AI & ML
AI agent turns plain-English lattice-QCD requests into expert-grade computations
LQCDMaster is a skill-guided scientific computing agent that converts natural-language lattice quantum chromodynamics tasks into complete PyQUDA workflows — measurement scripts, job submissions, and validated numerical outputs — constraining the algebraically fragile parts of code generation with a deterministic Wick-contraction tool. On a 70-task research-grade benchmark it reproduced expert-written implementations at machine precision in 63 cases (three more differed only by convention), cutting implementation time from hours to minutes. The team also used it to compute light-cone distribution amplitudes with a diagonal Wilson line, a quantity accessible with standard methods but never previously calculated, alongside spectra from the proton up to hypertriton.
Quantum Computing
Neural network on an FPGA catches cosmic-ray charge jumps in qubits in real time
Ionizing radiation causes correlated charge jumps in superconducting qubits that today are only diagnosed offline, too slowly for in-the-loop control. A dilated causal convolutional network, trained on Ramsey scans from qubits at Fermilab's NEXUS underground site and compiled to FPGA firmware via hls4ml, now flags these events in 6.19 microseconds per inference on the QICK control platform — matching the established offline chi-squared algorithm's detection efficiency (0.843 vs 0.866 at matched false-positive rate) without per-qubit hyperparameter tuning. That turns charge-jump detection from a post-hoc diagnostic into a control-loop primitive for adaptive error mitigation, and doubles as a particle-detection signature for quantum sensing.
QAOA parameter tuning recast as signal processing, with certified global optima
Parameter optimization is a central bottleneck of the Quantum Approximate Optimisation Algorithm even at depth one. This Quantum paper shows the p=1 landscape for weighted Ising models is a partial Fourier series whose frequencies follow explicitly from the problem's couplings, yielding sampling-resolution bounds that explain why coarse-grid searches return spurious optima. Eliminating the mixer angle analytically reduces the search to one dimension, where a subdivision algorithm finds the globally optimal angle in polynomial time with a certificate when weights are commensurable. Validated inside Recursive QAOA on 128- and 256-qubit weighted instances, the method consistently beat both coarsely optimized RQAOA and semidefinite programming.
Lattice surgery on degree-three chips: fewer qubits, fewer gates, better fidelity
Building on recent results that degree-three qubit connectivity suffices for fault-tolerant error correction, this work introduces scalable circuit constructions for lattice surgery — the standard way to entangle logical qubits — on trivalent planar architectures. The trivalent protocol saves O(d) physical qubits from a total of O(d^2) and O(d) two-qubit gates from O(d^3) compared with the conventional four-valent measurement scheme. Realistic simulations targeting a fluxonium-based implementation show up to about 25% logical-fidelity improvement at distance three, supporting sparse-connectivity routes to surface-code logical processors.
Low-rank trick brings gradient-based pulse design to bigger open quantum systems
LROC observes that because quantum protocols are designed to preserve purity, the density matrix stays dominated by a few pure states and admits an accurate low-rank factorization. Propagating only that factorization — with an adjoint equation giving gradients at the same reduced cost — yields a quadratic time and memory improvement over full master-equation optimization. Demonstrations on five-qubit GHZ preparation, a CNOT gate, qubit readout, and an error-correction primitive with realistic multilevel transmons reached fidelities at the intrinsic dissipation limits.
One framework unifies how measurement shapes quantum reservoir computers
Measurement back-action in online quantum reservoir computing has been treated scheme-by-scheme. This general theory based on indirect measurements unifies projective, weak, partial, and dissipative monitoring, deriving when monitored dynamics satisfy the echo-state property and fading memory — including a necessary-and-sufficient condition for emergent strict contractivity. It shows back-action can supply the effective dissipation a reservoir needs even when the unmonitored dynamics are unsuitable, establishing measurement engineering as a design tool across platforms.
Stabilizer scars offer a yardstick for quantum-advantage simulation claims
Verifying quantum simulations that classical computers cannot reproduce is a core obstacle for quantum-advantage claims. This proposal uses stabilizer scars — a special class of quantum many-body scars whose structure permits both classical simulability and efficient direct fidelity estimation — and shows that, under a physically motivated error model, fidelity on these states bounds the fidelity of classically intractable non-equilibrium simulations on the same device.
Adaptive entanglement pre-generation speeds up multi-core quantum chips in simulation
Multi-core quantum processors move qubits between cores via entanglement-assisted teleportation, making entanglement supply a performance bottleneck. Comparing reactive on-demand generation, static continuous pre-generation, and a new adaptive scheme (ACGP) that adjusts generation probabilities to observed inter-core traffic, simulations in an extended SeQUeNCe on mesh architectures with real benchmark circuits show ACGP significantly reduces average teleportation latency, while entanglement purification recovers the fidelity lost to storage with minimal latency impact.
Smarter moment selection makes the NPA hierarchy go further on a budget
The NPA hierarchy's semidefinite relaxations are limited in practice by combinatorial growth in operator moments. Treating moment selection as combinatorial subset selection, this work shows parallel tempering, an RBM-based reinforcement-learning policy, and Bayesian optimization all substantially outperform greedy selection at around two orders of magnitude below brute-force cost on the I3322 Bell inequality benchmark. Applied to the 174 Bell inequalities of the (4,4,2,2) scenario and the 1D Heisenberg chain, a budget-aware search improves certified bounds on long-range correlations by nearly two orders of magnitude.
Backpropagation comes to Pauli-propagation circuit simulation
This work derives a backpropagation algorithm for evaluating parameter gradients in quantum circuits simulated via Pauli propagation, with computational cost comparable to standard sparse Pauli techniques and gradient accuracy on par with the expectation values themselves. Exploiting circuit reversibility cuts memory by a factor of the parameter count relative to conventional reverse-mode automatic differentiation. Demonstrations include optimizing low-energy state-preparation circuits for transverse-field Ising models in one to three dimensions and the 3D Heisenberg model, and compressing 2D time-evolution circuits.
Non-Abelian topological order enlisted to make logical magic states
Published in npj Quantum Information, this work proposes generating logical magic states — the key resource for universal fault-tolerant quantum computation — with the aid of non-Abelian topological order, offering an alternative to conventional magic-state distillation overheads.
From superradiance to superabsorption: exact results beyond the Markovian regime
Going beyond Markovian and mean-field approximations, this Quantum paper derives a complete analytical solution for two emitters and numerically exact dynamics for ensembles up to a thousand atoms in a lossy cavity. It maps three regimes — the standard superradiant burst, a non-Markovian phase where the ensemble spontaneously superabsorbs its own emitted field, and a critical regime of pulsed collective emission — and shows the critical spectral width grows with emitter number, meaning cooperativity enhances environmental memory effects relevant to quantum batteries and collective light-matter technologies.
Quantum Comms
Remote quantum operations that use two photonic degrees of freedom at once
Most quantum remote-control protocols built on hyperentangled photons use only one degree of freedom at a time. This proposal simultaneously exploits polarization and spatial modes of a two-qubit hyperentangled state to remotely implement an arbitrary hybrid operator on an unknown two-qubit hyperstate, constructed from linear optical elements and cross-Kerr interactions. The authors analyze how finite coherent-state distinguishability and dissipation degrade success probability, and show appropriate choices of Kerr phase shift and amplitude substantially improve performance for distributed quantum information processing.
Photonic qubit gates from scattering off emitter arrays, in npj Quantum Information
A new npj Quantum Information paper presents photonic qubit gates implemented through one-dimensional scattering of photons from an array of two-level emitters, a waveguide-QED route to photon-photon logic that avoids bulk nonlinear media.
Post-Quantum Crypto
AT&T and Palo Alto Networks pitch a quantum-resilient SASE fabric
AT&T Business is extending its quantum-safe networking portfolio, following its Cisco-powered quantum-safe SD-WAN, through a partnership with Palo Alto Networks that integrates Prisma SD-WAN with AT&T's global fiber and cellular networks into a natively quantum-resilient Secure Access Service Edge offering. The announcement signals post-quantum cryptography moving into mainstream enterprise networking products, though it is a vendor announcement without published technical or performance details.
AI & ML
Graph-based RAG lets a 1.7B model punch 18x above its weight on-device
SmartRAG decomposes an on-device assistant into Perception, Memory, Focus, and Thinking modules: a continually learnable named-entity recognizer absorbs unseen entity types without retraining the backbone, extracted knowledge lands in a three-layer provenance-preserving graph, and retrieval combines graph traversal, lexical matching, and dense search. On TriviaQA, Natural Questions, HotpotQA, and MultiHopQA, a quantized 1.7B-parameter backbone achieves multi-hop reasoning competitive with models up to 18 times larger, entirely within smartphone memory and latency envelopes.
ADC-free analog in-memory compute brings Transformers into FPGA crossbars
Conventional ReRAM in-memory computing accelerates only static-weight matrix multiplication, leaving Transformers' nonlinearities and dynamic matrix products to the FPGA fabric — and its ADCs consume over 70% of block area and power. NIFA replaces the ADC with analog content-addressable memories that perform nonlinear operations natively inside the block and maps dynamic matrix-matrix multiplication onto them, extending in-memory computing to attention. Reported gains reach 40x energy efficiency on CNN benchmarks and 1.9x on Transformer workloads, with 4.1x and 2.5x area-efficiency improvements.
A 7B model out-reasons its 72B teacher via execution-verified program distillation
Natural-language rationales are unreliable supervision for numerical reasoning because of arithmetic errors. This framework distills financial reasoning over hybrid tables and text using execution-verified Python programs: gold derivations guide teacher-side program synthesis, only programs that execute to the gold answer are kept, and an iterative recovery stage revisits teacher failures. The resulting 7B student scores 87.00 EM / 87.18 F1 on TAT-QA, substantially above its 72B teacher's 78.46 EM and strong baselines including TAT-LLM.
New agent benchmark spanning 354 domains keeps frontier models under 60%
OmniaBench evaluates general agents across consumer, business, and enterprise scenarios via a hierarchical taxonomy of 90 level-1 and 354 level-2 domains, with executable environments and tasks synthesized through four complementary routes. Across 1,431 tasks (with a 644-task challenge subset to limit contamination), frontier models top out at Overall Pass@1 scores of 58.54 for Claude-Sonnet-5 and 57.14 for GPT-5.6-Sol, with analyses pointing to persistent limitations in planning, constraint maintenance, and adaptive correction.
Robotics
A scaling recipe for humanoid behavior foundation models pays off on real hardware
Behavior Foundation Models promise general-purpose humanoid control, but how to scale them has been unclear. This work coordinates three ingredients — reformulating diverse control problems as global-frame whole-body motion tracking, balancing on-policy rollout quantity against reference-motion diversity, and an expressive Humanoid Transformer architecture — and reports over 10% reduction in Mean Per-Keypoint Position Error in local mode and 82% in global mode against existing humanoid controllers, with gains carrying through extensive simulation and real-world deployment.
Dual-timescale adapters let a robot keep learning without forgetting
LifelongVLA tackles the plasticity-stability trade-off in vision-language-action models with a dual-timescale LoRA gating module: a short-term adapter absorbs new tasks while a long-term adapter consolidates stable skills, integrated through a task-aware gate. A cache-efficient stochastic replay strategy preserves balanced retention signals without storing full trajectories. Experiments show better skill expansion and retention than existing lifelong-learning baselines, including real-world deployment on an xArm robot.
A humanoid that picks its own moves from what it hears
Most humanoid performances are pre-scripted or externally triggered. This framework processes continuous audio streams and routes them into music or speech branches: audio fingerprinting and semantic embeddings align musical segments to motion policies dynamically, while speech is grounded into a discrete library of imitation-learned skills, all scheduled over a reinforcement-learning whole-body control pipeline. The approach is validated in simulation and on a Unitree G1 humanoid with consistent audio-conditioned policy selection and robust sim-to-real transfer.