PaperlayerView exact revision on arXiv

Source-linked AI summary

Aligning Quantum Operators with Large Language Models

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

arXiv:2606.13811v2quant-phcs.AI

TL;DR

LLMs lack mechanisms to process and reason over the unitary matrices that define quantum operations. This paper maps quantum operators into an LLM’s latent space for 4-qubit Clifford+T synthesis, achieving strong synthesis results that scale with data and inference-time compute while supporting text-conditioned synthesis.

  • Problem

    LLMs process symbolic descriptions of quantum objects but lack mechanisms to ingest, interpret, or reason over unitary matrices, which define quantum operations.

  • Method

    The approach maps 4-qubit unitary operators into an LLM’s latent space and synthesizes circuits over a Pauli-rotation gate set.

  • Results

    99.4% overall success is achieved with best-of-80 sampling, while success rate improves more than 3× as training data grows from 145K to 9.2M circuits.

  • Takeaways & Limitations

    The work demonstrates a proof of concept for models that unify natural language and quantum representations for quantum circuit synthesis.

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper addresses LLMs’ inability to process quantum operators directly by aligning unitary operators with an LLM’s latent space. It validates this approach on 4-qubit Clifford+T synthesis, achieving data-scalable results and enabling language-conditioned quantum synthesis.

  • Motivation: Existing quantum-LLM systems use symbolic descriptions, but lack mechanisms to ingest, interpret, or reason over unitary matrices with complex-valued numerical structure.This blind spot limits LLM application to quantum tasks involving the mathematical objects that define quantum operations.
  • Method: The proposed alignment approach maps real-valued Pauli Transfer Matrices into a pretrained LLM’s latent space using a lightweight encoder and projector.Quantum operators are presented as visual inputs alongside textual context and translated into word embeddings understandable by the LLM.
  • Contribution: The work claims the first approach enabling an LLM to condition directly on quantum operators rather than textual or programmatic descriptions.This establishes direct operator conditioning as the paper’s central novelty.
  • Evaluation: The method is evaluated on 4-qubit unitary synthesis with a Pauli-rotation gate set, connecting the task to Clifford+T circuits while providing a uniform 256-way action space.The PTM representation scales as 4n×4n, like all full-matrix representations.
  • Results: More than 3× success-rate improvement occurs as training data grows from 145K to 9.2M circuits for 4-qubit Clifford+T synthesis with 1–15 gates.Best-of-N sampling further improves inference performance, surpassing simulated annealing and prior reinforcement-learning approaches.
  • Broader implications: The alignment framework suggests quantum-aware multimodal models that share an embedding space across natural language and modalities such as Clifford tableaux, Pauli lists, and tensor-network descriptions.Additional quantum modalities could be incorporated through modality-specific encoders projecting into the same LLM token space.

II. RELATED WORK · III. PRELIMINARIES: QUANTUM CIRCUIT SYNTHESIS · A. Quantum circuit synthesis.

Prior work spans LLM-based quantum code generation, text- and graph-conditioned synthesis, classical approximation, and reinforcement-learning methods, while this approach uses supervised fine-tuning without reinforcement learning. Quantum circuit synthesis decomposes a unitary operation into a short sequence of gates from a fixed set, an intractable search problem for long circuits.

  • II. RELATED WORK: LLM–quantum-computing research remains sparse, covering Qiskit and OpenQASM generation as well as text- and graph-conditioned circuit synthesis.Examples include Granite for Qiskit, Qiskit HumanEval, KetGPT, Agent-Q, and QUASAR.
  • II. RELATED WORK: Classical exact Clifford+T and approximate single-qubit synthesis methods offer optimality guarantees but scale poorly to multi-qubit unitaries.The Solovay–Kitaev theorem gives O(log^c(1/ϵ)) gate scaling for single-qubit approximation, and gridsynth achieves near-optimal gate counts in practice.
  • II. RELATED WORK: RL-based synthesis methods target unitary specifications, hardware constraints, Pauli-rotation representations, and reduced T-counts.The cited methods include Rietsch et al., Kremer et al., and AlphaTensor-Quantum.
  • II. RELATED WORK: RL-based approaches require reward shaping, extensive hyperparameter tuning, and substantial environment interaction.These requirements motivate contrast with the paper’s RL-free training strategy.
  • II. RELATED WORK: The proposed approach is RL-free and uses only supervised fine-tuning with a standard next-token prediction loss.This distinguishes its training procedure from the reinforcement-learning methods discussed above.
  • A. Quantum circuit synthesis.: An n-qubit operation is represented by a d × d unitary matrix U ∈ U(d), with d = 2^n.The unitary representation defines the target object for circuit synthesis.
  • A. Quantum circuit synthesis.: Quantum circuit synthesis decomposes a target unitary U into a sequence of elementary gates from a fixed gate set G, seeking the smallest possible K.The circuit satisfies g_K · · · g_2 g_1 ≈ U with each g_t ∈ G.
  • A. Quantum circuit synthesis.: The synthesis search space grows as |G|^K, making exhaustive search intractable except for very short circuits.This exponential growth is a central computational challenge in quantum circuit synthesis.

B. Clifford+T synthesis in a Pauli-rotation basis. · C. Pauli Transfer Matrix (PTM).

The paper formulates Clifford+T synthesis using universal, noncommuting n-qubit Pauli π/8 rotations and represents unitaries through real-valued Pauli Transfer Matrices. PTMs preserve global-phase invariance and compose multiplicatively, enabling stepwise circuit modeling.

  • B. Clifford+T synthesis in a Pauli-rotation basis.: Clifford gates are classically efficiently simulable, while adding the T gate yields a universal set that approximates any unitary arbitrarily precisely.The T gate is the π/8 phase gate.
  • B. Clifford+T synthesis in a Pauli-rotation basis.: The gate set is parameterized with n-qubit Pauli rotations built from tensor products of I, X, Y, and Z.The n-qubit Pauli group contains 4^n operators.
  • B. Clifford+T synthesis in a Pauli-rotation basis.: Each Pauli rotation applies a π/8 rotation associated with an n-qubit Pauli operator.The supplied passage introduces these gates as π/8-rotations; its displayed formula is not preserved in the excerpt.
  • B. Clifford+T synthesis in a Pauli-rotation basis.: For n = 4, the Pauli-rotation gate set contains |G| = 256 gates labeled by four-character Pauli strings, and circuit lengths are measured in this set.The rotations are universal, generally noncommuting, and their order therefore matters.
  • B. Clifford+T synthesis in a Pauli-rotation basis.: Each nonidentity Pauli rotation can be viewed as a single-qubit T gate conjugated by an n-qubit Clifford unitary, connecting the basis to standard Clifford+T.This relates the Pauli-rotation representation to the standard Clifford+T gate set.
  • C. Pauli Transfer Matrix (PTM).: Instead of complex unitary matrices, the method uses a real-valued 4^n × 4^n Pauli Transfer Matrix representation for a unitary U.The PTM is defined through its action on Pauli operators.
  • C. Pauli Transfer Matrix (PTM).: PTMs are invariant to global phase and compose multiplicatively, so a circuit PTM equals the product of its gates’ PTMs.This multiplicative composition property is central to the paper’s stepwise approach.

D. Fidelity.

Synthesis progress is measured by the channel fidelity between the residual Pauli Transfer Matrix and the identity. Exact synthesis occurs if and only if fidelity equals 1, with success defined as F ≥0.999.

  • Fidelity: Synthesis progress is measured using channel fidelity between the residual PTM and the identity.The fidelity evaluates how closely the residual operation matches the identity channel.
  • Fidelity: Fidelity equals 1 if and only if the target has been exactly synthesized.Exact synthesis is therefore characterized by unit channel fidelity.
  • Fidelity: Synthesis is deemed successful when F ≥0.999.This threshold operationalizes successful synthesis in the evaluation.

IV. APPROACH · A. PTM-Language Alignment

The approach aligns quantum unitary operators with a pretrained LLM’s latent space for autoregressive circuit compilation. It combines PTM encoding and projection with stepwise synthesis that predicts Pauli rotation gates approximating a target unitary.

  • IV. APPROACH: The framework maps quantum unitary operators into a pretrained LLM’s latent space for autoregressive circuit compilation.Given a target unitary, it iteratively predicts Pauli rotation gates whose composition approximates the target with high fidelity.
  • IV. APPROACH: The method combines a lightweight PTM encoder-projector with a stepwise synthesis procedure.The encoder-projector aligns the PTM representation with the LLM embedding space, while synthesis predicts gates iteratively.
  • A. PTM-Language Alignment: For n=4 qubits, the PTM is a 256×256 real matrix normalized element-wise to [−1, 1] by dividing by its maximum absolute value.The bounded PTM is treated as a single-channel image before neural encoding.
  • A. PTM-Language Alignment: The PTM encoder partitions the matrix into 16 × 16 non-overlapping patches, producing V = 256 patch vectors of dimension 256.Each patch is projected to hidden dimension hv = 768, followed by layer normalization and learned positional embedding.
  • A. PTM-Language Alignment: A two-layer MLP with GELU activation maps the PTM embeddings to the LLM dimension dLLM.The projector follows the patch encoder and prepares the representation for integration with text embeddings.
  • A. PTM-Language Alignment: The V =256 visual tokens are prepended to text token embeddings, forming the input processed by the LLM.The resulting sequence is [v1, . . . , vV , t1, . . . , tL].
  • A. PTM-Language Alignment: The encoder and projector introduce ∼14M parameters, less than 0.4% of the total, while leaving the LLM architecture unchanged.This keeps the alignment components lightweight relative to the pretrained language model.

B. Stepwise Autoregressive Synthesis

The method performs quantum compilation autoregressively, predicting one gate at a time from the residual portion of the target unitary. After each prediction, it updates the residual and repeats until the synthesized unitary is complete.

  • Stepwise synthesis: The model predicts one gate at a time conditioned on the residual PTM, rather than generating the full gate sequence simultaneously.The residual represents the portion of the target that remains to be synthesized.
  • Stepwise synthesis: Gates are predicted in reverse execution order by peeling off the leftmost remaining factor of the residual circuit decomposition.The residual PTM is initialized from the target unitary and the model predicts gT−1, gT−2, …, g0.
  • Residual update: After each predicted gate, the residual PTM is left-multiplied by the gate’s inverse PTM, ideally driving the residual toward identity and fidelity toward 1.The update is P(t+1) = PTM(ĝ)^−1P(t) = PTM(ĝ)^⊤P(t).
  • Model interface: Each step combines visual residual-PTM tokens, textual context containing fidelity and prior gates, and an instruction prefix, then outputs a single gate before external residual updating.This multimodal input-output cycle repeats throughout synthesis.
  • Training objective: Training uses causal language modeling to predict the next gate from visual, context, and instruction tokens, with gradients computed only at the gate position.Samples are generated on the fly from random Clifford+T circuits, including a final END prediction when the residual fidelity is approximately 1.

C. Two-Stage Training

Training proceeds in two stages: projector alignment with the LLM frozen, followed by joint fine-tuning with differential learning rates and a Warmup–Stable–Decay schedule.

  • Stage 1: Projector alignment: Stage 1 optimizes only the vision encoder and projector while freezing the LLM, establishing cross-modal alignment without perturbing pretrained representations.Training uses ∼7K steps with learning rate 10−3 and cosine decay.
  • Stage 2: Joint fine-tuning: Stage 2 jointly optimizes all parameters using a lower ηLLM and higher ηproj ≈4ηLLM for vision components.The WSD schedule warms up linearly, remains constant for 75% of training, then decays to 5% of peak over the final 25%.

V. EXPERIMENTS · A. Circuit Synthesis Results: Data Scaling · B. Inference-Time Scaling

Experiments evaluate Clifford+T synthesis for 4-qubit unitaries, showing consistent gains as training data increases and substantial accuracy improvements from best-of-N inference-time sampling. Sampling especially benefits longer circuits, indicating that stochastic exploration helps uncover synthesis paths missed by greedy decoding.

  • V. EXPERIMENTS: The evaluation targets Clifford+T synthesis of 4-qubit unitaries using Granite 4.0 Micro, a 3B-parameter LLM backbone.Models use 2 nodes × 8 GPUs, an effective batch size of 64, held-out circuits with zero overlap against training splits, and fidelity threshold τ = 0.999.
  • A. Circuit Synthesis Results: Data Scaling: Training varies circuit counts from 145K to 9.2M while keeping all other hyperparameters fixed.Models train on circuits containing 1–15 gates, and evaluation uses 2,000 held-out circuits.
  • A. Circuit Synthesis Results: Data Scaling: More than 3× improvement in success rate is achieved as training data grows from 145K to 9.2M circuits.Table I reports success rate and mean fidelity on the held-out set, with performance improving consistently with additional training data.
  • B. Inference-Time Scaling: Best-of-N sampling trades inference-time compute for accuracy by selecting the highest-fidelity result from N independent synthesis rollouts.The first rollout uses greedy decoding, the remaining N−1 use temperature 0.7, and sampling stops early once fidelity exceeds τ.
  • B. Inference-Time Scaling: 87.9% success is obtained with greedy decoding alone at N=1, rising to 97.1% at N=10 and 99.4% at N=80.The total improvement from sampling is 11.5 percentage points.
  • B. Inference-Time Scaling: The gains are roughly log-linear in N and concentrated on longer 11–15-gate circuits.Stochastic exploration discovers synthesis paths that greedy decoding misses, indicating a well-calibrated distribution over synthesis paths rather than collapse to one strategy.

C. Baselines · D. Haar Random Unitaries

The model outperforms four baselines on held-out 4-qubit Clifford+T circuits, reaching 99.4% overall success with best-of-80 sampling and retaining above 94% accuracy at 15 gates. On out-of-distribution Haar-random unitaries, longer-circuit training substantially improves fidelity, although results remain far below the threshold for exact synthesis.

  • C. Baselines: The evaluation compares greedy search, SynthetiQ, Rietsch et al.’s RL method, and Theißinger et al.’s MDL beam search on 2,000 held-out circuits spanning 1–15 gates.SynthetiQ was rerun using public code, whereas RL and MDL results are published approximate comparisons.
  • C. Baselines: 13.8% overall success is achieved by greedy search, which collapses beyond 3 gates.SynthetiQ reaches 62.7% overall, while RL and MDL beam search achieve 83.7% and 68.8%, respectively, with MDL surpassing RL on longer circuits.
  • C. Baselines: 87.9% success is achieved by the model with greedy decoding, surpassing all baselines.Best-of-80 sampling raises overall success to 99.4%, while accuracy remains above 94% at 15 gates.
  • C. Baselines: Approximately 1 s per sample is required for model inference on a single NVIDIA H100 GPU, or roughly 80 s for best-of-80 sampling.MDL beam search reports 22 s per sample as a reference.
  • D. Haar Random Unitaries: Haar-random unitaries lie outside the training distribution because they are generally not expressible as finite Clifford+T circuits.The study therefore treats their compilation as approximate synthesis and uses them as a stress test for meaningful progress toward arbitrary unitaries.
  • D. Haar Random Unitaries: The 1–15-gate model makes limited progress on 200 Haar-random unitaries, with mean fidelity plateauing below 0.02 over 800 synthesis steps.These fidelities remain far below the ≥0.999 threshold needed for exact synthesis.
  • D. Haar Random Unitaries: Training on 1–150-gate circuits with only 1M circuits achieves substantially higher mean fidelity than training on 1–15-gate circuits.The improvement over 800 steps suggests that extending the training gate range improves generalization to arbitrary unitaries.

E. Patch Size Ablation · F. Qualitative Results · G. Text-Conditioned Circuit Synthesis

The paper studies patch-size trade-offs, qualitative fidelity behavior, and text-conditioned synthesis under unseen gate-placement constraints. The model generalizes beyond training examples and uses natural-language instructions to control circuit generation, with pretrained initialization improving constraint compliance.

  • E. Patch Size Ablation: Smaller visual patches preserve finer PTM structure but increase the visual-token sequence length attended to by the language model.The ablation considers P ∈ 8, 16, 32, 64, 256, corresponding to 1024, 256, 64, 16, and 1 visual tokens.
  • E. Patch Size Ablation: The patch-size ablation trains all variants on 1.15M four-qubit circuits with 1–15 gates and evaluates them on 2000 held-out circuits using greedy autoregressive decoding.Models use identical hyperparameters, isolating the effect of patch size.
  • F. Qualitative Results: Test-time fidelity traces show predictions generalizing beyond imitation on trajectories entirely absent from the training distribution.The qualitative behavior suggests the model internalizes PTM structure rather than memorizing input–output mappings.
  • F. Qualitative Results: In a highlighted example, fidelity rises, drops sharply, and then recovers to 1.0 as the model reverses course.This trajectory demonstrates non-monotonic correction during generation.
  • G. Text-Conditioned Circuit Synthesis: Natural-language prompts restrict which qubits or qubit pairs particular gates may act on, supporting hardware-aware compilation and routing scenarios.The constraints correspond to physically realizable interactions and restrictions such as CNOT placement.
  • G. Text-Conditioned Circuit Synthesis: The text-conditioned experiment trains on 3M circuits of 1–15 gates using a ten-gate set, with randomly sampled placement constraints present in half the prompts.The gate set is {H, T, T†, S, S†, X, Y, Z, CNOT, CZ}, and constrained prompts specify one or two restrictions.
  • G. Text-Conditioned Circuit Synthesis: 91% compliance is achieved by the LLM-initialized model with constraint prompts on 250 circuits containing five unseen constraint combinations.The test set ranges from single-gate restrictions to dual constraints blacklisted during training.
  • G. Text-Conditioned Circuit Synthesis: Removing constraint text lowers compliance to 53%, while random initialization reaches 65%, showing active instruction conditioning and the value of pretrained language understanding.Both comparisons use the same unseen placement-constraint evaluation setting.

VI. CONCLUSION

The paper maps quantum unitary operators into an LLM’s latent space for native quantum reasoning and demonstrates the approach on Clifford+T circuit synthesis. The authors report text-conditioned synthesis, scaling with training data and inference-time compute, and several directions for future expansion.

  • Conclusion: The approach maps quantum unitary operators into an LLM’s latent space, enabling native reasoning over quantum representations.It is instantiated on Clifford+T circuit synthesis for 4-qubit unitaries.
  • Conclusion: The method achieves strong 4-qubit synthesis results, including text-conditioned synthesis, using only supervised fine-tuning.Performance scales consistently with both training data and inference-time compute without signs of saturation.
  • Conclusion: Future work includes scaling to larger qubit counts and deeper circuits, augmenting training with latent reasoning and GRPO, and incorporating different circuit and operator representations.The authors envision quantum–language models unifying textual context, instruction following, and direct operator reasoning within one system.
Loading 2606.13811v2…