Quantum Computing
Ultralow-loss silicon-nitride chip sets a four-photon GHZ fidelity record on a manufacturable platform
Researchers demonstrate a monolithic, ultralow-loss silicon-nitride (Si₃N₄) photonic platform that integrates narrowband photon-pair sources, low-loss qubit-fusion circuits and reconfigurable analyzers on CMOS-compatible 150-mm wafers. On-chip EPR sources reach 0.9875(3) fidelity with near-unity photon indistinguishability (0.990(6) Hong–Ou–Mandel visibility), and fusing two EPR states yields four-photon GHZ states at a record 0.943(8) fidelity and a 27 Hz fourfold rate—more than two orders of magnitude higher than previous silicon-photonic implementations. The work directly attacks the rate-loss barrier that has confined photonic quantum information processing to proof-of-principle scales.
Quantum Computing
Rydberg-atom hybrid algorithm pushes the Quantum Max Cut approximation ratio to 0.651
Quantum Max Cut—the antiferromagnetic Heisenberg Hamiltonian—is a QMA-complete benchmark for quantum approximation algorithms. The authors combine the natural dynamics of a Rydberg-atom system with semidefinite programming and randomized rounding to reach a conditional approximation ratio of 0.651, up from the best-known 0.614 obtained by SDP alone. The advantage persists even when the annealing reaches only 89% of the true ground-state energy, suggesting a practical route for hybrid quantum-classical optimization.
Trapped-ion gate synthesis scales to 1,000-ion chains with global control
As ion counts grow, the dense spectrum of collective motional modes makes multimode entangling-gate design a hard, high-dimensional non-convex problem. This numerical framework searches directly for control fields that realize target spin-spin interactions while suppressing residual spin-motion entanglement, using alternating minimization for stability. It synthesizes all-to-all and nearest-neighbor gates in chains of up to 1,000 ions with only global laser control—and a structured qLDPC target at N=512 with individual addressing—reporting control resources that scale gently with size.
Fixed-frequency 'lattice-patch' transmon design targets surface-code plaquettes
To ease frequency crowding and surface-code mapping in superconducting processors, the authors propose coupling four fixed-frequency transmons to a single fixed-frequency coupler, matching a surface-code plaquette and avoiding flux noise. Multi-level simulations show CNOT fidelities exceeding 0.98 across all six connectivity directions, with residual cross-resonance phase corrected by virtual R_z gates.
Learned quantum embeddings compress ImageNet to 13 qubits
Extending the autoencoder idea to quantum machine learning, this variational framework learns task-specific embeddings of classical data, compressing high-dimensional datasets including ImageNet into a 13-qubit representation that a learned decoder can reconstruct from only polynomially many measurements. It hits 98.5% validation accuracy on MNIST (3 vs 5)—within 1.2 points of a classical baseline and 30+ points above naive amplitude embedding—and remains stable under real IBM hardware noise.
Quantum graph neural network with Weisfeiler-Leman guarantees scales to 56-qubit simulations
The authors build a quantum GNN that performs message passing, is permutation-equivariant, and sits at a chosen level of the Weisfeiler-Leman hierarchy, the standard measure of graph-distinguishing power. Like classical GNNs it can be pre-trained on small graphs to ease variational trainability, with low-cost readout as graphs grow; large-scale simulations of up to 56 qubits cover synthetic graphs, molecular property prediction and the travelling-salesperson problem.
Valence-bond embeddings build shallow circuits for larger quantum-chemistry systems
Most quantum-chemistry simulations focus on small active spaces because of hardware noise and exponential bottlenecks. This work combines hybrid fermionic-bosonic encodings with Quantum Valence Bond Theory to construct a single quantum circuit for the whole molecular system, demonstrating circuit designs that outperform active-space counterparts and approximate exact solutions for comparatively large molecules.
A fidelity metric for quantum annealers, benchmarked to 100 million simulated atoms
Rather than judging annealers only by how well they solve an optimization problem, the authors propose ε, an accuracy for the annealer's equation of state analogous to gate fidelity. Quantum Monte-Carlo benchmarks on Rydberg systems—reaching 100,000,000 atoms on a single CPU—suggest that within ε~10^-2–10^-3 a quantum annealer is indistinguishable from its thermal classical counterpart, outpacing current Rydberg platforms in size and precision and tightening constraints on future hardware.
A thermodynamic Tsetlin machine learns from autonomous heat flow
Turning noise, dissipation and irreversibility into computational resources, this work builds interpretable rule-based classifiers from thermodynamic neurons—autonomous quantum thermal machines that implement AND/NOT/OR logic through heat flow. The resulting stochastic Tsetlin machine runs without external time-dependent control yet, thanks to thresholding and redundancy, reaches accuracy statistically comparable to the standard Tsetlin machine, positioning thermodynamic computation as a viable framework for physical machine learning.
Nonlinear cosφ coupling suppresses measurement-induced transmon transitions
Published in PRX Quantum, this experimental and numerical study explores a nonlinear cosφ coupling that suppresses measurement-induced state transitions during transmon readout, a known obstacle to high-fidelity qubit measurement. The analysis of the resulting dynamics points toward more reliable superconducting-qubit readout.
Configurable photonic simulator maps quantum field dynamics onto optical elements
Reported in PRX Quantum, this work shows that a large class of quantum field dynamics can be decomposed into standard optical elements, enabling a highly configurable and efficient photonic platform for executing field-simulation algorithms.
Taiyi Quantum raises ~$44M to commercialize ytterbium neutral-atom machines
Taiyi Quantum (Shanghai) closed a heavily oversubscribed 300 million yuan (~$44M USD) strategic round to commercialize ytterbium neutral-atom quantum computers, on top of a 100 million yuan angel round from March 2026. The raise—reported by both The Quantum Insider and Quantum Computing Report—was backed by investors including Gaorong Venture Capital and IDG Capital.
AI & ML
Analog oscillator hardware learns to generate images at ~23 µJ apiece
Analog platforms such as coupled oscillators and Ising machines solve differential equations cheaply but impose fixed, physics-determined dynamics that clash with software-defined generative models. The Analog Interaction Systems framework narrows this expressivity gap with time-varying piecewise parameters and hidden physical states, trained through a Wasserstein-GAN procedure that doesn't require following a fixed trajectory. The oscillator-based model reaches FID scores of 27.6 (MNIST) and 80.8 (Fashion-MNIST)—3-4× better than the best prior hardware-implementable analog generators—at an estimated 23 µJ per image, about two orders of magnitude below digital baselines.
LLM meta-optimization discovers a faster 3-SAT algorithm, ~67× on large instances
The authors argue automated scientific discovery should optimize the evaluation criterion alongside the solution. Their 'consensus objective aggregation' combines LLM-generated objective functions via correlation-weighted voting into a stable, self-correcting criterion that evolves as understanding deepens. Applied to algorithm discovery for 3-SAT on digital MemComputing machines, it reduces the empirical scaling from about N^2.51 to N^1.33, delivering roughly a 67× speedup on the largest instances tested.
A Transformer neural quantum state hits chemical accuracy on classical GPUs
HI-NQS embeds a classically trained autoregressive Transformer—with explicit spin-up/spin-down cross-attention as a fermionic inductive bias—inside the sample/diagonalize/update loop of Sample-Based Quantum Diagonalization, distilling each eigenvector back into the network via a spin-marginal teacher signal. Across small molecules and a nitrogen active-space series it reaches chemical accuracy with determinant-count scaling more favorable than conventional CIPSI-based selected configuration interaction, all on GPUs without quantum resources.
Robotics
Embodied test-time scaling lifts robot manipulation without retraining
E-TTS unifies reasoning and action scaling for robotic manipulation through history-aware, closed-loop iterative refinement with vision-language verifiers, using a history buffer to evaluate jointly sampled reasoning-action candidates. Across four benchmarks, six environments, three embodiments and four base vision-language-action models—without extra data or retraining—it improves task success by up to 33.14% in simulation and 26.62% in the real world.