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

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

IEEE Access

Instance-Wise Mixed Integer Programming Solver Configuration for Quantum Circuit Decomposition via Large Language Model Distillation

Yuheon Song, Sungbin Lim

Published Jan 1, 2026• 0 citations
Publisher

Abstract

Quantum Circuit Decomposition (QCD) aims to realize target unitary operations under stringent physical and logical constraints using limited native gate sets. However, standard data-driven approaches struggle to enforce these constraints explicitly, failing to guarantee physically valid and logically equivalent circuits. Mixed Integer Programming (MIP) formulations address this limitation by encoding such constraints, enabling globally optimal, constraint-satisfying solutions. MIP solvers expose large configuration spaces that affect search performance, and effective configurations vary substantially across instances. This motivates instance-specific approaches rather than a single global tuning strategy. We propose a two-stage instance-wise configuration optimization framework for QCD MIPs. Stage 1 trains a Graph Neural Network (GNN)–based reinforcement learning policy on bipartite graphs that capture MIP structure, using a novel instance-wise relative reward to stabilize learning under heterogeneous and timeout-dominated solver feedback. Stage 2 distills this graph-informed policy into a graph-free Large Language Model (LLM) student operating directly on QCD-level inputs. Both proposed policies outperform the solver’s built-in default configuration and remain competitive with automatic configurators. On the 4-qubit benchmark, LLM-Student solves an average of 53.4 instances versus 2.2 under the default, and its average geometric mean solving time on commonly solved instances is 25.2% and 28.5% lower than those of SMAC3 and irace, respectively. It remains competitive with the GNN-Teacher while avoiding approximately nine hours of graph construction per 4-qubit instance at inference time. These results suggest that combining graph-based policy learning with graph-free distillation provides a practical and promising direction for learning-based optimization of QCD MIPs and beyond.

DOI
10.1109/access.2026.3746321

Related papers

Rethinking Quantum Circuit Optimization as a Reinforcement Learning Problem

Arjun Mehta, Sven Richter

Quantum circuit optimization is a foundational challenge in realizing practical quantum computation, requiring the reduction of gate counts and circuit depth while preserving logical equivalence across noisy intermediate-scale quantum (NISQ) devices. Conventional rule-based and heuristic compilation approaches demonstrate diminishing effectiveness as circuit complexity grows, motivating the search for adaptive, data-driven alternatives. This paper reformulates quantum circuit optimization as a Markov decision process (MDP), enabling deep reinforcement learning (RL) agents to autonomously discover efficient transformation strategies. A graph neural network (GNN)-based actor-critic architecture is developed to encode circuit topology and guide transformation selection. Drawing on empirical evidence from phase-structured quantum control tasks, neural feedback architectures, and convergence behavior in active learning agents, we demonstrate that the proposed framework achieves an average two-qubit gate count reduction of 31.4% and a circuit depth reduction of 27.8% across standard benchmark circuits, outperforming established compilation baselines. This work provides a principled theoretical foundation for the RL-as-optimizer paradigm and contributes an extensible experimental platform suited to NISQ-era constraints.

American Journal Of Big Data•Apr 1, 2026•0 citations

RubriQ: Rubric-Guided Group Relative Policy Optimization for Constraint-Aware Quantum Circuit Synthesis

Ziqing Guo, Ziwen Pan

Designing fault-tolerant quantum circuits that are both algorithmically correct and hardware compatible remains a major bottleneck in the transition to scalable quantum computing. We introduce RubriQ, a scalable framework that formulates circuit synthesis as a large language model (LLM) code-generation task, optimized via group relative policy optimization (GRPO). Unlike conventional black-box neural critics, RubriQ employs a domain-grounded programmatic rubric as the reinforcement learning reward function, evaluating circuits for T-gate reduction, hardware topology compliance, and unitary fidelity. To support high-throughput training, RubriQ integrates GPU-accelerated CUDA-Q simulation directly into the reinforcement learning (RL) loop and is deployed on NERSC Perlmutter using DeepSpeed ZeRO2 across multinode NVIDIA A100 clusters. On benchmark tasks, RubriQ achieves a mean T-gate compression of 3.31x, significantly outperforming sparse-reward RL baselines (2.05x), converging 2-3x faster, and maintaining less than 1\% hardware-constraint violations. Validated on IBM and IonQ quantum processors, RubriQ establishes an automated, high-performance computing (HPC)-driven pipeline for generating hardware-ready, fault-tolerant quantum circuits at scale.

quant-pharXiv:2607.07554•Jul 9, 2026•0 citations

Architecture-Aware Reinforcement Learning for Communication-Efficient Distributed Quantum Circuit Compilation

Chien-Tung Kuo, Felix Burt, Samuel Yen-Chi Chen +2 more

Distributed quantum computing provides a scalable route for executing quantum circuits beyond the capacity limits of a single quantum processing unit (QPU), but it introduces a communication-aware compilation problem involving strict hardware constraints and circuit dependencies. This paper presents an architecture-aware reinforcement-learning framework that formulates distributed quantum compilation as a constrained Markov Decision Process (MDP). The compiler-level communication actions dynamically update logical-qubit placement and enable subsequent gate execution. A heterogeneous graph model represents interactions among hardware, logical qubits, and circuit operations, while a policy trained via Proximal Policy Optimization optimizes EPR-pair consumption and communication makespan. Evaluation across benchmark circuits shows that our policy matches state-of-the-art heuristics on structured workloads, with lookahead reward shaping yielding modest improvements on unstructured circuits. These results demonstrate that reinforcement learning is a flexible alternative to manual heuristics, though scalability remains a key bottleneck for practical use.

quant-pharXiv:2608.06892•Aug 10, 2026•1 citation

A mixed-integer program for circuit execution time minimization with precedence constraints

Mostafa Atallah, James Ostrowski, Rebekah Herrman

We present a mixed-integer programming (MIP) model for scheduling quantum circuits to minimize execution time. Our approach maximizes parallelism by allowing non-overlapping gates (those acting on distinct qubits) to execute simultaneously. Our methods apply to general circuits with precedence constraints. First, we derive closed-formulas for the execution time of circuits generated by ma-QAOA on star graphs for a layered, greedy, and MIP schedules. We then compare the MIP schedule against layered and greedy scheduling approaches on the circuits generated by ma-QAOA for solving the MaxCut problem on all non-isomorphic connected graphs with 3-7 vertices. These experiments demonstrate that the MIP scheduler consistently results in shorter circuit execution times than greedy and layered approaches, with up to 24\% savings.

arXiv:2504.09268•Apr 12, 2025•0 citations

QPLAN: Deep Reinforcement Learning Assisted Quantum Network Planning

Yuhang Gan, Shiyi Ling, Ruilin Zhou +2 more

Quantum networks attract increasing attention due to their unique advantages over classical networks. However, the challenge of efficiently and economically constructing reliable wide-area quantum networks remains unresolved. This problem, termed network planning, is typically modeled as an Integer Linear Programming (ILP) problem. Existing solver-based solutions for classical networks face scalability issues, while heuristic approaches that sacrifice optimality for scalability heavily rely on human expertise and lack generalizability. Recently, deep reinforcement learning (DRL) has demonstrated potential in adaptive decision-making for classical network planning. However, quantum networks introduce unique constraints due to their end-to-end entanglement distribution requirements. To address this, we propose a two-stage quantum network resource allocation framework, QPLAN, that combines multi-agent DRL with traditional optimization solvers. Our approach uses a novel multi-agent reinforcement learning strategy using graph neural networks (GNNs) and transformers for traffic flow and corresponding hardware resource allocation to minimize hardware cost while cooperatively satisfying problem constraints as much as possible. The output of the DRL algorithm is then used as an initial input to prune the solution space, "warm-starting" the ILP solver and further improving the solution. This hybrid approach effectively reduces manual effort and enhances scalability. Experimental results show that our method yields near-optimal solutions with better scalability than only using ILP solvers.

International Conference on Quantum Communications, Networking, and Computing (QCNC)•Mar 31, 2025•1 citation

Machine learning strategies for near-term quantum algorithms

Y.J. Patel

The central claim of this thesis is simple but underappreciated: the primary bottleneck in near-term quantum computing is not a shortage of algorithms; it is a lack of systematic decision-making. Which problem instances should we test? Which circuit architectures are worth training? Which hyperparameters actually drive performance, and which are distractions? Today’s quantum hardware sits in an uncomfortable middle ground. It is expressive enough to explore nontrivial quantum dynamics, yet it is too noisy and shallow for the theoretical speedups that motivate the field. This thesis argues that applying machine learning to classical control decisions in hybrid quantum classical methods is one of the most tractable ways to make these methods genuinely useful in this regime. This argument is developed through four research questions, each addressing a distinct decision-making layer and illustrating the thesis. The first asks whether the decisions inside a hybrid heuristic can be learned rather than hand-coded. Chapter 4 analyzes Recursive QAOA (RQAOA), a promising yet brittle approach to combinatorial optimization, and identifies regimes in which its fixed-elimination strategy fails. It then replaces that strategy with a reinforcement learning policy, RL-RQAOA, that adapts to instance structure. The resulting method performs better on the constructed failure cases while remaining competitive elsewhere. The broader point is that learned classical control policies can materially improve hybrid quantum heuristics without increasing quantum depth, which is a significant practical advantage on constrained hardware. The second question evaluates whether the quantum optimization algorithm benchmarking instances we use are principled. Chapter 5 observes that much benchmarking in this space is driven by convenience. Researchers test on whatever graphs are readily available and then report aggregate performance. Instead, the chapter proposes a generative pipeline that couples a permutation-invariant graph autoencoder with evolutionary optimization to deliberately construct Max-Cut instances that strongly discriminate between a quantum-inspired solver, RQAOA, and a strong classical baseline, Goemans-Williamson, in both directions. By connecting solver behavior to structural graph features, the chapter turns benchmarking from a static choice into an interpretable experimental tool. The generated instances also expose failure modes that random graphs often fail to reveal. The third question asks what changes when realistic noise constraints are taken seriously during circuit design, not only during execution. Chapter 6 develops RLQAS, a reinforcement learning framework for quantum architecture search that incorporates noise assumptions informed by real IBM device characteristics throughout training, rather than treating noise as a post hoc correction. Paired with an efficient Pauli transfer matrix simulation, the framework makes large-scale noisy circuit search computationally feasible. The circuits it discovers are compact, trainable, and achieve competitive accuracy on ground-state energy tasks. This suggests that hardware-aware circuit design is tractable and can yield materially different choices than noise-agnostic search. The fourth question asks: even when the circuit architecture is fixed, how much of what we attribute to quantum model design is actually due to hyperparameter sensitivity? Chapter 7 extends functional ANOVA to quantum neural networks to quantify which hyperparameter choices drive performance variance. The finding is clear. Learning rate, circuit depth, and data encoding/reuploading dominate, while many architectural micro-choices contribute far less than the literature often implies. Rather than stopping at diagnosis, the chapter turns these importance rankings into data-driven priors that accelerate hyperparameter optimization. It converts a critique into a practical improvement. Taken together, these contributions form a coherent toolkit for developing and evaluating hybrid quantum–classical methods in the NISQ regime. While the work does not eliminate the fundamental constraints of shallow, noisy hardware, it clarifies what can be improved immediately and how to do so in a reproducible manner. The broader goal is to help shift near-term quantum research from isolated demonstrations toward an engineering discipline, in which algorithms, benchmarks, and model-selection practices can be compared fairly, stress-tested rigorously, and iterated systematically as hardware matures.

•Nov 10, 2026•0 citations
Quantum Circuit Decomposition (QCD) aims to realize target unitary operations under stringent physical and logical constraints using limited native gate sets. However, standard data-driven approaches struggle to enforce these constraints explicitly, failing to guarantee physically valid and logically equivalent circuits. Mixed Integer Programming (MIP) formulations address this limitation by encoding such constraints, enabling globally optimal, constraint-satisfying solutions. MIP solvers expose large configuration spaces that affect search performance, and effective configurations vary substantially across instances. This motivates instance-specific approaches rather than a single global tuning strategy. We propose a two-stage instance-wise configuration optimization framework for QCD MIPs. Stage 1 trains a Graph Neural Network (GNN)–based reinforcement learning policy on bipartite graphs that capture MIP structure, using a novel instance-wise relative reward to stabilize learning under heterogeneous and timeout-dominated solver feedback. Stage 2 distills this graph-informed policy into a graph-free Large Language Model (LLM) student operating directly on QCD-level inputs. Both proposed policies outperform the solver’s built-in default configuration and remain competitive with automatic configurators. On the 4-qubit benchmark, LLM-Student solves an average of 53.4 instances versus 2.2 under the default, and its average geometric mean solving time on commonly solved instances is 25.2% and 28.5% lower than those of SMAC3 and irace, respectively. It remains competitive with the GNN-Teacher while avoiding approximately nine hours of graph construction per 4-qubit instance at inference time. These results suggest that combining graph-based policy learning with graph-free distillation provides a practical and promising direction for learning-based optimization of QCD MIPs and beyond.
Quantum circuit optimization is a foundational challenge in realizing practical quantum computation, requiring the reduction of gate counts and circuit depth while preserving logical equivalence across noisy intermediate-scale quantum (NISQ) devices. Conventional rule-based and heuristic compilation approaches demonstrate diminishing effectiveness as circuit complexity grows, motivating the search for adaptive, data-driven alternatives. This paper reformulates quantum circuit optimization as a Markov decision process (MDP), enabling deep reinforcement learning (RL) agents to autonomously discover efficient transformation strategies. A graph neural network (GNN)-based actor-critic architecture is developed to encode circuit topology and guide transformation selection. Drawing on empirical evidence from phase-structured quantum control tasks, neural feedback architectures, and convergence behavior in active learning agents, we demonstrate that the proposed framework achieves an average two-qubit gate count reduction of 31.4% and a circuit depth reduction of 27.8% across standard benchmark circuits, outperforming established compilation baselines. This work provides a principled theoretical foundation for the RL-as-optimizer paradigm and contributes an extensible experimental platform suited to NISQ-era constraints.
Designing fault-tolerant quantum circuits that are both algorithmically correct and hardware compatible remains a major bottleneck in the transition to scalable quantum computing. We introduce RubriQ, a scalable framework that formulates circuit synthesis as a large language model (LLM) code-generation task, optimized via group relative policy optimization (GRPO). Unlike conventional black-box neural critics, RubriQ employs a domain-grounded programmatic rubric as the reinforcement learning reward function, evaluating circuits for T-gate reduction, hardware topology compliance, and unitary fidelity. To support high-throughput training, RubriQ integrates GPU-accelerated CUDA-Q simulation directly into the reinforcement learning (RL) loop and is deployed on NERSC Perlmutter using DeepSpeed ZeRO2 across multinode NVIDIA A100 clusters. On benchmark tasks, RubriQ achieves a mean T-gate compression of 3.31x, significantly outperforming sparse-reward RL baselines (2.05x), converging 2-3x faster, and maintaining less than 1\% hardware-constraint violations. Validated on IBM and IonQ quantum processors, RubriQ establishes an automated, high-performance computing (HPC)-driven pipeline for generating hardware-ready, fault-tolerant quantum circuits at scale.
Distributed quantum computing provides a scalable route for executing quantum circuits beyond the capacity limits of a single quantum processing unit (QPU), but it introduces a communication-aware compilation problem involving strict hardware constraints and circuit dependencies. This paper presents an architecture-aware reinforcement-learning framework that formulates distributed quantum compilation as a constrained Markov Decision Process (MDP). The compiler-level communication actions dynamically update logical-qubit placement and enable subsequent gate execution. A heterogeneous graph model represents interactions among hardware, logical qubits, and circuit operations, while a policy trained via Proximal Policy Optimization optimizes EPR-pair consumption and communication makespan. Evaluation across benchmark circuits shows that our policy matches state-of-the-art heuristics on structured workloads, with lookahead reward shaping yielding modest improvements on unstructured circuits. These results demonstrate that reinforcement learning is a flexible alternative to manual heuristics, though scalability remains a key bottleneck for practical use.
We present a mixed-integer programming (MIP) model for scheduling quantum circuits to minimize execution time. Our approach maximizes parallelism by allowing non-overlapping gates (those acting on distinct qubits) to execute simultaneously. Our methods apply to general circuits with precedence constraints. First, we derive closed-formulas for the execution time of circuits generated by ma-QAOA on star graphs for a layered, greedy, and MIP schedules. We then compare the MIP schedule against layered and greedy scheduling approaches on the circuits generated by ma-QAOA for solving the MaxCut problem on all non-isomorphic connected graphs with 3-7 vertices. These experiments demonstrate that the MIP scheduler consistently results in shorter circuit execution times than greedy and layered approaches, with up to 24\% savings.
Quantum networks attract increasing attention due to their unique advantages over classical networks. However, the challenge of efficiently and economically constructing reliable wide-area quantum networks remains unresolved. This problem, termed network planning, is typically modeled as an Integer Linear Programming (ILP) problem. Existing solver-based solutions for classical networks face scalability issues, while heuristic approaches that sacrifice optimality for scalability heavily rely on human expertise and lack generalizability. Recently, deep reinforcement learning (DRL) has demonstrated potential in adaptive decision-making for classical network planning. However, quantum networks introduce unique constraints due to their end-to-end entanglement distribution requirements. To address this, we propose a two-stage quantum network resource allocation framework, QPLAN, that combines multi-agent DRL with traditional optimization solvers. Our approach uses a novel multi-agent reinforcement learning strategy using graph neural networks (GNNs) and transformers for traffic flow and corresponding hardware resource allocation to minimize hardware cost while cooperatively satisfying problem constraints as much as possible. The output of the DRL algorithm is then used as an initial input to prune the solution space, "warm-starting" the ILP solver and further improving the solution. This hybrid approach effectively reduces manual effort and enhances scalability. Experimental results show that our method yields near-optimal solutions with better scalability than only using ILP solvers.
The central claim of this thesis is simple but underappreciated: the primary bottleneck in near-term quantum computing is not a shortage of algorithms; it is a lack of systematic decision-making. Which problem instances should we test? Which circuit architectures are worth training? Which hyperparameters actually drive performance, and which are distractions? Today’s quantum hardware sits in an uncomfortable middle ground. It is expressive enough to explore nontrivial quantum dynamics, yet it is too noisy and shallow for the theoretical speedups that motivate the field. This thesis argues that applying machine learning to classical control decisions in hybrid quantum classical methods is one of the most tractable ways to make these methods genuinely useful in this regime. This argument is developed through four research questions, each addressing a distinct decision-making layer and illustrating the thesis. The first asks whether the decisions inside a hybrid heuristic can be learned rather than hand-coded. Chapter 4 analyzes Recursive QAOA (RQAOA), a promising yet brittle approach to combinatorial optimization, and identifies regimes in which its fixed-elimination strategy fails. It then replaces that strategy with a reinforcement learning policy, RL-RQAOA, that adapts to instance structure. The resulting method performs better on the constructed failure cases while remaining competitive elsewhere. The broader point is that learned classical control policies can materially improve hybrid quantum heuristics without increasing quantum depth, which is a significant practical advantage on constrained hardware. The second question evaluates whether the quantum optimization algorithm benchmarking instances we use are principled. Chapter 5 observes that much benchmarking in this space is driven by convenience. Researchers test on whatever graphs are readily available and then report aggregate performance. Instead, the chapter proposes a generative pipeline that couples a permutation-invariant graph autoencoder with evolutionary optimization to deliberately construct Max-Cut instances that strongly discriminate between a quantum-inspired solver, RQAOA, and a strong classical baseline, Goemans-Williamson, in both directions. By connecting solver behavior to structural graph features, the chapter turns benchmarking from a static choice into an interpretable experimental tool. The generated instances also expose failure modes that random graphs often fail to reveal. The third question asks what changes when realistic noise constraints are taken seriously during circuit design, not only during execution. Chapter 6 develops RLQAS, a reinforcement learning framework for quantum architecture search that incorporates noise assumptions informed by real IBM device characteristics throughout training, rather than treating noise as a post hoc correction. Paired with an efficient Pauli transfer matrix simulation, the framework makes large-scale noisy circuit search computationally feasible. The circuits it discovers are compact, trainable, and achieve competitive accuracy on ground-state energy tasks. This suggests that hardware-aware circuit design is tractable and can yield materially different choices than noise-agnostic search. The fourth question asks: even when the circuit architecture is fixed, how much of what we attribute to quantum model design is actually due to hyperparameter sensitivity? Chapter 7 extends functional ANOVA to quantum neural networks to quantify which hyperparameter choices drive performance variance. The finding is clear. Learning rate, circuit depth, and data encoding/reuploading dominate, while many architectural micro-choices contribute far less than the literature often implies. Rather than stopping at diagnosis, the chapter turns these importance rankings into data-driven priors that accelerate hyperparameter optimization. It converts a critique into a practical improvement. Taken together, these contributions form a coherent toolkit for developing and evaluating hybrid quantum–classical methods in the NISQ regime. While the work does not eliminate the fundamental constraints of shallow, noisy hardware, it clarifies what can be improved immediately and how to do so in a reproducible manner. The broader goal is to help shift near-term quantum research from isolated demonstrations toward an engineering discipline, in which algorithms, benchmarks, and model-selection practices can be compared fairly, stress-tested rigorously, and iterated systematically as hardware matures.