Quantum Computing
Trapped-ion architecture clears the bar for early fault tolerance — no postselection required
Quantinuum and collaborators introduce the [[20,2,6]] C4-Helix code and validate a complete early-fault-tolerant architecture on the 98-qubit Helios trapped-ion processor, rather than the bare logical memories that have dominated the field. Repeated error correction runs at 4.6e-5 error per logical qubit per QEC cycle, benchmarking the full Clifford group on two logical qubits in a codeblock gives 2.8e-4 per two-qubit logical Clifford, and a fault-tolerant chain-map interface to a distance-5 surface code prepares a heterogeneous three-logical-qubit GHZ state with a fidelity lower bound of 99.925%. In every case the encoded implementation outperforms the corresponding physical baseline without postselection — the caveat that has qualified most prior logical-advantage claims. Circuit-level simulations project the same architecture into the 1e-6 to 1e-8 logical error regime as physical fidelities improve.
arXiv quant-ph — Experimental validation of a compact fault-tolerant architecture for trapped ions
Quantum Computing
Magic-state distillation with hundreds of QEC rounds, simulated exactly in 40 seconds
Simulating error-correction circuits that contain non-Clifford gates is the bottleneck holding back fault-tolerance design work, and near-Clifford simulators only partly cover it. This work shows matrix-product-state methods handle many such circuits exactly with no restriction on gate type — provided the implementation is right. A series of targeted optimizations reduces bond dimensions and simulation time by several orders of magnitude over naive MPS: a 15-to-1 magic-state distillation circuit with hundreds of QEC rounds compresses to 11 logical qubits (187 physical) at a maximum bond dimension of 64 and runs in under 40 seconds, and rotated surface-code memory reaches distance 11. The headline finding is methodological — circuit-level optimization choices, not the MPS technique itself, determine whether the approach is viable.
arXiv quant-ph — Optimized Matrix-Product State Simulations of Quantum Error Correction Circuits
Neutral-atom compilation trick buys three orders of magnitude of fidelity on a 30-qubit QFT
Neutral-atom processors pay a heavy price for moving atoms around, so how a circuit is compiled onto the array matters as much as gate quality. This work adapts Parity Twine Networks — an efficient realization of the Parity Architecture — to different atom layouts, native entangling gates (CZ, CZSWAP, iSWAP) and shuttling capabilities on both static and mobile platforms, giving a general hardware-aware framework for encoding arbitrary interaction graphs. Using the quantum Fourier transform as the test case, the constructions cut two-qubit gate count, atom transport and depth by enough to give an estimated circuit fidelity three orders of magnitude higher than competing compilation strategies at 30 qubits. The estimate is analytic rather than measured on hardware.
arXiv quant-ph — Optimizing Atom Transport, Gate-Count and Depth with Parity Twine
Haar-like randomness from depth-seven circuits
Random quantum states are a workhorse for characterization, metrology and many-body physics, but exact Haar randomness is expensive. This construction — a sparse commuting Clifford phase layer plus independent single-qubit Cliffords acting on a product state — gives epsilon-approximate projective 2- and 3-designs in relative error, the strictest notion, with an interaction degree that is asymptotically optimal within the circuit family. It runs ancilla-free at O(log(n/epsilon)) depth on all-to-all hardware, or in adaptive constant depth — specifically depth seven — using O(n log(n/epsilon)) ancillas. The third-order analysis needed a new block decomposition that the authors argue points toward higher-order shallow designs.
arXiv quant-ph — Ultra-Precise Quantum Projective Designs in Constant Depth
Nuclear scattering priced at a few hundred logical qubits
Real-time nuclear dynamics is a canonical target for quantum simulation and a canonical source of vague advantage claims, so a hard resource count is worth something. Working in first quantization with the full leading-order pionless EFT Hamiltonian, and costing out both product formulas and quantum signal processing, this Quantum paper finds time evolution scales polynomially in particle number and only logarithmically in the number of single-particle basis states — exponentially better than the same model in second quantization. Interesting low-energy scattering lands at tens of millions of T gates and a few hundred logical qubits, putting simple nuclear reactions inside the plausible reach of early fault-tolerant machines rather than a distant one.
Quantum — Quantum Simulation of Nuclear Dynamics in First Quantization
The gauge ambiguity in noise learning turns out to be a free parameter worth tuning
Learning the Pauli noise afflicting a quantum processor comes with an irreducible 'gauge' ambiguity — several distinct noise models are consistent with any set of measurements. This PRX Quantum paper shows that ambiguity does not hinder error mitigation, and goes further: the leftover degree of freedom can be chosen deliberately, making mitigation dramatically more efficient. It converts a known limitation of noise characterization into a knob.
PRX Quantum 7, 033045 — Disambiguating Pauli Noise in Quantum Computers
Randomisation comes to open-system simulation
Randomisation techniques such as QDRIFT sharply cut the cost of Hamiltonian simulation, but had not been carried over to Markovian open systems, where any approximation must also keep the evolution physical. This Quantum paper builds non-probabilistic randomised first- and second-order Trotter-Suzuki formulas and a QDRIFT channel that do both, derives error bounds and step counts while bypassing the mixing lemma usually needed in the Hamiltonian case, and implements them via classical sampling to demonstrate gate-complexity advantages over deterministic product formulas.
Quantum — Faster Quantum Simulation Of Markovian Open Quantum Systems Via Randomisation
A circuit-level benchmark that never needs a classical reference
Circuit-level benchmarks capture how noise accumulates across interacting operations, but existing ones tend to demand structured gate sets, costly compilation, or classical simulation of the ideal output. EPCL sidesteps all three: apply the same random circuit to two disjoint registers, measure the overlap of their outputs as a function of depth, and fit the decay to an effective layer polarization. It works with arbitrary gate sets including non-Clifford gates. Simulations recover the predicted polarization under weak local stochastic noise, though coherent errors from fixed entangling layers may need Pauli twirling to produce clean decay, and inter-register correlations add a covariance term. Hardware runs on IBM devices show clear decay at 8 and 16 qubits.
arXiv quant-ph — Characterizing Large Scale Quantum Systems with Error Per Circuit Layer
Anyonic statistics, proposed on off-the-shelf transmon arrays
Analog simulation realizes a target Hamiltonian directly in hardware rather than compiling it into gates. This proposal drives lattices of capacitively coupled flux-tunable transmons in an alternating frequency-modulation pattern, resonantly addressing multiple many-body transitions to tune on-site interaction and density-dependent hopping for up to three bosons per site with no cap on total particle number. Adding phases to the modulation makes transition amplitudes complex-valued, which the authors use to give the first transmon protocol for the anyon-Hubbard model. Numerics at experimentally realistic parameters reproduce interaction-dependent localization and the statistics-dependent asymmetry of anyonic quantum walks. No hardware implementation yet.
arXiv quant-ph — Analog quantum simulation of bosonic and anyonic models with flux-driven transmons
Shallow QAOA as a hint generator for classical MIP solvers
Rigetti and Purdue researchers published joint work extending quantum preconditioning to hard-constrained combinatorial optimization. Rather than solving the problem on a quantum device, they extract two-point variable correlations from shallow QAOA circuits and use them to reshape the objective function handed to a commercial mixed-integer programming solver — quantum output as a heuristic prior for classical machinery. The trade write-up does not carry the head-to-head figures, and the vendor is a co-author, so the size of the benefit is not independently established here.
Quantum Comms
Two atomic species, less control overhead, better purification
Entanglement purification is the step that makes quantum networking and distributed computation tolerable in the presence of noisy links, and its usual cost is elaborate local control. Using two atomic species in one array lets one species act as ancillas, removing the need for local addressing and permitting nondestructive midcircuit measurement — technical overhead that has limited purification circuit design on single-species platforms. The PRX Quantum paper sets out circuit designs exploiting that structure for networking and distributed-computation applications.
Quantum Sensing
Resonance, not chaos, drives a t^6 sensitivity scaling
Non-KAM systems break their invariant phase-space tori abruptly under weak driving, and this work shows that sensitivity at those resonances is a metrological resource in its own right. The authors derive a transport bound for frequency estimation under Floquet unitary encodings: if mean excitation number grows as t^a, quantum Fisher information is bounded by t^(2a+2). The quantum kicked harmonic oscillator walks the full hierarchy — localized dynamics give quadratic growth, diffusion along stochastic webs quartic, and translationally invariant resonances an anomalous hexic t^6 that saturates the bound, established analytically at resonance R=2 and numerically at R=4. The enhancement traces to resonance-induced translational symmetry, distinguishing it from chaos-assisted and criticality-based sensing.
arXiv quant-ph — Non-Kolmogorov-Arnold-Moser Quantum Sensors for Quantum Parameter Estimation
AI & ML
OpenAI ships GPT-6 Astra and calls it 'Critical' on cyber capability
OpenAI unveiled GPT-6 Astra, describing it as state-of-the-art on computer use, browsing and software engineering, days after saying the model had crossed the 'Critical' cybersecurity threshold in its Preparedness Framework — the first model it has placed there. Trade coverage reports a 100% score on ExploitBench, and OpenAI is refusing proof-of-concept exploit requests as a deployment mitigation. The capability numbers come from OpenAI's own evaluations and its accompanying safety overview; there is no independent benchmark yet, so treat the specific figures as vendor-reported. The threshold classification itself is the news: a frontier lab has, on its own criteria, shipped a broadly deployed model at the top of its cyber risk scale.
The Hacker News — GPT-6 Astra Scores 100% on ExploitBench as OpenAI Blocks PoC Exploit Requests · OpenAI — Safety overview: GPT-6 Astra
Fixing the encoding, not the circuit, is what buys a quantum edge
Basis-encoding high-dimensional data into small quantum registers causes cross-class collisions — different-labelled samples mapping to identical bit-strings, which no downstream model can separate. DAFT adapts a pre-trained chemical foundation model so its representations survive quantization, using a differentiable soft collision loss. On blood-brain-barrier penetration prediction with ChemBERTa-77M it reduces collisions by several orders of magnitude and improves quantum classification accuracy by over 12 percentage points versus a frozen backbone. The framing is unusually honest: without DAFT, classical models beat the quantum ones on identical discretized inputs; with it, the comparison flips at 10 qubits (0.883 vs 0.855, p=0.026). The claim is about representation alignment, not quantum hardware.
AI moves from tuning knobs to drafting the apparatus
Nature publishes an examination of AI in physics experimental design, drawing the distinction that matters: the interesting systems search across vast spaces of hardware configurations and propose entirely new experimental layouts, rather than optimizing a handful of parameters in a design a human already fixed. The framing covers optima-finding and generative layout proposal as two faces of the same search problem.
Nature — Designing physics experiments with artificial intelligence
The complete male fruit fly connectome
Google Research announced the mapping of a complete male fruit fly brain, extending the connectomics program that has driven much of the machine-learning work on automated neuron segmentation and reconstruction at scale. The announcement is a vendor blog post without a technical summary attached, so reconstruction accuracy, dataset size and the accompanying paper are not corroborated here.
Google Research — A connectomics milestone: Mapping the complete male fruit fly brain
Gemini 3.8 Flash and a security-tuned sibling arrive
Google DeepMind released Gemini 3.8 Flash alongside Gemini 3.8 Flash Cyber, the latest turn of a Flash-plus-Cyber release pattern that ran through 3.5 and 3.6 earlier this year. The announcement carries no benchmarks, no comparison against the previous generation, and no detail on what the Cyber variant is tuned for beyond its name — a vendor release note rather than a result, and it lands the same week as OpenAI's GPT-6 Astra.
Google DeepMind — Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
A bill to ban something nobody can define
Science reports on Bernie Sanders's proposal to prohibit the development of AI 'superintelligence', backed by stiff penalties. The reporting's own emphasis is the drafting problem: researchers do not agree on what the term means, which makes the prohibited conduct hard to specify, and the bill's route through Congress is uncertain. Worth tracking as the first US attempt to legislate against a capability threshold rather than an application.
An agent that teaches itself traffic simulation, and reports where it doesn't help
SimSkill wraps the SUMO traffic simulator in a self-evolving agent loop: it finds its own capability gaps, generates and solves environment-grounded tasks, verifies solutions through an action-critic loop, and consolidates what worked into episodic, procedural and semantic memory — building a reusable skill library without ever updating the backbone model's weights. On two held-out benchmarks with three different backbones and independent artifact-based verification, it improves verified completion by up to 25 percentage points, with ablations separating procedural and semantic memory contributions. The paper is unusually forthright that the gains are backbone- and budget-dependent, and that memory neither helps every model nor uniformly reduces inference cost. Code and data are public.
arXiv cs.AI — SimSkill: A Lifelong Learning AI Agent for Autonomous Mastery of Traffic Simulation
Reorder the words, keep the meaning, lose the arithmetic
Free word-order languages provide a clean test of whether multilingual reasoning is compositional or keyed to surface form. The authors build IndicReStruct — GSM8K-Reordered and GSM8K-Voice, constructed from GSM8K in Hindi and Malayalam via constrained constituent reordering and active-passive transformation, both meaning-preserving — and find consistent, significant degradation in mathematical reasoning across six state-of-the-art models and multiple prompting strategies. Residual-stream activation patching localizes the failure: entity-quantity alignment breaks down, and intermediate transformer layers contribute most to restoring correct reasoning. A negative result with a mechanism attached, and a released benchmark.
A GNN that gives logic synthesis a sense of physical place
Logic synthesis traditionally guesses at physical reality through wire load models, and the resulting mismatch costs power, performance and area at nanometer nodes. LevelSyn predicts gate coordinates using a level-asynchronous graph neural network that respects the hierarchical logic depth and signal flow of And-Inverter Graphs — structure that spectral placement predictors discard — with level-aligned subgraph partitioning to keep industrial-scale designs inside memory. Those spatial estimates drive a physical-informed synthesis engine built into Berkeley ABC. On the EPFL benchmark suite it reports an average 6.89% power reduction and 27.48% timing improvement over state-of-the-art methods, and post-place-and-route validation shows 99.59% fewer design-rule-check violations.
arXiv cs.LG — LevelSyn: Physical-Aware Logic Synthesis via Level-Asynchronous Graph Neural Networks
AI-for-data-center-energy research can't rank its own methods
AI both optimizes data centers and loads them, and the literature treats those as separate problems — control studies model workload as exogenous arrivals, sustainability studies model infrastructure as a fixed multiplier. This coded review screens roughly 194 papers and codes 63. The findings are pointed: of 28 primary control-oriented studies, 18 are validated in simulation alone and only 5 reach physical hardware or a production facility; none account for water withdrawal or embodied carbon; and reported savings intervals across four technique families overlap so completely that the field cannot presently rank its own methods. The authors propose CLEAR-DC, a framework coupling control policy to workload demand through an explicit elasticity term and reporting net rather than direct benefit — explicitly an architectural proposal, not a trained system. The corpus analysis is the defended contribution.
arXiv cs.LG — Artificial Intelligence for Energy Optimization in Data Centers
Robotics
A 150-kg inflatable ball is the answer to lunar craters nobody can climb out of
Texas A&M's RoboBall III is a 1.8-metre-wide, 150-kg inflatable sphere designed to be dropped into permanently shadowed lunar craters such as the 21-km, 4-km-deep Shackleton — terrain NASA will not send astronauts near and tethered rovers cannot cover. It drives by swinging an internal pendulum to shift its centre of mass, needing just two actuators, both sealed inside the shell away from dust and 300-degree temperature swings, and it uses the same mechanism to brake on descents. In Texas quarry tests reported in IEEE Transactions on Field Robotics, an upgraded build with 2.5x more torque climbed 20-degree slopes, crossed gravel and wet clay, and launched hypothetical sample payloads back out by small rocket. At roughly $250,000 it is not flight-ready: the electronics are not space-grade and the shell material has never been tested at lunar extremes.
IEEE Spectrum — The Best Way to Explore Lunar Craters Is a Giant Robot Ball
Imagining less, and only where it helps, cuts VLA post-training compute 80%
World models let a robot policy evaluate candidate behaviours in imagined futures instead of expensive expert demonstrations or risky real-world exploration — but long rollouts accumulate prediction error and produce unreliable training signal. WISE treats imagination as something to schedule rather than always run: it fires only at interaction-relevant states, keeps multi-view rollouts inside a bounded horizon, scores candidate futures with progress and completion signals, and uses the relative outcomes to refine actions taken from real interaction contexts. With pi_0 and pi_0.5 backbones it improves manipulation across diverse tasks while using about 80% less GPU time than full imagination, with real-world gains under distribution shift.
Skipping the codebook lookup buys 33 points of closed-loop driving success
Vision-language models reason in discrete tokens; driving demands continuous, physics-respecting trajectories. LaPla trains a residual VQ-VAE action tokenizer to capture vehicle kinematics, then — instead of looking up discrete codebook entries, which injects quantization error — repurposes that latent space as a physical prior, projecting the VLM's hidden states straight into it in a single forward pass with concurrent action queries. A frozen decoder turns the latents into actions, bypassing autoregressive generation entirely. On nuScenes it reduces long-horizon L2 error by 15.52% against state-of-the-art VLA methods, and closed-loop evaluation in NVIDIA's AlpaSim shows a 33.34-percentage-point success-rate gain at lower inference latency.
Instance segmentation for lunar rovers, at 5.7 watts and radiation-aware
Autonomous lunar perception has to survive three constraints at once: near-darkness, tiny compute budgets, and radiation faults that silently corrupt inference. This deployment-oriented framework addresses quantization calibration and fault exposure together. AVIS picks calibration samples deterministically from activation-variance statistics with no labels; a YOLO-based segmentation model is modified to cut CPU fallback paths and compile statically onto a Deep Learning Processor Unit with bounded latency in low light. On a lunar micro-rover platform, AVIS with bias correction recovers 69.8% of the accuracy lost to quantization at 309 ms inference and 5.7 W, and a software-level criticality analysis directs mitigations that reduce global criticality by 31.7%.