Quantum Computing
Silicon spin processor runs itself: 18 qubits driven by a 4 K CMOS controller
HRL Laboratories — jointly owned by Boeing and General Motors — reports in Nature a silicon quantum processing unit that executes high-fidelity multiqubit circuits with every time-varying control signal produced by a digitally programmed cryogenic CMOS chip operating at 4 K, delivered to the low-noise exchange-only qubit device through a high-density superconducting ribbon cable. Trade coverage puts the device at 18 qubits and stresses that the integrated architecture replaces racks of room-temperature control instruments, attacking the I/O-wiring bottleneck rather than the qubit count. A companion Nature commentary frames it alongside a second silicon-processor demonstration as engineering progress toward scalable machines; the missing numbers to watch are the per-gate fidelities and the controller's power budget per qubit as the array grows.
A digitally controlled silicon quantum processing unit (Nature) · HRL Laboratories Demonstrates Self-Running Silicon QPU in Nature Benchmark (Quantum Computing Report) · How silicon-chip technology is being re-engineered for quantum computing (Nature)
Quantum Computing
74-qubit Floquet experiment resolves prethermal oscillations classical simulators can't pin down
On an IBM Heron r3 superconducting processor with the QESEM mitigation stack, a Floquet Ising magnet on the heavy-hex lattice was measured to percent-level precision across up to 74 qubits, resolving long-lived subharmonic prethermal oscillations. The claimed regime is one where leading tensor-network simulations do not converge and sparse Pauli-path simulations stay strongly truncation-dependent despite extensive GPU and Fugaku runs. Finite-size scaling shows an unexpectedly slow decay of oscillation amplitude with system size, evidence the response survives in the thermodynamic limit; the reliability case rests on unbiased mitigation, agreement between independent estimators, noise-model validation, and cross-platform corroboration at selected cycles on Quantinuum's H2 and Helios trapped-ion machines. Preprint, not yet peer reviewed — and the classical-hardness claim is the part most likely to be contested.
Resolving Structure in Prethermal Floquet Dynamics with Precision Quantum Computation (arXiv)
Error-detecting Iceberg blocks beat bare qubits on a 42-logical-qubit Ising simulation
Pairing fault-tolerant syndrome extraction with non-fault-tolerant logical operations, the authors encode 42 logical qubits in 21 [[4,2,2]] Iceberg blocks on up to 136 physical qubits of IBM's ibm_boston, and show encoded Ising simulations estimate local observables better than unencoded ones. Gains run 2-6% at intermediate times in 1+1D and exceed 200% in 2+1D at the latest times studied, where the code's square logical connectivity also cuts circuit depth. The enabling trick is Observable-Ranked Postselection, a syndrome-correlation filter that recovers accuracy without the shot loss of full syndrome postselection.
Hard sampling circuits get a fidelity certificate instead of a proxy benchmark
Sampling-based quantum-advantage demonstrations usually report fidelity proxies that rest on strong noise assumptions. This work introduces structured circuits that keep complexity-theoretic hardness guarantees while admitting an encoding in a quantum code, so an experimental fidelity can be certified from the circuit structure plus measured code syndromes under substantially weaker assumptions. The demonstration is a 70-qubit, depth-70 Clifford circuit doped with 468 T gates, encoded in spacetime codes across 97 physical qubits: syndrome post-selection suppresses gate errors roughly tenfold and yields a fidelity lower bound of 0.284 with 95% confidence. The certificate is device-dependent, and 0.284 is a floor rather than a headline fidelity.
Sampling hard circuits with verifiably high fidelity (arXiv)
An LLM-driven search turns up qLDPC codes that beat bivariate-bicycle baselines at equal qubit budget
OmniQEC frames error-correcting-code design as an iterative discovery process: an LLM orchestrator coordinates code generation, cheap code-level screening in a fast loop, and physically grounded syndrome-extraction synthesis plus decoder-based circuit evaluation in a slow loop that feeds evidence back into the search. Evaluated over four qLDPC families, three LLM backends and 14 total-physical-qubit budgets each, the discovered codes show logical-error suppression improving with budget and beat the standard [[72,12,6]] and [[144,12,12]] bivariate-bicycle codes at complete-implementation budgets of 98 and 240 physical qubits. Simulation-level result, with hardware-friendliness argued rather than demonstrated on a device.
OmniQEC: discovering practical quantum error-correcting codes by an AI scientist (arXiv)
Agents took papers to overnight neutral-atom QPU runs — and needed humans to catch two silent errors
An agentic pipeline drives Pasqal's cloud-accessible neutral-atom processors end to end, from protocol design through compilation, simulation and execution, with a researcher retained for validation. In three many-body-physics and optimization case studies the agent went from a published paper or patent to an overnight QPU campaign — but the honest part of the report is the failures: it picked an inadequate observable in one experiment and constructed a plausible but wrong hardware diagnosis in another, both caught only by domain-expert review. A second agent classified 633 Rydberg-array arXiv papers, judging nearly half implementable on today's hardware and naming the upgrades the rest would need.
Lowering the implementation barrier of neutral-atom quantum computing with agentic workflows (arXiv)
How to trust a quantum estimate with no classical answer key
As quantum simulation moves past the reach of approximate classical methods, the open question is what counts as evidence. This work assembles a validation framework — a designed series of quantum-heuristic experiments that test the underlying assumptions, plus a new observable called the operator Loschmidt echo — and applies it to a physical model whose semi-scrambling dynamics strain several leading classical simulators. It further shows how careful noise characterization converts the problem of validating observable estimation into the more tractable problem of validating a noise model, yielding accuracy bounds rather than bare confidence.
Observable Estimation in the Absence of Classical Verification (arXiv)
Distributed fault tolerance without distillation: one nucleus per node
Distributed architectures usually pay for noisy optical links by stocking each node with extra high-quality data qubits and burning them on Bell-pair or GHZ distillation. This proposal engineers a communication error bias instead: photonic Bell pairs suffer frequent phase errors but rare bit flips, and syndrome-extraction circuits are designed so that phase noise shows up solely as a measurement error that never propagates to data qubits and is suppressed by repeating the measurement — letting the code itself purify the link. Floquet codes then need only a single data qubit per node, general stabilizer codes one extra ancilla, with high thresholds throughout and lattice surgery identified as inherently robust in this regime.
Fault-tolerant distributed quantum computing with a single nucleus per node (arXiv)
LLM fuzzer finds 13 confirmed bugs in Qiskit, PennyLane and Cirq
KQFuzz grounds LLM test generation in codebase knowledge via a prompting scheme tailored to quantum programs, then couples it with fitness-guided evaluation and two-level mutations to push into complex execution paths. Run against Qiskit, PennyLane and Cirq, it improves coverage by up to 18.44% over prior state-of-the-art fuzzers and turned up 13 bugs, all confirmed by maintainers with 12 already fixed — the rare software-quality paper whose headline number is patches landed rather than a benchmark delta.
KQFuzz: Knowledge-Guided Fuzzing for Quantum Libraries via Large Language Models (arXiv)
Open-source atom-rearrangement simulator adds noise models and a dual-species algorithm
Atom rearrangement is the loading step every neutral-atom processor depends on, yet prior algorithm work targeted noiseless settings and shipped little reproducible code. atommovr provides an open-source framework covering three regimes — time-optimal noiseless rearrangement, noisy rearrangement under realistic error models, and noiseless dual-species rearrangement — extracting lower bounds for time-optimal rearrangement and mapping which strategies win in which error regime. It also introduces InsideOut, a dual-species algorithm that avoids 'blocked' configurations with near-unity success rate.
atommovr: An open-source simulation framework for rearrangement in atomic arrays (Quantum 10, 2177)
When readout is the noisiest operation, repeat the measurements — but only then
For hardware where measurement introduces more noise than unitary or idling operations, this work asks how best to schedule error-detecting measurements in dynamically condensed colour codes. Strategic repetition of measurements within the schedule improves performance under measurement-biased noise, but offers little for unbiased or Z-biased models — contrary to the authors' own expectation, and they dig into why. The analysis introduces the teraquop volume (qubits times measurement rounds required to push spacelike or timelike logical error below 1e-12), showing performance differences are driven mainly by rounds rather than qubit count, and finds belief matching rather than minimum-weight perfect matching can flip a worst-performing code into a best-performing one.
Dynamical codes for hardware with noisy readouts (Quantum 10, 2176)
Corrected product formulas cut Hamiltonian-simulation depth
Product formulas are the workhorse of Hamiltonian simulation on both quantum and classical simulators, and their error is usually bought down by taking more, smaller steps. This PRX Quantum paper instead injects carefully designed auxiliary correction terms into the formula, reporting significantly better accuracy for little cost increase — so a target accuracy is reached at substantially lower circuit depth. The work spans a large collaboration including Dominic Berry and Alán Aspuru-Guzik.
Warwick proposes phononic on-chip links for the million-qubit interconnect problem
A Warwick-and-NRC team describes in APL Quantum a chip architecture using phononic links to carry information between qubits over long on-chip distances, aimed squarely at the wiring and connectivity bottleneck that stands between current devices and million-qubit processors. Reported via press release and trade coverage: this is an architectural concept with a peer-reviewed paper behind it, not a fabricated device or measured link fidelity, so treat the million-qubit framing as the authors' motivation rather than a demonstrated capability.
Quantum Comms
AI-discovered heralding schemes lift path-entanglement success probability exponentially
Weak photon-photon interactions make deterministic assembly of large multiphoton entangled states a standing problem in quantum optics. Applying AI search to heralded linear-optical circuits, the authors found solutions that they then lift from individual circuits into a scalable family — one that contains previously known constructions as special cases while generally offering exponential and super-exponential improvements over them, and extends to broader classes of target states. The output is a set of compact experimental proposals for large path-entangled states; the demonstrations remain to be built.
Automated discovery of high-probability heralded schemes for path-entangled states (arXiv)
Quantum Sensing
An agent ran an NV-center experiment autonomously — and more reasoning made it hallucinate resonances
Built around persistent project records, quantitative analysis tools and deterministic hardware control, an LLM agent autonomously selected a single nitrogen-vacancy center, calibrated its resonant frequency, measured T2* with Ramsey sequences, and added a CPMG measurement to probe a weak feature possibly from a nearby 13C nucleus. Two offline benchmarks separate reasoning from lab execution across GPT-5.4, GPT-5.5 and GPT-5.6 Sol: more reasoning effort helped the agent spot a residual calibration offset in the Ramsey checkpoint, but produced more false-positive resonance calls on pulsed-ODMR data — unless the agent was required to compute an expected signal first, which kept false positives low across all models and settings. The authors' conclusion is a division of labor: agents hypothesize, deterministic code holds the hardware and the safety constraints.
Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments (arXiv)
AI & ML
Codec-native tokenizer cuts a streaming VLM's visual tokens by 75% and triples throughput
Mage-VL attacks the inefficiency of uniform frame sampling in vision-language models by tokenizing video the way a codec sees it: the Mage-ViT encoder selectively encodes dynamic, entropy-rich 16x16 regions using motion vectors and residual energy across sparse anchor and predicted frames, cutting visual token consumption by more than 75% while keeping spatiotemporal context. Trained from scratch on roughly 560M unlabeled images and 100M unlabeled video frames, the 4B model is reported to match Qwen3-VL-4B on static tasks, improve on video understanding and 2D/3D spatial reasoning with up to 3.5x wall-clock inference speedup, and beat the 15B Phi-4-reasoning-vision baseline. A dual-system design — a lightweight event gate plus a causal decoder — supports proactive streaming perception. Self-reported benchmarks, no independent replication.
Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model (arXiv)
Quantization forgets memorized text faster than capability — but not fast enough to be a privacy defense
Work claiming quantization reduces privacy risk has measured it almost entirely with membership inference; this study measures the thing people actually worry about — a model reproducing training data word for word — using Pythia models and their known-memorized sequence sets across five precision levels and three sizes, tracking perplexity at every point. Verbatim memorization does fall off faster than capability at every precision and size tested, under two unrelated quantization algorithms and two evaluation corpora. But the selectivity is not enough: at the largest model, four-bit quantization still reproduces most memorized sequences for a few percent of capability, and the surviving fraction grows with model size. The authors' recommendation is to stop treating compression as data removal and to watch extraction rates instead of membership inference. Code and per-configuration results released.
Bits and Memories: Measuring Verbatim Extraction Across LLM Quantization (arXiv)
Mode-division photonic accelerator drops transformer inference power by 64%
Photonic transformer accelerators have leaned on expensive multi-wavelength light generation and large active phase-shifter arrays. MDTransformer instead uses spatial-mode interference: inverse-designed multi-mode couplers, crossings and Mach-Zehnder IQ modulators form a compact tensor core in which each guided mode (TE0-TE3) is an independent computational lane, giving fourfold parallelism per waveguide with no spectral filtering and full compatibility with a single 1550 nm continuous-wave laser. Coherent detection plus IQ modulation encodes amplitude and phase for complex-valued arithmetic, at sub-4-bit effective precision with inter-modal crosstalk below -30 dB. Across DeiT-Tiny/Small/Base and BERT-Base/Large the authors report 40.4% area, 63.6% power and 40.6% energy reductions at comparable latency versus the prior state of the art.
OpenAI pitches GPT-5.6 on cost-efficiency, without publishing numbers
OpenAI's post on GPT-5.6 frames the release around efficiency gains spanning the model itself, inference, and agentic workflows — the stated goal being more useful intelligence per dollar rather than a new capability ceiling. The announcement carries no benchmark results, no pricing table and no third-party evaluation, so the efficiency claim is currently unverifiable; independent throughput and cost-per-task measurements are what would substantiate it.
How GPT-5.6 fuses frontier intelligence with frontier efficiency (OpenAI)
OpenAI opens its top models to 100,000 academic researchers for free
OpenAI says it is giving 100,000 academic researchers free access to its most advanced models through a ChatGPT for Academic Researchers program, positioned as accelerating research, collaboration and discovery. The number is the one concrete detail; eligibility criteria, duration, rate limits and data-handling terms are not in the announcement, and the program is a distribution move rather than a technical result.
Accelerating scientific discovery with ChatGPT for Academic Researchers (OpenAI)
Robotics
Prosthetic control learns combined gestures it was never trained on
Most sEMG prosthesis research recognizes basic movements, but daily tasks like opening a door need combined ones — and collecting data for every combination means retraining for each new pairing. Two methods for prototype networks, Compositional Prototype Interpolation and Synthetic Adaptation for Prototypes, exploit a linear-interpolation structure in embedding space to recognize unseen combined movements zero-shot after training on basic movements only. SAP improves combined-movement accuracy by over 20% versus prior zero-shot methods across the NearLab and NinaPro DB3 datasets and a newly recorded BasCom set, with the advantage retained in online inference during a user study.
Prototype Adaptation for Zero-Shot sEMG Movement Classification (arXiv)