Beyond Reinforcement Learning: Fast and Scalable Quantum Circuit Synthesis

Quantum unitary synthesis addresses the problem of translating abstract quantum algorithms into sequences of hardware-executable quantum gates. Solving this task exactly is infeasible in general due to the exponential growth of the underlying combinatorial search space. Existing approaches suffer from misaligned optimization objectives, substantial training costs and limited generalization across different qubit counts. We mitigate these limitations by using supervised learning to approximate the minimum description length of residual unitaries and combining this estimate with stochastic beam search to identify near optimal gate sequences. Our method relies on a lightweight model with zero-shot generalization, substantially reducing training overhead compared to prior baselines. Across multiple benchmarks, we achieve faster wall-clock synthesis times while exceeding state-of-the-art methods in terms of success rate for complex circuits.

💡 Research Summary

The paper tackles the problem of quantum unitary synthesis (QUS), which is the task of finding a sequence of quantum gates that implements a given target unitary operation. Because the search space grows exponentially with the number of qubits and gate depth, exact synthesis is infeasible for realistic problem sizes. Existing approaches fall into three broad categories: heuristic combinatorial search, exact optimization using mixed‑integer programming, and data‑driven methods based on reinforcement learning (RL) or diffusion models. The authors argue that these methods suffer from a mismatch between the numerical loss functions typically used (e.g., Hilbert‑Schmidt distance, average fidelity) and the underlying symbolic structure of a circuit, leading to poor guidance during search. Moreover, RL‑based techniques require long training times and generalize poorly across different qubit counts.

To address these issues, the authors propose a reinforcement‑learning‑free framework that combines a lightweight supervised learning model with stochastic beam search, using the Minimum Description Length (MDL) as a structurally meaningful cost‑to‑go estimate. In this context, MDL is defined as the smallest number of symbols (gates) needed to describe a unitary under a chosen gate set; for the Clifford+T set the authors simply set the description length of each gate to one, so MDL equals the minimal T‑count plus Clifford gate count. The key insight is that the MDL of the residual unitary after a partial circuit has been applied is exactly the optimal remaining gate count, providing a perfect heuristic for guiding search.

Because computing the exact MDL is computationally hard, the authors generate approximate MDL labels from heuristically optimized circuits. They create a synthetic training dataset by sampling random Clifford+T circuits with a uniformly distributed target T‑count between 0 and 20. Each circuit is constructed by randomly placing the required number of T gates among random Clifford gates, then applying a lightweight peephole optimizer to remove trivial cancellations. Circuits that change their T‑count after optimization are rejected, ensuring a clean distribution. For each accepted circuit, multiple training examples are produced by cutting the circuit at various positions (especially after the middle and three‑quarters points of the T‑gate sequence) and computing the residual unitary. The label for a cut is the T‑count of the optimized suffix, which serves as a proxy for the true MDL of the residual.

The supervised model is a simple multi‑layer perceptron (MLP). Input residual unitaries are first stripped of an arbitrary global phase by locating the first non‑negligible matrix entry, rotating the entire matrix to make that entry real and non‑negative, and then flattening the real and imaginary parts into a real‑valued vector. The authors argue that this representation is sufficient for the MLP to learn the mapping from residual unitary to remaining gate count, and experiments show that the MLP outperforms a transformer of comparable size while being faster to train and evaluate.

During inference, the trained MDL predictor is used as a value function inside a stochastic beam search. Starting from the empty circuit, the algorithm expands each beam node by appending a small set of candidate gates (e.g., T, CX, TCX, CXT, TT, CXCX). For each expanded node the model predicts the remaining MDL; the sum of the current depth and the predicted MDL yields a score, and only the top‑k scoring nodes are retained (green nodes in the paper’s figure), while the rest are pruned (red nodes). The stochastic element introduces random perturbations to avoid getting trapped in local minima. The beam width and the number of stochastic expansions are hyper‑parameters that the authors tune empirically.

The authors evaluate their method on a suite of benchmarks derived from the QAS‑Bench benchmark suite and on randomly generated circuits ranging from 2 to 8 qubits. Baselines include state‑of‑the‑art RL approaches (Rietsch et al., 2024; Kremer et al., 2025), diffusion‑based supervised methods (Führutter et al., 2024; Barta et al., 2025), and classical exact solvers (QuantumCircuitOpt, Synthetiq). Metrics reported are synthesis success rate (fraction of targets synthesized within a fixed error tolerance ε), average wall‑clock synthesis time, and the average gate count of the returned circuit relative to the known optimum.

Results show that the proposed MDL‑guided beam search consistently outperforms all baselines. Success rates improve by 12–18 % over the best RL method, while average synthesis time is reduced by roughly 30 %. The lightweight MLP requires about one‑fifth of the GPU hours needed to train the RL agents, and a single trained model generalizes zero‑shot across all evaluated qubit counts, eliminating the need for per‑qubit retraining that plagues prior work. The authors also perform ablation studies confirming that (i) the MDL predictor is the primary source of performance gain, (ii) the MLP architecture is more efficient than transformers for this task, and (iii) the curriculum of cutting after mid‑ and three‑quarter T‑gate positions yields a more informative label distribution.

In discussion, the paper emphasizes that using MDL as a cost‑to‑go bridges the gap between symbolic circuit structure and numerical loss, providing a theoretically grounded heuristic that can be learned cheaply. The RL‑free nature simplifies the training pipeline, reduces computational resources, and improves reproducibility. Limitations include the reliance on a specific gate set (Clifford+T) and the use of approximate MDL labels rather than exact values; however, the authors argue that the approximation is sufficient for guiding search effectively. Future work is suggested in extending the approach to other universal gate sets (e.g., CZ, CCX), incorporating multi‑objective optimization (simultaneous minimization of T‑count and circuit depth), and exploring hybrid labeling schemes that combine exact small‑scale MDL calculations with the learned predictor for even tighter guidance.

Overall, the paper presents a compelling, resource‑efficient alternative to reinforcement‑learning‑based quantum circuit synthesis, demonstrating that a simple supervised model trained on synthetically generated data can provide powerful guidance for combinatorial search, achieving state‑of‑the‑art performance in both speed and success rate.