Parameterized quantum random number generators for superconducting quantum hardware: novel architectures and NIST SP 800-22 evaluation under IBM noisy simulators Authors: Ramin Salehi, Asad W Malik, Mark A Novotny, Samee U Khan Published: 2026 ABSTRACT AbstractQuantum Random Number Generators (QRNGs) exploit intrinsic quantum mechanical randomness and are promising candidates for higher-quality randomness than classical RNGs. However, QRNGs on Noisy Intermediate-Scale Quantum (NISQ) platforms are constrained by decoherence, gate and measurement errors, and limited circuit depth, and the resulting impact on statistical randomness is not yet systematically characterized. This work introduces parameterized quantum random number generator (PQRNG) architectures built from parameterized quantum circuits (PQCs) to improve tunability and expressive power under realistic noise. We study three architectures, PQC-H-CH, H-PQC-CH, and H-CH-PQC, and evaluate their randomness under Transpiler optimization and circuit-level error mitigation as preprocessing steps, using NIST SP 800-22 for statistical validation. Utilizing default NIST settings, PQC-H-CH attains the largest number of fully passing preprocessing configurations (126 combinations passing all 15 tests), compared with 110 for H-PQC-CH and 69 for H-CH-PQC, indicating strong robustness across Transpiler optimization and mitigation settings, and we find that preprocessing choices influence randomness quality more strongly than adjusting NIST test parameters. Overall, these results demonstrate that the proposed approach provides a solid foundation for developing more reliable and practical QRNGs using PQCs in NISQ devices.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-51033-9.Keywords: Parameterized quantum circuits, Parameterized quantum random number generators, NIST test suiteSubject terms: Engineering, Mathematics and computing, Physics Introduction Motivation Random Number Generators (RNGs) have been integral to numerous fields for decades, supporting essential tasks in daily life, such as personal computing, secure financial transactions1,2, digital communications3, and beyond4. In addition, RNGs play a critical role in fields such as artificial intelligence5,6, cryptography7, simulations8, Monte Carlo methods9, hashing algorithms10, and other computational processes11. This widespread need for random numbers has led to the design of RNGs becoming a highly active area of research. Moreover, in the era of blockchain12 and cryptocurrency13, with increasing threats to cybersecurity14, the need to explore true and reliable randomness resources is inevitable. However, not all randomess sources are applicable and useful for cryptographic applications, as one of the most important applications of QRNGs. For example, pseudo-RNGs (PRNGs), while computationally efficient, are deterministic and therefore, vulnerable to prediction, making them unsuitable for security-critical applications. For example, A.M. Ferrenberg et al.15 identified hidden systematic errors in Monte Carlo simulations when using Pseudo Random Number Generators (PRNGs), such as CONG, R250, R1279, SWC and SWCW. Although these generators successfully passed standard statistical randomness tests, they exhibited hidden correlations and significant biases when applied to real-world simulations. This finding underscores the importance of evaluating RNGs within real-world applications to verify their true randomness, rather than relying solely on statistical test suites15. Conversely, classical RNGs rely on classical mechanics as a deterministic interpretation of nature. Therefore, any phenomenon described in this manner cannot be responsible for the true non-deterministic randomness required for true RNGs (TRNGs). Moreover, classical random number generators are prone to bias in their outputs and cannot ensure the level of unpredictability necessary for many applications11. The nondeterministic nature of quantum mechanics is a key feature in obtaining true randomness, which is essential for building reliable TRNGs. Therefore, QRNGs derive randomness from fundamental quantum-mechanical properties, such as quantum superposition. Accordingly, the answer to the question of the applicability of intrinsic quantum indeterminism for real-world applications lies at the core of the quantum random number generation research area. Therefore, new ideas are utilized in conjunction with various technologies, ranging from laser beams16 to superconducting circuits17, to address the need for reliable random number generation. That is, this field of research remains actively evolving. However, among various QRNG implementations, gate-based quantum circuits offer a unique advantage as they are programmable, tunable (for PQRNGs), and can be executed on publicly available quantum hardware, making them an attractive solution for scalable randomness generation. A glossary of abbreviations used in this research is provided in Appendix A. Problem statement Classical RNGs are based on deterministic phenomena in nature. Therefore, they are not capable of serving as true random number generators. For example, mouse motions, keyboard input, disk access, or sudden system interruptions are not guaranteed to be truly random1. Stojanovski et al. mentioned that chaotic systems exhibit deterministic behavior on the microscopic scale, while they appear random on the macroscopic scale18. That is, the nondeterministic nature of chaotic systems cannot be certified for use as a reliable True Random Number Generator (TRNG). This drawback causes insecure communications, vulnerable cryptographic algorithms, and imprecise simulations. Once randomness sources suffer unreliability, we face a lack of trusted secure keys and strong data encryption/decryption algorithms that have crucial applications. This significant limitation compels us to seek inherently reliable sources of randomness, which are naturally present in the quantum mechanical domain. Accordingly, one of the earliest and widely adopted approaches to exploiting quantum randomness is exploiting the quantum states of light. These quantum states provide truly reliable sources of randomness, offering QRNGs with throughput in the range of Gigabits per second. However, they are not publicly accessible and often require costly setup equipment. These limitations lead researchers to use publicly available quantum processors to design, test, and optimize their quantum circuits to work as a new generation of QRNGs. Existing cloud-based QRNGs implemented on NISQ devices often lack tunability and exhibit limited capacity to explore the Hilbert space, which directly reflects on their circuit expressibility. Moreover, as we will see in the subsequent sections of the paper, based on the gate decomposition of the previously proposed QRNGs, the circuit’s architectures limit the ability to investigate the effect of circuit depth on randomness. In PQRNGs, the depth effect can be analyzed thoroughly. However, since this work focuses on introducing the foundational architectures and default configurations intended to serve as a benchmark for future research, the impact of circuit depth is not explored thoroughly in this study. In this research, we investigate the impact of PQCs depth, L, on the NIST randomness performance test; that includes the number of passed tests and P-value distribution across the Transpiler optimization levels. This investigation is provided in the supplementary materials. Previously, the noise and error effects on various quantum algorithms were investigated in the field19,20. For example, Gonzalez-Garcia et al.19 studied the error effect in NISQ devices for solving optimization problems. However, to the best of our knowledge, the existing research on gate-based quantum random number generators has not accounted for the impact of noise and errors on the randomness quality of quantum circuits. Particularly, no study in the field has investigated the combined effects of circuit optimizations, error mitigation techniques, and randomness postprocessing. Moreover, as mentioned, the limited circuit expressibility affects the current QRNGs’ ability to fully utilize the circuit-based randomness potential by exploring the Hilbert space21. To address these gaps, our research introduces and evaluates parameterized quantum circuit-based QRNGs, which offer high expressibility and tunability, allowing for the generation of high-quality randomness while adapting to noise-prone quantum processors. Several studies have examined the NIST test to evaluate its quality and assess statistical errors, for example, the reader can refer to22–24. However, limited research has focused on investigating these studies in the context of QRNGs. Further, we offer a systematic investigation to examine the impact of preprocessing and postprocessing methods on the randomness of QRNGs in noisy quantum environments. Our contributions In this research, we focus on developing QRNGs using publicly available superconducting quantum processors. To achieve this, we employ the gate-based quantum computing model, which includes the aforementioned family of parameterized circuits. These circuits, characterized by an extensive set of rotation and controlled rotation gates, offer significant potential for extracting quantum randomness. Moreover, we extensively apply various error mitigation techniques, such as Dynamical Decoupling and Pauli Twirling25, in conjunction with Transpiler optimization methods, to improve the efficiency of quantum random number generation. Given the high practicality of PQCs, analyzing error mitigation techniques in these circuits is of great value. Our analysis not only aids in quantum random number generation, but also improves the understanding of error and noise effects on quantum circuits that use parameterized quantum circuits. For instance, PQC-based architectures employed in quantum machine learning and quantum chemistry will particularly benefit from this study. Similarly, careful attention to Transpiler optimization is crucial for improving the overall performance of quantum circuits. Further, we investigate the impact of calibrating the NIST test parameters to provide a comprehensive understanding of the overall performance of quantum-generated randomness. Thereafter, the diverse range of quantum circuits implemented in this work, along with their corresponding parameters, establishes a concrete benchmark for future advancements in the development of PQRNGs. This paper is organized as follows. In Section "Introduction", a brief introduction is presented, including a problem statement and our contribution to the topic, and related works. Section "Methods" outlines our methodology, including detailed explanations of different components of parameterized quantum random number generators and the associated quantum circuits. In Section "Results", we analyze the proposed circuits and examine the effects of various parameters on the desired randomness and the results are provided. In addition, Section "Conclusion" summarizes the results of the analysis in this investigation accompanied by future works. Related works Random number generators are broadly divided into pseudo random number generators (PRNGs) and true random number generators (TRNGs). PRNGs are algorithmic-based and depend on initial seeds to produce random-like sequences4, whereas TRNGs harness randomness based on physical phenomena, such as thermal noise26, vacuum fluctuations27, and radioactive decay28. This research focuses specifically on trusted, nonoptical QRNGs using gate-based quantum circuits. Although previous optical QRNGs heavily depend on photon detection (e.g., photon arrival time29 or photon counting numbers30), they primarily rely on specialized optical setups, such as lasers and photodetectors. Nonoptical QRNGs exploit other quantum phenomena, such as quantum tunneling31, radioactive decay28, or noise in electronic and CMOS circuits, to generate randomness32. These types of generators tend to produce lower bit rates compared to optical QRNGs, but benefit from integration into solid-state and gate-based platforms. More recently, superconducting quantum circuits have allowed the observation of macroscopic quantum uncertainty33, creating a distinct category of QRNGs. Cloud-based noisy intermediate-scale quantum random number generators In the recent years, the availability of Noisy Intermediate-Scale Quantum (NISQ) processors through cloud platforms has led to a surge in designing novel QRNG architectures based on quantum circuits. In Noisy Intermediate-Scale Quantum Random Number Generators (NISQRNGs), in contrast to the former groups of QRNGs, where researchers directly implement their proposed architecture, access is usually available only through cloud space. Moreover, most of the available quantum processors utilize gate-based models of quantum computation34,35 for realizing desired computations. That is, the QRNG circuits must be mapped to each quantum processor, which has its own predefined set of available quantum gates and qubits36. This different implementation categorizes NISQRNGs as a distinct category of random number generators. As mentioned, the intrinsic uncertainty of quantum mechanical phenomena is harnessed indirectly through quantum gates and qubits, rather than the direct interpretation of the observations. Correspondingly, this emerging branch of QRNGs, based on their physical implementation of qubits and quantum gates, can be categorized as optical-NISQRNG37 or non-optical NISQRNG38,39. According to the literature, Quantum Processing Units (QPUs) exhibit diverse physical implementations, including photonic systems40, superconducting circuits41, and ion-trapped devices42. This variety has facilitated the development of both optical and non-optical NISQRNGs, which are accessible through cloud-based platforms. Most of these platforms follow the gate-based (circuit-based) model of quantum computation. Specifically, as mentioned earlier, this model has become the preferred approach among researchers, especially for NISQ devices. Accordingly, IBM, a leader in NISQ superconducting quantum computers, which has provided public access to its quantum computers since 201743 utilizes this model of quantum computing. Since then, other companies, including Quantinuum44, and Rigetti45 have also contributed to the development and benchmarking of QRNGs using cloud-accessible quantum computers. As mentioned earlier, we are designing a gate-based QRNG based on non-optical qubits. We then briefly review representative gate-based quantum random number generators. In 2020, Kentaro Tamura et al. proposed a QRNG utilizing Hadamard gates to realize randomness generation46. In our previous work38, we developed a QRNG utilizing a hybrid architecture that incorporates Hadamard and Controlled-Hadamard gates. Our results demonstrated that combining a layer of Controlled-Hadamard gates, randomly controlled by a layer of Hadamard gates, increases both the randomness and entropy of the generated bit strings. Moreover, it is worth noting that the gate-based model of quantum computing offers significant interpretability. Accordingly, based on this model, several circuit variations have been introduced in recent year38,47–53. For example, in53, the authors use , z-basis measurement, and x-basis measurement for randomness generation, and then they utilize a Toeplitz-matrix hashing randomness extractor to improve the randomness of the generated bit strings. Whilst Kumar et al.52 apply odd numbers of Hadamard gates consecutively to generate randomness. Whereas the authors in49 introduced a Hybrid Hadamard CNOT gate using an XOR bit mask postprocessing scheme as part of an NISQRNG architecture. In their research, the XOR scheme is employed as one of the basic methods for randomness extraction. Moreover, de Souza et al. utilized a photonic QPU for implementing a cloud-based QRNG37. Further, the authors in54 implemented the QRNG with a well-known Bell States circuit for extracting the quantum randomness, which contains Hadamard and CNOT gates. After reviewing some previous works, we present our proposed model in detail in the following sections. In this work, we present various structures for random number generation based on different circuits of parameterized quantum circuits. PQCs are a family of circuits containing several single and two-qubit gates, including multiple rotation and controlled-rotation gates around the three coordinates (x, y, z) with the angles as tunable parameters. This family of circuits is well-known for its expressibility power in representing quantum states in Hybrid Quantum Classical algorithms (HQC). Sim et al. define the expressibility of quantum circuits as their capability to produce pure states that effectively represent the Hilbert space. This capability enables us to explore more points of Hilbert space (Bloch Sphere for single qubits)55. PQCs are comprised of single-rotation gates and controlled-rotation gates, which introduce a higher degree of tunability compared to other quantum circuits. This enhanced tunability explains PQCs’ significant expressibility. Accordingly, our goal in this research is to utilize the expressibility power of PQCs to generate the desired randomness. To this end, we will first discuss the necessary background in the following section. This foundation will facilitate a clearer understanding of parameterized quantum circuits. Methods PQC circuits Sime et al. define parameterized quantum circuits as tunable unitary operations on n qubits, , that is applied to a reference state , often set to . The resulting parameterized quantum state is , where is a vector of a polynomial number of circuit parameters. The circuit parameters correspond to angles of rotation gates; however, they can represent any tunable parameters in a quantum operation21. As mentioned earlier, parameterized quantum circuits are a family of circuits containing several single and two-qubit gates, including multiple rotation and controlled-rotation gates around the three coordinates (x, y, z) with the angles as tunable parameters. As PQCs are composed of a universal set of gates, they can express any arbitrary quantum state with corresponding parameters (angles) . According to the interpretation of the operation of quantum gates as rotation of quantum states on the Bloch sphere, PQCs have a significant ability to probe the Hilbert space21,54. Based on this capability, parameterized quantum circuits have been widely used in variational quantum algorithms56, which play a vital role in Quantum Chemistry57. Moreover, PQCs are used in hybrid quantum-classical machine learning methods58–60 as one of the promising areas of quantum computing. Further, as mentioned earlier, this group of circuits is also a necessary component of many well-known quantum algorithms, including those for optimization61,62. They are also used in hash function implementation, making their role worth highlighting10. In this paper, we design new circuits for QRNGs with parameterized quantum circuits. In previous studies, various research groups have concentrated on optimizing the parameters and/or adjusting the number of layers of parameterized quantum circuits to enhance circuit expressibility, thereby improving their ability to solve specific problems. In contrast, in this work, we focus on PQCs ability to produce randomness and operate as a QRNG. Moreover, the circuit parameters (angles) are not optimized in this study, as the objective is to present a baseline case for future research. All single-qubit rotation angles and controlled-rotation angles in the PQC blocks were fixed and selected from the discrete set . In upcoming works, parameter optimization will be explored to maximize the p-value function. The new circuits provided here differ slightly from traditional PQCs in the literature, as they include additional blocks of gates. Three blocks of gates are defined in this kind of QRNGs; so that PQRNGs are composed of Hadamard, PQC and Controlled-Hadamard blocks as depicted in Fig. 2. Parameterized QRNG Block Diagram. H and CH blocks encompass only H and CH gates. (a) PQRNG includes PQC, H, and CH blocks, respectively. (b) PQRNG includes H, PQC, and CH blocks, respectively. (c) We first apply CH gates, which result in no generation of randomness. (d) PQC block is implemented after H and CH blocks. In the measure lines, the classical output is measured, which is represented as C (for classical) in quantum circuits. Based on different permutations of these blocks, various PQRNG architectures are implemented. As expected, they produce different values of randomness in the output bit strings, which are discussed in more detail in subsequent sections of the paper. The Hadamard block contains only Hadamard gates, which are applied to even qubits (qubit zero, qubit two in the four-qubit circuit). Meanwhile, the Controlled-Hadamard (CH) block includes a Controlled-Hadamard gate in a manner that its control qubits are on the even qubits, and the target ones are applied to the odd qubits. For example, the output of the first Hadamard, which is applied to qubit zero, controls the operation of the CH gate, whose target qubit is qubit one. Moreover, the original PQC block encompasses single rotation gates and two-qubit controlled-rotation gates as mentioned earlier. Now, we examine several examples of PQRNG circuits to provide a better understanding of the architectures. In Fig. 3 the decomposed circuits of some of these PQRNGs are represented. In Fig. 3 (a), the Hadamard block is implemented first, then the CH block, and finally, the PQC block is added. Figure 3(b) shows that the PQC block is applied, and then the H block and CH block are the last blocks implemented in the circuit. The following circuit, Fig. 3 (c), is designed in a manner that H block is inserted into the circuit first, followed by the PQC and CH blocks, respectively. Additional architectural notes, including the discussion of the impracticality of the CH-PQC-H architecture, are provided in Appendix B. PQRNG circuit implementations. (a) H-CH-PQC circuit of PQRNG (see also Supplementary Fig. 3), (b) PQC-H-CH circuit of PQRNG (see also Supplementary Fig. 1), and (c) H-PQC-CH circuit of PQRNG (see also Supplementary Fig. 2). In each circuit, rotation gates around the - and -axes are implemented using pre-specified, fixed angles drawn from the discrete set . The output bits are measured and saved in classical registers (C), while qubits are represented as and . Applying preprocessing and postprocessing on PQRNGs In this research, two main processing steps are carried out, including the preprocessing stage and the post-processing stage. Preprocessing consists of manipulating PQC parameters, optimizing the circuit Transpiler for compilation on the backends (optimization levels, 0–3), and implementing error mitigation techniques. At the same time, post-processing involves calibrating the NIST SP 800-22 test parameters and post-selecting the bit string. According to Fig. 1, the influence of different types of parameters is considered. These parameters are categorized into three main categories: randomness test suite parameters, circuit parameters, and quantum circuit compilation parameters on QPUs. As a first step, QPU Transpiler optimization and error mitigation techniques are applied to the circuits proposed in Qiskit. Parameterized QRNG design workflow containing circuit parameters manipulation, Transpiler optimization, and error mitigation as preprocessing, bit post-selection, and randomness test analysis as postprocessing stage. Transpiler optimization and error mitigation in PQRNGs Given that both transpilation optimization and error mitigation significantly impact circuit performance, we examine their effects on randomness generation in subsequent sections. As described in IBM Quantum Documentation63, “Transpilation is the process of rewriting a given input circuit to match the topology of a specific quantum device, and/or to optimize the circuit for execution on present-day noisy quantum systems. Accordingly, many quantum circuits require a series of transformations that make them compatible with a given target device.” Consequently, based on the optimization level, various techniques, such as using swapping gates, resynthesizing two-qubit blocks, and Unitarity-breaking passes are utilized in the literature64 to modify circuit layouts and routings. These techniques are provided as internal, modifiable parameters in Qiskit, enabling users to implement them according to their desired schemes. The goal of Transpiler optimization is to reduce circuit depth and gate counts, thereby decreasing noise and gate errors, particularly for two-qubit gates in the circuit, to improve circuit functionality. In this paper, we establish a benchmark for future investigations using default values for Transpiler optimization levels and do not delve into the optimization of internal parameters embedded in the Qiskit Circuit Transpiler, which could be a topic for future research. Further, error mitigation techniques play a crucial role in NISQ devices in enhancing the performance of quantum circuits. The scalability of quantum computers is highly dependent on increasing the resistance of quantum processors to noise and error. Without addressing this issue in quantum circuits, the accessibility of reliable commercialized quantum computers cannot be expected. According to the fact that we use NISQ devices, which are not fault-tolerant, errors always accompany the circuits. Various mitigation techniques available in near-term devices can be used to improve circuit functionality and performance. Three major parts of any quantum circuit are susceptible to errors, including the operation of the gates, implementation of the qubits (decoherence), and the measurement procedure64–67. Therefore, different possible error mechanisms influence the operation of quantum circuits. In terms of qubit coherency, two parameters are mainly introduced, T1 (Amplitude Damping), which relates to qubit relaxation time, and T2 (Phase Damping), which represents phase loss of qubits64. The operation of quantum gates can also be susceptible to error. For example, in superconducting circuits, any imperfection in the pulse shaping of microwave pulses or interference from control line circuits can introduce error to the system in such a way that the final quantum state can be different from the expected one68. Moreover, given that superconductivity occurs at extremely low temperatures (milli-Kelvin), their operation can be influenced by thermal noise69. However, error also happens in the measurement step of quantum circuits67. As a result, various error correction and mitigation techniques have been introduced in both software and hardware parameters. Some of these error mitigation techniques are Dynamical Decoupling65, Pauli Twirling66, Twirled Readout Error Extinction (TREE)67, Zero Noise Extrapolation (ZNE)70, Probabilistic Error Amplification (PEA)71 and Probabilistic Error Cancellation (PEC)72. The next important set of parameters to consider is the circuit parameters, specifically rotation angles, which are covered in the next section. Detailed IBM Brisbane backend settings, that is used in this work, are provided in Appendix C. Circuit parameter manipulation and bit post-selection In this study, in addition to the parameters mentioned above that affect the generation of randomness in PQRNGs, a comprehensive investigation has been carried out to examine various factors influencing circuit quality. The gate parameters, specifically the rotation angles, influence the randomness of the circuits, which is reasonable, since these angles of the rotation gates determine the final state of the qubits. Moreover, the structure of the circuits, in terms of the sequence of gates implemented (which results in different PQCs), affects the quality of randomness in the circuits, which is clear. Further, as we showed in our previous work38, the randomness will increase significantly compared to previous works46 whenever you only measure the outputs of Controlled-Hadamard gates. In that paper, we discussed that the intrinsic randomness of Hadamard gates will propagate and intensify once CH gates are controlled randomly. Therefore, in this research, a post-selection (postprocessing) strategy is also implemented. This approach differs from using randomness extraction techniques, such as those proposed by Von Neumann73 or cryptographic methods74. Here, based on IBM QX features, measurement gates must be applied to all qubits, followed by the manual removal of undesired ones. After generating the desired bit sequences, we must evaluate their randomness to verify that they originate from a reliable QRNG. This is achieved by applying the NIST randomness test suite, which is explained in the next section. NIST randomness analysis for PQRNGs The test suite parameters are test-specific and must be determined before running the test. The NIST 800-SP 22 has six parameters, including Block Frequency test block length, Non-overlapping Template test block length, Overlapping Template block length, Approximate Entropy test block length, Serial test block length, and Linear Complexity test block length75. Although all of these parameters have predefined values, the user can modify these values. Here, four sets of parameters are utilized, specifically focusing on changing the nonoverlapping template test, overlapping template test, approximate entropy test, and serial test block lengths. This version of the NIST randomness tests concentrates specifically on the analysis of RNGs that are applicable for cryptographic applications. This means that NISQRNGs that pass all tests are valuable for cryptographic tasks and also beneficial for other applications discussed earlier. Detailed NIST parameter definitions and significance-level description are provided in Appendix D. To obtain a clear understanding of the proposed workflow, a summarized block diagram of randomness generation in this project is presented in Fig. 1. In the figure, in ① we have raw circuits, which means the circuits are transpiled without any Transpiler optimization and error mitigations. In ②, optimized circuits with suppressed errors are produced. In this step, the preprocessing stage is finished. Next, in ③, the measurement is made, and the quantum states collapse to classical states. The user initiates postprocessing after the measurements are taken. Then, in ④, the raw bit strings are generated. Finally, in ⑤, the desired random bit strings are available. If the P-values fall outside the range expected for reliable QRNGs, attention should be paid to the preprocessing step. This includes optimizing circuit parameters, adjusting Transpiler optimization settings, and refining error mitigation techniques. In the existing work, nineteen different PQC circuits are utilized in each of the three main architectures to build PQRNG circuits. The PQC block shown in Figure 3 is one of the simplest forms of this family of circuits. Different research groups have investigated this family using multiple layers of parameterized circuits to increase their expressibility and entangling capabilities, for example, see55. The entangling capability of PQCs will undoubtedly impact their ability to generate random numbers, and this requires thorough investigation in future work. However, in the following paragraph, we briefly explain the relationship between the entangling ability of PQCs and the generation of quantum randomness. Entanglement in quantum systems is achieved by implementing two-qubit and three-qubit gates in quantum circuits. The most widely used two-qubit gate for entanglement realization is the CNOT gate. However, due to the limited connectivity of quantum hardware, specifically superconducting QPUs, not all CNOT gates can be implemented directly. To solve this issue, several techniques, such as use of SWAP gates, are introduced76, which are beyond the scope of the current research. Nonetheless, as the number of two- and three-qubit gates increases, the circuits become more entangled. This is accompanied by higher levels of accumulated noise and error in the circuits77. Therefore, to understand the effect of entanglement on randomness generation, we suggest analyzing the number of tests passed for circuits with different numbers of two- and three-qubit gates. This experiment allows us to identify the relationship between the aforementioned factors. However, more in-depth investigation and analysis are required to uncover meaningful correlations. Building upon the discussion in the preceding paragraph and as demonstrated in the results section of this paper, circuits with a higher number of CNOT and other two-qubit gates tend to produce more random bit strings following the application of error mitigation techniques. That is, circuits with higher levels of entanglement, reflected in higher numbers of two-qubit gates, show a lower level of randomness. Consequently, it can be asserted that, unlike the expressibility of PQCs, their entangling power adversely affects the randomness generation capability of PQRNGs. Therefore, there is a trade-off between the expressibility and entangling capabilities of parameterized quantum circuits in terms of randomness production, which should be carefully considered in the design and implementation of PQRNGs. After exploring two key features of PQCs, expressibility and entanglement, in the previous paragraphs, we now dedicate our focus to a new aspect. In this research, as previously mentioned, we introduce and examine another feature of PQCs (in their modified form), that is, their inherent ability to generate randomness. Based on the comprehensive discussions from the previous sections, we are now equipped to discuss the results of testing the circuits, which demonstrate that PQRNGs are practical quantum random number generators. We will provide a detailed analysis of the PQRNG implementation in the following parts of the paper. Additional PQRNG circuits are provided in the supplementary information. Results We first examine the effects of Transpiler optimization on randomness production and then investigate error mitigation techniques. Finally, we evaluate how calibrating the NIST test parameters influences the quality of our PQRNGs. The detailed discussion is provided in Appendices F, G, and H. In this research, we analyze 1,368 distinct bit sequences generated from three primary architectures (19 circuits per architecture, 57 circuits total). Each circuit is evaluated in six error mitigation strategy (five error mitigation techniques plus WEM) and four Transpiler optimization levels. Taking into account four sets of NIST test parameters, this yields 5,472 total data points, where each point represents a unique configuration and the corresponding number of tests passed. The three architectures show clearly different randomness quality: PQC-H-CH provides the strongest overall performance, H-PQC-CH is the next candidate, and H-CH-PQC is the weakest among the three. PQC-H-CH (see Fig. 4) and H-PQC-CH (see Supplementary Figure 28) typically pass between 10 and 15 tests, while H-CH-PQC (see Supplementary Figure 29) (including Circuit 9) falls more often in the 7–13 range. Here, Case 1 is used as the reference NIST setting, with parameters (128, 9, 9, 10, 16, 500) for the Block Frequency, Non-overlapping Template, Overlapping Template, Approximate Entropy, Serial, and Linear Complexity block lengths. Under these default parameters, PQC-H-CH reaches 15/15 for all circuits except Circuit 9 (126 successful preprocessing combinations), while H-PQC-CH and H-CH-PQC achieve 15/15 for seventeen circuits (110 combinations) and eleven circuits (69 combinations), respectively. Across all architectures, the next most frequent outcome is passing 13/15 tests, with 17, 17, and 11 circuits achieving this level for PQC-H-CH, H-PQC-CH, and H-CH-PQC, respectively, still indicating considerable randomness. The corresponding figures for H-PQC-CH and H-CH-PQC are provided in Supplementary Figures 28, and 29. Number of passed tests for different circuits with four optimization levels (levels 0–3) and six error mitigation techniques for the PQC-H-CH architecture. Detailed preprocessing parameter analysis shows that the impact of Transpiler optimization and error mitigation on randomness quality is architecture-dependent rather than universal. Among the three proposed architectures, as mentioned, PQC-H-CH remains the most robust across preprocessing variations, whereas H-PQC-CH and H-CH-PQC exhibit greater sensitivity to the selected configuration. Similarly, calibration of the NIST SP 800−22 test parameters does not alter the overall ranking of the architectures, although it highlights different stability trends across them. Detailed architecture-specific discussions and the corresponding figures for Transpiler optimization, error mitigation, and calibration of NIST parameter are provided in Appendix F, Appendix G, and Appendix H, respectively. Moreover, the architecture-dependent trends observed here are also compatible with the expected role of PQC expressibility in randomness generation. Increased expressibility can bring the measured output distribution closer to the unbiased and weekly correlated behavior targeted by the NIST SP 800-22 tests. A theoretical discussion of the relationship between PQC expressibility and statistical randomness is provided in Appendix E with further details in section S2 in supplementary information. Conclusion In this work, we proposed and benchmarked three parameterized quantum random number generator (PQRNG) architectures, PQC-H-CH, H-PQC-CH, and H-CH-PQC, under systematic variations of Transpiler optimization levels, error mitigation techniques in IBM QPUs, and NIST SP 800-22 test parameter settings. Across the evaluated configurations, PQC-H-CH consistently delivered the strongest randomness performance, followed by H-PQC-CH and H-CH-PQC. Using the default NIST settings with preprocessing sweeps, PQC-H-CH achieved 126 successful preprocessing combinations that passed all 15 NIST tests, compared to 110 for H-PQC-CH and 69 for H-CH-PQC, demonstrating the strongest robustness among the three architectures under noisy conditions. In particular, PQC-H-CH yielded 16 and 57 more successful full-pass combinations than H-PQC-CH and H-CH-PQC, respectively. Moreover, the results indicate that preprocessing parameters related to Transpiler optimization and error mitigation techniques have a stronger impact on randomness quality than moderate calibration of NIST test parameters. Overall, these results establish a clear quantitative benchmark for evaluating PQC-based QRNG architectures under realistic noisy conditions. A limitation of the present study is that the benchmarking was conducted on IBM noisy simulator backend rather than real superconducting quantum hardware; therefore, the reported results should be interpreted as simulator-based performance trends that require validation on physical devices. Future work will focus on executing the proposed architectures on real quantum hardware and extending the design to deeper PQRNG circuits, supported by parameter optimization and enhanced post-processing, such as incorporating randomness extractors to further improve the quality of the generated bit strings. Moreover, we will extend the present benchmark by replacing fixed angles with Haar-random parameter sampling to study how fully random parameterizations affect PQRNG performance. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements The authors gratefully thank Dr. Mohammad Razaghi, Dr. Bashir Fotouhi, and Dr. Yaroslav Koshka for their valuable guidance, support, and insightful feedback throughout this research. A Glossary of abbreviations Table 1Glossary of Quantum Computing-Related Abbreviations. Abbreviation Description QRNGQuantum Random Number GeneratorTRNGTrue Random Number GeneratorCTRNGClassical True Random Number GeneratorRNGRandom Number GeneratorPRNGPseudo Random Number GeneratorNISQRNGNoisy Intermediate Scale Quantum Random Number GeneratorQPUQuantum Processing UnitPQCParameterized Quantum CircuitHQCHybrid Quantum ClassicalQAOAQuantum Approximate Optimization AlgorithmCHControlled-HadamardNISTNational Institute of Standards and TechnologyPQRNGParameterized Quantum Random Number GeneratorWEMWithout Error MitigationDDEMDynamical Decoupling Error MitigationPTEMPauli Twirling Error MitigationTREETwirled Readout Error ExtinctionZNEMZero Noise Extrapolation MethodPECMProbabilistic Error Cancellation Method Glossary of Quantum Computing-Related Abbreviations. B Additional architectural notes This appendix provides the additional architectural discussion referenced in Section Methods. According to Fig. 2, in the block representation, the case in which CH is first implemented is not a practical architecture. Remember that qubits are initialized in the zero state on IBM quantum processors. Therefore, the CH-PQC-H architecture is not suitable for designing PQRNGs. In this setup, the cascade architecture of Hadamard and Controlled-Hadamard gates, which is intended to generate randomness, is not implemented correctly. Specifically, the randomness produced by the Hadamard gate does not propagate to the CH gate. That is, the CH gates are not applied to the target qubits ( and ), and as a result, this layer does not produce random outputs. C IBM Brisbane experimental details This appendix provides the detailed IBM Brisbane backend information referenced used in this work with in the fake backend mode. We implement our circuits on an IBM Fake Brisbane noisy simulator that mimics real hardware performance. IBM Brisbane consists of 127 qubits with EagleR3 processor type. For this processor, the values of (relaxation time) and (decoherence time), which indicate the duration for which a qubit preserves its quantum state before undergoing energy loss or decoherence, can be obtained from the calibration data. The values range from (for qubit 120) to (for qubit 38), while the values vary between (for qubit 120) and (for qubit 2). Further, the qubit frequencies range from (for qubit 33) to (for qubit 125). On the other hand, the anharmonicities are around , which affects the ability to drive qubit transitions63 selectively. The readout assignment errors, which indicate the probability of misidentifying qubit states during measurement, range from 0.0061 (for qubit 6) to 0.1564 (for qubit 114). Lower values indicate more reliable measurements. Measurement errors also include Prob meas0 prep1 (the probability of measuring 0 when 1 was prepared), which varies from 0.0009 (for qubit 120) to 0.8320 (for qubit 15), and Prob meas1 prep0 (the probability of measuring 1 when 0 was prepared), ranging from 0.0029 (qubit 31) to 1 (for qubit 120). The difference between readout errors and measurement errors guides us in selecting suitable qubits to implement PQRNG circuits. Gate errors are also included in the calibration data, covering the Z-axis rotation error, (SX) error, and Pauli-X error, reflecting the precision of quantum gate operations. The Z-axis rotation error is reported as zero across qubits, while the (SX) error and Pauli-X error range from to for qubits 17 and 116, respectively. The calibration data set also contains ECR (Echoed Cross-Resonance) error rates for specific qubit pairs, along with gate times (ns) for two-qubit operations. These calibration data are important for evaluating the reliability of quantum computations on IBM Brisbane, refining transpiler optimization, and implementing error mitigation techniques in quantum experiments. Understanding these parameters improves the performance of quantum algorithms by reducing noise and improving the accuracy of quantum computations63. In these experiments, dynamical decoupling used the XpXm sequence. We carried out Pauli twirling and measurement mitigation with 32 randomizations and 100 shots per randomization. Zero-noise extrapolation involved noise factors (1, 3, 5), along with an exponential extrapolator. Probabilistic error cancellation was enabled, with a maximum overhead of 100. All circuits were implemented and run using Qiskit (version 2.1.1) and Qiskit Aer (version 0.17.2) with the number of shots equal to shots=20,000 in Fake Brisbane noisy simulator, which yields raw output bit strings with length per each PQRNG run. We conducted the randomness evaluation using the official NIST Statistical Test Suite on a Linux operating system with four sets of parameters as Case 1, Case 2, Case 3, and Case 4. All single-qubit rotation angles in the PQC blocks were fixed and selected from the discrete set (with the corresponding controlled-rotation angles also chosen from the same set) to enable deterministic, reproducible benchmarking across all circuit architectures and preprocessing configurations. D NIST statistical parameters This appendix provides the detailed NIST statistical parameter discussion referenced in Section NIST Randomness Analysis for PQRNGs. In NIST SP 800-22, the P-value is used as a criterion to quantify the degree of randomness. Each P-value represents the probability that a perfectly random number generator would generate a sequence exhibiting less randomness than the tested sequence, based on the specific type of non-randomness the test evaluates75. Moreover, to present a meaningful analysis, a significance level () must be defined as a threshold to decide whether the bit string produced is random or not. Then the P-value and the significance level are compared. If the P-value , the bit sequence is random. However, if the P-value , the bit sequence is not random. For example, means that one sequence out of 100 sequences would be expected to be rejected. A P-value indicates that the sequence is considered random with 99% confidence, while a P-value suggests that the sequence is considered nonrandom with 99% confidence75. For many cryptographic applications, is the standard significance level. Considering that the default value for is in NIST SP 800-22, we also select this value for our analysis. E PQC expressibility and statistical randomness NIST SP 800-22 models an ideal random bitstream as independent unbiased coin flips, i.e., independent and identically distributed (i.i.d.) Bernoulli bits with , and many tests explicitly detect deviations from the ideal bias and simple correlations (such as, Frequency, Block-frequency, Runs, and Serial). From a quantum-information perspective, an ensemble that approaches the Haar measure (or a low-order unitary or state design) suppresses low-order local statistics. For traceless Pauli observables, such as and , Haar- or design-like behavior implies and . As a result, marginal bit bias and two-point correlations in computational-basis measurements are reduced. The parameterized quantum circuit (PQC) insertion enlarges and enables tunable sampling of the circuit ensemble. Increasing expressibility–defined as the closeness of the PQC-induced state statistics to Haar- or design-level statistics55, therefore drives the output distribution closer to the i.i.d. unbiased benchmark targeted by NIST, improving empirical randomness quality. F Transpiler optimization effect on PQRNG randomness This appendix provides the detailed Transpiler optimization analysis referenced in the Results section. Across the three architectures, Transpiler optimization has a clear impact on randomness, but the best optimization level is not universal and depends on both the circuit family and the accompanying mitigation technique. For PQC-H-CH (see Fig. 5), optimization Level-1 most consistently maximizes 15/15 outcomes across the main settings (WEM, DDEM, PTEM, and ZNEM), making it the most reliable default choice when the objective is perfect-pass performance, while TREE and PECM are the main exceptions that shift the best behavior toward Level-0. For H-PQC-CH (see Supplementary Figures 34-37), the optimal level is more configuration-dependent: Level-0 is preferred for WEM and PTEM (seven successful combinations each), Level-3 becomes best for DDEM (five successful combinations) and Level-1 is preferred for TREE, ZNEM, and PECM (five, seven, and six successful combinations, respectively). In contrast, H-CH-PQC shows the weakest overall performance (Check Supplementary Figures 38-41), where the baseline is best at Level-0 (five successful combinations), while the strongest mitigation impact on the results tend to favor Level-3 for TREE and ZNEM (five and six successful combinations) and Level-1 for PECM (five successful combinations). Therefore, these results show that the selection of the optimization level has a meaningful impact on randomness quality, but it must be chosen jointly with architecture and error mitigation rather than treated as a fixed global setting. Transpiler optimization and error mitigation effect on the number of passed tests for PQC-H-CH architecture. Each spot belongs to different combinations of Transpiler optimization and error mitigation techniques. G Error mitigation effect on PQRNG randomness This appendix provides the detailed error mitigation analysis referenced in the Results section. Error mitigation is not a universal improvement and its effectiveness depends on the architecture and Transpiler optimization level. For PQC-H-CH (Fig. 5), strong randomness is achievable across mitigation methods, but the best choice shifts with optimization: at Level-0, PECM25 yields the most successful combinations (eight), while at Level-1, DDEM65 becomes dominant (ten), indicating the strongest mitigation benefit when paired with Level-1 optimization; at Level-2, PTEM66 is the best (seven), and at Level-3, DDEM again leads (seven), although randomness tends to weaken at optimization levels 2 and 3 as circuits become increasingly constrained by device-based restrictions of QPUs. For H-PQC-CH (see Supplementary Figures 33-37), the best result at Level-0 occurs without mitigation (seven), indicating that error mitigation can deteriorate this family, whereas at Level-1, ZNEM70 and WEM provide the strongest outcome (seven); at Level-2, PTEM66, ZNEM70, and PECM25 each yield five successful combinations (with DDEM performing better when comparing across different numbers of tests passed); and at Level-3, DDEM and PECM achieve five successful combinations. For H-CH-PQC (see Supplementary Figures 38-41), error mitigation techniques worsen the randomness at Level-0 optimization (WEM has five successful combinations, and all other mitigation techniques except PTEM further worsen performance), while improvements mainly appear at higher optimization, where ZNEM and TREE67 with Level-3 provide the best results (six and five successful combinations, respectively). H NIST test parameter calibration effect on randomness This appendix provides the detailed NIST parameter calibration analysis referenced in the Results section. Figure 6 evaluates the calibration of NIST parameter for Case 1 (128, 9, 9, 10, 16, 500), while the figures for the other architectures including Case 2 (128, 18, 18, 20, 32, 500), Case 3 (128, 18, 18, 20, 16, 500), and Case 4 (128, 15, 18, 20, 32, 500) are presented in the supplementary information, noting that other values may also be tested75. This calibration does not change the main ranking of the architectures, but it reveals different sensitivity levels. PQC-H-CH (Fig. 6) is the most robust, with successful combinations remaining nearly constant (126, 124, 124, 131), and Case 4 performs best. H-PQC-CH (check supplementary information Figures 46-49) shows greater shifts and a net decrease in successful combinations under calibration (110, 100, 100, 105), making Case 1 the best choice. H-CH-PQC (check supplementary informationl Figures 50-53) changes only slightly (69, 66, 68, 70), with stronger configurations (13–15 passed tests) remaining largely stable; Case 4 is best for this family. The full violin-swarm distributions for H-PQC-CH and H-CH-PQC are provided in Supplementary Figures 46-49, Supplementary Figures 50-53. For higher-resolution figures, the corresponding plots are also provided as separate figures in the supplementary information that includes individual figures for each optimization level, allowing a clearer investigation of the effects of each error mitigation technique within a fixed Transpiler optimization level. We also provide separate figures for each NIST test calibration parameters cases for each architecture. NIST test parameter calibration effect on randomness for PQC-H-CH Circuits. The number of tests passed is portrayed on the Y-axis. At the same time, the X-axis categorizes their distribution across optimization levels, with Level-0 positioned on the left side of each subplot. Different colors distinguish each error mitigation technique. (a) Case 1: (128, 9, 9, 10, 16, 500); (b) Case 2: (128, 18, 18, 20, 32, 500); (c) Case 3: (128, 18, 18, 20, 16, 500); (d) Case 4: (128, 15, 18, 20, 32, 500). Author contributions All authors contributed equally to the conception, design, analysis, and writing of this manuscript. Data availability The data generated and analyzed during the current study are available from the corresponding author on reasonable request. Code availability The codes developed and used in this study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data generated and analyzed during the current study are available from the corresponding author on reasonable request. The codes developed and used in this study are available from the corresponding author on reasonable request.