Quantum Computing
Grid-state qubit SPAM errors fall below 10⁻³ — a hundredfold leap for bosonic error correction
Grid (GKP) states encode a qubit in a harmonic oscillator and already run quantum error correction beyond break-even, but state preparation and measurement (SPAM) errors have lagged badly. This work uses high-performance QEC for repeat-until-success preparation of both cardinal and magic states, plus a measurement protocol that corrects finite-energy envelope and auxiliary-readout errors while tolerating photon loss, achieving combined SPAM error below 10⁻³ — two orders of magnitude better than the previous state of the art and on par with transmon qubits. That removes a key practical objection to the hardware-efficient bosonic route to fault tolerance.
Quantum Computing
Good qLDPC codes get a speed upgrade: constant qubit overhead with low time overhead
Good quantum LDPC codes promise fault tolerance with constant qubit overhead, but logical operations on them have been slow. Zhang, Zhu and Li show in PRX Quantum that parallelized code surgery combined with locally testable state preparation delivers fault-tolerant computation that keeps the constant qubit overhead while also achieving low time overhead — chipping away at the main practical objection to qLDPC-based architectures versus surface codes.
Quantum computer simulates real-world nonlinear dynamics — including Gulf Stream currents — via Koopman embedding
Unitary quantum evolution is a poor fit for nonlinear dynamics, so this work learns Koopman observables from trajectory data, projects the lifted dynamics onto a finite linear subspace, and decomposes the non-unitary propagator into parallel spectral channels executable as shallow circuits. On a superconducting processor with up to 32 parallel 10-qubit circuits, it captures the dominant multiscale patterns of reaction-diffusion dynamics, fluid motion on a sphere, and satellite-derived Gulf Stream currents, and maps a practical boundary — hardware-noise-limited for weakly nonlinear systems, representation-limited as nonlinearity grows.
RubriQ trains an LLM with rubric-based RL to synthesize hardware-ready fault-tolerant circuits
RubriQ formulates quantum circuit synthesis as an LLM code-generation task optimized with group relative policy optimization, replacing black-box neural critics with a domain-grounded programmatic rubric scoring T-gate reduction, topology compliance and unitary fidelity. With GPU-accelerated CUDA-Q simulation in the RL loop on NERSC's Perlmutter, it reaches 3.31x mean T-gate compression (versus 2.05x for sparse-reward RL baselines), converges 2-3x faster, keeps hardware-constraint violations under 1%, and is validated on IBM and IonQ processors.
RISC-V vector extension scales quantum control to 128 qubits per instruction
Rather than another custom quantum-control ISA, this work extends the RISC-V Vector engine with quantum-oriented instructions that address up to 128 qubits at once and embed parameterized rotations for hybrid programs. A hardware halt-resume protocol resumes the pipeline within 80 ns of a mid-circuit measurement result, and FPGA prototypes show up to 2.52x speedup over the baseline — reusing the mainstream RISC-V toolchain along the way.
Trapped-ion computer targets neutron-scattering observables frequency by frequency
Dynamical structure factors — the quantity neutron-scattering experiments measure — are computed on Quantinuum's Reimei trapped-ion machine using a pumping approach that time-evolves under a Hamiltonian perturbed by a source oscillating at the target frequency. Demonstrated on a 20-site 1D Heisenberg model and a copper sulfate crystal, the method cuts shot overhead significantly when only a few frequencies are needed.
Spectral Born machines: 190-qubit generative models trained entirely on classical hardware
Spectral Born machines view IQP Born machines through group Fourier analysis, using the quantum Fourier transform as an inductive bias for integer-structured data while remaining classically hard to sample in general. They train efficiently on classical hardware via a spectral maximum-mean-discrepancy loss — a 190-qubit model with over a million parameters learns 93-nucleotide rRNA distributions — and appear robust to overfitting; the tooling ships as PennyLane's new tcdq module.
Metastable helium-3 arrays proposed as a platform from CV quantum computing to materials simulation
A PRX Quantum paper maps out quantum science with arrays of metastable helium-3 atoms, showing the species' unusually small mass enables tunable hopping between tweezer potentials and advantages spanning continuous-variable quantum computing and quantum simulation of materials — a new contender among neutral-atom platforms.
Quantum ADMM promises better scaling for large semidefinite programs
Published in Quantum, QADMM builds an inexact ADMM framework that tolerates block-encoding and measurement errors, handles semidefinite conic constraints with a polynomial proximal operator, and accelerates the costly projection updates with quantum singular value transformation. The authors prove convergence to an ε-optimal solution under strong duality and show favorable dimension scaling versus classical ADMM and quantum interior-point methods.
Xanadu plants a U.S. flag in Albany's semiconductor corridor
Photonic quantum computing company Xanadu (NASDAQ/TSX: XNDU) is expanding U.S. operations with a dedicated Albany, New York office, positioning itself inside Upstate New York's advanced semiconductor research and packaging corridor. Covered by both The Quantum Insider and Quantum Computing Report; a strategic footprint move with no financial terms disclosed.
Quantum Comms
Germany funds €3M TruQuaC consortium for trustworthy distributed-quantum control
TruQuaC (Trustworthy Quantum Control and Communication), backed primarily by a €2.46M grant from Germany's Federal Ministry for Research, Technology, and Space, is a €3.06M ($3.5M) project to build a secure control-plane and gateway architecture for distributed quantum systems, with Qoro Quantum and XeedQ among the participants — infrastructure plumbing for networked quantum computing.
Quantum Sensing
Shanghai quantum-sensing startup lands angel round for navigation and gas-detection sensors
The Quantum Insider reports a Shanghai startup raised an angel round worth tens of millions of yuan to scale commercial production of quantum sensors for navigation and industrial monitoring — another data point that sensing remains the most commercially mature corner of quantum tech, though the report names no exact figure.
AI & ML
Theory pins down when LLM self-reflection actually beats parallel sampling
Modeling in-context search as approximate inference over reasoning traces, this analysis shows a sharp dichotomy: when self-reflections reliably localize early mistakes, iterative generate-critique-revise loops solve problems with exponentially small zero-shot pass rates in polynomially many sequential attempts; when they don't, conditioning on past attempts offers no asymptotic gain over parallel sampling. The gains are shown to be learnable via cross-entropy training on search rollouts, and qualitative predictions hold on real large reasoning models.
Danus coordinates parallel math-reasoning agents through a verified fact graph
Danus pairs a planning main agent with parallel proof-search workers and a stateless verifier that checks every proposed claim before it enters a shared fact graph storing proofs and logical dependencies. Six research-level case studies in algebraic geometry, singularity theory and combinatorics illustrate how the fact-graph memory lets the system build long, detailed proofs incrementally; the system is open source, though no quantitative benchmark is reported.
Microsoft's Aurora 1.5 adds hourly resolution and ensembles to its open weather model
Aurora 1.5 extends Microsoft's open Earth-system foundation model with 22 more variables, hourly temporal resolution, and probabilistic ensemble forecasting, targeting real-world weather, climate and energy applications. Vendor announcement — concrete feature deltas but no independent evaluation yet.
Robotics
ABot-C0 claims first motion-tracking scaling law for quadruped robots
The ABot-C0 technical report tackles the quadruped data problem — animal motion capture is scarce and cross-embodiment retargeting fragile — by building a 16,074-clip corpus from conditional video-generation synthesis, annotated mocap, teleoperation and human design. A flow-matching generalist policy trained on it shows, the authors say for the first time, a consistent scaling law for quadruped motion tracking with zero-shot tracking of unseen motions, and a privileged-to-perceptive pipeline with temporal LiDAR memory supports all-terrain locomotion and real-world urban navigation.