DiffQAS-QLSTM: Quantum Sequence Learning

• The paper presents a quantum sequence learning framework that integrates differentiable quantum architecture search into QLSTM for automated circuit design optimization.
• It applies a differentiable relaxation over quantum modules, enabling joint optimization of both circuit parameters and architecture selection during training.
• Empirical evaluations demonstrate that DiffQAS-QLSTM consistently outperforms handcrafted baselines on multiple sequence prediction tasks, achieving lower loss and higher stability.

DiffQAS-QLSTM is a quantum sequence learning framework that integrates Differentiable Quantum Architecture Search (DiffQAS) directly into the Quantum LSTM (QLSTM) model, so that both the quantum circuit parameters and the architecture selection are optimized simultaneously during training. It was proposed to address a central difficulty in quantum recurrent modeling: designing effective variational quantum circuits (VQCs) remains challenging and often task-specific. In the reported evaluations, DiffQAS-QLSTM consistently outperforms handcrafted baselines and achieves lower loss across diverse test settings, positioning architecture search itself as a primary design variable in quantum sequence modeling (Chen et al., 20 Aug 2025).

1. Terminology and research lineage

In the DiffQAS-QLSTM literature, QLSTM denotes Quantum Long Short-Term Memory, not quaternion long-short term memory. This distinction matters because earlier recurrent-network literature also used the acronym QLSTM for Quaternion LSTM in speech recognition, where the model operated in quaternion algebra rather than with quantum circuits (Parcollet et al., 2018). DiffQAS-QLSTM belongs to the quantum line of work inaugurated by hybrid quantum-classical QLSTM models in which LSTM gate computations are replaced by VQCs (Chen et al., 2020).

The underlying motivation is inherited from prior QLSTM research. The original hybrid QLSTM formulation replaces the classical neural networks inside each LSTM cell with VQCs, preserving the recurrent memory structure while using quantum circuits for gate computation (Chen et al., 2020). Subsequent work showed that QLSTM can be competitive in several settings, but also that performance depends strongly on circuit design, parameterization, and task structure. DiffQAS-QLSTM targets precisely this design bottleneck by moving VQC construction from manual specification to differentiable search (Chen et al., 20 Aug 2025).

A common misconception is to treat “quantum LSTM” as a single fixed architecture. The broader literature does not support that view. Existing QLSTM studies vary in encoding layers, entanglement patterns, qubit counts, parameter-sharing schemes, and the degree of classical post-processing (Chen et al., 2020). DiffQAS-QLSTM is therefore best understood not as a new recurrent update rule, but as an automated method for selecting the quantum substructures embedded inside that recurrent rule (Chen et al., 20 Aug 2025).

2. QLSTM substrate and recurrent cell formulation

DiffQAS-QLSTM retains the standard QLSTM recurrent semantics and applies architecture search to the quantum neural networks inside the gates. The recurrent cell is defined by replacing each gate with a QNN acting on the concatenated previous hidden state and current input, $v_t = [h_{t-1}, x_t]$:

$$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$

Within each gate, the QNN is built from three components: a data-encoding map $U(\vec{x})$, a variational quantum circuit $W(\Theta)$, and a measurement stage that returns classical values by measuring observables $\hat{B}_k$. The paper describes the QNNs as quantum circuits whose candidate modules can differ in encoding choices and variational substructures, including alternatives based on different Pauli rotations or entanglement structures (Chen et al., 20 Aug 2025).

This formulation follows the broader QLSTM pattern established in earlier work. In the original hybrid QLSTM, the classical gate mappings are replaced by separate VQCs for the forget, input, candidate, and output transformations, while the cell-state update remains classical (Chen et al., 2020). DiffQAS-QLSTM preserves that separation of responsibilities but turns the internal circuit design into an optimization problem rather than a handcrafted choice (Chen et al., 20 Aug 2025).

The consequence is architectural modularity at the gate level. Instead of fixing a single quantum ansatz for all recurrent computations, DiffQAS-QLSTM treats each gate-local QNN as a searchable composition of quantum modules. This makes the recurrent cell simultaneously a temporal model and a differentiable architecture-search container.

3. Differentiable architecture search mechanism

The search space is defined by decomposing a quantum circuit $\mathcal{C}$ into a sequence of modules $\mathcal{S}_i$, where each module is chosen from a candidate set $\mathcal{B}_i$. The total number of discrete architectures is

$$N = |\mathcal{B}_1| \times |\mathcal{B}_2| \times \dots \times |\mathcal{B}_n|.$$

The paper gives a simple example: if encoding circuits have 2 options and parameterized gates have 3 options, then each stage has $2 \times 3 = 6$ choices, and with two stages there are $$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$0 total circuit realizations (Chen et al., 20 Aug 2025).

Rather than search this space combinatorially, DiffQAS-QLSTM uses a differentiable relaxation inspired by DARTS. Each candidate configuration $$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$1 receives a trainable structural weight $$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$2, and the effective circuit output is the weighted combination

$$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$3

The paper notes that the weights $$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$4 are often normalized via softmax in differentiable NAS literature, though this is not explicitly specified. The loss $$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$5 is then differentiated with respect to both the structural weights and the circuit parameters:

$$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$6

This yields true end-to-end optimization: the model learns both what to compute and how to compute it in a single gradient-based training loop. The same loss used for the downstream task, such as MSE for regression, drives both ordinary parameter fitting and architecture selection (Chen et al., 20 Aug 2025).

The principal methodological departure from handcrafted QLSTM is therefore not the recurrent update itself, but the replacement of manual ansatz design by a continuous relaxation over circuit families. This is the defining feature of DiffQAS-QLSTM.

4. Configurations, baselines, and benchmark protocol

The paper evaluates four searched variants in addition to handcrafted baselines. The searched variants differ in whether candidate circuits have separate trainable parameters, shared parameters, or fixed random “reservoir” parameters.

Configuration	Parameter rule	Training behavior
DiffQAS-QLSTM-NonShared	Each candidate circuit has its own trainable parameters	Structural weights and circuit parameters are optimized
DiffQAS-QLSTM-Shared	All candidates share the same circuit parameters	Structural weights and shared parameters are optimized
Reservoir-NonShared	Circuit parameters are randomly initialized and not updated	Only structural weights or output projections are trainable
Reservoir-Shared	Circuit parameters are randomly initialized and not updated	Only structural weights or output projections are trainable

The baselines consist of six manually designed, fixed QNN/VQC architectures with handpicked encoding and rotation gates (Chen et al., 20 Aug 2025). This baseline design is important, because DiffQAS-QLSTM is explicitly evaluated against human-chosen circuit structures rather than against a single canonical QLSTM.

The benchmark suite covers five sequence-prediction tasks: Bessel function ($$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$7), Damped Simple Harmonic Motion (SHM), Delayed Quantum Control, NARMA 5, and NARMA 10. For all tasks, the sequence length is fixed at $$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$8, and the predictive objective is to infer $$\begin{align} f_t &= \sigma(\text{QNN}_1(v_t)) \ i_t &= \sigma(\text{QNN}_2(v_t)) \ \tilde{C}_t &= \tanh(\text{QNN}_3(v_t)) \ c_t &= f_t * c_{t-1} + i_t * \tilde{C}_t \ o_t &= \sigma(\text{QNN}_4(v_t)) \ h_t &= o_t * \tanh(c_t) \end{align}$$9 from $U(\vec{x})$0. The main evaluation metric is Mean Squared Error (MSE) on test data, with lower values indicating better performance (Chen et al., 20 Aug 2025).

This protocol isolates short-horizon recurrent prediction and makes the architecture-search comparison direct. It does not test varying context lengths or hardware noise in the DiffQAS-QLSTM paper itself; those questions are taken up elsewhere in the QLSTM literature.

5. Empirical performance and observed behavior

Across the reported benchmarks, DiffQAS-QLSTM-NonShared achieves the lowest or near-lowest MSE on most tasks. Its test MSE values are 0.000229 on Bessel, 0.000019 on Damped SHM, 0.001859 on Delayed Quantum Control, 0.000030 on NARMA 5, and 0.000094 on NARMA 10 (Chen et al., 20 Aug 2025).

DiffQAS-QLSTM-Shared is strongest on the Bessel task, where it reaches 0.000117, and also performs well on Damped SHM with 0.000036, but it is worse than the NonShared variant on Delayed Quantum Control (0.002486), NARMA 5 (0.000472), and NARMA 10 (0.000385) (Chen et al., 20 Aug 2025).

The reported handcrafted reference, identified as Best Baseline (Config 4), gives 0.007316 on Bessel, 0.010588 on Damped SHM, 0.001931 on Delayed Quantum Control, 0.000025 on NARMA 5, and 0.000101 on NARMA 10. These numbers show three distinct regimes. First, the searched models are dramatically better on Bessel and Damped SHM. Second, on Delayed Quantum Control and NARMA 10, DiffQAS-QLSTM-NonShared is slightly better than the best handcrafted baseline. Third, on NARMA 5, the best handcrafted baseline remains slightly better than DiffQAS-QLSTM-NonShared (Chen et al., 20 Aug 2025).

The qualitative analysis is equally important. The paper reports that DiffQAS-QLSTM visualizations show rapid convergence and high stability of predictions, even on complex and delayed-feedback tasks such as NARMA 10. By contrast, manually designed QLSTM baselines frequently show higher drift, prediction lag, or instability. It also reports that the Reservoir variants perform significantly worse, especially on hard tasks. This directly supports one of the paper’s central conclusions: structural search alone is not sufficient; joint optimization of both architecture and circuit parameters is crucial (Chen et al., 20 Aug 2025).

The empirical profile therefore does not reduce to a simple “searched is always best” narrative. The strongest factual summary is narrower and more precise: differentiable architecture search substantially improves QLSTM performance relative to handcrafted baselines on most tested settings, but the gain is task-dependent and does not eliminate all baseline advantages.

6. Position within the broader QLSTM literature

DiffQAS-QLSTM addresses a problem that is visible across the wider QLSTM literature: quantum enhancements are highly sensitive to architectural choice. In urban telecommunication forecasting, for example, handcrafted QLSTM outperforms classical LSTM only for short sequences $U(\vec{x})$1, while classical LSTM regains superiority for $U(\vec{x})$2, and the paper concludes that quantum enhancements are not universally advantageous (Chen et al., 6 Aug 2025). In air-quality sensor calibration, QLSTM slightly outperforms LSTM on the test set, with 66 weights versus 482 weights, but cross-validation shows that the two models have very similar performance overall (Bergadano et al., 2024). These results make the design problem that DiffQAS-QLSTM targets more consequential rather than less.

The broader field is also exploring orthogonal scaling directions. A distributed QLSTM framework partitions VQCs across multiple quantum processing units to address scalability challenges on NISQ devices, emphasizing modular execution and improved training dynamics in a distributed setting (Chen et al., 18 Mar 2025). DiffQAS-QLSTM does not solve that hardware-level scaling problem directly; instead, it automates circuit selection inside the recurrent cell. A plausible implication is that architecture search and distributed execution are complementary rather than competing directions.

The framework is also part of a more general shift from static quantum ansatz design toward adaptive quantum model construction. In the DiffQAS-QLSTM paper, the claimed significance is threefold: it demonstrates that differentiable quantum architecture search can meaningfully improve sequence learning performance in quantum neural models; it lowers the barrier for non-quantum experts by automating circuit design; and it supports broader use in sequence domains such as time-series prediction, NLP, and reinforcement learning (Chen et al., 20 Aug 2025).

Two clarifications help prevent overstatement. First, DiffQAS-QLSTM is not evidence that quantum recurrence is universally superior to classical recurrence; the QLSTM literature remains mixed, with advantages depending on task, sequence regime, and architectural integration (Chen et al., 6 Aug 2025). Second, DiffQAS-QLSTM is not merely a search over fixed random quantum features; the reservoir results show that non-updated circuit parameters are insufficient, and that the practical gain comes from joint optimization of structure and parameters (Chen et al., 20 Aug 2025). Within those bounds, the framework establishes a technically specific contribution: it converts QLSTM circuit design from manual ansatz selection into a differentiable component of recurrent-model training.