Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2603.12303

Evaluation artifacts in reversible quantum reservoir protocols: a withdrawal of our own claims, and an entangled channel that survives the controls

Hikaru Wakaura, Taiki Tanimae

quant-phSubmitted Mar 12, 2026• Announced Mar 13, 2026• 0 citations• Revised Aug 13, 2026 (v2)
v1 · Mar 12, 2026v2 · Aug 13, 2026v3 · Aug 27, 2026
arXivHTMLPDF

Abstract

This replaces v1 of this manuscript, which reported that a quantum reservoir autoencoder on reset-noise channels suppresses shot-noise sensitivity by ten orders of magnitude and that a two-phase protocol reconstructs unseen inputs at MSE ~ 10^-4. We withdraw both claims. Under one consistent evaluation the suppression reverses sign: reset channels degrade the reconstruction 144x, the apparent gain being an in-sample number compared with an out-of-sample one. The two-phase target is an element-wise map with a public key, invertible in closed form without training at MSE = 7.4x10^-16, and a polynomial using no quantum features matches or beats the full model in all 14 conditions. We trace these to four evaluation artifacts; the fourth also affects the companion work relied on here, whose latent code never leaves its input-independent initialization, so substituting a constant changes nothing. We then report a construction passing the same controls: an entangled channel between two reservoir halves, injected on one side and read out on the other, decoded by a readout calibrated once and applied to unseen messages. Over 32 realizations the held-out error is below chance in 31-32 (permutation p<10^-4) while a shuffled control stays at chance, and non-Clifford resources prove necessary. A decoder-free response-Gram spectrum exposes an access threshold invisible to the decode error, though not a universal one: its median spans f^*=0.20-0.50, growing with the number of multiplexed components. Linear visibility is necessary yet insufficient, diverging from decodability at finite input amplitude-itself a design variable worth an order of magnitude. A variational ansatz trained to convergence on the same task is beaten by the fixed untrained circuit by 2.6-3.6x. We close with the controls, seed convergence and untrained baselines included, separating transport from interpolation.

Related papers

Quantum Reservoir Autoencoder: Conditions, Protocol, and Noise Resilience

Hikaru Wakaura, Taiki Tanimae

Quantum reservoir computing exploits fixed quantum dynamics and a trainable linear readout to process temporal data, yet reversing the transformation -- reconstructing the input from the reservoir output -- has been considered intractable due to the recursive nonlinearity of sequential quantum state evolution. We introduce the quantum reservoir autoencoder, a four-equation encode--decode protocol with cross-key pairing, and constructively empirically demonstrate that satisfying reservoir--key combinations can be found using a full XYZ Hamiltonian reservoir (10~data qubits, feature dimension~76, 16~random Hamiltonian realizations). Under ideal conditions the mean-squared error (MSE) reaches ${\sim}10^{-17}$ for data lengths up to 30; under shot noise (1\,000~shots) and depolarizing noise ($p = 0.005$), the MSE degrades to $10^{-3}$--$10^{-1}$. Asymmetric resource allocation -- 10~shots for encoding, $10^5$ for decoding -- yields a 102-fold MSE improvement (16~seeds $\times$ 3~trials). Comparison of single-body features (dimension~31) with the full feature set and six baselines identifies the iterative protocol structure -- not the feature dimension -- as the dominant noise bottleneck: baselines solving the linear system in a single step retain machine precision under identical noise, whereas per-iteration noise inconsistency in the coupled solver limits the MSE to ${\sim}10^{-1}$. The current protocol requires plaintext access during decoder training, restricting practical deployment. These results establish a proof-of-concept for bidirectional information transformation within quantum reservoir computing and identify iterative noise mismatch and blind decryption as the principal open challenges.

quant-pharXiv:2602.19700•Feb 24, 2026•0 citations

Coherent-feedback and dissipative training of quantum reservoir computers

Hikaru Wakaura, Taiki Tanimae

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 $\sim\!10^{4}$ 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 $\ge0.97$). All code, seeds, and job identifiers are released.

•Oct 6, 2026•0 citations

Trainable Quantum-Native Language Models: Dissipative Memory, Capacity Limits, and Hardware Validation

Deniz Yunus Göğüş

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.

quant-pharXiv:2610.12806•Oct 12, 2026•0 citations

OSIRIS-HP-001: Hayden–Preskill Quantum Information Recovery Benchmark (Preregistration Protocol v1.0)

Devin Phillip Davis

OSIRIS-HP-001 Preregistration Protocol v1.0 This protocol establishes a formal, pre-registered, and falsifiable quantum information recovery benchmark based on the Hayden-Preskill thought experiment (Hayden & Preskill 2007; Yoshida & Kitaev 2017). Designed for execution on IBM Quantum Heron r2 processors (ibm_marrakesh) following deterministic Stage 1 CPU simulation. Core Thesis: Investigates whether an unknown quantum state can be reconstructed by a restricted-access decoder operating strictly on early radiation and newly emitted late radiation after scrambling across an entangled register, and quantifies recovery degradation under calibrated physical noise. Three Controlled Experimental Arms: Arm A (Identity Recovery Control): Full-access reversible control establishing the SPAM ceiling and hardware noise baseline (unit fidelity prediction). Arm B (Restricted-Access Recovery): Message entangled with reference R; black hole B entangled with early radiation E. Scrambler U acts on M + B. Remnant C and Reference R are strictly quarantined. Decoder V_ED operates exclusively on E + D, outputting decoded message M'. Evaluates entanglement fidelity F_rec = <Phi+| rho_RM' |Phi+> reconstructed via XX, YY, and ZZ tomography. Arm C (Noise Degradation Sweep): Calibrated physical noise sweep (depolarizing, dephasing, relaxation) evaluating recovery resilience. Formal Pre-Registered Hypotheses: H1 (Reversible Control): Full-access inverse achieves F >= 0.999 in noiseless simulation. H2 (Restricted Recovery): Decoder on E + D achieves F_rec > 0.50 (quantum entanglement witness) with >10 pp advantage over unentangled controls. H3 (Page Transition): Monotonic fidelity scaling as late radiation size crosses the Page threshold (kD = 1 < kD = 2 < kD = 3 = 1.0). H4 (Noise Sensitivity): Monotonic degradation under physical noise channels. Stage 1 CPU simulation passed all acceptance criteria across N = 20 Haar-random trials (Arm B F = 0.9097 vs. unentangled control F = 0.7223, advantage +18.74 pp, p = 9.05e-12).

•Nov 5, 2026•0 citations

Limitations of Error Model Approximations in Quantum Network Simulation

Julia Freund, Jorge Miguel-Ramiro, Julius Wallnöfer +1 more

Efficient classical simulation of large-scale quantum networks frequently relies on noise approximations, which consider a restricted set of operators to describe noisy channels and operations. In this work, we demonstrate how such simplified error models, such as Pauli twirling or reset channels, can lead to severe quantitative and qualitative discrepancies in protocol performance predictions. We analyze, in particular, how small differences can accumulate in iterative and sequential protocols such as entanglement purification, entanglement swapping, and repeater chains. Our results reveal that neglected error contributions can lead to important performance under- and over-estimations, measurement-outcome dependency, and oscillations in the fidelity, which are entirely overlooked by the simplified error model approximations. These results show that rigorous validation of complete noise architectures is indispensable for accurately predicting operational thresholds in future quantum technologies.

quant-pharXiv:2607.00998•Jul 2, 2026•0 citations

CircuitJEPA: a fully circuit-based Joint-Embedding Predictive Architecture for quantum dynamics

Steve Roques

CircuitJEPA is a Joint-Embedding Predictive Architecture (JEPA) whose three parametrized components (encoder, exponential-moving-average target encoder and horizon-conditioned predictor) are parametrized quantum circuits. Training is done in exact simulation with classical optimization; the trained circuits are then evaluated on shot-based simulators and a real QPU, with embeddings reconstructed from Pauli measurements by classical post-processing. The data are states of a 3-qubit transverse-field Ising chain evolved by Trotter steps; from a context state the model learns, without reconstruction or negative pairs, to predict the latent embedding of the state reached 1, 2 or 3 steps later. Over ten training seeds in exact simulation, the full model reaches a retrieval top-1 accuracy of 0.539 ± 0.341 (chance 1/64) on the original test set, which was also used to select the configuration, and 0.530 ± 0.365 on an independent held-out test set never used for any choice, with large variance across seeds. On the IBM Heron QPU ibm_marrakesh, the pre-registered seed-0 model retrieves 15 of 16 targets and passes the pre-registered criterion against an untrained baseline; in exploratory, not pre-registered runs, two further seeds retrieve 10/16 and 6/16. Negative and inconclusive findings are reported with the same weight: without the EMA and the stop-gradient, 2 of 10 seeds end with near-zero retrieval, not flagged by the formal collapse criterion, and a linear dynamics probe does not separate the full model from the no-predictor ablation. This is a feasibility study on 5 qubits, with no claim of quantum advantage or scalability. Code, results and raw QPU counts: https://github.com/roquess/circuitjepa (software archive: doi:10.5281/zenodo.24259038).

•Nov 15, 2026•0 citations
This replaces v1 of this manuscript, which reported that a quantum reservoir autoencoder on reset-noise channels suppresses shot-noise sensitivity by ten orders of magnitude and that a two-phase protocol reconstructs unseen inputs at MSE ~ 10^-4. We withdraw both claims. Under one consistent evaluation the suppression reverses sign: reset channels degrade the reconstruction 144x, the apparent gain being an in-sample number compared with an out-of-sample one. The two-phase target is an element-wise map with a public key, invertible in closed form without training at MSE = 7.4x10^-16, and a polynomial using no quantum features matches or beats the full model in all 14 conditions. We trace these to four evaluation artifacts; the fourth also affects the companion work relied on here, whose latent code never leaves its input-independent initialization, so substituting a constant changes nothing.
We then report a construction passing the same controls: an entangled channel between two reservoir halves, injected on one side and read out on the other, decoded by a readout calibrated once and applied to unseen messages. Over 32 realizations the held-out error is below chance in 31-32 (permutation p<10^-4) while a shuffled control stays at chance, and non-Clifford resources prove necessary. A decoder-free response-Gram spectrum exposes an access threshold invisible to the decode error, though not a universal one: its median spans f^*=0.20-0.50, growing with the number of multiplexed components. Linear visibility is necessary yet insufficient, diverging from decodability at finite input amplitude-itself a design variable worth an order of magnitude. A variational ansatz trained to convergence on the same task is beaten by the fixed untrained circuit by 2.6-3.6x. We close with the controls, seed convergence and untrained baselines included, separating transport from interpolation.
Quantum reservoir computing exploits fixed quantum dynamics and a trainable linear readout to process temporal data, yet reversing the transformation -- reconstructing the input from the reservoir output -- has been considered intractable due to the recursive nonlinearity of sequential quantum state evolution. We introduce the quantum reservoir autoencoder, a four-equation encode--decode protocol with cross-key pairing, and constructively empirically demonstrate that satisfying reservoir--key combinations can be found using a full XYZ Hamiltonian reservoir (10~data qubits, feature dimension~76, 16~random Hamiltonian realizations). Under ideal conditions the mean-squared error (MSE) reaches ∼10−17{\sim}10^{-17}∼10−17 for data lengths up to 30; under shot noise (1\,000~shots) and depolarizing noise (p=0.005p = 0.005p=0.005), the MSE degrades to 10−310^{-3}10−3--10−110^{-1}10−1. Asymmetric resource allocation -- 10~shots for encoding, 10510^5105 for decoding -- yields a 102-fold MSE improvement (16~seeds ×\times× 3~trials). Comparison of single-body features (dimension~31) with the full feature set and six baselines identifies the iterative protocol structure -- not the feature dimension -- as the dominant noise bottleneck: baselines solving the linear system in a single step retain machine precision under identical noise, whereas per-iteration noise inconsistency in the coupled solver limits the MSE to ∼10−1{\sim}10^{-1}∼10−1. The current protocol requires plaintext access during decoder training, restricting practical deployment. These results establish a proof-of-concept for bidirectional information transformation within quantum reservoir computing and identify iterative noise mismatch and blind decryption as the principal open challenges.
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−12F-12F−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 2.52.52.5-fold in a like-for-like comparison. Third, a fault-tolerant realization---arithmetic and synthesis costs taken from real circuits---needs only ∼ ⁣104\sim\!10^{4}∼104 physical qubits and is runtime-, not qubit-bound. On three free IBM Quantum 156156156-qubit Heron backends at three error-mitigation levels the simulation-optimized state survives device noise (RMSE 0.0160.0160.016--0.0570.0570.057, i.e.$7$--26%26\%26% of the signal spread, correlation ≥0.97\ge0.97≥0.97). All code, seeds, and job identifiers are released.
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.
OSIRIS-HP-001 Preregistration Protocol v1.0 This protocol establishes a formal, pre-registered, and falsifiable quantum information recovery benchmark based on the Hayden-Preskill thought experiment (Hayden & Preskill 2007; Yoshida & Kitaev 2017). Designed for execution on IBM Quantum Heron r2 processors (ibm_marrakesh) following deterministic Stage 1 CPU simulation. Core Thesis: Investigates whether an unknown quantum state can be reconstructed by a restricted-access decoder operating strictly on early radiation and newly emitted late radiation after scrambling across an entangled register, and quantifies recovery degradation under calibrated physical noise. Three Controlled Experimental Arms: Arm A (Identity Recovery Control): Full-access reversible control establishing the SPAM ceiling and hardware noise baseline (unit fidelity prediction). Arm B (Restricted-Access Recovery): Message entangled with reference R; black hole B entangled with early radiation E. Scrambler U acts on M + B. Remnant C and Reference R are strictly quarantined. Decoder V_ED operates exclusively on E + D, outputting decoded message M'. Evaluates entanglement fidelity F_rec = <Phi+| rho_RM' |Phi+> reconstructed via XX, YY, and ZZ tomography. Arm C (Noise Degradation Sweep): Calibrated physical noise sweep (depolarizing, dephasing, relaxation) evaluating recovery resilience. Formal Pre-Registered Hypotheses: H1 (Reversible Control): Full-access inverse achieves F >= 0.999 in noiseless simulation. H2 (Restricted Recovery): Decoder on E + D achieves F_rec > 0.50 (quantum entanglement witness) with >10 pp advantage over unentangled controls. H3 (Page Transition): Monotonic fidelity scaling as late radiation size crosses the Page threshold (kD = 1 < kD = 2 < kD = 3 = 1.0). H4 (Noise Sensitivity): Monotonic degradation under physical noise channels. Stage 1 CPU simulation passed all acceptance criteria across N = 20 Haar-random trials (Arm B F = 0.9097 vs. unentangled control F = 0.7223, advantage +18.74 pp, p = 9.05e-12).
Efficient classical simulation of large-scale quantum networks frequently relies on noise approximations, which consider a restricted set of operators to describe noisy channels and operations. In this work, we demonstrate how such simplified error models, such as Pauli twirling or reset channels, can lead to severe quantitative and qualitative discrepancies in protocol performance predictions. We analyze, in particular, how small differences can accumulate in iterative and sequential protocols such as entanglement purification, entanglement swapping, and repeater chains. Our results reveal that neglected error contributions can lead to important performance under- and over-estimations, measurement-outcome dependency, and oscillations in the fidelity, which are entirely overlooked by the simplified error model approximations. These results show that rigorous validation of complete noise architectures is indispensable for accurately predicting operational thresholds in future quantum technologies.
CircuitJEPA is a Joint-Embedding Predictive Architecture (JEPA) whose three parametrized components (encoder, exponential-moving-average target encoder and horizon-conditioned predictor) are parametrized quantum circuits. Training is done in exact simulation with classical optimization; the trained circuits are then evaluated on shot-based simulators and a real QPU, with embeddings reconstructed from Pauli measurements by classical post-processing. The data are states of a 3-qubit transverse-field Ising chain evolved by Trotter steps; from a context state the model learns, without reconstruction or negative pairs, to predict the latent embedding of the state reached 1, 2 or 3 steps later. Over ten training seeds in exact simulation, the full model reaches a retrieval top-1 accuracy of 0.539 ± 0.341 (chance 1/64) on the original test set, which was also used to select the configuration, and 0.530 ± 0.365 on an independent held-out test set never used for any choice, with large variance across seeds. On the IBM Heron QPU ibm_marrakesh, the pre-registered seed-0 model retrieves 15 of 16 targets and passes the pre-registered criterion against an untrained baseline; in exploratory, not pre-registered runs, two further seeds retrieve 10/16 and 6/16. Negative and inconclusive findings are reported with the same weight: without the EMA and the stop-gradient, 2 of 10 seeds end with near-zero retrieval, not flagged by the formal collapse criterion, and a linear dynamics probe does not separate the full model from the no-predictor ablation. This is a feasibility study on 5 qubits, with no claim of quantum advantage or scalability. Code, results and raw QPU counts: https://github.com/roquess/circuitjepa (software archive: doi:10.5281/zenodo.24259038).