Quantum Recurrent Neural Networks (QRNN)

• QRNNs are quantum-enhanced sequence models that implement recurrence via quantum memory, variational circuits, and optical delay lines.
• They integrate classical recurrent structures with quantum operations, often achieving faster convergence and improved data efficiency.
• Architectures range from persistent quantum memory models to hybrid gate-substitution and recurrent embedding approaches, addressing various practical tasks.

Quantum recurrent neural networks (QRNNs) are a heterogeneous family of sequence-learning models in which recurrence is realized through quantum states, quantum channels, or variational quantum circuits rather than solely through a classical hidden vector. Across the literature, the term does not denote a single canonical architecture: some QRNNs use a persistent quantum memory register that is carried across timesteps, some replace classical recurrent submodules such as LSTM or GRU gates with variational quantum circuits, and some realize recurrence physically in photonic delay lines or via repeated controlled data embeddings (Siemaszko et al., 2022). The field is also terminologically unstable: one influential paper explicitly states that there was still a “lack of canonical model of quantum recurrent neural network (QRNN)” (Li et al., 2023), while another “QRNN” paper concerns Quaternion rather than Quantum recurrent neural networks and therefore belongs to a different, entirely classical lineage (Parcollet et al., 2018).

1. Conceptual scope and terminological boundaries

The recurrent idea shared by QRNN proposals is that information from previous timesteps is retained and re-used at later timesteps, but the object that stores this information varies substantially. In continuous-variable photonic QRNNs, the hidden state is a set of unmeasured register modes or delay-line memory modes that persist physically from one layer application to the next (Siemaszko et al., 2022). In hybrid gate-substitution models such as QLSTM and QGRU, the recurrent state update remains structurally classical, but the affine gate computations are replaced by variational quantum circuits (Chen et al., 19 Feb 2025). In fully quantum process-learning models, the hidden state is a quantum memory subsystem $H^{\text m}$ that is fed forward between applications of a dissipative quantum neural network (Bondarenko et al., 2023). In quantum-optical online models, recurrence is embodied in optical feedback loops and delay lines rather than simulated through a classical hidden vector (Prins et al., 2023).

This plurality of constructions matters because different papers attach different claims to “QRNN.” Some works study classical-input/classical-output temporal learning, such as time-series prediction or image classification after sequentialization (Siemaszko et al., 2022, Li et al., 2023). Others target quantum-input/quantum-output tasks, such as short-term quantum memory, temporal entanglement generation, bosonic-channel coding, and quantum channel equalization (Prins et al., 2023). Still others frame recurrence as repeated data re-embedding across circuit slots rather than as an explicit timestep-wise hidden-state update, as in the Quantum Recurrent Embedding Neural Network (QRENN) (Jing et al., 16 Jun 2025). This suggests that “QRNN” is best treated as an umbrella term for quantum recurrent architectures rather than a single standardized model class.

A persistent source of confusion is the abbreviation “QRNN” itself. “Quaternion Recurrent Neural Networks” (Parcollet et al., 2018) introduces quaternion-valued RNNs and QLSTMs based on Hamilton products in quaternion algebra; it does not use quantum states, quantum circuits, or quantum hardware. By contrast, the QRNN literature in the quantum-computing sense relies on qubits, qumodes, unitary or symplectic evolution, measurement, and hybrid quantum-classical optimization.

2. Recurrent mechanisms and major architectural families

A useful way to organize QRNN research is by the physical or algorithmic location of memory.

Family	Representative mechanism	Representative papers
Persistent quantum-memory models	Unmeasured memory qubits or modes are carried forward across timesteps	(Siemaszko et al., 2022, Bondarenko et al., 2023, Li et al., 2023, Xu, 29 Oct 2025)
Hybrid gate-substitution models	Classical recurrent equations are retained, but gate computations use VQCs	(Chen et al., 19 Feb 2025, Chen, 2023, Chen et al., 2022)
Recurrent photonic/optical models	Memory is implemented physically via delay lines or looped optical modes	(Prins et al., 2023, Siemaszko et al., 2022)
Recurrent embedding/query models	Recurrence is realized by repeated controlled embeddings across slots	(Jing et al., 16 Jun 2025)

In the continuous-variable QRNN of “Rapid training of quantum recurrent neural networks” (Siemaszko et al., 2022), the network acts on $n=n_1+n_2$ qumodes, with $n_1$ register modes storing memory and $n_2$ input modes reset each step. Classical input $\mathbf{x}\in\mathbb{R}^{n_2}$ is encoded by displacement,

$$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$

followed by a trainable photonic circuit

$$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$

where the interferometers are compiled via the Clements decomposition. After homodyne measurement of the input modes, the register modes remain unmeasured and become the next-step quantum memory. The architecture is therefore recurrent in a directly physical sense.

The photonic quantum optical recurrent neural network (QORNN) of (Prins et al., 2023) similarly partitions an $m$-mode Gaussian circuit into $io$ input/output modes and $mem=m-io$ memory modes. At each discrete step, fresh quantum input enters the upper modes, the full state passes through the same symplectic circuit $n=n_1+n_2$0, the upper modes are emitted as output, and the lower modes are looped back through delay lines. The short-term quantum memory analysis makes the recurrent structure explicit by block-partitioning an orthogonal symplectic matrix

$n=n_1+n_2$1

with desired delayed-recall behavior

$n=n_1+n_2$2

This is a linear-dynamical-systems formulation of recurrent memory in Gaussian phase space.

A different family keeps the classical recurrent scaffold but replaces gate computations by PQCs. In the encoder–decoder QRNN for time-dependent PDEs (Chen et al., 19 Feb 2025), the QLSTM cell is defined by

$n=n_1+n_2$3

$n=n_1+n_2$4

$n=n_1+n_2$5

with $n=n_1+n_2$6. The corresponding QGRU uses three VQCs: $n=n_1+n_2$7

$n=n_1+n_2$8

$n=n_1+n_2$9

Here the hidden state is classical, but the transition map is quantum-enhanced.

The canonical-model proposal of (Li et al., 2023) instead uses quantum recurrent blocks (QRBs) with a data register Reg. D and a history register Reg. H. Each timestep angle-encodes the current element into Reg. D, applies a hardware-efficient ansatz over Reg. D and Reg. H, measures the first qubit of Reg. D for output, resets Reg. D, and passes Reg. H forward. The paper distinguishes a plain QRNN (pQRNN) from a staggered QRNN (sQRNN), where qubits are assigned to recurrent roles sequentially so as to reduce the coherent-time burden on any single physical qubit.

A closely related but symmetry-aware variant is the time-warping-invariant QRNN (TWI-QRNN) (Nikoloska et al., 2023). Starting from the deterministic QRNN recurrence

$n_1$0

it introduces a quantum-classical adaptive gate

$n_1$1

where

$n_1$2

The gate chooses whether to preserve or update the quantum state, thereby importing a classical gating principle into a quantum recurrent memory.

Finally, the fully quantum QRNN for learning quantum processes with memory (Bondarenko et al., 2023) builds recurrence from dissipative feed-forward QNNs. It splits total input and total output into external and memory subsystems,

$n_1$3

and iterates a one-step quantum channel so that $n_1$4 becomes the next-step memory. The authors explicitly frame this as learning general causal quantum automata, and also describe the architecture as trainable matrix-product quantum channels.

3. Training, differentiation, and recurrent optimization

Training methodologies in the QRNN literature are as diverse as the architectures themselves. Continuous-variable photonic QRNNs can be trained through classical autodiff over simulated circuits. In (Siemaszko et al., 2022), the CV-QRNN is implemented in Strawberry Fields with a TensorFlow backend, optimized with ADAM, using learning rate $n_1$5 for time-series tasks and $n_1$6 for MNIST classification, with batch size $n_1$7. The loss for sequence prediction is mean squared error,

$n_1$8

and binary classification uses

$n_1$9

Because the model is simulated in truncated Fock space, the paper adds

$n_2$0

to penalize trace loss caused by the cutoff.

The density-matrix emulator for multivariate time series (Viqueira et al., 2023) develops a QRNN-specific recurrent differentiation framework. With exchange register $n_2$1 and memory register $n_2$2, the recurrent hidden-state update is

$n_2$3

and, after decomposing $n_2$4, the tensor-form recurrence becomes

$n_2$5

For noisy or shot-based settings, the paper derives analytical recurrent parameter-shift formulas. The first derivative of the output at time $n_2$6 with respect to parameter $n_2$7 is

$n_2$8

and the Hessian entries are given by exact shift formulas for $n_2$9 and $\mathbf{x}\in\mathbb{R}^{n_2}$0. This is effectively a QRNN-specific analogue of backpropagation through time in which parameter perturbations at earlier recurrent blocks contribute to later outputs.

In Bausch’s recurrent quantum neural network (Bausch, 2020), the recurrent cell is built from parametrized quantum neurons plus amplitude amplification. The paper supplies a PyTorch implementation and benchmarks optimization hyperparameters for simple memorization and Elman-style temporal structure learning. The emphasis is on trainable unitary recurrence and on whether unitarity counteracts vanishing gradients. By contrast, the NISQ-oriented canonical QRNN of (Li et al., 2023) uses finite-difference gradients in experiments, although it also presents the parameter-shift rule

$\mathbf{x}\in\mathbb{R}^{n_2}$1

together with the chain rule for $\mathbf{x}\in\mathbb{R}^{n_2}$2 loss. Its reported experiments used gradient descent rather than Adam, even though Adam improved training speed by about $\mathbf{x}\in\mathbb{R}^{n_2}$3 with a slight decrease in accuracy (Li et al., 2023).

Hybrid recurrent reinforcement-learning work pushes training simplification further by freezing the quantum recurrent core. “Reservoir Computing via Quantum Recurrent Neural Networks” (Chen et al., 2022) treats QRNN, QGRU, and QLSTM cells as fixed random reservoirs and trains only a final classical readout. The recurrent equations are unchanged—for example,

$\mathbf{x}\in\mathbb{R}^{n_2}$4

for QRNN, and the six-VQC QLSTM equations for QLSTM—but the quantum parameters remain fixed after random initialization. A related QLSTM-reservoir RL framework (Chen, 2023) similarly uses a frozen 8-qubit QLSTM inside A3C, training only surrounding classical layers.

4. Benchmark tasks and empirical behavior

The empirical QRNN literature spans synthetic sequence learning, real-valued time-series forecasting, image classification under sequentialization, reinforcement learning, PDE latent dynamics, and quantum-process learning.

The CV-QRNN of (Siemaszko et al., 2022) reports that on the Bessel $\mathbf{x}\in\mathbb{R}^{n_2}$5 task the quantum model reaches cost values below $\mathbf{x}\in\mathbb{R}^{n_2}$6 after roughly 10 epochs and below $\mathbf{x}\in\mathbb{R}^{n_2}$7 within 50 epochs, while a parameter-matched TensorFlow LSTM requires about 150 epochs to reach the same order of magnitude. The paper describes this as a “300%” speed gain in epochs-to-convergence. On binary MNIST $\mathbf{x}\in\mathbb{R}^{n_2}$8-vs-$\mathbf{x}\in\mathbb{R}^{n_2}$9, using 1000 downsampled $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$0 images represented as 49-pixel sequences and a 3-qumode architecture with 1 input mode and 2 register modes, the CV-QRNN achieves over $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$1 accuracy after about 10 epochs and exceeds $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$2 by 25 epochs, while the comparable LSTM reaches only about $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$3 and saturates (Siemaszko et al., 2022).

The canonical QRNN of (Li et al., 2023) evaluates pQRNN and sQRNN on three classical sequential tasks. For meteorological prediction using seven previous days to predict the eighth, with Reg. D = 3 qubits, Reg. H = 3 qubits, and 7 QRBs, the pQRNN achieves $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$4 on atmospheric pressure, $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$5 on minimum temperature, $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$6 on maximum temperature, $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$7 on relative humidity, and $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$8 on wind speed, all exceeding the classical RNN baseline; the staggered version remains close while reducing coherence-time requirements (Li et al., 2023). On stock-price prediction, pQRNN similarly outperforms RNN across opening, highest, lowest, closing price, and volume. On a binary sentence-classification task with 130 sentences, pQRNN and sQRNN both reach $$D(\alpha) := \exp\bigl\{ \alpha\hat{a}^\dagger - \alpha^{*}\hat{a} \bigr\}, \qquad D(\mathbf{v}) \equiv \bigotimes_i D(v_i),$$9, matching QSANN and exceeding DisCoCat’s $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$0 (Li et al., 2023).

Bausch’s earlier QRNN (Bausch, 2020) evaluates sequence memorization, Elman’s temporal structure learning tasks, pixel-by-pixel MNIST classification, and an MNIST variant using data augmentation. The paper also analyzes to what extent the unitary recurrent transition mitigates vanishing gradients, making this one of the earliest works to explicitly connect QRNN design to classical unitary-RNN motivations.

The density-matrix-emulated QRNN for multivariate time series (Viqueira et al., 2023) reports three tasks. Using up to 5 qubits total and a hardware-efficient ansatz, the best analytical-gradient RMSEs are $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$1 train / $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$2 val / $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$3 test / $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$4 full-test extrapolation on a dimmed triangular signal, $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$5 on a forced Van der Pol signal, and $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$6 on a multivariate two–Van der Pol task (Viqueira et al., 2023). The paper’s comparison is primarily between analytical and numerical gradients rather than between QRNNs and external classical baselines.

The image-sequentialization study in (Tran et al., 28 Apr 2026) uses a staggered QRNN with a quantum recurrent block (QRB) as core and reports, for binary classification, $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$7 accuracy, $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$8 loss, and generalization bound $$I(\mathbf{\theta}_1,\mathbf{\varphi}_1)\; \rightarrow\; S(\mathbf{r}_{1,2})\; \rightarrow\; I(\mathbf{\theta}_2,\mathbf{\varphi}_2)\; \rightarrow\; D(\mathbf{\alpha}_{1,2}),$$9 on MNIST $m$0-vs-$m$1, but only $m$2 accuracy, $m$3 loss, and generalization bound $m$4 on CIFAR-10 cat-vs-dog. This supports the paper’s broader claim that QRNNs perform well on low-feature data but degrade on high-feature datasets. The same study finds that QRNN reaches maximum accuracy with just 500 data samples and 100 training epochs and that larger datasets did not further improve performance (Tran et al., 28 Apr 2026). This suggests strong data efficiency in that specific binary-sequentialized setting, though the paper also reports that clean-data accuracy remains well below QViT on CIFAR-10.

The encoder–decoder QRNNs for latent PDE evolution (Chen et al., 19 Feb 2025) report strong gains over a classical LSTM baseline. On Burgers’ equation, final test MAE/RMSE are $m$5 for LSTM, $m$6 for QLSTM, and $m$7 for QGRU. On Gray–Scott, QLSTM is clearly best, while QGRU is worse than classical LSTM. On Hamilton–Jacobi–Bellman, both quantum models vastly outperform LSTM, with QLSTM best at $m$8 MAE and $m$9 RMSE. On 3D Michaelis–Menten, both quantum models again beat LSTM, and QGRU is numerically best (Chen et al., 19 Feb 2025).

For reinforcement learning, the QLSTM-reservoir paper (Chen, 2023) evaluates MiniGrid-Empty under fixed and random starts. Its main result is comparative: QLSTM-RC can achieve performance comparable to fully trained QLSTM RL in many settings with the same architecture and training conditions, despite freezing the recurrent quantum parameters. The earlier reservoir-computing paper (Chen et al., 2022) similarly finds that QRNN-RC, QGRU-RC, and QLSTM-RC can approach the performance of fully trained recurrent quantum models, often after around 15 epochs on temporal tasks such as NARMA5 and NARMA10.

Quantum-output QRNNs benchmark a different problem class altogether. The QORNN of (Prins et al., 2023) solves short-term quantum memory perfectly whenever $io$0 for orthogonal-symplectic circuits, can generate temporal entanglement whose logarithmic negativity increases with memory size, achieves communication gain $io$1 up to $io$2 for 30 bosonic memory-channel uses at $io$3, and equalizes memory channels without redundant encoding when decoder delay $io$4 is allowed. A small passive-linear-optical version is demonstrated on Borealis (Prins et al., 2023).

Finally, the fully quantum process-learning QRNN (Bondarenko et al., 2023) learns delay channels, time-dependent quantum-state evolution, and high-/low-frequency noise mitigation. The paper reports that even 8 training pairs can suffice for a one-memory-qubit QRNN to learn and generalize well on a delay-by-one task, while a feed-forward QNN fails to generalize. This is presented as evidence of “striking” generalization from small training sets (Bondarenko et al., 2023).

5. Hardware realizability, resource constraints, and noise

Hardware considerations are central to QRNN design, and different architectures are explicitly tailored to different platforms.

The CV-QRNN of (Siemaszko et al., 2022) is presented as photonic-hardware compatible because its primitive operations—displacements, squeezers, beam splitters, phase shifters, and homodyne detectors—are standard in continuous-variable optics. The paper stresses that the effective nonlinearity is induced by entanglement, partial measurement, and feedback, rather than by strong inline non-Gaussian interactions such as Kerr nonlinearities. At the same time, all of its results are simulation-based, with a Fock cutoff of 6 and exponential classical resource growth, and 50 epochs can take from 1 to 24 hours depending on sequence length. Twenty-five epochs on 1000 downsampled MNIST images take about two days on a cluster (Siemaszko et al., 2022).

The sQRNN of (Li et al., 2023) is motivated directly by NISQ coherence limitations. By staggering the recurrent blocks, it lowers the duration any individual qubit must remain coherent, while angle encoding, hardware-efficient ansätze, partial measurement, and qubit reset make the design more accessible to current devices. This is a hardware-aware architectural claim rather than a hardware experiment.

The QORNN of (Prins et al., 2023) is the most explicit physical recurrence model. It uses optical delay lines to loop memory modes back into the same Gaussian circuit. Borealis is used to realize the simplest QCE setting with $io$5. Because Borealis lacks phase shifters inside loops, the authors implement virtual loop phases through dynamic modulation of pre-loop phase shifters, but the phase range is restricted to $io$6, making some desired settings unreachable without artificial $io$7-shifts. The reported hardware demonstration is therefore a feasibility result rather than a large-scale performance study (Prins et al., 2023).

The alternating-feature-register QRNN of (Morgan et al., 22 Aug 2025) targets a different hardware bottleneck: the circuit depth of amplitude encoding. The paper proposes an amplitude-encoded QRNN using EnQode, a shallow approximate amplitude-encoding method, and an alternating-register recurrent layout claimed to be mathematically equivalent to the original model on an ideal processor while reducing circuit depth. Exact amplitude encoding with Qiskit state preparation gives MSE $io$8 in noiseless simulation, while EnQode gives $io$9, but under IBM Torino noise EnQode slightly improves noisy MSE ($mem=m-io$0 vs $mem=m-io$1) by shortening the circuit (Morgan et al., 22 Aug 2025). This suggests that shallow approximate state preparation can dominate exact state preparation once hardware noise is included.

The robustness study (Tran et al., 28 Apr 2026) distinguishes adversarial perturbations from quantum noise. Against APGD on CIFAR-10, QRNN has the lowest Lipschitz bound and the strongest adversarial robustness among QCNN, QRNN, and QViT, with QRNN accuracy dropping from $mem=m-io$2 clean to $mem=m-io$3 at $mem=m-io$4 and $mem=m-io$5 at $mem=m-io$6, while fidelity changes only from $mem=m-io$7 to $mem=m-io$8 (Tran et al., 28 Apr 2026). However, the same paper reports that QRNN is vulnerable to channel noise, especially amplitude damping, and that on MNIST accuracy under amplitude damping falls to approximately $mem=m-io$9 and then plateaus. This suggests a split robustness profile: strong against classical adversarial perturbation, weaker against physically motivated quantum decoherence.

6. Limitations, controversies, and open directions

Several open issues recur across the QRNN literature.

The first is definitional. Multiple papers either implicitly or explicitly acknowledge that there is no single standard QRNN formalism (Li et al., 2023). Some architectures have genuinely quantum hidden states, such as CV-QRNN, QORNN, TWI-QRNN, the fully quantum process-learning QRNN, and the 2025 hybrid quantum-classical recurrent core model (Xu, 29 Oct 2025). Others only quantize part of the recurrent map, as in QLSTM/QGRU gate substitution (Chen et al., 19 Feb 2025). QRENN (Jing et al., 16 Jun 2025) broadens the concept further by treating repeated controlled data embeddings as a recurrent motif. This suggests that QRNN is still a family resemblance term.

The second is trainability. Several works are motivated by the high cost of recurrent quantum gradient computation (Chen et al., 2022, Chen, 2023). Reservoir-style approaches freeze the recurrent quantum core precisely to avoid this bottleneck. QRENN (Jing et al., 16 Jun 2025) is unusual in that it provides a dynamical-Lie-algebra argument that a recurrent quantum embedding architecture can avoid barren plateaus under specific overlap and locality assumptions, with

$n=n_1+n_2$00

This is a rigorous trainability statement, but it applies to a specialized recurrent embedding architecture rather than to QRNNs in general.

The third is benchmark fairness. Many papers compare against a parameter-matched LSTM or RNN and report better accuracy, lower loss, or faster convergence in epochs (Siemaszko et al., 2022, Chen et al., 19 Feb 2025, Li et al., 2023). But several papers also note that matching parameter count does not guarantee matched expressivity, wall-clock cost, or representational power (Siemaszko et al., 2022). Some studies benchmark against only a limited classical baseline set, and image tasks are often binary or sequentialized rather than standard large-scale vision settings (Tran et al., 28 Apr 2026). This suggests caution when interpreting claims of superiority.

The fourth is simulation dependence. A large fraction of the literature remains classical-simulation based, even when hardware compatibility is emphasized. The 2025 hybrid quantum-classical recurrent-core model (Xu, 29 Oct 2025) scales to 14 qubits in simulation and reports competitive performance on IMDB, MNIST, pMNIST, copying memory, PTB, and Multi30K, but it also states that the experiments are simulation-only and idealized with respect to mid-circuit readout and persistent coherent memory. Likewise, CV-QRNN, encoder–decoder QRNNs for PDEs, and most gate-substitution QLSTM/QGRU studies are simulation-based.

The fifth is expressivity versus implementability. Fully quantum process-learning QRNNs (Bondarenko et al., 2023) can represent general causal quantum automata with finite input and output systems, but the main simulations remain small and mostly use product inputs and outputs across time. QORNNs are physically recurrent and detector-free during online processing (Prins et al., 2023), but remain limited to Gaussian-state processing unless extended with non-Gaussian resources. The entangling/disentangling-power proposal (Daskin, 10 May 2025) provides a conceptually rich view of quantum memory and forgetting, but does not yet furnish a gate-by-gate quantum LSTM formalism.

A final open issue is whether there is a practically meaningful quantum advantage for sequential learning. The literature repeatedly distinguishes empirical advantages—fewer training epochs, small-parameter improvements, better robustness on certain tasks—from complexity-theoretic quantum speedups (Siemaszko et al., 2022, Chen, 2023, Xu, 29 Oct 2025). At present, the field supports several narrower conclusions more strongly than a general advantage claim: that quantum memory can be formulated in several physically meaningful ways; that recurrent quantum models can be trained, at least in simulation and at modest scale; that they can be competitive on some sequence-learning tasks; and that hardware-aware recurrent designs are emerging across qubit, continuous-variable, and photonic platforms.

In that sense, QRNN research presently occupies a transitional position. It has progressed beyond isolated analogies to classical RNNs and now includes persistent-memory qubit models, photonic delay-line processors, trainability-aware recurrent embedding circuits, gate-substitution QLSTM/QGRU systems, reservoir-style recurrent quantum models, and fully quantum learners of memory-bearing quantum processes (Bondarenko et al., 2023). What remains unsettled is not whether quantum recurrence is possible, but which formulation will become the standard reference architecture for broad sequential learning.