AI & ML
SpiNNaker2 puts hard numbers on the neuromorphic promise: 4.5 TOPS, 2.7 TOPS/W, 150,000 neurons
Neuromorphic hardware has been promised as the energy-efficient alternative to deep networks for two decades, usually without the measurements to back it. The SpiNNaker2 paper is the exception: a real chip, 152 processing elements each pairing an ARM M4F core with dedicated accelerators, an extended SpiNNaker routing fabric for event-based communication, and conventional system interfaces (Gbit Ethernet, LPDDR4) so it can sit in an ordinary machine. The measured envelope is what matters — up to 4.5 TOPS for INT8 deep-network workloads in high-performance mode, up to 2.7 TOPS/W in high-efficiency mode, support for over 150,000 spiking neurons and more than 1.8 billion synaptic events per second at a 1 ms simulation step, and a baseline power under 250 mW that keeps efficiency intact across varying load. The significance is less any single figure than the positioning: the same silicon runs mainstream dense deep-learning workloads and sparse event-based ones, which is the bridge the field has been arguing about rather than building.
Quantum Computing
Qubit-efficient encoding pushes variational optimization to 10,296 variables — and survives real hardware
Variational quantum optimization usually stalls on encoding: one qubit per binary variable puts anything industrially interesting out of reach. This work encodes binary variables in expectation values of Pauli correlation operators, a continuous relaxation that packs many variables into few qubits, and applies it to electric power demand portfolio optimization with a two-stage scheme where a time-averaged problem initializes the time-resolved one. Simulations span m=18 to m=10,296 variables with normalized cost gaps around 1e-4 versus certified-optimal solutions, and the authors show performance is governed by the interplay between the continuous relaxation and the discretization step — larger instances behave more consistently. A trapped-ion run recovers good solutions despite noise and finite sampling. Note the caveat the field is now converging on: a companion line of work (arXiv:2607.20409, covered 2026-07-24) dequantizes PCE entirely, so the encoding's value may prove classical.
A transmon readout that protects itself: nonperturbative cross-Kerr coupling builds in Purcell suppression
Fast, high-fidelity qubit readout and long qubit lifetime pull in opposite directions — coupling a transmon strongly enough to measure it quickly also opens a decay channel into the readout line, conventionally patched with an added Purcell filter. This PRX Quantum work engineers the readout around nonperturbative cross-Kerr coupling instead, so Purcell protection is intrinsic to the architecture rather than bolted on. The result is peer-reviewed and experimental, from the Blais and Scarlino groups; the published abstract does not carry the fidelity and duration figures, so the size of the improvement over filtered designs is worth checking in the paper before drawing conclusions.
Semi-tensor-product exact synthesis outruns SAT-based quantum circuit optimization
Exact synthesis re-implements small circuit shards optimally and is a workhorse kernel inside quantum compilers, but existing methods bottleneck on encoding overhead, memory, and poor parallel scaling. Reformulating the problem with semi-tensor products — enumerating undirected partial-gate topologies, then deciding gate directions by right-to-left factorization — makes both stages parallel, yielding up to 12.8x speedup on 32 workers for this NP-hard task. Against a SAT baseline it runs 100x-1000x faster on small random targets and stays competitive on hard ones; dropped into a real optimization workflow it wins on 89% of QASMBench circuits with a 1.91x median speedup.
Quantum Krylov diagonalization gets practical parameter guidelines — and holds up on IBM hardware
Quantum Krylov diagonalization estimates ground-state properties on near-term devices, but its stability depends on parameters practitioners have mostly tuned by feel. Building on a low-depth Jordan-Wigner implementation that cuts CNOT count for time evolution, this study sweeps Krylov dimension, evolution parameters, system size, interaction strength and singular-value truncation for the 1D periodic Hubbard model, showing that near-closing energy gaps demand longer evolution times while truncation and Trotter error must be balanced against them to avoid numerical instability. The resulting guidelines are validated on IBM hardware, where experiment reproduces ideal-simulation convergence trends with only readout-error mitigation and a modest shot budget.
Partition functions estimate logical error curves without waiting for decoders to fail
Measuring a code's logical error rate by counting decoder failures gets brutally sample-hungry exactly where you care most — the low-noise regime. This paper leans on the stabilizer-code-to-disordered-spin-model mapping to define a ratio of partition functions whose value is the success probability of maximum-partition-function decoding, equal to maximum-likelihood decoding at the Nishimori temperature, and separates it from the order probability, which describes a distinct probabilistic decoding strategy that reduces to maximum-probability decoding at zero temperature. The gap between the two quantifies how much ensembling can recover by exploiting degeneracy among maximum-probability errors, and raises the possibility of a decodability boundary outside the order-disorder phase boundary. Worked out in detail for the toric code under bitflip noise (the Random Bond Ising Model), including qubit-wise error rates, and sketched for the color code.
Bosonic barren plateaus mapped: where linear optics escapes concentration, classical simulation mostly follows
In qubit architectures, barren plateaus and classical simulability have turned out to be two faces of the same coin. This work asks whether the same trade-off holds for bosonic systems, using a representation-theoretic framework for moments of random passive linear-optical circuits: concentration of expectation values is governed by how badly the input state's and the observable's projections onto irreps of the unitary group are misaligned, giving a unified reading of generalized entanglement and locality in the bosonic setting. Broad classes of trainable observables turn out classically simulable. The interesting residue — Fock-state inputs and observables that appear to dodge exponential concentration while keeping a polynomially large signal — is honestly reported as a partial separation: most of the signal stays classically tractable and the rest admits a truncated classical surrogate with polynomially small error.
Grover search, generalized to qudits of mixed dimension
Grover's quadratic speedup is written for qubits and Hilbert spaces of dimension 2^n, which is awkward now that qudit platforms — including heterogeneous systems mixing qudits of different dimensions — are real hardware. This paper builds the general construction: oracles and diffusion operators with and without ancilla qubits, deterministic and fixed-point variants guaranteeing exact or bounded success probabilities, phase-matching analysis, explicit circuit decompositions per platform, and Bloch-sphere trajectory comparisons showing how different phase choices amplify the target. The claimed payoff is reduced circuit depth and higher success probability on multilevel systems; it is a toolkit paper rather than a demonstration, with no hardware results.
A DFT-to-ADAPT-VQE pipeline finds exactly where the classical simulation wall sits
Quantum-embedding workflows for strongly correlated materials involve a long classical preamble before any quantum algorithm runs, and the cost of that preamble is usually left vague. This paper assembles the full chain — DFT geometry relaxation, NSCF calculation, Wannier-based orbital localization producing an active-space Hamiltonian — and hands it to ADAPT-VQE augmented with greedy-operator commutativity partitioning and a Taylor-expanded O(5) operator evolution. Benchmarking Li-hBN while expanding the active space from 8 to 14 spatial orbitals quantifies how computational demand tracks active-space size and locates a sharp 'scaling wall' where classical simulation of the hybrid workflow flips from manageable to intractable. It is a baseline measurement rather than a speedup claim, which is the useful kind of contribution here.
Qutrit neural quantum kernels beat their qubit-model baselines — but the gains saturate
Neural quantum kernels pretrain a quantum neural network and reuse the trained circuit as a task-adapted embedding. Pushing this to qudits with local unitaries in SU(d) gives richer embeddings from more local degrees of freedom and a natural interface for multiclass classification. Focusing on qutrits, the authors sweep the design space — number of encoded features, number of qutrits, 1-to-n versus n-to-n kernel construction, and SU(3) parameterization — across four benchmark datasets on binary and three-class tasks. Qutrit NQKs beat the corresponding QNN baselines in nearly all settings and benefit from scaling both feature budget and system size, but the paper is careful that gains saturate, vary by dataset, and depend materially on the unitary parameterization, which an ablation shows affects both optimization behaviour and final performance.
One classical bit is enough to fake symmetric qubit measurements, a neural network reveals
Reproducing qubit prepare-and-measure statistics classically is known to need exactly two bits of communication in the general case — but that bound says nothing about restricted measurement families. Training a neural network on the simulation task shows a single bit gets high average accuracy for specific families, and that the winners are symmetric measurements with uniformly weighted elements, of the kind that form regular polyhedra. The more interesting step is what follows: reading the patterns the network learned lets the authors write down an analytical protocol, extremely accurate for finite informationally complete symmetric configurations and exact in the limit of a continuous isotropic measurement. Machine learning used as a route to a closed-form result rather than as the result itself.
npj Quantum Information: data-induced randomness in QML classification
A peer-reviewed study in npj Quantum Information examines how randomness induced by the data itself shapes quantum machine learning classification tasks — a question that sits underneath the recurring finding that quantum kernels and classifiers often owe their behaviour to the embedding rather than to anything quantum. The feed entry carries no abstract, so the specific claims and any quantitative results need to be read from the paper.
npj Quantum Information: QML advantage traced to general computational advantage
A peer-reviewed paper in npj Quantum Information argues for grounding quantum machine learning advantage in general computational advantages rather than treating it as a separate phenomenon — the direction the field has drifted as one QML speedup after another has been dequantized. As with the companion npj entry, the feed provides only the title, so the strength and scope of the result require the paper itself.
AI & ML
VQVLA claims 6.5x over an A100 by quantizing robot policies to the robot's own motion state
VLA models are too slow on GPUs for real-time embodied deployment, and existing accelerators treat them as full-precision workloads, leaving redundancy on the table. VQVLA attacks both halves: MotionVQ varies quantization precision dynamically with the robot's execution state, cutting memory traffic while holding task success, and a merged-centroid vectorized GEMM works directly on the codebook-index representation, eliminating redundant multiplications through spatial aggregation and temporal centroid reuse. A matching accelerator design supports dynamic precision selection and centroid reuse. Reported speedups are 6.5x over an A100 and 1.9x-4.3x over the LUT-DLA, CodeGEMM, ShiftAddLLM and Dadu-Corki baselines, with accuracy degradation described as negligible.
Turning verified runs into durable skill exposes an abstraction-execution gap between model sizes
LLMs solve scientific-computing tasks and then forget how: executable feedback from one problem rarely becomes capability on the next. SciConsolidate contrasts verified successes against failures to induce cross-task procedures rather than source-specific patches, filters them through a development-validation gate, and grows the consolidation set with failure-informed, answer-free query synthesis that needs no reference answers. The most useful finding is negative: injecting procedures at runtime gives Qwen3.6-27B +3.85 sub-step and +6.26 main-problem points on SciCode but almost nothing to Qwen3.5-9B — direct evidence that a weaker model cannot operationalize a valid abstract procedure. Having a stronger model concretize those abstractions into executable code supervision for ordinary SFT closes the gap: the 9B student gains +3.89/+6.25 over a matched no-procedure control and +5.62/+11.25 over the base model, deployed with no procedures at all.
Sketch2DES puts inspectable intermediates between an LLM and a simulation model
Generating simulation code straight from a natural-language description leaves non-programmers with nothing to verify. Sketch2DES inserts checkable artefacts: a multimodal LLM turns a queuing-network diagram into semi-structured text, a second LLM pass with a reflection-based verification loop converts that into schema-validated JSON, and a deterministic software adapter compiles the JSON into an executable discrete-event model. Every intermediate can be inspected and automatically validated before anything runs. Across eight diagrams of varying complexity all three stages were highly reliable and outputs were statistically indistinguishable from human-coded and analytical benchmarks — using open-weight models throughout. Scope is restricted to queuing networks and the pipeline still depends on correct visual interpretation.
AI-assisted exploit development turns a kernel traffic-control race into root
STAR Labs published a working local-privilege-escalation exploit for CVE-2026-53264 (CVSS 7.8), a use-after-free race condition in the Linux kernel's network traffic-control subsystem, taking an ordinary local user to root on the targeted CentOS Stream 9 build. The claim worth tracking is the researcher's account of the process: AI assistance contributed both to finding the bug and to speeding up weaponization. That is a specific, attributable data point in the ongoing argument about whether AI shortens the vulnerability-to-exploit pipeline — a claim usually made in the abstract. The assistance is self-reported and not independently measured.
Robotics
ARCHITECT swaps black-box VLAs for synthesized robot programs you can actually correct
Vision-language-action models manipulate impressively zero-shot and fail opaquely — when distribution shift hits, the failure cascades with nothing to inspect. ARCHITECT reframes policy acquisition as interactive program synthesis: an LLM coding agent composes modular programs from a suite of perception and control tools, so a failure localizes to a module rather than smearing across a network. A human supervisor gives natural-language corrections, which are grounded in the policy code by execution traces and distilled into a persistent skill library — long-term in-context learning that accumulates reusable, readable behaviors. On a Franka Panda benchmark it beats both state-of-the-art VLA models and program-synthesis baselines on long-horizon tasks including articulated-object manipulation and cloth folding, with human intervention decreasing as the skill library grows. The headline numbers are not in the abstract, which is the main thing keeping this off the lead.
A robot fingertip that senses in colour reaches 100-micrometre tactile resolution with no compute in the loop
Most tactile skins read pressure at discrete taxels and pay a computational cost to reconstruct a surface. This fingertip moves the sensing into the material: a light-sensitive film exposed for seven minutes to a 5 mW 635 nm laser polymerizes into alternating-density layers forming a Bragg reflector, which shifts its reflected wavelength — red through green to blue — as contact deforms it. An LED and camera inside the transparent silicone finger read that colour field directly into topology, strain and contact-pressure maps at 100-micrometre resolution with no computational latency, demonstrated on a human fingertip, a US penny and a leaf. Unlike most soft sensors it extracts quantitative depth and size, not just relative feature scale. Outside experts (Shadow Robot, Daimon Robotics) call the approach genuinely different while flagging the standing weakness of soft sensing — durability under repeated use, and integration into a working hand, both untested.
One hypernetwork-generated policy flies four different orbital GNC missions
Orbital free-flyers usually carry a separate policy per mission profile, which is brittle as requirements change. HYPER-GNC trains a hypernetwork that maps physics-informed task embeddings to the weights of a shared actor-critic policy, so a single compact controller covers velocity tracking, docking, inspection, and navigation with obstacle avoidance — and because the embedding space is continuous, new mission configurations can be specified at deployment without retraining. Reported sample efficiency matches single-task specialists, with stability held under significant inertial perturbations and external body wrenches, and the sim-to-real gap is crossed on a physical satellite emulator across all four profiles. Code, trained models and deployment scripts are released.
Non-invasive grasp-force decoding hits R^2 = 0.82 across unseen subjects
Continuous grasp-force decoding from scalp EEG is the hard case for non-invasive brain-machine interfaces: the temporal dynamics are messy and inter-subject variability usually destroys generalisation. EEGForceFusion combines a convolutional-recurrent encoder for fine-grained neural structure with quantisation-based tokenisation and transformer temporal modelling, fusing both representations in a single regression architecture. Under strict leave-one-subject-out evaluation on WAY-EEG-GAL it reports R^2 = 0.817 offline and R^2 = 0.793 in simulated real-time operation, with latency the authors describe as deployment-viable — the leave-one-subject-out protocol is what makes the numbers worth noting for assistive robotics and neuro-rehabilitation.