We study byte-level autoregressive language models whose hidden state is a small quantum register, whose transitions are learned token-conditioned quantum channels, and whose next-byte probabilities come directly from Born-rule measurements without a learned classical vocabulary projection or softmax. In the fixed-width recurrent setting, unitary dynamics cannot contract or reset stored distinctions; experimentally, unitary recurrence loses to its retrained one-token control in every tested variant. Introducing learned dissipation, then token-conditioned dissipative gating, yields useful memory beyond one token. The gated model improves test negative log-likelihood by 0.20 +/- 0.03 nats/byte over retrained parameter-matched one-token controls at context length 12, and by 0.29-0.33 at longer contexts, but never beats retrained two-token controls. A key-value recall benchmark shows a width-dependent threshold: one stored pair meets the pre-registered accuracy bar at 4-5 memory qubits, two pairs at 6-7, and three pairs remain unsolved. This is consistent with, but does not establish, a capacity explanation for the short effective memory horizon. As a secondary result, one-qubit models reach lower stationary memory entropy than a heuristically optimized classical predictive frontier at matched error in 9/9 seeds. On IBM Heron hardware, computational-basis memory-register distributions show signal against a uniform null in 12/12 circuits across prefixes of up to 12 tokens, next-bit probabilities pass pre-registered accuracy bars at four-token prefixes, and bytes are generated autoregressively by sampling physically measured bit probabilities. The tested models are classically simulable and substantially worse than classical n-gram baselines. We claim no computational quantum advantage.
We study byte-level autoregressive language models whose hidden state is a small quantum register, whose transitions are learned token-conditioned quantum channels, and whose next-byte probabilities come directly from Born-rule measurements without a learned classical vocabulary projection or softmax. In the fixed-width recurrent setting, unitary dynamics cannot contract or reset stored distinctions; experimentally, unitary recurrence loses to its retrained one-token control in every tested variant. Introducing learned dissipation, then token-conditioned dissipative gating, yields useful memory beyond one token. The gated model improves test negative log-likelihood by 0.20 +/- 0.03 nats/byte over retrained parameter-matched one-token controls at context length 12, and by 0.29-0.33 at longer contexts, but never beats retrained two-token controls. A key-value recall benchmark shows a width-dependent threshold: one stored pair meets the pre-registered accuracy bar at 4-5 memory qubits, two pairs at 6-7, and three pairs remain unsolved. This is consistent with, but does not establish, a capacity explanation for the short effective memory horizon. As a secondary result, one-qubit models reach lower stationary memory entropy than a heuristically optimized classical predictive frontier at matched error in 9/9 seeds. On IBM Heron hardware, computational-basis memory-register distributions show signal against a uniform null in 12/12 circuits across prefixes of up to 12 tokens, next-bit probabilities pass pre-registered accuracy bars at four-token prefixes, and bytes are generated autoregressively by sampling physically measured bit probabilities. The tested models are classically simulable and substantially worse than classical n-gram baselines. We claim no computational quantum advantage.
Abstract We introduce and test two measurement-free training primitives for quantum reservoir computers (QRC). Protocol~A encodes the loss gradient in an auxiliary phase that is fed back coherently before a dissipative reset; Protocol~B realizes gradient descent as relaxation of an engineered open system. Both match exact-gradient training in simulation---the ancilla and reset simulated as channels in a de-risk stage, their validated error model applied at the benchmark stage---on a minimal transverse-field Ising reservoir, with the coherent cost oracle emulated classically and its realization costed. A fully coherent update loop (quantum parameter register, unitary controlled update, no readout anywhere) descends the true landscape, its feedback drift attenuated by exactly 2F−1 with the fixed point unchanged. Benchmarked against classical reservoir computing--- including the memory-augmented hybrid---they improve one-step chaotic prediction, though a thirty-seed study finds no autonomous-prediction advantage; we therefore make no quantum-advantage claim. The study instead addresses three questions that gate this route. First, within the channel model studied, the reset fidelity already demonstrated in measurement-free quantum error correction [Butt \textit{et al.}, Nat.\Commun.\\textbf{17}, 995 (2026)] suffices: reset is not the bottleneck. Second, the binding constraint is coherent cost-oracle precision, set by a finite-difference catastrophic cancellation that the exact parameter-shift rule removes, loosening the oracle-precision target 2.5-fold in a like-for-like comparison. Third, a fault-tolerant realization---arithmetic and synthesis costs taken from real circuits---needs only ∼104 physical qubits and is runtime-, not qubit-bound. On three free IBM Quantum 156-qubit Heron backends at three error-mitigation levels the simulation-optimized state survives device noise (RMSE 0.016--0.057, i.e.$7$--26% of the signal spread, correlation ≥0.97). All code, seeds, and job identifiers are released.
Quantum language models have shown competitive performance on sequential tasks, yet whether trained quantum circuits exploit genuinely quantum resources -- or merely embed classical computation in quantum hardware -- remains unknown. Prior work has evaluated these models through endpoint metrics alone, without examining the memory strategies they actually learn internally. We introduce the first mechanistic interpretability study of quantum language models, combining causal gate ablation, entanglement tracking, and density-matrix interchange interventions on a controlled long-range dependency task. We find that single-qubit models are exactly classically simulable and converge to the same geometric strategy as matched classical baselines, while two-qubit models with entangling gates learn a representationally distinct strategy that encodes context in inter-qubit entanglement -- confirmed by three independent causal tests (p < 0.0001, d = 0.89). On real quantum hardware, only the classical geometric strategy survives device noise; the entanglement strategy degrades to chance. These findings open mechanistic interpretability as a tool for the science of quantum language models and reveal a noise-expressivity tradeoff governing which learned strategies survive deployment.
We present a hybrid quantum-classical recurrent neural network (QRNN) architecture in which the recurrent core is realized as a parametrized quantum circuit (PQC) controlled by a classical feedforward network. The hidden state is the quantum state of an n-qubit PQC in an exponentially large Hilbert space C2n, which serves as a coherent recurrent quantum memory. The PQC is unitary by construction, making the hidden-state evolution norm-preserving without external constraints. At each timestep, mid-circuit Pauli expectation-value readouts are combined with the input embedding and processed by the feedforward network, which provides explicit classical nonlinearity. The outputs parametrize the PQC, which updates the hidden state via unitary dynamics. The QRNN is compact and physically consistent, and it unifies (i) unitary recurrence as a high-capacity memory, (ii) partial observation via mid-circuit readouts, and (iii) nonlinear classical control for input-conditioned parametrization. We evaluate the model in simulation with up to 14 qubits on sentiment analysis, MNIST, permuted MNIST, copying memory, and language modeling. For sequence-to-sequence learning, we further devise a soft attention mechanism over the mid-circuit readouts and show its effectiveness for machine translation. To our knowledge, this is the first model (RNN or otherwise) grounded in quantum operations to achieve competitive performance against strong classical baselines across a broad class of sequence-learning tasks.
Simulating the showers that particles deposit in a calorimeter, the detector that records their energy across many cells, is one of the largest computing costs in high-energy physics, and fast generative surrogates are needed. Quantum circuits have been proposed as such surrogates, but demonstrations on calorimeter data have either required a register that grows with the image, roughly one qubit per cell, or have passed the information between parts of the image through a classical channel. We remove both constraints for this problem, using a construction drawn from the sequential-generation and hidden-quantum-Markov-model literature. A shower image is generated block by block on a fixed register of six qubits. Three of them, the memory, are never measured. They hold what the circuit knows about the blocks already generated as a quantum state, so the correlations between distant parts of the image cross each block boundary as unmeasured amplitudes. The other three are measured and reset once per block, and each outcome selects the energy pattern of one block. The register size is set by the block, so adding cells to the image adds steps to the sequence, not qubits. We call the model a coherent-memory Born machine (CoMB), a Born machine being a circuit whose measurement statistics are the generated distribution. Its training loss grows only linearly with the number of blocks and never requires enumerating the image distribution, and a depth-two instance runs on an IBM superconducting processor. On a twelve-cell benchmark the model reproduces the energy distribution of every cell, the correlation matrix between cells, and the total-energy spectrum. Removing the three memory qubits, with everything else unchanged, produces independent blocks. Every correlation between blocks is carried by the unmeasured qubits and by nothing else.
Abstract Generative models are a class of machine learning models that aim to learn the underlying probability distribution of data. Unlike discriminative models, generative models focus on capturing the data's inherent structure, allowing them to generate new samples that resemble the original data. To fully exploit the potential of modeling probability distributions using quantum physics, a quantum-inspired generative model known as the Born machines have shown great advancements in learning classical and quantum data over matrix product state(MPS) framework. The Born machines support tractable log-likelihood, autoregressive and mask sampling, and have shown outstanding performance in various unsupervised learning tasks. However, much of the current research has been centered on improving the expressive power of MPS, predominantly embedding each token directly by a corresponding tensor index. In this study, we generalize the embedding method into trainable quantum measurement operators that can be simultaneously honed with MPS. Our study indicated that combined with trainable embedding, Born machines can exhibit better performance and learn deeper correlations from the dataset.
We propose a compact NISQ-compatible quantum transformer architecture for synthetic QNLP sequence modelling. The model preserves the autoregressive next-token interface of a classical transformer, but replaces attention and feed-forward sublayers with variational quantum encoder blocks, connector circuits, decoder blocks and a direct two-qubit measurement readout. Token contexts are angle-encoded into small quantum registers, processed by parallel variational heads and encoder integration circuits and conditioned through decoder ancillae to produce a distribution over a four-token vocabulary. We evaluate several architecture variants on deterministic and lexicographic grammar-generation tasks against a compact classical transformer baseline. The quantum models are trainable end-to-end and learn nontrivial grammar structure, including perfect deterministic generation in individual runs and high lexicographic validity in the strongest variant. The classical baseline remains more accurate and stable and the quantum models are sensitive to initialization. The contribution is therefore not a claim of quantum advantage, but a concrete architecture and evaluation of transformer-inspired QNLP sequence modelling under near-term quantum constraints.