Aligning Quantum Operators with Large Language Models

David Kremer; Hang Hua; Pengyuan Li; Rogerio Feris; Yunchao Liu

REVIEW 3 major objections 6 minor 25 references

An LLM can condition directly on a quantum operator, encoded as a patch matrix, and compile it into a circuit, outperforming specialized solvers on 4-qubit Clifford+T synthesis.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 11:34 UTC pith:RBJFCLPY

load-bearing objection A clever first step toward LLMs that read quantum operators, but the 'beats all baselines' claim rests on apples-to-oranges comparisons and the autoregressive loop has an unanalyzed train/test mismatch. the 3 major comments →

arxiv 2606.13811 v2 pith:RBJFCLPY submitted 2026-06-11 quant-ph cs.AI

Aligning Quantum Operators with Large Language Models

Rogerio Feris , Yunchao Liu , Pengyuan Li , Hang Hua , David Kremer This is my paper

classification quant-ph cs.AI MSC 81P6868T50

keywords quantum circuit synthesislarge language modelsPauli transfer matrixClifford+T circuitsmultimodal alignmentunitary synthesistext-conditioned synthesisautoregressive decoding

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a large language model can reason about a quantum operation when the operation is presented to it directly as a matrix, not as a textual description. It does this by slicing the operator's real-valued Pauli Transfer Matrix into patches, projecting those patches into the model's token embedding space, and fine-tuning the model to peel off one gate at a time from the remaining matrix. On 4-qubit Clifford+T synthesis the model reaches 87.9% success with greedy decoding, exceeding specialized solvers, and its performance keeps improving as training data grows. Because the same model also follows natural-language constraints on which gates may be used, the paper argues this is a step toward quantum-aware foundation models.

Core claim

The paper claims this is the first approach that lets an LLM condition directly on quantum operators rather than their textual or programmatic descriptions. The mechanism: any n-qubit unitary is represented as a 4^n × 4^n Pauli Transfer Matrix (a real-valued, global-phase-invariant matrix that multiplies across circuit composition); the matrix is normalized and cut into non-overlapping patches, a lightweight encoder plus MLP projects these patches into the LLM's word-embedding space, and the LLM is fine-tuned to autoregressively predict Pauli-rotation gates. At each step the residual PTM, the part of the target not yet compiled, is re-encoded and shown again, so the model conditions each pre

What carries the argument

The central object is the Pauli Transfer Matrix (PTM), the real-valued 4^n × 4^n matrix representing a unitary's action on Pauli operators, which multiplies like the unitary and discards global phase. The paper treats the PTM as a single-channel image, partitions it into 16×16 patches, and routes each patch through a linear layer, layer norm, positional embedding, and a two-layer MLP that projects into the LLM's embedding space; these 'visual tokens' are prepended to text tokens encoding current fidelity and previous gates. The load-bearing mechanism is stepwise autoregressive synthesis: after each predicted gate, the residual PTM is updated externally by multiplying by the inverse PTM of th

Load-bearing premise

The model is trained on residuals computed from the ground-truth circuit, and the whole approach assumes it will keep predicting correctly when the residual is instead the product of its own previous (possibly imperfect) predictions, so that errors do not compound during autoregressive synthesis.

What would settle it

Present a held-out target unitary with its PTM patches randomly permuted at inference time: if greedy success stays near the reported 87.9% instead of collapsing, the model is reading the operator's content rather than its structure, and the central alignment claim is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

LLMs gain a first-class route to quantum operators, so compilation, transpilation, and verification tasks can be conditioned on the operator itself rather than a symbolic proxy.
Data scaling shows no saturation, so adding more synthetic training circuits should keep raising synthesis success within this gate-set and qubit-count regime.
Best-of-N sampling converts inference compute into accuracy, lifting success from 87.9% to 99.4%, so a deployed system can choose a reliability/compute trade.
Natural-language constraint following works for gate placements unseen in training, pointing to hardware-aware compilation where connectivity rules are specified at inference time.
Because the alignment is representation-agnostic, the same recipe could accept Clifford tableaux, Pauli lists, or tensor-network descriptions, which the paper identifies as the path to larger-scale compilation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The patch-encoding recipe is likely transferable to any structured linear-algebra array, so the same alignment trick could be applied to matrix factorization or tensor-decomposition tasks that call for conditioning on the actual array.
The external residual loop is a form of scratchpad memory; letting the LLM backtrack or revise earlier gates could recover from the mid-trajectory fidelity drops the paper shows qualitatively.
The 4^n × 4^n PTM cost caps direct use at small qubit counts; a concrete next test is whether encoding a factorized or stabilizer-based representation instead of the full PTM preserves the gains.
The paper does not separate single-step prediction errors from error propagation in its 12% failure rate; measuring per-step accuracy under self-conditioned rollouts would pinpoint which one dominates.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A clever first step toward LLMs that read quantum operators, but the 'beats all baselines' claim rests on apples-to-oranges comparisons and the autoregressive loop has an unanalyzed train/test mismatch.

read the letter

The short version: this paper is the first to feed a quantum operator directly into an LLM, via patch-encoded Pauli Transfer Matrices, and it does a decent job on 4-qubit Clifford+T synthesis. The core idea is worth knowing about. But the headline 'surpasses all baselines with 87.9% success' is not established, because the RL and MDL baselines were run on different unitary distributions (their circuits are Clifford+T, ours are Pauli rotations). The paper admits this in Sec. V-C, calling the numbers 'approximate comparisons,' which is honest but should not be presented as a win in the abstract. Only SynthetiQ was run on your held-out set, and you beat it. That's real.

What's genuinely new: the representation of a 256×256 PTM as a 16×16 patch grid, the projection into the LLM token space, and the stepwise residual loop where the model peels gates one by one while the residual is recomputed externally. That's a clean formulation and the scaling table (71% at 9.2M, 87.9% after continued training) suggests the approach is not saturated. The text-conditioned synthesis experiment is a nice bonus — the 91% vs 53% compliance drop when the constraint text is removed is convincing evidence that the model uses the language.

Soft spots beyond the baseline issue: (1) Training uses teacher-forced residuals (from ground-truth gate orders); inference uses self-generated residuals. The paper never analyzes this distribution shift. The greedy results imply the model copes, but it doesn't show how often early errors compound. (2) Success is defined as fidelity ≥0.999, not exact synthesis. Acceptable for a proof of concept, but it flatters the numbers relative to the 'exact synthesis' language used elsewhere. (3) No error bars, and patch size was selected on the held-out set. Minor. (4) The 4^n × 4^n PTM scaling limits this to 4 qubits; acknowledged, and fine for a proof of concept.

Overall, the central claim — that an LLM can condition directly on a quantum operator and synthesize circuits — is likely sound. The evaluation just needs to be tightened before the numbers are taken at face value. This deserves serious peer review, but the authors should be asked to fix or restrict the baseline comparison, analyze the autoregressive mismatch, and add variance estimates.

If you're working on quantum compilation or LLM interfaces, this is a useful read. I'd bring it to a reading group for the idea, not for the benchmark claims.

Referee Report

3 major / 6 minor

Summary. The paper proposes a multimodal framework that maps 4-qubit unitary operators into the latent space of a pretrained LLM. A unitary is encoded as a real-valued Pauli Transfer Matrix (PTM), partitioned into patches, and projected into the LLM's token embedding space. The model then performs stepwise autoregressive circuit synthesis: at each step it observes the residual PTM, predicts one Pauli rotation gate, and the residual is updated externally by left-multiplying the inverse of the predicted gate. Training is supervised on synthetically generated Clifford+T-like circuits (Pauli rotation gate sets), and inference is greedy or Best-of-N sampling. On a held-out set of 2,000 circuits with 1–15 gates, the model reaches 87.9% greedy success and 99.4% with Best-of-N=80. The paper also demonstrates text-conditioned synthesis with gate-placement constraints that were unseen during training, and reports data-scaling behavior with no saturation.

Significance. The central idea—conditioning an LLM directly on an operator-level representation rather than on a textual/programmatic description—is a meaningful step for quantum-aware foundation models. The PTM patch encoding is a natural choice given the multiplicative composition property of PTMs, and the stepwise residual formulation is a clean way to reduce a hard search problem to conditional single-token prediction. The data-scaling and inference-time-scaling results are suggestive, and the language-conditioned constraint experiment is a genuinely useful demonstration that the LM prior can steer synthesis. However, the headline comparison to prior RL and MDL baselines is not like-for-like, and the training/inference mismatch in how residuals are produced is not analyzed. If the comparison is properly qualified and the exposure-bias question is addressed empirically, the contribution would be solid as a proof of concept.

major comments (3)

[Section V-C / Abstract / Table II] The claim that 'our model surpasses all baselines with 87.9% success using greedy decoding alone' is not supported by a like-for-like comparison. As the paper itself states, RL [8] and MDL [24] results are 'approximate comparisons' taken from their publications, and 'the underlying unitary distributions may differ slightly: our circuits are products of Pauli rotations while theirs are generated directly in Clifford+T.' This is not merely a minor distributional mismatch: a 15-gate Pauli-rotation circuit and a 15-gate Clifford+T circuit are different synthesis problems, so the superior numerical success rates may reflect task difficulty rather than method quality. In addition, the paper's success criterion is fidelity F≥0.999 (Section III-D), whereas the published baselines may target exact synthesis. I request either (a) running the RL and MDL methods on the same held-out set and evaluati
[Section IV-B.a / Eq. (6) / Section IV-B] The training and inference procedures differ in a way that is not analyzed. In training, 'Training samples are generated on the fly: at each iteration, a random Clifford+T circuit of length K is sampled ... and the K+1 stepwise decompositions ... are used as training examples.' The residual PTM at each step is therefore computed from the ground-truth gate sequence (teacher forcing). At inference, the residual is updated with the model's own previously predicted gates: P^{(t+1)} = PTM(hat g)^{-1} P^{(t)}. If the model makes an early error, all subsequent residuals lie off the training distribution, and greedy autoregressive synthesis can diverge. The paper does not measure this exposure bias, nor does it report any diagnostic (e.g., fraction of rollouts that first deviate at step k, or a comparison of greedy success when conditioned on oracle residuals vs. self-generated residuals). Table
[Section V-C / Figure 2 / Section III-D] The evaluation protocol conflates 'success' with 'fidelity at least 0.999' and exact synthesis. Because the gate set R(P) = exp(-i pi/8 P) is universal only in the approximate sense, F=0.999 means the synthesized circuit is not exactly the target unitary. The paper reports a mean predicted-to-oracle gate ratio of 1.007 for successful circuits, which is informative, but the threshold itself should be justified. More importantly, when comparing to RL/MDL baselines, the reader needs to know whether those baselines report exact synthesis rates or also allow an approximation tolerance. If the baselines report exact success and this paper reports approximate success, the headline comparison is systematically biased in the authors' favor. I recommend reporting both exact match (or a much tighter threshold, e.g., F=1-10^{-8}) and F≥0.999, or at least clearly stating the approximation tolerance f

minor comments (6)

[Abstract / Section V-C] Please make the strength of the claim consistent: the Abstract says 'competitive with state-of-the-art methods,' while Section V-C says 'surpasses all baselines.' After the major comparison issue is addressed, the language should be aligned.
[Tables I and II] No error bars or multiple-seed results are reported. The held-out set is 2,000 circuits; reporting confidence intervals or results over several splits would make the scaling and Best-of-N improvements more robust.
[Section IV-A, Eq. (4)] The description of the patch encoder is slightly underspecified: it is ambiguous whether LayerNorm is applied before or after adding the positional embedding e_j. Please clarify the order of operations and the initialization of the learned positional embedding.
[Section V-G] The 'unseen constraint combinations' are blacklisted during training, but the LLM backbone has been pretrained on large amounts of text and may have seen similar constraint language. The random-init comparison is a good control, but the wording 'unseen' should be qualified as 'unseen in the fine-tuning data' rather than 'never seen by any component of the model.'
[Section V-F / Figure 4] The sentence 'They clearly show the model generalize beyond imitation' contains a grammar error ('generalize' should be 'generalizes'). Also, Figure 4(b) is described qualitatively; it would help to include the exact gate sequences or a quantitative measure (e.g., edit distance to the oracle) to support the claim of generalization rather than memorization.
[Section VI / References] The reference to DeepSeek-V3 [25] for 'wait' and 'aha' moments is not the most direct citation for deliberative reasoning in LLMs; consider citing the original 'wait'/'aha' reasoning papers instead.

Circularity Check

0 steps flagged

No significant circularity: the synthesis results are produced by supervised learning on synthetic circuits with independent held-out evaluation; self-citations are attribution, not load-bearing derivation.

full rationale

The paper's central derivation chain is not circular. The model is trained with a standard next-token prediction loss (Eq. 6) on synthetic ground-truth circuit decompositions; at inference it predicts a gate from the current residual PTM, updates the residual by PTM multiplicativity (Eq. 2), and repeats. Success is measured externally on 2,000 held-out circuits with a fidelity threshold τ=0.999 (Eq. 3). No fitted parameter is defined in terms of the reported success metric, and no prediction is statistically forced by a prior fit. The PTM representation and Pauli-rotation gate set are introduced via the paper's own Eqs. 1-2; the citation to [20] for the Pauli-rotation parameterization is attribution, not an imported uniqueness theorem or an unverified ansatz that carries the argument. The self-citations in the paper are related-work context or attribution for standard representations, and none is the load-bearing justification for the experimental claims. The approximate RL/MDL baseline comparisons in Section V-C are explicitly disclosed as approximate and may differ in unitary distribution; this is a benchmark-validity caveat, not circularity. The scaling behavior, held-out success, Best-of-N gains, Haar-random generalization, and text-conditioned constraint compliance are all empirical outcomes of a learned model rather than consequences of an input definition. Therefore, no circular step is present.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard PTM algebra and on supervised training over a synthetic distribution. No fitted law or invented physical entity is introduced. The main hand-chosen quantity is the success threshold τ, which directly defines the headline numbers.

free parameters (1)

Fidelity threshold τ = 0.999
Hand-chosen success criterion. All reported success rates depend on it; exact synthesis (τ=1.0) would yield lower numbers. Introduced in Sec. III-D.

axioms (4)

domain assumption PTM transforms multiplicatively under circuit composition
Used in Sec. IV-B to update the residual PTM by left-multiplying the inverse of the predicted gate. Standard property of the representation, not proven in the paper.
domain assumption The Pauli-rotation gate set e^{-iπ/8 P} is universal and equivalent to Clifford+T up to Clifford conjugation
Invoked in Sec. III-B to frame the task as Clifford+T synthesis; the specific equivalence to standard Clifford+T is stated, not derived.
domain assumption Randomly sampled circuits from the gate set provide a representative distribution for evaluating synthesis
All experiments rely on synthetic data drawn from products of Pauli rotations; the paper does not justify that this distribution matches practical synthesis workloads or the baselines' distributions.
domain assumption Held-out circuits have no overlap with training data
The paper states 'verified zero overlap against all training splits' but does not describe the verification method; since training data is generated on the fly, this is an assertion without a clear procedure.

pith-pipeline@v1.3.0-alltime-deepseek · 9942 in / 11764 out tokens · 125085 ms · 2026-08-02T11:34:00.646009+00:00 · methodology

0 comments

read the original abstract

Can Large Language Models (LLMs) understand and reason about quantum operators? Despite their remarkable capabilities in mathematics and symbolic reasoning, LLMs remain inherently blind to quantum representations such as unitary matrices. In this work, we take a step toward bridging this gap by introducing an approach that maps unitary operators into the latent space of an LLM, enabling unified modeling over quantum and linguistic inputs. We instantiate this idea on Clifford+T circuit synthesis over a Pauli rotation gate set, where our model achieves results competitive with state-of-the-art methods and scales consistently with training data, with no signs of saturation. Our approach further enables language-conditioned synthesis, allowing gate constraints unseen during training to be specified directly in natural language. This work suggests a path toward quantum--aware foundation models that can natively interpret and reason about quantum operations, which could have broader implications reaching across quantum compilation and algorithm discovery.

Figures

Figures reproduced from arXiv: 2606.13811 by David Kremer, Hang Hua, Pengyuan Li, Rogerio Feris, Yunchao Liu.

**Figure 3.** Figure 3: Fidelity over 800 synthesis steps on 200 Haar-random 4-qubit [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Fidelity traces for training oracle sequences and model predictions at test time. The model’s predictions generalize beyond the training data, most [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Text-conditioned synthesis on unseen constraint configurations. (a) Overall success rate (similar for the three settings) and gate-level constraint [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 7 linked inside Pith

[1]

Quantum verifiable rewards for post-training qiskit code assistant,

N. Dupuis, A. Tiwari, Y . Mroueh, D. Kremer, I. Faro, and J. Cruz-Benito, “Quantum verifiable rewards for post-training qiskit code assistant,” arXiv preprint arXiv:2508.20907, 2025

Pith/arXiv arXiv 2025
[2]

Granite-3.2-8B-Qiskit,

Qiskit Team, “Granite-3.2-8B-Qiskit,” https://huggingface.co/Qiskit/ granite-3.2-8b-qiskit, 2025

2025
[3]

Exploring llm-driven expla- nations for quantum algorithms,

G. d’Aloisio, S. Fortz, C. Hanna, D. Fortunato, A. Bensoussan, E. Mendiluze Usandizaga, and F. Sarro, “Exploring llm-driven expla- nations for quantum algorithms,” inProceedings of the 18th ACM/IEEE international symposium on empirical software engineering and mea- surement, 2024, pp. 475–481

2024
[4]

Qagent: An llm-based multi-agent system for autonomous openqasm programming,

Z. Fu, F. Chen, and L. Jiang, “Qagent: An llm-based multi-agent system for autonomous openqasm programming,”arXiv preprint arXiv:2508.20134, 2025

Pith/arXiv arXiv 2025
[5]

Granite vision: a lightweight, open-source multimodal model for enterprise intelligence,

G. V . Team, L. Karlinsky, A. Arbelle, A. Daniels, A. Nassar, A. Alfassi, B. Wu, E. Schwartz, D. Joshi, J. Kondicet al., “Granite vision: a lightweight, open-source multimodal model for enterprise intelligence,” arXiv preprint arXiv:2502.09927, 2025

Pith/arXiv arXiv 2025
[6]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, pp. 34 892– 34 916, 2023

2023
[7]

Synthetiq: Fast and versatile quantum circuit synthesis,

A. Paradis, J. Dekoninck, B. Bichsel, and M. Vechev, “Synthetiq: Fast and versatile quantum circuit synthesis,”Proceedings of the ACM on Programming Languages, vol. 8, no. OOPSLA1, pp. 55–82, 2024

2024
[8]

Unitary synthesis of clifford+ t circuits with reinforcement learning,

S. Rietsch, A. Y . Dubey, C. Ufrecht, M. Periyasamy, A. Plinge, C. Mutschler, and D. D. Scherer, “Unitary synthesis of clifford+ t circuits with reinforcement learning,” in2024 IEEE international conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2024, pp. 824–835

2024
[9]

Qiskit humaneval: An evaluation benchmark for quantum code generative mod- els,

S. Vishwakarma, F. Harkins, S. Golecha, V . S. Bajpe, N. Dupuis, L. Buratti, D. Kremer, I. Faro, R. Puri, and J. Cruz-Benito, “Qiskit humaneval: An evaluation benchmark for quantum code generative mod- els,” in2024 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2024, pp. 1169–1176

2024
[10]

Ketgpt–dataset augmenta- tion of quantum circuits using transformers,

B. Apak, M. Bandic, A. Sarkar, and S. Feld, “Ketgpt–dataset augmenta- tion of quantum circuits using transformers,” inInternational Conference on Computational Science. Springer, 2024, pp. 235–251

2024
[11]

Agent-q: fine-tuning large language models for quantum circuit generation and optimization,

L. Jern, V . Uotila, C. Yu, and B. Zhao, “Agent-q: fine-tuning large language models for quantum circuit generation and optimization,” in2025 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2025, pp. 1621–1632

2025
[12]

Quasar: Quantum assembly code generation using tool-augmented llms via agentic rl,

C. Yu, V . Uotila, S. Deng, Q. Wu, T. Shi, S. Jiang, L. You, and B. Zhao, “Quasar: Quantum assembly code generation using tool-augmented llms via agentic rl,”arXiv preprint arXiv:2510.00967, 2025

arXiv 2025
[13]

Quantumllminstruct: A 500k llm instruction-tuning dataset with problem-solution pairs for quantum computing,

S. Kashani, “Quantumllminstruct: A 500k llm instruction-tuning dataset with problem-solution pairs for quantum computing,”arXiv preprint arXiv:2412.20956, 2024

Pith/arXiv arXiv 2024
[14]

Synthesis of unitaries with Clifford+T circuits,

V . Kliuchnikov, D. Maslov, and M. Mosca, “Synthesis of unitaries with Clifford+T circuits,”Quantum Information & Computation, vol. 13, no. 7–8, pp. 607–630, 2013

2013
[15]

Representation of quantum circuits with Clifford andπ/8gates,

K. Matsumoto and K. Amano, “Representation of quantum circuits with Clifford andπ/8gates,”arXiv preprint arXiv:0806.3834, 2008

Pith/arXiv arXiv 2008
[16]

Optimal ancilla-free Clifford+T approxima- tion of z-rotations,

N. J. Ross and P. Selinger, “Optimal ancilla-free Clifford+T approxima- tion of z-rotations,”Quantum Information & Computation, vol. 16, no. 11–12, pp. 901–953, 2016

2016
[17]

Practical approximation schemes for single-qubit unitaries,

V . Kliuchnikov, A. Bocharov, and K. M. Svore, “Practical approximation schemes for single-qubit unitaries,”IEEE Transactions on Information Theory, vol. 69, no. 6, pp. 3912–3925, 2023

2023
[18]

The Solovay–Kitaev algorithm,

C. M. Dawson and M. A. Nielsen, “The Solovay–Kitaev algorithm,” Quantum Information & Computation, vol. 6, no. 1, pp. 81–95, 2006

2006
[19]

Practical and efficient quantum circuit synthesis and transpiling with reinforcement learning,

D. Kremer, V . Villar, H. Paik, I. Duran, I. Faro, and J. Cruz-Benito, “Practical and efficient quantum circuit synthesis and transpiling with reinforcement learning,”arXiv preprint arXiv:2405.13196, 2024

Pith/arXiv arXiv 2024
[20]

Optimizing the non-Clifford-count in unitary synthesis using reinforcement learning,

D. Kremer, A. Javadi-Abhari, and P. Mukhopadhyay, “Optimizing the non-Clifford-count in unitary synthesis using reinforcement learning,” arXiv preprint arXiv:2509.21709, 2025

arXiv 2025
[21]

Quantum circuit optimization with alphatensor,

F. J. Ruiz, T. Laakkonen, J. Bausch, M. Balog, M. Barekatain, F. J. Heras, A. Novikov, N. Fitzpatrick, B. Romera-Paredes, J. Van De We- teringet al., “Quantum circuit optimization with alphatensor,”Nature Machine Intelligence, vol. 7, no. 3, pp. 374–385, 2025

2025
[22]

Quantum circuit synthesis with diffusion models,

F. F ¨urrutter, G. Mu˜noz-Gil, and H. J. Briegel, “Quantum circuit synthesis with diffusion models,”Nature Machine Intelligence, vol. 6, no. 5, pp. 515–524, 2024

2024
[23]

Granite-4.0-micro,

IBM Granite Team, “Granite-4.0-micro,” https://huggingface.co/ ibm-granite/granite-4.0-micro, 2025

2025
[24]

Beyond reinforcement learning: Fast and scalable quantum circuit synthesis,

L. Theißinger, T. Gerlach, D. Berghaus, and C. Bauckhage, “Beyond reinforcement learning: Fast and scalable quantum circuit synthesis,” arXiv preprint arXiv:2602.15146, 2026

arXiv 2026
[25]

Deepseek-v3 technical report,

A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical report,”arXiv preprint arXiv:2412.19437, 2024

Pith/arXiv arXiv 2024

[1] [1]

Quantum verifiable rewards for post-training qiskit code assistant,

N. Dupuis, A. Tiwari, Y . Mroueh, D. Kremer, I. Faro, and J. Cruz-Benito, “Quantum verifiable rewards for post-training qiskit code assistant,” arXiv preprint arXiv:2508.20907, 2025

Pith/arXiv arXiv 2025

[2] [2]

Granite-3.2-8B-Qiskit,

Qiskit Team, “Granite-3.2-8B-Qiskit,” https://huggingface.co/Qiskit/ granite-3.2-8b-qiskit, 2025

2025

[3] [3]

Exploring llm-driven expla- nations for quantum algorithms,

G. d’Aloisio, S. Fortz, C. Hanna, D. Fortunato, A. Bensoussan, E. Mendiluze Usandizaga, and F. Sarro, “Exploring llm-driven expla- nations for quantum algorithms,” inProceedings of the 18th ACM/IEEE international symposium on empirical software engineering and mea- surement, 2024, pp. 475–481

2024

[4] [4]

Qagent: An llm-based multi-agent system for autonomous openqasm programming,

Z. Fu, F. Chen, and L. Jiang, “Qagent: An llm-based multi-agent system for autonomous openqasm programming,”arXiv preprint arXiv:2508.20134, 2025

Pith/arXiv arXiv 2025

[5] [5]

Granite vision: a lightweight, open-source multimodal model for enterprise intelligence,

G. V . Team, L. Karlinsky, A. Arbelle, A. Daniels, A. Nassar, A. Alfassi, B. Wu, E. Schwartz, D. Joshi, J. Kondicet al., “Granite vision: a lightweight, open-source multimodal model for enterprise intelligence,” arXiv preprint arXiv:2502.09927, 2025

Pith/arXiv arXiv 2025

[6] [6]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, pp. 34 892– 34 916, 2023

2023

[7] [7]

Synthetiq: Fast and versatile quantum circuit synthesis,

A. Paradis, J. Dekoninck, B. Bichsel, and M. Vechev, “Synthetiq: Fast and versatile quantum circuit synthesis,”Proceedings of the ACM on Programming Languages, vol. 8, no. OOPSLA1, pp. 55–82, 2024

2024

[8] [8]

Unitary synthesis of clifford+ t circuits with reinforcement learning,

S. Rietsch, A. Y . Dubey, C. Ufrecht, M. Periyasamy, A. Plinge, C. Mutschler, and D. D. Scherer, “Unitary synthesis of clifford+ t circuits with reinforcement learning,” in2024 IEEE international conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2024, pp. 824–835

2024

[9] [9]

Qiskit humaneval: An evaluation benchmark for quantum code generative mod- els,

S. Vishwakarma, F. Harkins, S. Golecha, V . S. Bajpe, N. Dupuis, L. Buratti, D. Kremer, I. Faro, R. Puri, and J. Cruz-Benito, “Qiskit humaneval: An evaluation benchmark for quantum code generative mod- els,” in2024 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2024, pp. 1169–1176

2024

[10] [10]

Ketgpt–dataset augmenta- tion of quantum circuits using transformers,

B. Apak, M. Bandic, A. Sarkar, and S. Feld, “Ketgpt–dataset augmenta- tion of quantum circuits using transformers,” inInternational Conference on Computational Science. Springer, 2024, pp. 235–251

2024

[11] [11]

Agent-q: fine-tuning large language models for quantum circuit generation and optimization,

L. Jern, V . Uotila, C. Yu, and B. Zhao, “Agent-q: fine-tuning large language models for quantum circuit generation and optimization,” in2025 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2025, pp. 1621–1632

2025

[12] [12]

Quasar: Quantum assembly code generation using tool-augmented llms via agentic rl,

C. Yu, V . Uotila, S. Deng, Q. Wu, T. Shi, S. Jiang, L. You, and B. Zhao, “Quasar: Quantum assembly code generation using tool-augmented llms via agentic rl,”arXiv preprint arXiv:2510.00967, 2025

arXiv 2025

[13] [13]

Quantumllminstruct: A 500k llm instruction-tuning dataset with problem-solution pairs for quantum computing,

S. Kashani, “Quantumllminstruct: A 500k llm instruction-tuning dataset with problem-solution pairs for quantum computing,”arXiv preprint arXiv:2412.20956, 2024

Pith/arXiv arXiv 2024

[14] [14]

Synthesis of unitaries with Clifford+T circuits,

V . Kliuchnikov, D. Maslov, and M. Mosca, “Synthesis of unitaries with Clifford+T circuits,”Quantum Information & Computation, vol. 13, no. 7–8, pp. 607–630, 2013

2013

[15] [15]

Representation of quantum circuits with Clifford andπ/8gates,

K. Matsumoto and K. Amano, “Representation of quantum circuits with Clifford andπ/8gates,”arXiv preprint arXiv:0806.3834, 2008

Pith/arXiv arXiv 2008

[16] [16]

Optimal ancilla-free Clifford+T approxima- tion of z-rotations,

N. J. Ross and P. Selinger, “Optimal ancilla-free Clifford+T approxima- tion of z-rotations,”Quantum Information & Computation, vol. 16, no. 11–12, pp. 901–953, 2016

2016

[17] [17]

Practical approximation schemes for single-qubit unitaries,

V . Kliuchnikov, A. Bocharov, and K. M. Svore, “Practical approximation schemes for single-qubit unitaries,”IEEE Transactions on Information Theory, vol. 69, no. 6, pp. 3912–3925, 2023

2023

[18] [18]

The Solovay–Kitaev algorithm,

C. M. Dawson and M. A. Nielsen, “The Solovay–Kitaev algorithm,” Quantum Information & Computation, vol. 6, no. 1, pp. 81–95, 2006

2006

[19] [19]

Practical and efficient quantum circuit synthesis and transpiling with reinforcement learning,

D. Kremer, V . Villar, H. Paik, I. Duran, I. Faro, and J. Cruz-Benito, “Practical and efficient quantum circuit synthesis and transpiling with reinforcement learning,”arXiv preprint arXiv:2405.13196, 2024

Pith/arXiv arXiv 2024

[20] [20]

Optimizing the non-Clifford-count in unitary synthesis using reinforcement learning,

D. Kremer, A. Javadi-Abhari, and P. Mukhopadhyay, “Optimizing the non-Clifford-count in unitary synthesis using reinforcement learning,” arXiv preprint arXiv:2509.21709, 2025

arXiv 2025

[21] [21]

Quantum circuit optimization with alphatensor,

F. J. Ruiz, T. Laakkonen, J. Bausch, M. Balog, M. Barekatain, F. J. Heras, A. Novikov, N. Fitzpatrick, B. Romera-Paredes, J. Van De We- teringet al., “Quantum circuit optimization with alphatensor,”Nature Machine Intelligence, vol. 7, no. 3, pp. 374–385, 2025

2025

[22] [22]

Quantum circuit synthesis with diffusion models,

F. F ¨urrutter, G. Mu˜noz-Gil, and H. J. Briegel, “Quantum circuit synthesis with diffusion models,”Nature Machine Intelligence, vol. 6, no. 5, pp. 515–524, 2024

2024

[23] [23]

Granite-4.0-micro,

IBM Granite Team, “Granite-4.0-micro,” https://huggingface.co/ ibm-granite/granite-4.0-micro, 2025

2025

[24] [24]

Beyond reinforcement learning: Fast and scalable quantum circuit synthesis,

L. Theißinger, T. Gerlach, D. Berghaus, and C. Bauckhage, “Beyond reinforcement learning: Fast and scalable quantum circuit synthesis,” arXiv preprint arXiv:2602.15146, 2026

arXiv 2026

[25] [25]

Deepseek-v3 technical report,

A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical report,”arXiv preprint arXiv:2412.19437, 2024

Pith/arXiv arXiv 2024