Quantum Computing
Dynamic 'LUCI' error-correction circuits hold their own against the surface code in first IBM hardware benchmark
Researchers ran a reset-free version of the LUCI framework — flexible syndrome-extraction circuits designed to route around broken or noisy qubits — on IBM quantum hardware, the first experimental test of the scheme. Despite needing two subroutine rounds for full syndrome extraction (nearly half the syndrome density in time), LUCI delivered error suppression of 1.75(10) for logical X and 1.93(12) for logical Z, versus 1.58(13) and 2.44(7) for the standard rotated surface code. The result is concrete evidence that dynamic codes can dodge defective components without giving up error suppression, loosening the grip of static fault-tolerant architectures.
Quantum Computing
Constant-copy stabilizer testing collapses without full quantum memory
With unrestricted memory, stabilizer states can famously be tested with just 6 copies of an unknown n-qubit state. This work proves that advantage is fragile: with only k qubits of coherent memory between measurements, testing needs Theta(n-k) copies — so even 0.99n qubits of memory forbids any constant-copy tester — and non-adaptive learning needs Theta(n^2/k). The results, built on a new connection to the hidden-shift problem plus combinatorics of the stochastic orthogonal group, pin coherent quantum memory as the exact resource separating testing from learning, and include an exponential lower bound for purity testing.
Algebraic compilation halves addressing and transport overhead on neutral-atom arrays
A compilation framework for 2D neutral-atom quantum computers converts hardware scheduling bottlenecks into tractable algebra and graph problems: rank-one decompositions of structured single-qubit gate families cut addressing layers by up to 2x versus row/column-wise implementations, while graph-coloring-based transport scheduling reduces atom moves for CZ gates by roughly 50%. Applied to QAOA MaxCut circuits, transport cost drops more than 30% on average, and the methods come with provable performance guarantees backed by extensive numerical experiments.
New 'CAD' permutation-invariant codes cut correlated amplitude-damping errors by an order of magnitude
A quantum error recovery approach tailors coherent recovery circuits to permutation-invariant codes under correlated amplitude-damping noise — a non-Pauli process where stabilizer codes need extra overhead. The new CAD code family's 9-qubit instance outperforms many existing codes by more than an order of magnitude, and the 4-qubit code corrects one global symmetric damping error with a compiled recovery circuit of just 10 gates realizable from linear geometric phase gates.
Gate scheduling alone suppresses idling errors — no extra pulses needed
Instead of inserting dynamical-decoupling pulses, this work suppresses idling decoherence purely by re-timing gates that have scheduling slack in the circuit. Both numerical simulation and hardware experiments show computational accuracy is significantly influenced — and in many cases improved — by scheduling choices, and the paper derives the density-matrix evolution under idling noise to explain the effect.
Neural networks inverse-design SRF cavities and transmons to within a few percent
Inverse design of cavity-transmon systems is a one-to-many problem that normally demands costly iterative electromagnetic simulation. Two deep-network approaches map desired behavior straight to candidate geometries — one proposing SRF cavity shapes for target cavity observables, the other proposing transmon designs for target coupling rate, qubit frequency, and anharmonicity — with re-simulated designs matching targets to ~5% and ~2% respectively.
Thermodynamic limits derived for quantum reservoir computers
By mapping Holevo capacities onto the Bogoliubov-Kubo-Mori geometric manifold, this theory work proves the predictive peak of quantum reservoir computers in the critical region comes from a strict spectral resonance, and derives a generalized Landauer bound for continuous temporal processing. The catch: the same critical resonance that unlocks optimal prediction maximizes irreversible dissipation, while dynamic quantum coherences amplify predictive capacity at no extra mechanical work — design principles for energy-efficient quantum neuromorphic hardware.
Measured tensor network states yield state designs — and an emergent confinement mechanism
New peer-reviewed work in npj Quantum Information studies what happens when tensor network states are subjected to measurement, characterizing when the resulting ensembles form quantum state designs and identifying an emergent confinement mechanism that governs their structure — a contribution to the growing theory of measurement-induced phases of quantum information.
Quantum Comms
Blueprint for quantum-secured machine learning over a prototype quantum network
This proposal combines single-qubit classifiers with blind quantum computation so a remote server can classify a client's data without ever seeing the data or the outcome, delivered over a heterogeneous quantum network link using entanglement swapping and a multiplexed solid-state memory. Simulations with realistic hardware parameters on a real credit-card fraud dataset approach the accuracy of an equivalent classical deep-belief network, and a two-qubit variant adds verification of the computation.
Quantum Sensing
Single qubit senses the tipping point where networks of oscillators break down
The aging transition — where a network's collective dynamics collapse once too many nodes go inactive — can be pinpointed by a single qubit probe coupled to a small subset of oscillators. As the inactive fraction nears the critical threshold, the probe's excited-state population becomes highly sensitive and the Fisher information spikes, enabling high-precision estimation of the breakdown point; remarkably, the enhancement survives even when the oscillators are classical.
AI & ML
LLM pipeline goes from a corpus of 11,083 physics papers to a publication-grade manuscript
An autonomous-research pipeline conceived its own research direction by mapping 11,083 recent condensed-matter arXiv papers, calibrated its methodology by reproducing published references, ran novel first-principles computations, and wrote a manuscript reporting three substantive findings on altermagnetic piezomagnetism. The system ran across 47 fresh-context LLM sessions sharing only on-disk state, with 2,162 literature consultations; ablations isolate enforced numerical confrontation at calibration checkpoints as the mechanism that prevents the hallucinated-but-plausible results that sink unscaffolded agents. Human intervention was bounded and needed only at reproduction failures.
Verified LLM recovery plans cut cloud fault recovery time by 40%
PASE reframes cloud fault recovery as neuro-symbolic program synthesis: an LLM generates structured recovery plans from semantic primitives, a neural-symbolic world model checks their feasibility in simulation, and a DRL-trained meta-prompt optimizer steers the planning. On a real-world cloud fault-injection dataset the reason-plan-verify-adapt loop reduces average recovery time by more than 40% and improves detection accuracy on unknown fault types over state-of-the-art baselines.
Robotics
Embodied.cpp brings llama.cpp-style portability to robot foundation models
Deployment of embodied AI models is fragmented across model-specific Python stacks and robot glue code. Embodied.cpp organizes a shared execution path into five layers (input adapters, sequence builders, backbone execution, head plugins, deployment adapters) for latency-first, batch-1, multi-rate inference on heterogeneous edge hardware. Two VLA deployments reach 100% and 91% closed-loop task success, and a world-action-model benchmark cuts block memory from 312.2 MiB to 88.1 MiB.
Preference learning replaces reward engineering for socially compliant robot navigation
SPLC sidesteps the notoriously hard problem of designing rewards for socially compliant navigation by generating preference data automatically from principled evaluation criteria over pedestrian dynamics. Integrated with offline RL methods it consistently beats state-of-the-art baselines on standard metrics, and real-world TurtleBot4 experiments plus released code back the claim.
Cyborg cockroaches breathe underwater for three hours in printed diving suits
Nature's briefing highlights biohybrid-robot work in which cockroaches carrying electronics were fitted with a printed suit that traps an air supply, letting the insects operate underwater for up to three hours — extending the operating envelope of insect-scale cyborg platforms toward amphibious inspection and search tasks.