](https://arxiv.org/html/2611.21701#S3.T3)ran on three156156\-qubit IBM Heron devices, with one shot per circuit and no simulator sample in any decision; the largest circuit had4,284,2664\{,\}284\{,\}266two\-qubit gates\. As on the score tables of Section[3\.1](https://arxiv.org/html/2611.21701#S3.SS1), most Grover iterations went into stopping:1,2071\{,\}207of the1,3811\{,\}381\(87%87\\%\) came after the last improvement of each search\. All four benchmark accuracies of the trained model lie within the five\-seed range \(Table[2](https://arxiv.org/html/2611.21701#S3.T2)\), under the single\-seed qualification of Section[3\.2](https://arxiv.org/html/2611.21701#S3.SS2)\. Appendices[K](https://arxiv.org/html/2611.21701#A11)and[L](https://arxiv.org/html/2611.21701#A12)give the device limits, the threshold paths and the IBM turnaround\.

Table 3:Hardware\-sampled searches at training checkpoints\.One search per checkpoint\. Returned schedule from the first layer group to the last;Φ\\Phi, final integer interaction score;RR, rounds;QQ, Grover iterations;UU, threshold updates; largestjj, the most Grover iterations in one circuit\.

## 4Summary, Discussion and Limitations

Quantum minimum finding was joined to language\-model fine\-tuning in two parts: a study of termination criteria, and the Interaction\-Guided Schedule Oracle\. On a lookup oracle over measured score tables, four termination criteria select with comparable quality and differ in stopping effort\. The Interaction\-Guided Schedule Oracle scores all joint schedules in superposition without a precomputed score table; its selected schedules drove three closed\-loop fine\-tunes, one sampled on IBM hardware, with benchmark accuracies within the range of uniform\-rate seeds\.111Each closed\-loop fine\-tune used a single seed, and the scores of unmeasured schedules are modeled\. Establishing training benefits and computational advantage requires further evaluation; Appendix[J](https://arxiv.org/html/2611.21701#A10)gives the modeling and simulation details\.Together, these studies demonstrate how classical checkpoint measurements can support quantum selection of joint learning\-rate schedules within a language\-model training loop\.

Three directions follow\. First, the interaction score can be exchanged while the quantum search is kept\. The two simulated runs used single\-batch and five\-batch checkpoint evaluations to supply different interaction coefficients to the same quantum scoring and marking structure \(Appendix[J](https://arxiv.org/html/2611.21701#A10)\)\. As with the kernel of a support vector machine\([Cortes & Vapnik, 1995](https://arxiv.org/html/2611.21701#bib.bib8)\), studies of dynamic learning\-rate control can supply coefficients for other groupings, rate menus, models or training phases\. Measurements can also be added where evaluations disagree, as in active learning\([Cohn et al\., 1996](https://arxiv.org/html/2611.21701#bib.bib7)\)or the reweighting of boosting\([Freund & Schapire, 1997](https://arxiv.org/html/2611.21701#bib.bib11)\)\. The score compiled into the circuit offers an interface for quantum neural\-network research\.

Second, quantum selection can refine a baseline schedulezbasez\_\{\\mathrm\{base\}\}proposed by a classical adaptive method, such as population\-based training\([Jaderberg et al\., 2017](https://arxiv.org/html/2611.21701#bib.bib21)\)or hypergradient descent\([Baydin et al\., 2018](https://arxiv.org/html/2611.21701#bib.bib3)\)\. With the checkpoint evaluation taken aroundzbasez\_\{\\mathrm\{base\}\}, the interaction score keeps the form of Eq\. \([15](https://arxiv.org/html/2611.21701#S2.E15)\):

Y^t​\(z∣zbase\)=∑g=1GYt​\(g,ηg\)⏟single\-group gains\+∑g<g′νt,g​g′​\(ηg,ηg′\)⏟pair interactions\+∑g<g′<g′′νt,g​g′​g′′​\(ηg,ηg′,ηg′′\)⏟triple interactions,\\widehat\{Y\}\_\{t\}\(z\\mid z\_\{\\mathrm\{base\}\}\)=\\underbrace\{\\sum\_\{g=1\}^\{G\}Y\_\{t\}\(g;\\eta\_\{g\}\)\}\_\{\\text\{single\-group gains\}\}\+\\underbrace\{\\sum\_\{g<g^\{\\prime\}\}\\nu\_\{t,gg^\{\\prime\}\}\(\\eta\_\{g\},\\eta\_\{g^\{\\prime\}\}\)\}\_\{\\text\{pair interactions\}\}\+\\underbrace\{\\sum\_\{g<g^\{\\prime\}<g^\{\\prime\\prime\}\}\\nu\_\{t,gg^\{\\prime\}g^\{\\prime\\prime\}\}\(\\eta\_\{g\},\\eta\_\{g^\{\\prime\}\},\\eta\_\{g^\{\\prime\\prime\}\}\)\}\_\{\\text\{triple interactions\}\},\(18\)where every gain and interaction is measured aroundzbasez\_\{\\mathrm\{base\}\}and vanishes when one of its groups stays at its baseline rate\. The baseline scores zero and is the first threshold, so the search looks for joint refinements beyond it\. The present runs usezbase=zrefz\_\{\\mathrm\{base\}\}=z\_\{\\mathrm\{ref\}\}; other baselines will be studied in future research \(Appendix[J](https://arxiv.org/html/2611.21701#A10)\)\.

Third, stopping\-policy design is a route to quantum\-search efficiency\. Most Grover iterations came after the last improvement in both the IBM hardware searches and the lookup\-oracle study\. In that study, criterion A used about8\.78\.7times the stopping effort of criterion D\. Criteria B and D are practical choices, and D is used here\. Deciding when enough search has been done can avoid unnecessary rounds and use quantum resources more effectively\.

This paper presents an early\-stage architecture that integrates quantum learning\-rate selection with classical machine learning\. Any practical benefit will depend on advances in quantum hardware and further evaluation\. Developing the algorithms and interfaces now provides a foundation for future research in machine learning and medical imaging\.

## References

- Amy et al\. \(2017\)Matthew Amy, Olivia Di Matteo, Vlad Gheorghiu, Michele Mosca, Alex Parent, and John Schanck\.Estimating the cost of generic quantum pre\-image attacks on SHA\-2 and SHA\-3\.In*Selected Areas in Cryptography \(SAC 2016\)*, volume 10532 of*LNCS*, pp\. 317–337, 2017\.doi:10\.1007/978\-3\-319\-69453\-5\_18\.
- Babbush et al\. \(2021\)Ryan Babbush, Jarrod R\. McClean, Michael Newman, Craig Gidney, Sergio Boixo, and Hartmut Neven\.Focus beyond quadratic speedups for error\-corrected quantum advantage\.*PRX Quantum*, 2:010103, 2021\.doi:10\.1103/PRXQuantum\.2\.010103\.
- Baydin et al\. \(2018\)Atılım Güneş Baydin, Robert Cornish, David Martínez Rubio, Mark Schmidt, and Frank Wood\.Online learning rate adaptation with hypergradient descent\.In*International Conference on Learning Representations \(ICLR\)*, 2018\.URL[https://openreview\.net/forum?id=BkrsAzWAb](https://openreview.net/forum?id=BkrsAzWAb)\.
- Boyer et al\. \(1998\)Michel Boyer, Gilles Brassard, Peter Høyer, and Alain Tapp\.Tight bounds on quantum searching\.*Fortschritte der Physik*, 46\(4\-5\):493–505, 1998\.
- Cade et al\. \(2023\)Chris Cade, Marten Folkertsma, Ido Niesen, and Jordi Weggemans\.Quantifying Grover speed\-ups beyond asymptotic analysis\.*Quantum*, 7:1133, 2023\.doi:10\.22331/q\-2023\-10\-10\-1133\.
- Chen et al\. \(2024\)Jwo\-Sy Chen, Erik Nielsen, Matthew Ebert, Volkan Inlek, Kenneth Wright, Vandiver Chaplin, Andrii Maksymov, Eduardo Páez, Amrit Poudel, Peter Maunz, and John Gamble\.Benchmarking a trapped\-ion quantum computer with 30 qubits\.*Quantum*, 8:1516, 2024\.doi:10\.22331/q\-2024\-11\-07\-1516\.
- Cohn et al\. \(1996\)David A\. Cohn, Zoubin Ghahramani, and Michael I\. Jordan\.Active learning with statistical models\.*Journal of Artificial Intelligence Research*, 4:129–145, 1996\.doi:10\.1613/jair\.295\.URL[https://www\.jair\.org/index\.php/jair/article/view/10158](https://www.jair.org/index.php/jair/article/view/10158)\.
- Cortes & Vapnik \(1995\)Corinna Cortes and Vladimir Vapnik\.Support\-vector networks\.*Machine Learning*, 20\(3\):273–297, 1995\.doi:10\.1007/BF00994018\.
- Dürr & Høyer \(1996\)Christoph Dürr and Peter Høyer\.A quantum algorithm for finding the minimum\.*arXiv preprint quant\-ph/9607014*, 1996\.URL[https://arxiv\.org/abs/quant\-ph/9607014](https://arxiv.org/abs/quant-ph/9607014)\.
- Franceschi et al\. \(2018\)Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil\.Bilevel programming for hyperparameter optimization and meta\-learning\.In*Proceedings of the 35th International Conference on Machine Learning \(ICML 2018\)*, volume 80 of*Proceedings of Machine Learning Research*, pp\. 1568–1577, 2018\.URL[https://proceedings\.mlr\.press/v80/franceschi18a\.html](https://proceedings.mlr.press/v80/franceschi18a.html)\.
- Freund & Schapire \(1997\)Yoav Freund and Robert E\. Schapire\.A decision\-theoretic generalization of on\-line learning and an application to boosting\.*Journal of Computer and System Sciences*, 55\(1\):119–139, 1997\.
- Gilliam et al\. \(2021\)Austin Gilliam, Stefan Woerner, and Constantin Gonciulea\.Grover adaptive search for constrained polynomial binary optimization\.*Quantum*, 5:428, 2021\.doi:10\.22331/q\-2021\-04\-08\-428\.URL[https://quantum\-journal\.org/papers/q\-2021\-04\-08\-428/](https://quantum-journal.org/papers/q-2021-04-08-428/)\.
- Google Quantum AI and Collaborators \(2025\)Google Quantum AI and Collaborators\.Quantum error correction below the surface code threshold\.*Nature*, 638:920–926, 2025\.doi:10\.1038/s41586\-024\-08449\-y\.
- Grover \(1996\)Lov K\. Grover\.A fast quantum mechanical algorithm for database search\.In*Proceedings of the 28th Annual ACM Symposium on Theory of Computing \(STOC\)*, pp\. 212–219, 1996\.doi:10\.1145/237814\.237866\.
- Gundlach et al\. \(2026\)Hans Gundlach, Hrvoje Kukina, Jayson Lynch, and Neil Thompson\.Position: Quantum deep learning still needs a quantum leap\.In*Proceedings of the 43rd International Conference on Machine Learning \(ICML\)*, volume 306 of*Proceedings of Machine Learning Research*, pp\. 170357–170374, 2026\.URL[https://proceedings\.mlr\.press/v306/gundlach26a\.html](https://proceedings.mlr.press/v306/gundlach26a.html)\.
- Hayou et al\. \(2024\)Soufiane Hayou, Nikhil Ghosh, and Bin Yu\.LoRA\+: Efficient low rank adaptation of large models\.In*Proceedings of the 41st International Conference on Machine Learning \(ICML\)*, volume 235 of*Proceedings of Machine Learning Research*, pp\. 17783–17806, 2024\.URL[https://proceedings\.mlr\.press/v235/hayou24a\.html](https://proceedings.mlr.press/v235/hayou24a.html)\.
- He et al\. \(2026\)Di He, Songjun Tu, Keyu Wang, Lu Yin, and Shiwei Liu\.One LR doesn’t fit all: Heavy\-tail guided layerwise learning rates for LLMs\.In*Proceedings of the 43rd International Conference on Machine Learning \(ICML\)*, volume 306 of*Proceedings of Machine Learning Research*, pp\. 41946–41960, 2026\.URL[https://proceedings\.mlr\.press/v306/he26ak\.html](https://proceedings.mlr.press/v306/he26ak.html)\.
- Hoefler et al\. \(2023\)Torsten Hoefler, Thomas Häner, and Matthias Troyer\.Disentangling hype from practicality: On realistically achieving quantum advantage\.*Communications of the ACM*, 66\(5\):82–87, 2023\.doi:10\.1145/3571725\.
- Howard & Ruder \(2018\)Jeremy Howard and Sebastian Ruder\.Universal language model fine\-tuning for text classification\.In*Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pp\. 328–339, 2018\.doi:10\.18653/v1/P18\-1031\.URL[https://aclanthology\.org/P18\-1031/](https://aclanthology.org/P18-1031/)\.
- Hu et al\. \(2022\)Edward J\. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen\-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen\.LoRA: Low\-rank adaptation of large language models\.In*International Conference on Learning Representations \(ICLR\)*, 2022\.URL[https://openreview\.net/forum?id=nZeVKeeFYf9](https://openreview.net/forum?id=nZeVKeeFYf9)\.
- Jaderberg et al\. \(2017\)Max Jaderberg, Valentin Dalibard, Simon Osindero, Wojciech M\. Czarnecki, Jeff Donahue, Ali Razavi, Oriol Vinyals, Tim Green, Iain Dunning, Karen Simonyan, Chrisantha Fernando, and Koray Kavukcuoglu\.Population based training of neural networks\.*arXiv preprint arXiv:1711\.09846*, 2017\.URL[https://arxiv\.org/abs/1711\.09846](https://arxiv.org/abs/1711.09846)\.
- Jeffery et al\. \(2013\)Stacey Jeffery, Robin Kothari, and Frédéric Magniez\.Nested quantum walks with quantum data structures\.In*Proceedings of the 24th Annual ACM\-SIAM Symposium on Discrete Algorithms \(SODA\)*, pp\. 1474–1485, 2013\.doi:10\.1137/1\.9781611973105\.106\.
- Jin et al\. \(2021\)Yuchen Jin, Tianyi Zhou, Liangyu Zhao, Yibo Zhu, Chuanxiong Guo, Marco Canini, and Arvind Krishnamurthy\.AutoLRS: Automatic learning\-rate schedule by Bayesian optimization on the fly\.In*International Conference on Learning Representations \(ICLR\)*, 2021\.URL[https://openreview\.net/forum?id=SlrqM9\_lyju](https://openreview.net/forum?id=SlrqM9_lyju)\.
- Pang et al\. \(2026\)Tianyu Pang, Vignesh Kothapalli, Shenyang Deng, Haohui Wang, Dawei Zhou, and Yaoqing Yang\.Balancing learning rates across layers: Exact two\-step dynamics and optimal scaling in linear neural networks\.In*Proceedings of the 43rd International Conference on Machine Learning \(ICML\)*, volume 306 of*Proceedings of Machine Learning Research*, pp\. 96356–96432, 2026\.URL[https://proceedings\.mlr\.press/v306/pang26d\.html](https://proceedings.mlr.press/v306/pang26d.html)\.
- Roulet et al\. \(2024\)Vincent Roulet, Atish Agarwala, Jean\-Bastien Grill, Grzegorz Swirszcz, Mathieu Blondel, and Fabian Pedregosa\.Stepping on the edge: Curvature aware learning rate tuners\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, volume 37, pp\. 47708–47740, 2024\.doi:10\.52202/079017\-1511\.URL[https://proceedings\.neurips\.cc/paper\_files/paper/2024/hash/555479a201da27c97aaeed842d16ca49\-Abstract\-Conference\.html](https://proceedings.neurips.cc/paper_files/paper/2024/hash/555479a201da27c97aaeed842d16ca49-Abstract-Conference.html)\.
- Sarah D\. \(2024\)Sarah D\.On the practical cost of Grover for AES key recovery\.Presentation, 5th NIST PQC Standardization Conference, Rockville, MD, 10–12 April 2024, 2024\.URL[https://csrc\.nist\.gov/Presentations/2024/practical\-cost\-of\-grover\-for\-aes\-key\-recovery](https://csrc.nist.gov/Presentations/2024/practical-cost-of-grover-for-aes-key-recovery)\.
- Sundararajan et al\. \(2020\)Mukund Sundararajan, Kedar Dhamdhere, and Ashish Agarwal\.The Shapley Taylor interaction index\.In*Proceedings of the 37th International Conference on Machine Learning \(ICML\)*, volume 119 of*Proceedings of Machine Learning Research*, pp\. 9259–9268, 2020\.URL[https://proceedings\.mlr\.press/v119/sundararajan20a\.html](https://proceedings.mlr.press/v119/sundararajan20a.html)\.
- Tsang et al\. \(2018\)Michael Tsang, Dehua Cheng, and Yan Liu\.Detecting statistical interactions from neural network weights\.In*International Conference on Learning Representations \(ICLR\)*, 2018\.URL[https://openreview\.net/forum?id=ByOfBggRZ](https://openreview.net/forum?id=ByOfBggRZ)\.
- Vaswani et al\. \(2017\)Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N\. Gomez, Lukasz Kaiser, and Illia Polosukhin\.Attention is all you need\.In*Advances in Neural Information Processing Systems \(NIPS\)*, volume 30, 2017\.URL[https://papers\.nips\.cc/paper\_files/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa\-Abstract\.html](https://papers.nips.cc/paper_files/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html)\.
- Wu et al\. \(2018\)Yuhuai Wu, Mengye Ren, Renjie Liao, and Roger B\. Grosse\.Understanding short\-horizon bias in stochastic meta\-optimization\.In*International Conference on Learning Representations \(ICLR\)*, 2018\.
- You et al\. \(2017\)Yang You, Igor Gitman, and Boris Ginsburg\.Large batch training of convolutional networks\.*arXiv preprint arXiv:1708\.03888*, 2017\.URL[https://arxiv\.org/abs/1708\.03888](https://arxiv.org/abs/1708.03888)\.

## Appendix AProtocols of the replay\-horizon study

To determine the replay horizon, the study measures how well a shorter replay preserves the candidate order obtained after a longer one\. Lethmaxh\_\{\\max\}be the longest horizon replayed,𝒜\\mathcal\{A\}a measured subset of the profiles of𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}, andρS\\rho\_\{\\mathrm\{S\}\}the Spearman rank correlation\. At checkpointtt, the rank agreement of horizonhhwithhmaxh\_\{\\max\}is

τt,𝒜​\(h\)=ρS​\(\(ft\(h\)​\(z\)\)z∈𝒜,\(ft\(hmax\)​\(z\)\)z∈𝒜\),hmax=200\.\\tau\_\{t,\\mathcal\{A\}\}\(h\)=\\rho\_\{\\mathrm\{S\}\}\\Big\(\\big\(f\_\{t\}^\{\(h\)\}\(z\)\\big\)\_\{z\\in\\mathcal\{A\}\},\\;\\big\(f\_\{t\}^\{\(h\_\{\\max\}\)\}\(z\)\\big\)\_\{z\\in\\mathcal\{A\}\}\\Big\),\\qquad h\_\{\\max\}=200\.\(19\)Table[A1](https://arxiv.org/html/2611.21701#A1.T1)reportsτt,𝒜16​\(h\)\\tau\_\{t,\\mathcal\{A\}\_\{16\}\}\(h\)forh∈\{1,10,25,50,100,150\}h\\in\\\{1,10,25,50,100,150\\\}at checkpointst∈\{200,600,1000\}t\\in\\\{200,600,1000\\\}, on the same measured subset𝒜16\\mathcal\{A\}\_\{16\}of1616profiles at every horizon and checkpoint\.

The one\-step ranking is unstable across the three checkpoints, whereas every horizon from1010to150150agrees far more strongly with the200200\-step order\. The literature reports the same disagreement between one step and the long run: short\-horizon meta\-objectives are biased toward small step sizes\([Wu et al\., 2018](https://arxiv.org/html/2611.21701#bib.bib30)\), one\-step gains are overtaken by a constant rate in the long term\([Roulet et al\., 2024](https://arxiv.org/html/2611.21701#bib.bib25)\), and a schedule search that probes each candidate over a short window fixes that window as a design constant\([Jin et al\., 2021](https://arxiv.org/html/2611.21701#bib.bib23)\)\.

At checkpoints200200and600600, measured before any𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}of this paper was built,h=25h=25is the shortest horizon at whichτt,𝒜16​\(h\)\\tau\_\{t,\\mathcal\{A\}\_\{16\}\}\(h\)reaches its maximum; the later replay at checkpoint10001000agrees with the200200\-step order above0\.970\.97at every horizon from1010to150150, and the setting was kept\. Table[A1](https://arxiv.org/html/2611.21701#A1.T1)was measured on one run under the uniform profile, on the same1616candidate profiles at every horizon and checkpoint\. Forh≥10h\\geq 10, each profile was replayed for200200optimizer steps from the saved model and optimizer state of the checkpoint, on the same sequence of training batches, and the held\-out evaluation loss was recorded at the marksh∈\{10,25,50,100,150,200\}h\\in\\\{10,25,50,100,150,200\\\}; the five entries of a row ath≥10h\\geq 10therefore come from one replay per profile, each compared with the200200\-step mark of the same replay\. Theh=1h=1entry follows a different protocol: one optimizer step under each profile, scored on the two held\-out sequences of the scoring batch rather than on the held\-out evaluation set\. Its value enters Table[A1](https://arxiv.org/html/2611.21701#A1.T1)for its rank only; the same one\-step protocol, on a different design of rows, gives the checkpoint evaluation of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4)\.

Table A1:Rank agreementτt,𝒜16​\(h\)\\tau\_\{t,\\mathcal\{A\}\_\{16\}\}\(h\)of Eq\. \([19](https://arxiv.org/html/2611.21701#A1.E19)\) at three checkpointsttof a run under the uniform profile \(one seed\), over the same measured subset𝒜16\\mathcal\{A\}\_\{16\}of1616candidate profiles\.∗Theh=1h=1entry uses a one\-step evaluation protocol; the entries ath≥10h\\geq 10use the replay protocol defined above\. Both protocols are described above\.

## Appendix BCriteria A–D from the cited bounds

#### Criterion A\.

Criterion A, Eq\. \([10](https://arxiv.org/html/2611.21701#S2.E10)\), is the interrupt of the minimum\-finding algorithm of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)written in the two counts that a run of this paper records\. The convention of that paper charges one time step per Grover iteration of the exponential search andlg⁡N\\lg Ntime steps per execution of the initialization stage, which prepares the uniform state over theNNrows and marks the rows below the current threshold; that stage runs once per threshold visit, and the observation and the threshold update are not counted\. A run that has spentQQGrover iterations overVVthreshold visits, the initial one included, has therefore usedQ\+V​lg⁡NQ\+V\\lg Ntime steps under that convention, the left side of Eq\. \([10](https://arxiv.org/html/2611.21701#S2.E10)\)\.

The right side is the interrupt of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9), twice the bound of Lemma 2 of that paper on the expected total time before the threshold holds the minimum,m0=454​N\+710​lg2​Nm\_\{0\}=\\tfrac\{45\}\{4\}\\sqrt\{N\}\+\\tfrac\{7\}\{10\}\\lg^\{2\}N\. The first term bounds the expected number of Grover iterations, Eq\. \([7](https://arxiv.org/html/2611.21701#S2.E7)\)\. The second term is the initialization cost: by Lemma 1 of the same paper the row of rankrris ever chosen as the threshold with probability1/r1/r, so the expected number of threshold updates is∑r=2N1/r=HN−1\\sum\_\{r=2\}^\{N\}1/r=H\_\{N\}\-1, withHNH\_\{N\}theNNth harmonic number, and the expected initialization time is at most\(HN−1\)​lg⁡N≤\(ln⁡N\)​lg⁡N=\(ln⁡2\)​lg2​N≤710​lg2​N\(H\_\{N\}\-1\)\\lg N\\leq\(\\ln N\)\\lg N=\(\\ln 2\)\\lg^\{2\}N\\leq\\tfrac\{7\}\{10\}\\lg^\{2\}N\. Stopping at2​m0=22\.5​N\+1\.4​lg2​N2m\_\{0\}=22\.5\\sqrt\{N\}\+1\.4\\lg^\{2\}Nreturns the minimum with probability at least12\\tfrac\{1\}\{2\}\(Theorem 1 of the same paper\)\.

#### Criterion B\.

Criterion B, Eq\. \([11](https://arxiv.org/html/2611.21701#S2.E11)\), limits one threshold visit\. Within the current visit the iteration countsjjaccumulate across rounds; when the visit has spent the budget that the proof of Theorem 3 in[Boyer et al\. \(1998\)](https://arxiv.org/html/2611.21701#bib.bib4)assigns to one exponential search,12​λλ−1​N\\tfrac\{1\}\{2\}\\tfrac\{\\lambda\}\{\\lambda\-1\}\\sqrt\{N\}, without finding a better profile, the run stops and returns the threshold\. Atλ=65\\lambda=\\tfrac\{6\}\{5\}the budget is3​N3\\sqrt\{N\}, and it restarts at every threshold update\.

#### Criterion C\.

Criterion C, Eq\. \([12](https://arxiv.org/html/2611.21701#S2.E12)\), counts failed rounds after the range of Eq\. \([6](https://arxiv.org/html/2611.21701#S2.E6)\) has reached its capmBBHT=Nm\_\{\\mathrm\{BBHT\}\}=\\sqrt\{N\}, wherejjis drawn uniformly from\{0,…,⌈N⌉−1\}\\\{0,\\dots,\\lceil\\sqrt\{N\}\\rceil\-1\\\}\. ForMMmarked rows andjjuniform on\{0,…,m−1\}\\\{0,\\dots,m\-1\\\},[Boyer et al\. \(1998, Lemma 2\)](https://arxiv.org/html/2611.21701#bib.bib4)give the probability that one round returns a marked row asPm=12−sin⁡\(4​m​θ\)/\(4​m​sin⁡2​θ\)P\_\{m\}=\\tfrac\{1\}\{2\}\-\\sin\(4m\\theta\)/\(4m\\sin 2\\theta\)withsin2⁡θ=M/N\\sin^\{2\}\\theta=M/N, and showPm≥14P\_\{m\}\\geq\\tfrac\{1\}\{4\}wheneverm≥1/sin⁡2​θm\\geq 1/\\sin 2\\theta\. Sincesin⁡2​θ=2​M⁡\(N−M\)/N\\sin 2\\theta=2\\sqrt\{M\(N\-M\)\}/N, the condition atm=Nm=\\sqrt\{N\}reads4​M​\(N−M\)≥N4M\(N\-M\)\\geq N; the productM⁡\(N−M\)M\(N\-M\)is smallest atM=1M=1andM=N−1M=N\-1, where the condition becomes4​\(N−1\)≥N4\(N\-1\)\\geq N, so it holds for every1≤M≤N−11\\leq M\\leq N\-1onceN≥2N\\geq 2, andm=⌈N⌉≥Nm=\\lceil\\sqrt\{N\}\\rceil\\geq\\sqrt\{N\}satisfies it a fortiori\. The threshold row is never marked, soM≤N−1M\\leq N\-1in every visit, and atM=0M=0every round fails\. Within one visit the threshold, and with itMM, is fixed; each round prepares the uniform state afresh, draws its ownjjand measures once, so the saturated rounds of a visit are independent trials with failure probability1−Pm≤341\-P\_\{m\}\\leq\\tfrac\{3\}\{4\}wheneverM≥1M\\geq 1\. The probability thatkkconsecutive saturated rounds all fail while a better row exists is therefore at most\(3/4\)k\(3/4\)^\{k\}, andkstop​\(δ\)=⌈ln⁡δ/ln⁡\(3/4\)⌉k\_\{\\mathrm\{stop\}\}\(\\delta\)=\\lceil\\ln\\delta/\\ln\(3/4\)\\rceilis the smallestkkwith\(3/4\)k≤δ\(3/4\)^\{k\}\\leq\\delta:1717,2525and3333forδ=10−2\\delta=10^\{\-2\},10−310^\{\-3\}and10−410^\{\-4\}\. The bound holds for everyM≥1M\\geq 1and every visit, and it bounds the event that a visit with a better row present ends without finding one; a probability overMMitself requires a prior, which criterion D supplies\. Rounds before saturation do not enter the count, and the count restarts at every threshold update\.

#### Criterion D\.

Criterion D, Eq\. \([13](https://arxiv.org/html/2611.21701#S2.E13)\), keeps a distribution over the rankrrof the current threshold in𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}, rank11being the lowest score, and updates it from the outcome and the iteration countjjof each round\. The first threshold is a row chosen uniformly at random, step 1 of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9), so its rank is uniform,π0​\(r\)=1/N\\pi\_\{0\}\(r\)=1/N\. A threshold of rankrrhas exactlyr−1r\-1rows below it,M=r−1M=r\-1, so a round withjjiterations fails with probability1−pj​\(r−1\)1\-p\_\{j\}\(r\-1\)by Eq\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\); a failed round multipliesπ⁡\(r\)\\pi\(r\)by this likelihood, which is Bayes’ rule for that observation, and the factor equals one atr=1r=1becausepj​\(0\)=0p\_\{j\}\(0\)=0\. Afterjjiterations from the uniform state theMMmarked rows carry one common amplitudesin⁡\(\(2​j\+1\)​θ\)/M\\sin\(\(2j\+1\)\\theta\)/\\sqrt\{M\}\([Boyer et al\., 1998](https://arxiv.org/html/2611.21701#bib.bib4), Eq\. \(3\)\), so a successful round from rankssreturns each of thes−1s\-1better rows with probability1/\(s−1\)1/\(s\-1\)whateverjjwas\([Dürr & Høyer, 1996](https://arxiv.org/html/2611.21701#bib.bib9), proof of Lemma 1\), and the new rank is uniform on\{1,…,s−1\}\\\{1,\\dots,s\-1\\\}\. Summing this transition over the old rankss\>rs\>rgives the success mapπu\+1​\(r\)∝∑s\>rπu​\(s\)/\(s−1\)\\pi\_\{u\+1\}\(r\)\\propto\\sum\_\{s\>r\}\\pi\_\{u\}\(s\)/\(s\-1\), applied to the distributionπu\\pi\_\{u\}held at the start of the visit, soπu\\pi\_\{u\}is the rank distribution of theuu\-th threshold in the descent of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)from the uniform start and depends on nojj\. The failed rounds of the current visit multiplyπu\\pi\_\{u\}in turn, andcπc\_\{\\pi\}restores total mass one after every update; at a successcπ=∑s≥2πu​\(s\)c\_\{\\pi\}=\\sum\_\{s\\geq 2\}\\pi\_\{u\}\(s\), the mass on the ranks from which a success is possible\. The mass on ranksr≥2r\\geq 2is the probability, given the number of threshold updates so far and the failed rounds of the current visit, that at least one row of𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}lies below the current threshold; the run stops when this mass is at mostδ\\deltaand returns the threshold\.

## Appendix CCost accounting from the cited analysis

#### The three clocks\.

Figure[1](https://arxiv.org/html/2611.21701#S2.F1)separates each search round into state preparation and measurement, Grover iterations, and a classical check\. Summed overR\(x\)R^\{\(x\)\}rounds andQ\(x\)Q^\{\(x\)\}Grover iterations, one quantum selection takes

Tsel\(x\)=Q\(x\)​titer\+R\(x\)​tsetup\+R\(x\)​tchk\.T\_\{\\mathrm\{sel\}\}^\{\(x\)\}=Q^\{\(x\)\}\\,t\_\{\\mathrm\{iter\}\}\+R^\{\(x\)\}\\,t\_\{\\mathrm\{setup\}\}\+R^\{\(x\)\}\\,t\_\{\\mathrm\{chk\}\}\.\(20\)Heretitert\_\{\\mathrm\{iter\}\}is the time of one Grover iteration,tsetupt\_\{\\mathrm\{setup\}\}the device time for one state preparation and measurement, andtchkt\_\{\\mathrm\{chk\}\}the time of one classical check\.

#### Training totals\.

With the termination criteria defined, the training cost separates into the ordinary training timeTCT\_\{C\}of Eq\. \([1](https://arxiv.org/html/2611.21701#S2.E1)\) and the cost of thenseln\_\{\\mathrm\{sel\}\}profile selections made during the run\. At each selection, exhaustive classical search evaluates allNNprofiles at a check timetchkt\_\{\\mathrm\{chk\}\}each, costingN​tchkN\\,t\_\{\\mathrm\{chk\}\}, whereas quantum minimum finding incurs a selection timeTsel\(x\)T\_\{\\mathrm\{sel\}\}^\{\(x\)\}that depends on the termination criterionx∈\{A,B,C,D\}x\\in\\\{\\mathrm\{A\},\\mathrm\{B\},\\mathrm\{C\},\\mathrm\{D\}\\\}\. The two total costs are therefore

TCexh=TC\+nsel​N​tchk,TQ\(x\)=TC\+nsel​Tsel\(x\)\.T\_\{C\}^\{\\mathrm\{exh\}\}=T\_\{C\}\+n\_\{\\mathrm\{sel\}\}\\,N\\,t\_\{\\mathrm\{chk\}\},\\qquad T\_\{Q\}^\{\(x\)\}=T\_\{C\}\+n\_\{\\mathrm\{sel\}\}\\,T\_\{\\mathrm\{sel\}\}^\{\(x\)\}\.\(21\)Thus, the simulations compareTCexhT\_\{C\}^\{\\mathrm\{exh\}\}withTQ\(x\)T\_\{Q\}^\{\(x\)\}, with the difference determined by the per\-selection costN​tchkN\\,t\_\{\\mathrm\{chk\}\}versusTsel\(x\)T\_\{\\mathrm\{sel\}\}^\{\(x\)\}\.

#### Cost convention\.

In the simulation, the classical work used to construct𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}is not included in the quantum selection timeTsel\(x\)T\_\{\\mathrm\{sel\}\}^\{\(x\)\}\. The exhaustive classical selection remainsN​tchkN\\,t\_\{\\mathrm\{chk\}\}, Eq\. \([21](https://arxiv.org/html/2611.21701#A3.E21)\), and the two are compared through the ratioS=N​tchk/Tsel\(x\)S=N\\,t\_\{\\mathrm\{chk\}\}/T\_\{\\mathrm\{sel\}\}^\{\(x\)\}, priced with equal unit costs or with hardware\-informed clocks \(below and Appendix[D](https://arxiv.org/html/2611.21701#A4)\)\.

#### Counts\.

Every round of Eq\. \([6](https://arxiv.org/html/2611.21701#S2.E6)\) has one state preparation,jjGrover iterations, and one measurement followed by one classical check, so the number of roundsR\(x\)R^\{\(x\)\}of Eq\. \([14](https://arxiv.org/html/2611.21701#S2.E14)\) equals the number of preparations and the number of checks\. The four criteria share the improvement phase on common random numbers \(Appendix[E](https://arxiv.org/html/2611.21701#A5)\), so the visits that improve the threshold are the same under everyxxin a trial that reaches the minimum andQimp\(x\)Q\_\{\\mathrm\{imp\}\}^\{\(x\)\}is the same count; a trial that a criterion ends before the minimum truncates it\. The number of threshold updatesUUis recorded as a diagnostic and never priced: by Lemma 1 of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)the element of rankrrbecomes the threshold at some point with probability1/r1/r, so𝔼⁡\[U\]=HN−1\\mathbb\{E\}\[U\]=H\_\{N\}\-1withHNH\_\{N\}theNNth harmonic number\.

#### The time convention of the cited analysis\.

[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)count one time step per Grover iteration andlg⁡N\\lg Nsteps per initialization of a threshold visit, and leave uncounted the stage in which the measured row is compared with the threshold; Criterion A transcribes that convention into Eq\. \([10](https://arxiv.org/html/2611.21701#S2.E10)\) \(Appendix[B](https://arxiv.org/html/2611.21701#A2)\)\. The accounting of Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\) prices that comparison, the classical check, at onetchkt\_\{\\mathrm\{chk\}\}per round, and absorbs the initialization, since every state preparation is one of theR\(x\)R^\{\(x\)\}rounds\.

#### The two ratios\.

The unit\-cost pricing is the convention of query complexity in which the setup, query and checking costs of a search are counted in one unit\([Boyer et al\., 1998](https://arxiv.org/html/2611.21701#bib.bib4);[Jeffery et al\., 2013](https://arxiv.org/html/2611.21701#bib.bib22)\)\. Attiter=tsetup=tchkt\_\{\\mathrm\{iter\}\}=t\_\{\\mathrm\{setup\}\}=t\_\{\\mathrm\{chk\}\}the right\-hand side of Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\) is\(Q\(x\)\+2​R\(x\)\)​tchk\(Q^\{\(x\)\}\+2R^\{\(x\)\}\)\\,t\_\{\\mathrm\{chk\}\}, so

Sunit=N​tchk\(Q\(x\)\+2​R\(x\)\)​tchk=NQ\(x\)\+2​R\(x\),S\_\{\\mathrm\{unit\}\}=\\frac\{N\\,t\_\{\\mathrm\{chk\}\}\}\{\(Q^\{\(x\)\}\+2R^\{\(x\)\}\)\\,t\_\{\\mathrm\{chk\}\}\}=\\frac\{N\}\{Q^\{\(x\)\}\+2R^\{\(x\)\}\},\(22\)a count ratio that holds for any per\-row scoring time\. Under the hardware\-informed pricing the two quantum terms of Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\) are non\-negative, soTsel\(x\)≥R\(x\)​tchkT\_\{\\mathrm\{sel\}\}^\{\(x\)\}\\geq R^\{\(x\)\}\\,t\_\{\\mathrm\{chk\}\}and

Shw=N​tchkTsel\(x\)≤NR\(x\),S\_\{\\mathrm\{hw\}\}=\\frac\{N\\,t\_\{\\mathrm\{chk\}\}\}\{T\_\{\\mathrm\{sel\}\}^\{\(x\)\}\}\\leq\\frac\{N\}\{R^\{\(x\)\}\},\(23\)with equality in the limitQ\(x\)​titer\+R\(x\)​tsetup≪R\(x\)​tchkQ^\{\(x\)\}\\,t\_\{\\mathrm\{iter\}\}\+R^\{\(x\)\}\\,t\_\{\\mathrm\{setup\}\}\\ll R^\{\(x\)\}\\,t\_\{\\mathrm\{chk\}\}; the termR\(x\)​tchkR^\{\(x\)\}\\,t\_\{\\mathrm\{chk\}\}is the checking cost of[Jeffery et al\. \(2013\)](https://arxiv.org/html/2611.21701#bib.bib22), the cost of testing whether a returned candidate is marked, which no device clock removes\. The hardware\-informed pricing follows the literature on the cost of Grover search:[Cade et al\. \(2023\)](https://arxiv.org/html/2611.21701#bib.bib5)quantify Grover speed\-ups beyond the asymptotic worst\-case analysis;[Babbush et al\. \(2021\)](https://arxiv.org/html/2611.21701#bib.bib2)conclude that a quadratic speedup alone does not give an early fault\-tolerant device a runtime advantage;[Amy et al\. \(2017\)](https://arxiv.org/html/2611.21701#bib.bib1)price a quantum computation in units of classical work; and[Hoefler et al\. \(2023\)](https://arxiv.org/html/2611.21701#bib.bib18)argue that only small\-data problems with super\-quadratic speedups are useful in practice\.[Gundlach et al\. \(2026\)](https://arxiv.org/html/2611.21701#bib.bib15)list the conditions an end\-to\-end claim would need in deep learning\.

#### Scope of the timing model\.

Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\) prices the lookup study, in which one classical check is one replay ofh=25h=25optimizer steps\. A round of the Interaction\-Guided Schedule Oracle of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4)also consists of a preparation and measurement,jjGrover iterations and one classical check, but its check is one evaluation of Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\) for the measured schedule rather than a replay, so the clocks of Appendix[D](https://arxiv.org/html/2611.21701#A4)do not transfer to it\. On hardware the compiled iteration bodies differ in gate count from body to body and from threshold to threshold \(Appendix[K](https://arxiv.org/html/2611.21701#A11)\), sotitert\_\{\\mathrm\{iter\}\}in Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\) can only be an effective average\. The cost of a hardware\-sampled selection further contains the checkpoint evaluation of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4), compilation and upload encoding, queueing, every physical execution, including diagnostic shots and circuits submitted but not used by the sequential search, and the training itself; these are accounted per stage in Appendix[L](https://arxiv.org/html/2611.21701#A12)\. No timing comparison is made for the closed\-loop runs\.

## Appendix DProvenance of the fitted hardware quantities

#### Grover\-iteration decomposition\.

Figure[A1](https://arxiv.org/html/2611.21701#A4.F1)shows the block decomposition of one Grover iteration\. Its scheduled time is

titer\(qsrch\)=tlayer\(q,j\)​\[Ddiff\(qsrch,1\)\+Dmark\(qres,1\)\+2​Dcompute\]\.t\_\{\\mathrm\{iter\}\}^\{\(q\_\{\\mathrm\{srch\}\}\)\}=t\_\{\\mathrm\{layer\}\}^\{\(q,\\,j\)\}\\Big\[D\_\{\\mathrm\{diff\}\}^\{\(q\_\{\\mathrm\{srch\}\},\\,1\)\}\+D\_\{\\mathrm\{mark\}\}^\{\(q\_\{\\mathrm\{res\}\},\\,1\)\}\+2\\,D\_\{\\mathrm\{compute\}\}\\Big\]\.\(24\)The block timings are estimated from circuits transpiled and scheduled against the calibration of an IBM Heron r2 device\. The superscript\(q,j\)\(q,j\)records the register sizeqqand Grover\-chain settingjjused for the corresponding compile or fit, withj=1j=1denoting a single compiled block\. The compilation details and the resulting values follow\. The clocks are anchored by one job on the device: the scheduled duration of a circuit is the preparation with its readout plusjjiteration times,

D⁡\(j\)=t0\(4,3\)\+titer\(4,3\)​j,D\(j\)=t\_\{0\}^\{\(4,3\)\}\+t\_\{\\mathrm\{iter\}\}^\{\(4,3\)\}\\,j,\(25\)fitted for each threshold, and the job’s executed time givestsetupt\_\{\\mathrm\{setup\}\}\.

#### Superscripts and bands\.

Every quantity of Eq\. \([24](https://arxiv.org/html/2611.21701#A4.E24)\) that comes from a compile\-and\-fit carries a superscript\(q,j\)\(q,j\):qqis the register size the circuit was compiled at \(qsrchq\_\{\\mathrm\{srch\}\}for the search register,qresq\_\{\\mathrm\{res\}\}for the resolution register\) andjjthe largest chain length of the fit, a straight line over1,…,j1,\\dots,j;j=1j=1means one block compiled and its depth read directly\. A quantity without a superscript is a declared choice or a calibration constant of the device\. The layer time is the ratio of the duration slope to the depth slope of the scheduled circuits, a compiler statistic of the layer mix bounded by the gate durations of the device, so it does not scale with circuit size\. Each fitted depth and the layer time carry a band, the minimum and maximum over three transpiler seeds and two calibration snapshots around the mean\. The setup clock is the executed wall\-clock of one round of the job per shot less its iteration content, so it contains the preparation circuit, the readout and the device’s inter\-shot reset delay\. Compiling the diffuser and the comparator as one circuit is the cross\-check of the model form; its depth exceeds the sum of the two blocks by the routing the compiler adds\. The break\-even of the estimate is the depthDcomputeD\_\{\\mathrm\{compute\}\}at which the share of the two quantum terms of Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\) inTsel\(x\)T\_\{\\mathrm\{sel\}\}^\{\(x\)\}reaches a given value; below it,ShwS\_\{\\mathrm\{hw\}\}of Eq\. \([23](https://arxiv.org/html/2611.21701#A3.E23)\) sits at its floor\.

#### The hardware job\.

The job ran onibm\_marrakesh: seven circuits \(the prepared state alone, andj=1,2,3j=1,2,3iterations for two thresholds that mark 4 and 1 of the 16 rows\),4,0964\{,\}096shots each, transpiled at optimization level 3 and scheduled against the device calibration of the day \(median two\-qubit gate 68 ns, single\-qubit gate 36 ns, readout2\.68​μ2\.68~\\mus\)\. Eq\. \([25](https://arxiv.org/html/2611.21701#A4.E25)\) is fitted separately for the two thresholds, since the depth of a lookup oracle depends on which rows it marks; the slope is the iteration time at this size and the interceptt0\(4,3\)t\_\{0\}^\{\(4,3\)\}the preparation with its readout\. The same circuits are scheduled again against a later calibration to check the slopes, and the job’s own execution metric, the QPU time over all shots, gives the executed cost of one round including reset and repetition delay, from whichtsetupt\_\{\\mathrm\{setup\}\}is taken\. These four\-qubit circuits use a lookup oracle, a diagonal phase over the 16 stored scores, whose depth grows withNN; the job validates the layer time, the diffuser depth atqsrch=4q\_\{\\mathrm\{srch\}\}=4and the additivity of iterations, and says nothing aboutDcomputeD\_\{\\mathrm\{compute\}\}\.

#### Validation of the oracle action\.

Run on the device and in the noise\-free simulation, the oracle and the diffuser of these circuits act as designed on a 16\-profile instance of the checkpoint\-200 table\. With four profiles marked, one iteration raises the probability of a marked profile from0\.250\.25to0\.850\.85on the device, against1\.001\.00in the simulation\. With the minimizer alone marked, three iterations raise its probability from0\.070\.07to0\.630\.63on the device, against0\.960\.96in the simulation; the deeper the circuit, the wider the gap, so depth is the quantity a future circuit has to keep in hand\.

Figure A1:Decomposition of one Grover iteration into score computation, threshold marking, score uncomputation, and diffusion\. Dashed score blocks denote the declared compute component\.
#### The clocks as measured\.

Table[A2](https://arxiv.org/html/2611.21701#A4.T2)summarizes the three clocks of Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\)\. The layer time of Eq\. \([24](https://arxiv.org/html/2611.21701#A4.E24)\) istlayer\(4\.\.9,3\)=33t\_\{\\mathrm\{layer\}\}^\{\(4\.\.9,3\)\}=33ns, the median of3434schedules fromqsrch=4q\_\{\\mathrm\{srch\}\}=4to99, stable at33±433\\pm 4ns across register sizes and constant injj; the two oracle families of the executed job give31\.831\.8and34\.534\.5ns\. The two known depths at the design size areDdiff\(18,1\)=730D\_\{\\mathrm\{diff\}\}^\{\(18,1\)\}=730layers \(706706–747747\) andDmark\(16,1\)=942D\_\{\\mathrm\{mark\}\}^\{\(16,1\)\}=942layers \(896896–969969\), a known part of1,6721\{,\}672layers, so

titer\(18\)=55​μ​s​\(53​–​57\)\+66​ns×Dcompute\.t\_\{\\mathrm\{iter\}\}^\{\(18\)\}=55~\\mu\\mathrm\{s\}\\ \(53\\text\{\-\-\}57\)\+66~\\mathrm\{ns\}\\times D\_\{\\mathrm\{compute\}\}\.Compiling the two blocks as one circuit gives2,0572\{,\}057layers and68​μ68~\\mus, about20%20\\%of routing depth above the model form, which remains the printed number\. The resolution register follows from a loss range of44and a seed\-to\-seed standard deviation of0\.00170\.0017along the reference run,log2⁡\(4/0\.0017\)≈12\\log\_\{2\}\(4/0\.0017\)\\approx 12bits, andqres=16q\_\{\\mathrm\{res\}\}=16is the margin above it\. The executed job gives the slopes of Eq\. \([25](https://arxiv.org/html/2611.21701#A4.E25)\) astiter\(4,3\)=2\.56​μt\_\{\\mathrm\{iter\}\}^\{\(4,3\)\}=2\.56~\\mus for the threshold that marks 4 of the 16 rows and5\.16​μ5\.16~\\mus for the one that marks 1, with interceptst0\(4,3\)=2\.7t\_\{0\}^\{\(4,3\)\}=2\.7to3\.1​μ3\.1~\\mus dominated by readout; scheduled again against a later calibration, the slopes reproduce within about5%5\\%\. The setup clock istsetup\(4,3\)=270​μt\_\{\\mathrm\{setup\}\}^\{\(4,3\)\}=270~\\mus: the job’s28,67228\{,\}672shots took7\.927\.92s of QPU time,276\.2​μ276\.2~\\mus per shot, less the6\.6​μ6\.6~\\mus of iteration content pooled over its seven circuits; the preparation circuit’s own share ist0\(4,3\)t\_\{0\}^\{\(4,3\)\}and the rest is the device’s inter\-shot reset delay, a backend setting taken as independent of the circuit\. One classical check coststchk=355\.5t\_\{\\mathrm\{chk\}\}=355\.5s, the median of389389timed replays \(10th to 90th percentile354\.4354\.4to357\.0357\.0s\); a later campaign of the same replay gave336\.6336\.6s, and the slower value is used\. One replay restores the checkpoint with its optimizer state and then runsh​ttr=25×12\.4​s=310h\\,t\_\{\\mathrm\{tr\}\}=25\\times 12\.4~\\mathrm\{s\}=310s of optimizer steps\. With these clocks and the counts atN=218N=2^\{18\}of Appendix[H](https://arxiv.org/html/2611.21701#A8), the classical checks of one selection takeR\(D\)​tchk=70,000R^\{\(\\mathrm\{D\}\)\}\\,t\_\{\\mathrm\{chk\}\}=70\{,\}000s,19\.519\.5h, and the two quantum terms of Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\) total less than0\.40\.4s whenDcompute=0D\_\{\\mathrm\{compute\}\}=0\. The share of the two terms reaches1%1\\%only ifDcomputeD\_\{\\mathrm\{compute\}\}exceeds about10610^\{6\}layers at the Heron clock, or about10410^\{4\}layers at a hypothetical fault\-tolerant layer of7\.7​μ7\.7~\\mus, seven cycles of a distance\-7 surface code at the1\.1​μ1\.1~\\mus cycle of[Google Quantum AI and Collaborators \(2025\)](https://arxiv.org/html/2611.21701#bib.bib13); below that,ShwS\_\{\\mathrm\{hw\}\}of Eq\. \([23](https://arxiv.org/html/2611.21701#A3.E23)\) sits at its floorN/R\(x\)N/R^\{\(x\)\}\. At the measured sizesN≤256N\\leq 256the two quantum terms total less than0\.10\.1s againstR\(x\)​tchkR^\{\(x\)\}\\,t\_\{\\mathrm\{chk\}\}of2\.82\.8to4\.14\.1h, so the choice of clock changes no measured row; it matters only for the continuation to largeNN\.

Table A2:The three clocks of one selection, Eq\. \([20](https://arxiv.org/html/2611.21701#A3.E20)\): the value as printed, how it was obtained, its source, and a literature comparison\.titert\_\{\\mathrm\{iter\}\}follows Eq\. \([24](https://arxiv.org/html/2611.21701#A4.E24)\) withDcomputeD\_\{\\mathrm\{compute\}\}declared;tsetupt\_\{\\mathrm\{setup\}\}is taken from the executed job described above;tchkt\_\{\\mathrm\{chk\}\}is a GPU clock\. The compile\-and\-fit quantities behind the first row are listed in the provenance table above\.

## Appendix EExperimental and implementation details

Figure A2:𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}mapped to the quantum search register, at an illustrative size \(K=2K=2admissible ratesηg∈ℰ\\eta\_\{g\}\\in\\mathcal\{E\},G=3G=3groups,N=8N=8rows,qsrch=3q\_\{\\mathrm\{srch\}\}=3qubits\)\. Each row of𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}is one profilezzof Eq\. \([3](https://arxiv.org/html/2611.21701#S2.E3)\), and its row index written in binary is the basis state\|z⟩\|z\\rangleof the register; every row enters the search with the same amplitude, the state\|ψ0⟩\|\\psi\_\{0\}\\rangleof Eq\. \([4](https://arxiv.org/html/2611.21701#S2.E4)\)\. The right\-hand column shows the amplitudes after the oracle has flipped the sign of the selected rows\.Figure A3:Threshold oracle: the proposed coherent circuit and what the experiments simulate\. A quantum search over𝑇𝑎𝑏𝑙𝑒prof\\mathit\{Table\}\_\{\\mathrm\{prof\}\}needs an oracle that flips the sign of every row whose score lies below the threshold,f⁡\(z\)<f⁡\(y\)f\(z\)<f\(y\)\. The proposed circuit evaluates the scoref⁡\(z\)f\(z\)on the device and compares it withf⁡\(y\)f\(y\)there\. The experiments simulate this behavior: the scores are computed classically, one replay ofh=25h=25optimizer steps per row, and stored in𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}, and a lookup applies the same phase flip from the stored values\. The searches of Section[3\.1](https://arxiv.org/html/2611.21701#S3.SS1)and Appendix[D](https://arxiv.org/html/2611.21701#A4)run on this simulated oracle, and the rounds of those searches are the recorded data; the closed\-loop searches of Section[3\.2](https://arxiv.org/html/2611.21701#S3.SS2)use the oracle of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4)\.Figure A4:Construction of the score tables: fixed designs ofN=81N=81,128128and256256profiles from the measured five\-level cube, scored at the five checkpoints of the uniform\-rate run by the 25\-step replay of Eq\. \([8](https://arxiv.org/html/2611.21701#S2.E8)\)\.#### The tables\.

Figure[A4](https://arxiv.org/html/2611.21701#A5.F4)summarizes the experimental setup: the reference run and its five checkpoints, the tables ofN=81N=81,128128and256256rows drawn from the measured cube, and the 25\-step replay that scores every row, sonsel=5n\_\{\\mathrm\{sel\}\}=5\. The model, training, table\-construction and replay details are given below\.

Figure A5:Trial design: the four criteria run on shared seeds at every checkpoint; the cross\-check run supports criterion D\.
#### The search\.

The search runs in a noise\-free statevector simulation; Figure[A5](https://arxiv.org/html/2611.21701#A5.F5)shows the trial design\. The declared\-rule cross\-check run supports criterion D: validated at the three measured sizes \(Appendix[H](https://arxiv.org/html/2611.21701#A8)\), it carries the rule to largerNN\. State preparation, simulation versions, the oracle implementation, the rank bookkeeping and the cross\-check run are given below\.

#### Model and training configuration\.

The model is DeepSeek\-R1\-Distill\-Qwen\-1\.5B \(28 transformer layers\), fine\-tuned with LoRA adapters of rank 16 on every attention and MLP projection, on the peS2o scientific\-text corpus, with AdamW and16,38416\{,\}384tokens per optimizer step on one NVIDIA RTX 5050 GPU \(8 GB\)\. The layers formG=6G=6groups of five \(three in the last\)\. The reference run trains the uniform profileηg=5×10−4\\eta\_\{g\}=5\\times 10^\{\-4\}for allgg, and its checkpoints at stepst∈\{200,400,600,800,1000\}t\\in\\\{200,400,600,800,1000\\\}are the five tables’ origins, sonsel=5n\_\{\\mathrm\{sel\}\}=5\.

#### Profile\-table construction\.

The design table of Sections[2\.1](https://arxiv.org/html/2611.21701#S2.SS1)–[2\.3](https://arxiv.org/html/2611.21701#S2.SS3)and Appendix[D](https://arxiv.org/html/2611.21701#A4)hasK=8K=8admissible rates,N=86=218=262,144N=8^\{6\}=2^\{18\}=262\{,\}144rows andqsrch=18q\_\{\\mathrm\{srch\}\}=18, with a resolution register ofqres=16q\_\{\\mathrm\{res\}\}=16qubits for the score; the measured eta space is theK=5K=5subsetℰ=\{0\.5,0\.7,1\.0,1\.4,2\.0\}×5×10−4\\mathcal\{E\}=\\\{0\.5,\\,0\.7,\\,1\.0,\\,1\.4,\\,2\.0\\\}\\times 5\\times 10^\{\-4\},56=15,6255^\{6\}=15\{,\}625rows, from which tables ofN=81N=81,128128and256256rows are drawn by fixed designs \(anchor profiles, two orthogonal arrays, a random set, and stride grids\); the row lists are part of the data record\. A profile is written by its levels,z=⟨k1​k2​k3​k4​k5​k6⟩z=\\langle k\_\{1\}k\_\{2\}k\_\{3\}k\_\{4\}k\_\{5\}k\_\{6\}\\ranglewith groupggat rateη\(kg\)\\eta^\{\(k\_\{g\}\)\}; the uniform\-rate baselinezrefz\_\{\\mathrm\{ref\}\}is⟨333333⟩\\langle 333333\\rangle\.

#### Replay and checkpoint protocol\.

At each checkpoint every row is scored by Eq\. \([8](https://arxiv.org/html/2611.21701#S2.E8)\):2525optimizer steps replayed from the frozen checkpoint under the row’s profile, with the same batch sequence and a cloned optimizer state for every row\. Replaying the uniform profile itself from the checkpoint at step200200reproduces the logged loss of the reference run exactly at the restart step; over the following4949steps the replay drifts by1\.9×10−41\.9\\times 10^\{\-4\}in loss on average, and the evaluation at step5050of the replay lands within1\.7×10−41\.7\\times 10^\{\-4\}of the logged value\. On theh=25h=25tables of Figure[A4](https://arxiv.org/html/2611.21701#A5.F4), the resolution of the score is the drift of the replay itself, of order10−410^\{\-4\}in loss\. Along the run, the score tables of adjacent checkpoints have a rank agreement of0\.9190\.919on the8181profiles \(the mean Spearman correlation of the four adjacent pairs\), so the candidate order changes slowly as training proceeds\.

#### Quantum simulation configuration\.

The search runs in Qiskit 2\.4\.2 with Qiskit Aer 0\.17\.2 \(statevector method\), noise\-free\. ForN=128N=128and256256the uniform state over the rows isH⊗qsrchH^\{\\otimes q\_\{\\mathrm\{srch\}\}\}on theqsrch=log2⁡Nq\_\{\\mathrm\{srch\}\}=\\log\_\{2\}Nqubits of the search register and the diffuser is the standard reflection; forN=81N=81the uniform state over the 81 valid rows is prepared by a dense Householder unitary with the matching reflection, so no padding state is ever marked or measured\. The oracle is a diagonal phase over𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}, thresholded at the scoref⁡\(y\)f\(y\)of the current threshold row\. Trial seeds are shared across criteria \(common random numbers\), so the four criteria share the visits that improve the threshold and differ only in the terminal visit, a paired comparison\. Ranks are attached after each trial for evaluation only and are never read by the run\. The seed\-stability check of Appendix[I](https://arxiv.org/html/2611.21701#A9)uses four further disjoint seed blocks of1,0001\{,\}000trials\. The declared\-rule cross\-check run replays criterion D, withδ\\deltaandλ\\lambdaunchanged, on ranks alone: the rankrrof the initial threshold is drawn uniformly from11toNN, a round withjjiterations succeeds with the analytic probability of Eq\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\) atM=r−1M=r\-1rows below the threshold, and a success moves the threshold to a rank drawn uniformly below the old one; no score table and no circuit enter\. It is run atN=81N=81,128128and256256against the simulations \(Appendix[H](https://arxiv.org/html/2611.21701#A8)\) and then continued to largerNN\.

## Appendix FRecorded outcome of the executed full runs

Every run of this section is a full run of Eq\. \([17](https://arxiv.org/html/2611.21701#S2.E17)\),ntr=1,221n\_\{\\mathrm\{tr\}\}=1\{,\}221steps under the setup of Section[2\.5](https://arxiv.org/html/2611.21701#S2.SS5), either under a seriesZ∗​\(table\)Z^\{\*\}\(\\text\{table\}\)of Section[2\.5](https://arxiv.org/html/2611.21701#S2.SS5), whose members are the minimizers of one table at the five selection instants, or under one profile held from the first step; the seed fixes the adapter initialization and the batch order, andFR⁡\(z,seed\)\\mathrm\{FR\}\(z,\\text\{seed\}\)denotes the run that holds one profilezzfrom the first step\. The series executed areZ∗​\(table​81\)=\(⟨131111⟩,⟨111111⟩,⟨111111⟩,⟨111111⟩,⟨111111⟩\)Z^\{\*\}\(\\text\{table \}81\)=\(\\langle 131111\\rangle,\\langle 111111\\rangle,\\langle 111111\\rangle,\\langle 111111\\rangle,\\langle 111111\\rangle\),Z∗​\(table​128\)=\(⟨211221⟩,⟨111112⟩,⟨111111⟩,⟨111111⟩,⟨111112⟩\)Z^\{\*\}\(\\text\{table \}128\)=\(\\langle 211221\\rangle,\\langle 111112\\rangle,\\langle 111111\\rangle,\\langle 111111\\rangle,\\langle 111112\\rangle\)andZ∗​\(table​256\)=\(⟨112111⟩,⟨111112⟩,⟨111111⟩,⟨111111⟩,⟨111112⟩\)Z^\{\*\}\(\\text\{table \}256\)=\(\\langle 112111\\rangle,\\langle 111112\\rangle,\\langle 111111\\rangle,\\langle 111111\\rangle,\\langle 111112\\rangle\); every member of a series was also trained as a constant profile\. The series were measured on the checkpoints of one uniform run, so in a series run every member after the first acts on a state it was not selected on: the loop is open, and the tables record an outcome, not a winner\.

Each run is read by its held\-out evaluation lossℓ⁡\(FR,tstep\)\\ell\(\\mathrm\{FR\},t\_\{\\mathrm\{step\}\}\), the mean cross\-entropy over a fixed set of6464held\-out sequences of1,0241\{,\}024tokens, written attstep∈\{0,50,…,1200,1221\}t\_\{\\mathrm\{step\}\}\\in\\\{0,50,\\dots,1200,1221\\\}, and compared at equal steps with the meanℓ¯​\(tstep\)\\bar\{\\ell\}\(t\_\{\\mathrm\{step\}\}\)of the uniform runsFR⁡\(⟨333333⟩,seed\)\\mathrm\{FR\}\(\\langle 333333\\rangle,\\text\{seed\}\)over the reference seeds\. The statistic is the step\-matched excess loss and its mean over the tail window𝒲=\{800,850,…,1200,1221\}\\mathcal\{W\}=\\\{800,850,\\dots,1200,1221\\\}of the last ten evaluations,

Δ⁡\(FR,tstep\)=ℓ⁡\(FR,tstep\)−ℓ¯​\(tstep\),Δ¯​\(FR\)=1\|𝒲\|​∑tstep∈𝒲Δ⁡\(FR,tstep\),\\Delta\(\\mathrm\{FR\},t\_\{\\mathrm\{step\}\}\)=\\ell\(\\mathrm\{FR\},t\_\{\\mathrm\{step\}\}\)\-\\bar\{\\ell\}\(t\_\{\\mathrm\{step\}\}\),\\qquad\\bar\{\\Delta\}\(\\mathrm\{FR\}\)=\\frac\{1\}\{\|\\mathcal\{W\}\|\}\\sum\_\{t\_\{\\mathrm\{step\}\}\\in\\mathcal\{W\}\}\\Delta\(\\mathrm\{FR\},t\_\{\\mathrm\{step\}\}\),together withn−n\_\{\-\}, the number of evaluations in𝒲\\mathcal\{W\}at which the run lies below the reference\. The reference is the mean of the five uniform runs, seeds00to44; each uniform run is scored leave\-one\-out, against the mean of the other four, and the five uniform rows give the seed\-to\-seed scale ofΔ¯\\bar\{\\Delta\}\. The uniform run of seed22is the run whose checkpoints are the origin of the tables\. The second table compares each series run with runs of its own seed,ℓ⁡\(FRa,tstep\)−ℓ⁡\(FRb,tstep\)\\ell\(\\mathrm\{FR\}\_\{a\},t\_\{\\mathrm\{step\}\}\)\-\\ell\(\\mathrm\{FR\}\_\{b\},t\_\{\\mathrm\{step\}\}\)averaged over𝒲\\mathcal\{W\}: against the uniform run of that seed, and against each of its members held from the first step\.

The last column of the first table is a consistency check on public question\-answering sets, run zero\-shot with the EleutherAI evaluation harness on the saved adapter of each run \(SciQ,1,0001\{,\}000questions; PubMedQA,1,0001\{,\}000; ARC\-Easy,2,3762\{,\}376\)\. A run is flagged when its accuracy on any of the three falls below the lowest accuracy of the uniform runs by more than two standard errors of that accuracy\. This is a screening rule fixed before the runs, not a hypothesis test at a stated level: the band minimum is itself the lowest of a few noisy scores, every run is checked on three sets, and the seed\-to\-seed spread of an accuracy is of the same size as its question\-sampling standard error\. A test at a stated level would need of the order of eight seeds per profile on the noisiest set and is left to later work\. No run is flagged\.

## Appendix GTermination criteria per checkpoint

Table[A3](https://arxiv.org/html/2611.21701#A7.T3)repeats Table[1](https://arxiv.org/html/2611.21701#S3.T1)at full width, with the two cost ratios of Appendix[C](https://arxiv.org/html/2611.21701#A3)\.

Table A3:Four termination criteria for checkpoint\-local Dürr–Høyer selection on theh=25h=25tables,N=81N=81, pooled over five checkpoints,1,0001\{,\}000Aer trials per checkpoint and criterion, common random numbers\. RoundsR\(x\)R^\{\(x\)\}count state preparations and classical checks alike \(Eq\. \([14](https://arxiv.org/html/2611.21701#S2.E14)\)\); queries split into the descent workQimp\(x\)Q\_\{\\mathrm\{imp\}\}^\{\(x\)\}, identical across criteria, and the stopping workQstop\(x\)Q\_\{\\mathrm\{stop\}\}^\{\(x\)\}, where the criteria differ\. The two speedups areSunitS\_\{\\mathrm\{unit\}\}of Eq\. \([22](https://arxiv.org/html/2611.21701#A3.E22)\) andShwS\_\{\\mathrm\{hw\}\}of Eq\. \([23](https://arxiv.org/html/2611.21701#A3.E23)\), the latter drawn with the interim clockstiter=1t\_\{\\mathrm\{iter\}\}=1ms,tsetup=40​μt\_\{\\mathrm\{setup\}\}=40~\\mus; the estimate of Eq\. \([24](https://arxiv.org/html/2611.21701#A4.E24)\) is smaller on both and leaves every entry at the floorN/R\(x\)N/R^\{\(x\)\}to two digits\.Per\-checkpoint detail behind Table[1](https://arxiv.org/html/2611.21701#S3.T1): four criteria, five checkpoints, 20 Aer trials each,δ=10−3\\delta=10^\{\-3\}for C and D\. The minimizer column names the row of Eq\. \([9](https://arxiv.org/html/2611.21701#S2.E9)\) at that checkpoint; the threshold\-update count is a diagnostic \(Appendix[C](https://arxiv.org/html/2611.21701#A3)\); the rank percentile is11when the minimum was returned\.

Figure[A6](https://arxiv.org/html/2611.21701#A7.F6)prices the same counts per checkpoint under the two conventions of Appendix[C](https://arxiv.org/html/2611.21701#A3)\. Under the unit\-cost query model no criterion reaches parity at this size:Q\(x\)\+2​R\(x\)Q^\{\(x\)\}\+2R^\{\(x\)\}is at least9090againstN=81N=81, givingSunitS\_\{\\mathrm\{unit\}\}of0\.210\.21,0\.840\.84,0\.350\.35and0\.900\.90for A to D\. The dotted line marks the bound of Eq\. \([7](https://arxiv.org/html/2611.21701#S2.E7)\) in the same units\. Under the hardware\-informed estimate every bar collapses onto its checking\-cost floorR\(x\)​tchkR^\{\(x\)\}\\,t\_\{\\mathrm\{chk\}\}, andShwS\_\{\\mathrm\{hw\}\}is1\.051\.05,2\.772\.77,1\.551\.55and2\.862\.86; the quantum terms are invisible at this scale, as follows from Eq\. \([23](https://arxiv.org/html/2611.21701#A3.E23)\)\. Both readings are conditional on the oracle convention of Appendix[C](https://arxiv.org/html/2611.21701#A3)\. Criterion D carries the scaling study\.

Figure A6:Selection time per checkpoint for the four termination criteria on theh=25h=25tables,N=81N=81, against exhaustive evaluation \(dashed,81×355\.581\\times 355\.5s=8\.00=8\.00h\)\. One round is one state preparation,jjGrover iterations, one measurement and one classical check, so preparations and checks both appear as the round countR\(x\)R^\{\(x\)\}\. \(a\) Unit\-cost query model,titer=tsetup=tchkt\_\{\\mathrm\{iter\}\}=t\_\{\\mathrm\{setup\}\}=t\_\{\\mathrm\{chk\}\}: bars stack the improvement and stopping iterations, the preparations and the checks; the right axis reads the bar in common cost units, and the dotted line is454​N≈101\\tfrac\{45\}\{4\}\\sqrt\{N\}\\approx 101units\. \(b\) Hardware\-informed estimate, drawn with the interim clockstiter=1t\_\{\\mathrm\{iter\}\}=1ms andtsetup=40​μt\_\{\\mathrm\{setup\}\}=40~\\mus and the checks attchkt\_\{\\mathrm\{chk\}\}; the clocks of Eq\. \([24](https://arxiv.org/html/2611.21701#A4.E24)\) are smaller and change no bar: every bar equals its checking\-cost floorR\(x\)​tchkR^\{\(x\)\}\\,t\_\{\\mathrm\{chk\}\}\. All timings are conditional on the lookup\-oracle convention; panel \(b\) is never a measured speedup\.
## Appendix HDeclared\-rule continuation and its agreement with the measured rows

The declared\-rule cross\-check of Section[2\.5](https://arxiv.org/html/2611.21701#S2.SS5)replays criterion D with the analytic probability of Eq\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\), no table and no circuit\. The first table repeats the measured rows at full width; the second continues the rule beyond the measured range; every row of it is a cross\-check of the rule and not a measurement, and the times assumetchk=355\.5t\_\{\\mathrm\{chk\}\}=355\.5s and the interim clocks11ms,40​μ40~\\mus \(Eq\.[24](https://arxiv.org/html/2611.21701#A4.E24)lowers both and moves no entry\)\. The second table is the validation gate at the three measured sizes: the pass band is\|Aer−model\|≤max⁡\(3​s\.e\.,3%\)\|\\text\{Aer\}\-\\text\{model\}\|\\leq\\max\(3\\,\\text\{s\.e\.\},\\,3\\%\)on bothR\(D\)R^\{\(\\mathrm\{D\}\)\}andQ\(D\)Q^\{\(\\mathrm\{D\}\)\}, and all three sizes pass\.

Part A, measured rows \(full width\)

Part B, declared\-rule continuation \(not a measurement\)

Part C, agreement at the measured sizes

Table A4:Selection cost against table size for criterion D \(δ=10−3\\delta=10^\{\-3\}\), measured rows only: ideal Grover circuits simulated on Qiskit Aer over the realh=25h=25scores\. Trials are checkpoints times trials per checkpoint; rounds and queries carry standard errors; the speedups follow Eqs\. \([22](https://arxiv.org/html/2611.21701#A3.E22)\) and \([23](https://arxiv.org/html/2611.21701#A3.E23)\) withtchk=355\.5t\_\{\\mathrm\{chk\}\}=355\.5s and the interim clocks11ms,40​μ40~\\mus \(Eq\. \([24](https://arxiv.org/html/2611.21701#A4.E24)\) lowers both and moves no entry\)\. The full\-width rows \(threshold updates, classical and quantum times\), the declared\-rule continuation beyondN=256N=256and its agreement with these rows are given in this appendix\.Criterion D is followed across the three measured sizes in Figure[A7](https://arxiv.org/html/2611.21701#A8.F7)\. At unit cost, one per iteration and two per round, the speedup over exhaustive evaluation crosses one betweenN=81N=81and128128:SunitS\_\{\\mathrm\{unit\}\}is\(0\.90,1\.13,1\.65\)\(0\.90,\\,1\.13,\\,1\.65\)and the break\-even sits nearN≈100N\\approx 100\. Under the hardware\-informed estimate the classical checks dominate andShwS\_\{\\mathrm\{hw\}\}is\(2\.86,3\.85,6\.14\)\(2\.86,\\,3\.85,\\,6\.14\), the floorN/RN/Rat every size: the quantum terms are negligible, one selection costs itsRRclassical checks, and the gain comes from needing fewer rounds rather than from a faster circuit\. Both ratios exclude the construction of𝑇𝑎𝑏𝑙𝑒score\\mathit\{Table\}\_\{\\mathrm\{score\}\}: with each table charged at one replay per profile, a selection over the lookup oracle takes longer than exhaustive evaluation at every measured size, the ratio of exhaustive to quantum time being0\.740\.74,0\.790\.79and0\.860\.86atN=81N=81,128128and256256, so the lookup route carries no end\-to\-end advantage\.

Figure A7:Runtime speedup versus table sizeNNfor criterion D withδ=10−3\\delta=10^\{\-3\}\. Dashed curves are the declared\-rule cross\-check obtained from the analytic success probability of Eq\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\), and are continued to the design sizeN=262,144N=262\{,\}144, where no corresponding table or circuit experiment is run\. At the measured sizes, the cross\-check agrees with the measured results within0\.5%0\.5\\%\. The construction of the score table is not included\.The declared\-rule cross\-check of Section[2\.5](https://arxiv.org/html/2611.21701#S2.SS5)reproduces the measured rows within0\.5%0\.5\\%in bothR\(D\)R^\{\(\\mathrm\{D\}\)\}andQ\(D\)Q^\{\(\\mathrm\{D\}\)\}at all three sizes, which is the license to draw the same rule beyond the measured range as the dashed curves of Figure[A7](https://arxiv.org/html/2611.21701#A8.F7)\. Continued to the design sizeN=86=262,144N=8^\{6\}=262\{,\}144the rule givesR\(D\)≈197R^\{\(\\mathrm\{D\}\)\}\\approx 197rounds andQ\(D\)≈3,250Q^\{\(\\mathrm\{D\}\)\}\\approx 3\{,\}250iterations, anSunitS\_\{\\mathrm\{unit\}\}of about7272and anShwS\_\{\\mathrm\{hw\}\}of about1,3301\{,\}330\. Along the ladderQ\(D\)Q^\{\(\\mathrm\{D\}\)\}grows asN\\sqrt\{N\}\(Q\(D\)/NQ^\{\(\\mathrm\{D\}\)\}/\\sqrt\{N\}from4\.14\.1to6\.46\.4, inside the11\.2511\.25of Eq\.[7](https://arxiv.org/html/2611.21701#S2.E7)\) andR\(D\)R^\{\(\\mathrm\{D\}\)\}as\(ln⁡N\)2\(\\ln N\)^\{2\}, so the floorN/R\(D\)N/R^\{\(\\mathrm\{D\}\)\}grows almost linearly inNN\. Nothing beyondN=256N=256is a measurement: the continuation assumes distinct scores and a uniform rank of the initial threshold, and, like every selection\-time estimate of the lookup study, it is conditional on the lookup\-oracle convention\.

## Appendix ISeed\-stability check at checkpoint 200

Criterion D at checkpoint 200 forN=81N=81,128128and256256, with the paper’s seed block \(block 0\) and four further disjoint blocks of1,0001\{,\}000trials\. The spread of the five block means is0\.62/0\.94/1\.190\.62/0\.94/1\.19rounds and0\.63/1\.02/1\.500\.63/1\.02/1\.50queries at the three sizes, i\.e\.2\.62\.6to3\.73\.7standard errors, the range expected for five independent blocks; no block stands out\. AtN=256N=256block 0 is the5,0005\{,\}000\-trial file and therefore contains blocks 1–4; blocks 1–4 alone carry the independent statement\.

## Appendix JInteraction\-Guided Schedule Oracle: score construction and closed\-loop runs

#### Checkpoint evaluation\.

At each checkpointtt, the checkpoint evaluation of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4)measures change patterns around the uniform\-rate baselinezrefz\_\{\\mathrm\{ref\}\}\. A change patternγ\\gammamoves one, two or three groups away fromzrefz\_\{\\mathrm\{ref\}\}and keeps the others at1\.01\.0\. For each pattern, the parameters are moved once from the saved model and optimizer state along the optimizer’s update direction for one fixed batch of1616training sequences, the update of each group scaled by its rate, andΛt​\(γ\)\\Lambda\_\{t\}\(\\gamma\)is the loss of the moved model on two fixed held\-out sequences; the baseline takes the same step from the same state\. The gain ofγ\\gammaisYt​\(γ\)=Λt​\(zref\)−Λt​\(γ\)Y\_\{t\}\(\\gamma\)=\\Lambda\_\{t\}\(z\_\{\\mathrm\{ref\}\}\)\-\\Lambda\_\{t\}\(\\gamma\), andYt​\(g,η\)Y\_\{t\}\(g;\\eta\)is the gain of moving groupggalone to the rateη\\eta\. Every term of Eq\. \([15](https://arxiv.org/html/2611.21701#S2.E15)\) vanishes when one of its groups stays at1\.01\.0, so the baseline scores zero\.

#### Interaction coefficients\.

For a change patternγ\\gamma, the interaction term is the discrete derivative of the gain over the moved groups\([Sundararajan et al\., 2020](https://arxiv.org/html/2611.21701#bib.bib27)\),

νt​\(γ\)=∑γ′⊆γ\(−1\)\|γ\|−\|γ′\|​Yt​\(γ′\),\\nu\_\{t\}\(\\gamma\)=\\sum\_\{\\gamma^\{\\prime\}\\subseteq\\gamma\}\(\-1\)^\{\|\\gamma\|\-\|\\gamma^\{\\prime\}\|\}\\,Y\_\{t\}\(\\gamma^\{\\prime\}\),\(26\)whereγ′\\gamma^\{\\prime\}runs over the sub\-patterns ofγ\\gamma, each moved group keeping its rate inγ\\gamma, the empty one with gain zero\. For one group, a pair and a triple, with the rates omitted,

νt,g\\displaystyle\\nu\_\{t,g\}=Yt,g,νt,g​g′=Yt,g​g′−Yt,g−Yt,g′,\\displaystyle=Y\_\{t,g\},\\qquad\\nu\_\{t,gg^\{\\prime\}\}=Y\_\{t,gg^\{\\prime\}\}\-Y\_\{t,g\}\-Y\_\{t,g^\{\\prime\}\},νt,g​g′​g′′\\displaystyle\\nu\_\{t,gg^\{\\prime\}g^\{\\prime\\prime\}\}=Yt,g​g′​g′′−Yt,g​g′−Yt,g​g′′−Yt,g′​g′′\+Yt,g\+Yt,g′\+Yt,g′′,\\displaystyle=Y\_\{t,gg^\{\\prime\}g^\{\\prime\\prime\}\}\-Y\_\{t,gg^\{\\prime\}\}\-Y\_\{t,gg^\{\\prime\\prime\}\}\-Y\_\{t,g^\{\\prime\}g^\{\\prime\\prime\}\}\+Y\_\{t,g\}\+Y\_\{t,g^\{\\prime\}\}\+Y\_\{t,g^\{\\prime\\prime\}\},whereYt,g​g′Y\_\{t,gg^\{\\prime\}\}is the measured gain of movingggandg′g^\{\\prime\}together: a pair or triple interaction is the part of a measured joint gain that its lower\-order terms do not explain, derived after the joint measurement rather than measured on its own\. Summed over the sub\-patterns of a measured pattern, the terms of Eq\. \([15](https://arxiv.org/html/2611.21701#S2.E15)\) therefore return its measured gain before rounding\. For a schedule with0\.70\.7in a pair or triple, or with more than three moved groups,Y^t​\(z\)\\widehat\{Y\}\_\{t\}\(z\)is a modeled gain rather than a measured joint gain\. Taken at the uniform\-rate baseline only, these terms are not Shapley–Taylor indices, whose highest\-order indices average over orderings\.

#### Design and coefficients\.

At each checkpoint the measurement design has249249rows: five repeats of the uniform\-rate baseline, the1212single\-group changes to0\.50\.5and1\.41\.4times the uniform rate,1212single\-group changes to0\.70\.7and2\.02\.0times that rate, the6060pairs and the160160triples of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4)at0\.50\.5and1\.41\.4\. The rows at2\.02\.0enter no coefficient, and the search menu does not contain that rate\. Over the moved rates0\.50\.5,0\.70\.7and1\.41\.4, the three sums of Eq\. \([15](https://arxiv.org/html/2611.21701#S2.E15)\) can draw on6×3=186\\times 3=18single,\(62\)×32=135\\binom\{6\}\{2\}\\times 3^\{2\}=135pair and\(63\)×33=540\\binom\{6\}\{3\}\\times 3^\{3\}=540triple cases,693693in all; these count possible rate\-specific cases, not stored coefficients, of which the design holds the232232single, pair and triple patterns at0\.50\.5and1\.41\.4and at most six corrections at0\.70\.7\. A group at0\.70\.7keeps its own measured single gain, through the correction of Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\), and shares the pair and triple interactions of its lowering to0\.50\.5\. The15×5=7515\\times 5=75pair and20×19=38020\\times 19=380triple cases that contain0\.70\.7have no measured row: sharing the interactions at0\.50\.5avoids these455455joint measurements per checkpoint, which would raise the design from249249to704704rows, and lets each pair or triple coefficient act on the direction of its groups, lowered or raised, rather than on separate rate\-specific terms\. The unitε\\varepsilonis the larger of the spread of the uniform repeats and the declared2×10−42\\times 10^\{\-4\}; the repeats are identical within one measurement, soε=2×10−4\\varepsilon=2\\times 10^\{\-4\}in both runs\. The five\-batch run measures the same rows on five fixed batches of1616training sequences, the first being the single batch, averages the losses row by row and records the spread of each row without using it\. The single\-batch measurements were taken with unseeded dropout and non\-deterministic GPU kernels and do not repeat exactly: a repeat of the same measurement on one checkpoint changes the losses by6\.8×10−46\.8\\times 10^\{\-4\}on average; the five\-batch run used seeded dropout and deterministic kernels\.

#### Integer score\.

Each coefficient is divided byε\\varepsilonbefore it is rounded, which enlarges the small gains before the rounding:at​\(γ\)=round⁡\(νt​\(γ\)/ε\)a\_\{t\}\(\\gamma\)=\\mathrm\{round\}\\big\(\\nu\_\{t\}\(\\gamma\)/\\varepsilon\\big\)for each of the232232patterns, and the0\.70\.7correction of groupggisbt​\(g\)=round⁡\(\(Yt​\(g,0\.7\)−Yt​\(g,0\.5\)\)/ε\)b\_\{t\}\(g\)=\\mathrm\{round\}\\big\(\(Y\_\{t\}\(g;0\.7\)\-Y\_\{t\}\(g;0\.5\)\)/\\varepsilon\\big\), so that a group at0\.70\.7carries its own measured single gain\. Withz⊇γz\\supseteq\\gammawhenzzmoves every group ofγ\\gammain the same direction asγ\\gamma\(0\.50\.5and0\.70\.7counting as lowered\) andΓ\\Gammathe set of the232232patterns, the score evaluated by the circuit and by the classical check of each sample is

Φt\(z\)=∑γ∈Γat\(γ\)1\[z⊇γ\]\+∑g=1Gbt\(g\)1\[ηg=0\.7\]\.\\Phi\_\{t\}\(z\)=\\sum\_\{\\gamma\\in\\Gamma\}a\_\{t\}\(\\gamma\)\\,\\mathbf\{1\}\[z\\supseteq\\gamma\]\+\\sum\_\{g=1\}^\{G\}b\_\{t\}\(g\)\\,\\mathbf\{1\}\[\\eta\_\{g\}=0\.7\]\.\(27\)At checkpoint200200of the single\-batch run, for example, the second layer group has the single gain3\.69×10−43\.69\\times 10^\{\-4\}at0\.50\.5, which rounds to zero directly but toat=2a\_\{t\}=2after division byε\\varepsilon\(1\.851\.85\), and the gain1\.14×10−31\.14\\times 10^\{\-3\}at0\.70\.7givesbt=4b\_\{t\}=4, so the group contributes66at0\.70\.7\. Patterns whose rounded coefficient is zero add nothing and are left out of the circuit\. Because the coefficients and corrections are rounded separately,Φt​\(z\)\\Phi\_\{t\}\(z\)need not equalY^t​\(z\)/ε\\widehat\{Y\}\_\{t\}\(z\)/\\varepsilonrounded once\.

#### Around another baseline\.

Eq\. \([18](https://arxiv.org/html/2611.21701#S4.E18)\) of Section[4](https://arxiv.org/html/2611.21701#S4)writes the score around a baseline schedulezbasez\_\{\\mathrm\{base\}\}as single\-group gains plus pair and triple interactions\. Withzbase⊕γz\_\{\\mathrm\{base\}\}\\oplus\\gammathe baseline with only the groups of a patternγ\\gammamoved to the rates ofzz, and𝒫⁡\(z∣zbase\)\\mathcal\{P\}\(z\\mid z\_\{\\mathrm\{base\}\}\)the patterns of one to three groups on whichzzdiffers fromzbasez\_\{\\mathrm\{base\}\}, the inclusion–exclusion of Eq\. \([26](https://arxiv.org/html/2611.21701#A10.E26)\) around the baseline gives

Y^t​\(z∣zbase\)=∑γ∈𝒫⁡\(z∣zbase\)∑γ′⊆γ\(−1\)\|γ\|−\|γ′\|​\[Λt​\(zbase\)−Λt​\(zbase⊕γ′\)\]\.\\widehat\{Y\}\_\{t\}\(z\\mid z\_\{\\mathrm\{base\}\}\)=\\sum\_\{\\gamma\\in\\mathcal\{P\}\(z\\mid z\_\{\\mathrm\{base\}\}\)\}\\ \\sum\_\{\\gamma^\{\\prime\}\\subseteq\\gamma\}\(\-1\)^\{\|\\gamma\|\-\|\\gamma^\{\\prime\}\|\}\\big\[\\Lambda\_\{t\}\(z\_\{\\mathrm\{base\}\}\)\-\\Lambda\_\{t\}\(z\_\{\\mathrm\{base\}\}\\oplus\\gamma^\{\\prime\}\)\\big\]\.\(28\)For a pattern of one, two or three groups the inner sum is the single\-group gain, the pair interaction or the triple interaction aroundzbasez\_\{\\mathrm\{base\}\}, so Eq\. \([28](https://arxiv.org/html/2611.21701#A10.E28)\) is Eq\. \([18](https://arxiv.org/html/2611.21701#S4.E18)\) term by term: every group outside a pattern keeps its baseline rate, every term vanishes when one of its groups stays at its baseline rate, and the baseline scores zero\. The terms are rate\-specific effects measured aroundzbasez\_\{\\mathrm\{base\}\}; the sharing of the0\.50\.5interactions at0\.70\.7is a choice of the reported runs and is not part of this expression\.

#### Circuit and search\.

The register holds two qubits per group and an accumulator of⌈log2⁡\(span\+1\)⌉\+2\\lceil\\log\_\{2\}\(\\text\{span\}\+1\)\\rceil\+2qubits, where the span is the range ofΦt\\Phi\_\{t\}, so2222to2525qubits in all\. The accumulator, in the Fourier basis, receives eachat​\(γ\)a\_\{t\}\(\\gamma\)andbt​\(g\)b\_\{t\}\(g\)as a phase rotation controlled on the qubits of the groups involved; subtractingΦt​\(y\)\+1\\Phi\_\{t\}\(y\)\+1, flipping the phase of the sign qubit and uncomputing the arithmetic realize Eq\. \([16](https://arxiv.org/html/2611.21701#S2.E16)\) on every schedule at once, and the circuit returns no score as a readable value\. The search is the minimum finding of[Dürr & Høyer \(1996\)](https://arxiv.org/html/2611.21701#bib.bib9)with a larger score being better\. Each round prepares the joint schedules in superposition, applies the oracle and diffusion the number of times drawn by Eq\. \([6](https://arxiv.org/html/2611.21701#S2.E6)\), and samples one schedule; a classical evaluation of Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\) on the sample updates the best schedule and the threshold after an improvement, and after an unsuccessful sample criterion D, the criterion chosen in Section[3\.1](https://arxiv.org/html/2611.21701#S3.SS1), decides whether to continue\. The uniform\-rate baseline is returned if nothing beats it\. Before each run, the circuit was checked on the checkpoint\-200200coefficients: at the initial threshold its marks equal those of Eq\. \([16](https://arxiv.org/html/2611.21701#S2.E16)\) evaluated with Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\) on all4,0964\{,\}096schedules, and the accumulator returns to zero\. This enumeration exists only in the test; the search never reads it\. Each search is one trial with its random seed set to the checkpoint step\. Figure[A8](https://arxiv.org/html/2611.21701#A10.F8)shows the workflow of one checkpoint at this level of detail\. Criterion D is reused here as a stopping rule\. Its rank distribution \(Appendix[B](https://arxiv.org/html/2611.21701#A2)\) assumes a first threshold drawn uniformly at random and distinct scores, whereas this search starts from the fixed uniform\-rate baseline and the integer scores of Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\) can tie; the rule was applied unchanged, soδ=10−3\\delta=10^\{\-3\}is a nominal setting and not a calibrated error level, and the returned schedule is not certified as the maximizer ofΦt\\Phi\_\{t\}\. The single\-batch searches ran on a laptop CPU in single precision \(4,4654\{,\}465to16,36216\{,\}362s of simulator time per search\), the five\-batch searches on a GPU in double precision \(158158to281281s\); these are times of the classical simulator, not of a quantum device\.

Figure A8:Detailed workflow of the Interaction\-Guided Schedule Oracle at one checkpoint\. Classical loss measurements define fixed integer coefficients, and the best schedule is initialized at the uniform\-rate baseline\. Each quantum round prepares4,0964\{,\}096schedules, applies the chosen number of oracle and diffusion iterations, and measures one schedule\. Classical checking updates the best schedule and threshold after an improvement; otherwise criterion D determines whether to continue\. After stopping, the best schedule controls the next training interval, and the next checkpoint rebuilds the evidence\.
#### Decisions\.

Table[A5](https://arxiv.org/html/2611.21701#A10.T5)lists, for each leg, the schedule returned \(group order from the first layers to the last\), its interaction scoreΦ\\Phiin units ofε\\varepsilon, the roundsRRand Grover iterationsQQof the search, the register width, and, for the five\-batch run, the number of the232232interaction terms whose sign differs between the first batch and the five\-batch average\. Both runs start from the checkpoint at step200200of the uniform\-rate run of seed22; each interval is200200steps \(221221for the last\)\. Besides the evidence, the two runs also differ in the GPU, the simulator precision and the determinism of the training kernels\.

Table A5:The five decisions of each closed\-loop run\. The scoresΦ\\Phiof the two runs are not comparable with each other, since the averaged coefficients are smaller\.

## Appendix KHardware\-constrained Grover iteration counts

#### Compiled gate accounting\.

For the hardware\-sampled searches of Section[3\.3](https://arxiv.org/html/2611.21701#S3.SS3), the circuit of a round withjjGrover iterations is assembled from compiled pieces: a prefix that prepares the uniform superposition on theqsrch=12q\_\{\\mathrm\{srch\}\}=12search qubits, thenjjiteration bodies, then the measurement of the search register\. Bodyj′j^\{\\prime\}is one Grover iteration, the oracle of Eq\. \([16](https://arxiv.org/html/2611.21701#S2.E16)\) followed by the diffuser, routed on the device starting from the qubit placement that bodyj′−1j^\{\\prime\}\-1left, so the bodies are not identical and differ in gate count; the submitted circuit is the concatenation of the cached compiled pieces, with no further compilation\. For each gate type,

CCZ​\(j\)=CCZ​\(0\)\+∑j′=1jΔ​CCZ​\(j′\),C\_\{\\mathrm\{CZ\}\}\(j\)=C\_\{\\mathrm\{CZ\}\}\(0\)\+\\sum\_\{j^\{\\prime\}=1\}^\{j\}\\Delta C\_\{\\mathrm\{CZ\}\}\(j^\{\\prime\}\),\(29\)and likewise for RZ and SX, the components of𝐂\\mathbf\{C\}in Eq\. \([30](https://arxiv.org/html/2611.21701#A11.E30)\)\. Each Hadamard gate of the prefix is compiled as RZ, SX, RZ, so𝐂⁡\(0\)=\(0,24,12\)\\mathbf\{C\}\(0\)=\(0,\\,24,\\,12\); the measurement adds no gate of the three types\. The counts depend on the checkpoint, through the integer coefficients of Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\), on the threshold, and on the backend, synthesis, layout and routing used at compile time, and they are recomputed for each\. On assembled circuits at checkpoint200200, with the threshold scoreΦt​\(y\)=42\\Phi\_\{t\}\(y\)=42onibm\_fez, the gate counts equal the sums of Eq\. \([29](https://arxiv.org/html/2611.21701#A11.E29)\) exactly:1,799,8101\{,\}799\{,\}810CZ,2,348,6322\{,\}348\{,\}632RZ and3,520,8863\{,\}520\{,\}886SX gates atj=28j=28, and3,478,8533\{,\}478\{,\}853,4,525,5114\{,\}525\{,\}511and6,797,2186\{,\}797\{,\}218atj=54j=54\.

#### The limit\.

IBM specifies per\-circuit limits of5×1065\\times 10^\{6\}two\-qubit gates,3×1073\\times 10^\{7\}RZ gates and2×1072\\times 10^\{7\}SX gates \(IBM Quantum documentation, “Job limits”,[https://quantum\.cloud\.ibm\.com/docs/en/guides/job\-limits](https://quantum.cloud.ibm.com/docs/en/guides/job-limits), read 11 November 2026\)\. With𝐂⁡\(0\)\\mathbf\{C\}\(0\)the numbers of CZ, RZ and SX gates of the preparation andΔ​𝐂​\(j′\)\\Delta\\mathbf\{C\}\(j^\{\\prime\}\)those added by thej′j^\{\\prime\}\-th body, the largest iteration count within the per\-circuit gate limits of the service is

Jdev=max⁡\{j∈ℤ≥0:𝐂⁡\(0\)\+∑j′=1jΔ​𝐂​\(j′\)≤\(5,30,20\)×106\},J\_\{\\mathrm\{dev\}\}=\\max\\Big\\\{\\,j\\in\\mathbb\{Z\}\_\{\\geq 0\}:\\ \\mathbf\{C\}\(0\)\+\\sum\_\{j^\{\\prime\}=1\}^\{j\}\\Delta\\mathbf\{C\}\(j^\{\\prime\}\)\\leq\(5,\\,30,\\,20\)\\times 10^\{6\}\\Big\\\},\(30\)the inequality holding for each gate type, recomputed for every checkpoint, threshold and backend\. The compiled circuits are CZ\-native, so Eq\. \([30](https://arxiv.org/html/2611.21701#A11.E30)\) applies these limits to the counts of Eq\. \([29](https://arxiv.org/html/2611.21701#A11.E29)\)\. The maximum in Eq\. \([30](https://arxiv.org/html/2611.21701#A11.E30)\) runs over all non\-negative integers, soJdevJ\_\{\\mathrm\{dev\}\}is a property of the compiled circuit family and the device limits and does not depend on the range of the BBHT draw\. Every body adds gates, so the counts never decrease injj, the admissible counts form the interval\{0,…,Jdev\}\\\{0,\\dots,J\_\{\\mathrm\{dev\}\}\\\}, andJdevJ\_\{\\mathrm\{dev\}\}is found by compiling bodies until the first one crosses a limit\. Two evaluations exist in the program and are kept apart\. A measurement stage compiles until a limit is crossed and returnsJdevJ\_\{\\mathrm\{dev\}\}exactly\. During the search, bodies are compiled only as far as the current block of rounds can draw, at most⌈N⌉−1=63\\lceil\\sqrt\{N\}\\rceil\-1=63atN=4,096N=4\{,\}096; if all of them fit, the search records onlyJdev≥63J\_\{\\mathrm\{dev\}\}\\geq 63and the bound is inactive\. The largest count that the schedule can draw,6363, is a property of the schedule and a different number fromJdevJ\_\{\\mathrm\{dev\}\}\.

#### Estimate before compilation\.

Before the whole range is compiled, a first\-order projection follows from the fixed counts and a representative per\-iteration countΔ​C¯\\overline\{\\Delta C\}of each gate type,

J^dev=min⁡\{⌊5×106−CCZ​\(0\)Δ​C¯CZ⌋,⌊3×107−CRZ​\(0\)Δ​C¯RZ⌋,⌊2×107−CSX​\(0\)Δ​C¯SX⌋\}\.\\widehat\{J\}\_\{\\mathrm\{dev\}\}=\\min\\bigg\\\{\\bigg\\lfloor\\frac\{5\\times 10^\{6\}\-C\_\{\\mathrm\{CZ\}\}\(0\)\}\{\\overline\{\\Delta C\}\_\{\\mathrm\{CZ\}\}\}\\bigg\\rfloor,\\ \\bigg\\lfloor\\frac\{3\\times 10^\{7\}\-C\_\{\\mathrm\{RZ\}\}\(0\)\}\{\\overline\{\\Delta C\}\_\{\\mathrm\{RZ\}\}\}\\bigg\\rfloor,\\ \\bigg\\lfloor\\frac\{2\\times 10^\{7\}\-C\_\{\\mathrm\{SX\}\}\(0\)\}\{\\overline\{\\Delta C\}\_\{\\mathrm\{SX\}\}\}\\bigg\\rfloor\\bigg\\\}\.\(31\)It is a planning estimate only; every decision of the search uses the exact cumulative counts of Eqs\. \([29](https://arxiv.org/html/2611.21701#A11.E29)\) and \([30](https://arxiv.org/html/2611.21701#A11.E30)\)\. Table[A6](https://arxiv.org/html/2611.21701#A11.T6)lists the compilations of the same oracle at checkpoint200200with the threshold score4242onibm\_fez\. In both measured cases the estimate equals the exact value, and CZ is the binding gate type\. For the shared\-AND synthesisJdev=77J\_\{\\mathrm\{dev\}\}=77, above the schedule maximum of6363, so every circuit the schedule can draw at this threshold is inside the three counted limits, the largest,j=63j=63, holding4,062,9404\{,\}062\{,\}940CZ,5,287,6205\{,\}287\{,\}620RZ and7,940,3247\{,\}940\{,\}324SX gates\. The shared\-AND synthesis is the same phase oracle built from shared AND indicators and compute–diagonal–uncompute relative\-phase Toffoli operations\. It agrees in amplitude with the validated circuit on three\- and four\-group instances \(largest difference1\.5×10−151\.5\\times 10^\{\-15\}\), marks all4,0964\{,\}096basis states of the full register at the threshold score4242as the validated circuit does \(largest deviation1\.0×10−121\.0\\times 10^\{\-12\}\), and its chained routing reproduces the marginal of the logical circuit on the search register on two\- and three\-group instances atj=1j=1to44\(largest probability difference6\.1×10−146\.1\\times 10^\{\-14\}\)\.

Table A6:The largest executable iteration count at checkpoint200200, threshold score4242, backendibm\_fez: CZ gates per compiled iteration body, the estimate of Eq\. \([31](https://arxiv.org/html/2611.21701#A11.E31)\) and the exactJdevJ\_\{\\mathrm\{dev\}\}of Eq\. \([30](https://arxiv.org/html/2611.21701#A11.E30)\), with the cumulative CZ counts on either side of the limit\.
#### Restricted draw\.

WithmBBHTm\_\{\\mathrm\{BBHT\}\}the range of Eq\. \([6](https://arxiv.org/html/2611.21701#S2.E6)\) at a round, the hardware search draws

j∼𝒰⁡\{0,…,min⁡\(⌈mBBHT⌉−1,Jdev\)\}\.j\\sim\\mathcal\{U\}\\big\\\{0,\\dots,\\min\\big\(\\lceil m\_\{\\mathrm\{BBHT\}\}\\rceil\-1,\\,J\_\{\\mathrm\{dev\}\}\\big\)\\big\\\}\.\(32\)WhenJdev≥⌈mBBHT⌉−1J\_\{\\mathrm\{dev\}\}\\geq\\lceil m\_\{\\mathrm\{BBHT\}\}\\rceil\-1, Eq\. \([32](https://arxiv.org/html/2611.21701#A11.E32)\) is the draw of Eq\. \([6](https://arxiv.org/html/2611.21701#S2.E6)\) and leaves the random\-number stream unchanged; with the rule switched on but inactive, a simulated search reproduces the sequential reference round for round\. When the bound is active, the support of the draw changes, and the search is reported as hardware\-constrained BBHT rather than as an unmodified run of the original algorithm\. Criterion D updates its rank distribution after each failed round with the success probability of Eq\. \([5](https://arxiv.org/html/2611.21701#S2.E5)\) at the iteration count actually executed, so its update is computed from the executed counts whether or not the bound is active; statements that rely on the unconstrained draw, such as expected numbers of rounds or of Grover iterations before termination, do not carry over\. No out\-of\-range draw is redrawn, no drawn circuit is truncated and no simulator outcome replaces a device measurement\.

#### Restricted rounds\.

Over theRRexecuted rounds of a search, the observed fraction of rounds whose draw was restricted is

p^cap=1R∑rounds𝟏\[Jdev<⌈mBBHT⌉−1\],\\widehat\{p\}\_\{\\mathrm\{cap\}\}=\\frac\{1\}\{R\}\\sum\_\{\\text\{rounds\}\}\\mathbf\{1\}\\big\[J\_\{\\mathrm\{dev\}\}<\\lceil m\_\{\\mathrm\{BBHT\}\}\\rceil\-1\\big\],\(33\)a count over the executed rounds rather than a probability estimated from hypothetical ones; the search logs the indicator for every round together withjj, the upper end of its draw andJdevJ\_\{\\mathrm\{dev\}\}\. The executed values ofjj, the totalQQof Eq\. \([14](https://arxiv.org/html/2611.21701#S2.E14)\),p^cap\\widehat\{p\}\_\{\\mathrm\{cap\}\}and every active bound with the threshold and the counts that caused it are reported per checkpoint with the run, in Appendix[L](https://arxiv.org/html/2611.21701#A12)\.

#### Service\-side checks and evidence\.

The service also limits each qubit to26\.826\.8million low\-level control\-system instructions, to which dynamical decoupling adds after the gate count, so Eq\. \([30](https://arxiv.org/html/2611.21701#A11.E30)\) is a necessary screen on the client side and not a guarantee that a circuit is accepted or executed faithfully; a rejection is logged with the service’s error and stops the stage, and the circuit is neither replaced nor given a smallerjj\. Two jobs of an earlier compilation onibm\_fezshow that the limit binds this workload: one circuit was rejected for5,010,8525\{,\}010\{,\}852CZ gates against the limit of5,000,0005\{,\}000\{,\}000, and a circuit of4,116,0574\{,\}116\{,\}057CZ gates, run with dynamical decoupling, was accepted and finished; that circuit, withj=23j=23, was round3939of the first search\. The completed run then executed shared\-AND circuits of up to4,284,2664\{,\}284\{,\}266CZ gates \(j=60j=60\), all accepted\.

#### Status and scope\.

Three objects are kept apart: the BBHT schedule of Eq\. \([6](https://arxiv.org/html/2611.21701#S2.E6)\), the feasibility boundJdevJ\_\{\\mathrm\{dev\}\}of the compiled circuits, and the trace of the executed countsjj\. The accounting of Eqs\. \([29](https://arxiv.org/html/2611.21701#A11.E29)\) and \([30](https://arxiv.org/html/2611.21701#A11.E30)\), the draw of Eq\. \([32](https://arxiv.org/html/2611.21701#A11.E32)\) and the logging of Eq\. \([33](https://arxiv.org/html/2611.21701#A11.E33)\) are implemented; Eq\. \([29](https://arxiv.org/html/2611.21701#A11.E29)\) is checked on assembled circuits,JdevJ\_\{\\mathrm\{dev\}\}is measured at one checkpoint, threshold and backend, and the equivalence and routing checks above are passed\. In the completed run, the restriction was logged on the191191rounds from round4040of the first search onward, when the search changed to the shared\-AND compilation, and was active in none of them, so every search used the unmodified draw of Eq\. \([6](https://arxiv.org/html/2611.21701#S2.E6)\); the largest count executed,j=60j=60, stays below the schedule maximum of6363\(Section[3\.3](https://arxiv.org/html/2611.21701#S3.SS3), Appendix[L](https://arxiv.org/html/2611.21701#A12)\)\. The hardware searches use the full Interaction\-Guided Schedule Oracle of Section[2\.4](https://arxiv.org/html/2611.21701#S2.SS4): the integer coefficients from the checkpoint evaluation enter the circuit as constants, the circuit evaluates the score of every schedule coherently and marks the schedules above the threshold, no table of schedule scores is computed or loaded, and each measured schedule is checked by one classical evaluation of Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\)\. The hardware validation of Appendix[D](https://arxiv.org/html/2611.21701#A4), with a1616\-row lookup oracle, is a separate experiment\. These counts describe feasibility and search behavior on the device and make no statement about end\-to\-end training acceleration\.

## Appendix LHardware\-sampled closed loop: search paths and IBM turnaround

Each search of Section[3\.3](https://arxiv.org/html/2611.21701#S3.SS3)ran on the least busy of three156156\-qubit IBM Heron devices at the start of its checkpoint, with one shot per circuit and up to four circuits per job\. Before each search, the compiled oracle was checked on the constants of that checkpoint, in amplitude on a three\-group instance and on all4,0964\{,\}096basis states of the full register\. Rounds11to3939of the first search used earlier compilations of the same oracle; every later round used the shared\-AND compilation of Appendix[K](https://arxiv.org/html/2611.21701#A11)\. A round draws its iteration count, runs one circuit and checks the sampled schedule classically with Eq\. \([27](https://arxiv.org/html/2611.21701#A10.E27)\); when a job holds several pre\-drawn rounds, its results are read in order and the rounds after an improvement or a stop are not used\. Every sample that entered a decision was measured on the device: no simulator sample replaced a device sample, no instance was reduced and no draw was changed or truncated\.

Table[A7](https://arxiv.org/html/2611.21701#A12.T7)gives the threshold path of each search, the rounds and Grover iterations after its last threshold update, and its circuit with the largestjj\. Table[A8](https://arxiv.org/html/2611.21701#A12.T8)gives the IBM turnaround of the jobs, from submission to completion at IBM with the queue included, split into queue and device time, and the QPU time charged\. Over the five searches,9494jobs took37,30037\{,\}300s \(10\.410\.4h\) of turnaround for188188s of charged QPU time\. Three further jobs at the first checkpoint, whose rounds were set aside, took1,7681\{,\}768s and66s of QPU time, and an acceptance test before the run took3030s and22s\.

Table A7:Threshold path, stopping phase and circuit with the largestjjof each hardware\-sampled search\. The path lists the threshold scoreΦ\\Phiafter each update with the round of the update, starting from the uniform\-rate baseline at00\. The deepest circuit of the run, round3939at checkpoint200200\(j=23j=23,4,116,0574\{,\}116\{,\}057CZ gates, depth7,787,2037\{,\}787\{,\}203\), used an earlier compilation\.Table A8:IBM turnaround of the hardware\-sampled searches: jobs with a result, turnaround from submission to completion at IBM \(queue included\), queue and device time, and QPU time charged\.

相似文章

量子增强扩散语言模型的电路超网络

Hugging Face Daily Papers

HyperQ 在冻结的掩码扩散语言模型中引入了令牌条件化的量子残差分支,使用电路超网络动态生成量子电路参数。这种方法在基准测试中提高了性能,并且计算效率高,与经典基线相比需要更少的微调示例。

用于学习路由的混合量子-经典神经网络

arXiv cs.LG

本文研究了用于车辆路径问题的混合量子-经典神经网络,发现编码器前馈替代可以将模型参数减少56.6%,同时在中小规模实例上保持接近基线性能。

混合架构中的竞争学习

Hugging Face Daily Papers

该论文在混合语言模型中发现了一种"竞争学习"现象,即循环层和注意力层会竞争学习相同的功能,并提出了一种分阶段训练方法,在参数量高达 1B 的模型上提升了语言建模损失、常识推理和检索任务的表现。

TempQuant:为量化语言模型学习自适应温度

arXiv cs.LG

TempQuant 提出自适应学习量化语言模型的温度与 top-p 设置,通过校准使量化模型的下一词元概率分布更接近全精度模型,在 3-bit Qwen3-4B 上将 MATH 准确率从 34% 提升至 51%。