Quantum Computing
Laptop dequantizes the flagship 77-qubit quantum chemistry run
The unitary cluster Jastrow (UCJ) ansatz has been the workhorse of the most-cited recent quantum-advantage-adjacent chemistry experiments — including a 77-qubit, 10,570-gate circuit on IBM hardware whose classical post-processing consumed up to 6400 nodes of Fugaku. A new preprint gives a polynomial-time classical algorithm for the energy of any single-layer UCJ circuit, with no locality constraints, and reproduces that experiment in under a minute on a laptop. Enabled by the speed of the simulation, the authors then optimize the circuit further and land on a lower ground-state energy than the hardware run reported. It is a clean dequantization: the result does not refute the hardware, but it removes the classical-intractability claim that gave the experiment its significance, and it raises the bar for what a single-layer UCJ demonstration has to show.
arXiv quant-ph — Efficient classical simulation of large-scale unitary cluster Jastrow circuits
Quantum Computing
Nuclear structure priced for a fault-tolerant quantum computer — and it costs about as much as FeMoco
Quantum resource estimation has been dominated by chemistry and condensed matter; atomic nuclei have been largely ignored despite the structural similarity to the electronic-structure problem. This work compiles fault-tolerant algorithms for effective shell-model and no-core-shell-model Hamiltonians with three-body interactions from chiral effective field theory, and reports what the authors believe are the first Toffoli and qubit estimates for nuclear simulation. The headline finding is that Mg-32 and At-219 shell-model Hamiltonians land in the same cost band as FeMoco — the field's reference hard problem — making them plausible early fault-tolerant targets. No-core-shell-model Hamiltonians for light nuclei up to around Ca-40 come out significantly more expensive, and the authors argue those need bespoke strategies rather than off-the-shelf compilation.
arXiv quant-ph — Fault-tolerant quantum algorithms for simulating atomic nuclei
Compact CNNs automate spin-qubit tuneup across 32 real silicon devices
Tuning up semiconductor spin qubits still means a human reading charge stability maps, and the isolated-mode regime that matters most for scaling has received little automation attention. Two convolutional networks under a million parameters each — CSMClassifier for instability and sensor artifacts, ChargeLineNet for transition-line localization and electron occupancy — were trained on maps from 16 SiMOS double-quantum-dot devices and evaluated on 16 held-out ones against hand-labeled ground truth, hitting 94% macro-averaged accuracy on 2,407 images and 95.3% exact line counts on 1,131 images. Cross-device generalization is the substantive part: the models transfer to devices they never saw, and synthetic pre-training keeps accuracy above 90% when experimental labels are scarce. At 6.5 MB total and under 60 ms per image on ordinary lab hardware, this is deployable rather than aspirational.
arXiv cs.LG — Machine Learning for Charge State Characterization of Isolated Double Quantum Dots
What the control electronics must deliver to run a fault-tolerant Shor circuit
Error-corrected quantum computing's next milestone is a non-Clifford logical circuit, and this Quantum paper works backwards from that target to the classical hardware. Taking Shor's factorization of 21 as a representative circuit, the authors convert logical to surface-code to physical level using typical superconducting parameters and extract system-level requirements on the controller-decoder loop. The binding constraint is closed-loop latency: tens of microseconds, achievable only by distributing decoding across several decoders with fast decoder-to-decoder and decoder-to-controller communication. Simulating the complete factorization at the physical level, they find near-term hardware at roughly 0.1% physical error rates and 1000 qubits is enough to execute it — a concrete, checkable spec sheet rather than a roadmap.
Neutral atoms are locked out of the standard quantum SDKs by superconducting-era device models
Universal quantum ecosystems abstract hardware behind a device model, and that model was shaped by superconducting processors: static qubit positions, fixed coupling maps. Neutral-atom machines have neither — they rearrange atoms dynamically and support zoned operations — so specialized neutral-atom compilers cannot retrieve the hardware information they need through Qiskit, Cirq or PennyLane. The authors show this produces suboptimal compilation and can exclude devices outright, then propose a device model that represents the platform faithfully; evaluated within the Quantum Device Management Interface, it improves routing-overhead fidelity by up to 100,000x on a 16-qubit, 600-gate circuit.
ARGON compiles neutral-atom circuits in under 10 seconds by splitting space from time
Neutral-atom compilers have to schedule atom transport alongside parallel entangling gates, and joint spatiotemporal search blows up exponentially as circuits grow. ARGON's move is to push static geometric conflict resolution offline, precomputing a library of hardware-certified high-parallelism layouts, then use a GNN predictor to pick among them against deep temporal horizons before a heuristic router turns the choice into collision-free transport. Compilation finishes in under 10 seconds, with up to 10^4x and 600x average speedup over prior methods, and fewer Rydberg stages translate into up to 100x better execution fidelity on dense circuits.
arXiv quant-ph — ARGON: A GNN-Empowered Compilation Framework for Scalable Neutral Atom Computing
Inverse-designed transmons hit target frequencies within 0.73% at 2,100x the speed
Designing a superconducting qubit to specified Hamiltonian parameters normally means a slow forward loop: pick a geometry, simulate, extract capacitances, refine. This work trains an inverse model paired with a frozen forward surrogate, evaluating loss in Hamiltonian space rather than layout space, and validates against a conventional EM solver: 97% of generated designs are usable, with mean errors of 0.73% on qubit frequency and 1.58% on anharmonicity — at or below the fabrication and simulation-to-measurement uncertainty expected for academic-process devices. A query takes ~56 ms on CPU against ~2 minutes for EM extraction, and batching drops that to 0.24 ms per sample on CPU. Notably it works from datasets on the order of 1,000 samples.
arXiv quant-ph — Component-Level Inverse Design of Transmon Qubits Using Neural Networks
Which photonic variational circuits are actually trainable, and which are not
Barren plateaus are usually framed as a gradient-magnitude problem; this paper reframes photonic trainability around the ratio of sample variance to circuit variance, which sets how many circuit samples are needed to resolve local loss differences and gradients to proportional accuracy. Applied to photon-number observables in passive linear-optical circuits, it cleanly separates the regimes: fixed-order photon-number polynomials need only polynomially many samples as the system grows, while high-order polynomials and observables built on output probabilities generally need exponentially many. Within the trainable regime the authors identify observable classes — including a practical neural-network-observable construction — where quantum estimation gives a polynomial speed-up over several classical methods.
arXiv quant-ph — The trainability of photonic quantum circuits
Work is a quantum observable — once you stop treating the controller as external
Canonical quantum-thermodynamics treatments assume classical control via a time-dependent Hamiltonian, quietly omitting a control cost that can exceed the system's own energy scale by orders of magnitude. This paper autonomizes the transfer, embedding the driven dynamics in an energy-conserving evolution on a larger system so that work is unambiguously the energy moved between the two — which picks out a unique work operator on the driven system. That operator evades the well-known no-go theorem and supports a quantum fluctuation theorem that reduces to Jarzynski classically; extended to open systems it gives an operatorial first law with distinct operators for work, heat and internal energy. The claimed upshot is that a long-standing debate is settled in the affirmative.
arXiv quant-ph — Autonomization of Quantum Systems and the Emergence of the Work Operator
A provable-guarantee algorithm for the closest Slater determinant, with a 2/3 fidelity cliff
Extracting a compact, interpretable description of a fermionic many-body state — the Slater determinant closest to it — sits at the intersection of Hartree-Fock practice and agnostic tomography, and has lacked guarantees. The authors give classical and quantum algorithms reaching within epsilon of maximal fidelity in m^poly(n,1/epsilon) time, prove matching hardness lower bounds along some parameter axes under standard complexity conjectures, and show poly(m,n,1/epsilon) copies suffice with quantum access. The optimization landscape has a sharp structure: above fidelity 2/3 any stationary point is the unique global maximum, below it spurious stationary points can appear. They apply the algorithm to the Fermi-Hubbard model, pulling the closest Slater determinant out of neural quantum state solutions.
Deeper parameterized quantum circuits show double descent, not degraded generalization
Formal generalization bounds for quantum models exist but are known to describe practice poorly, leaving the field with a working assumption that bigger PQCs generalize worse. This paper argues otherwise: gradient-based PQCs display double descent, with performance on unseen data improving again as parameter count grows. The analysis leans on add-one-in perturbation techniques and spectral properties of random matrices, and numerical experiments on re-uploading PQCs reproduce the predicted behavior across several datasets and training-set sizes. The framing is deliberately modest — other obstacles to practical quantum machine learning remain — but one specific worry about depth looks misplaced.
arXiv quant-ph — Cautious optimism for deep parameterized quantum circuits
A lightweight Hoare logic for quantum programs, built out of the Heisenberg picture
Gottesman's 1998 stabilizer semantics turns out to support a Hoare-like program logic that efficiently characterizes a useful slice of quantum programs. The applications are practical verification questions: whether auxiliary qubits can be safely discarded, whether a system is separable across a given bipartition, whether a gate is transversal with respect to a stabilizer code, and what the post-measurement state is for computational-basis measurements. The logic extends beyond Clifford circuits through Hoare triples for the T gate, multiply-controlled unitaries such as Toffoli, and magic-state gate-injection circuits — and one side effect is a lower bound on the number of T gates needed for a multiply-controlled Z.
Quantum — Hoare meets Heisenberg: A Lightweight Logic for Quantum Programs
IBM puts $50M of QPU time behind the DOE Genesis Mission
IBM is committing $50 million in quantum computing access over five years to the Department of Energy's Genesis Mission, opening direct access to its utility-scale systems for researchers at DOE national laboratories and collaborating universities. It follows DOE's Quantum Genesis initiative, announced in late June, which targets a fault-tolerant, scientifically relevant quantum capability by 2028. As a compute-donation commitment rather than a technical result, the thing to watch is what comes out of the allocation rather than the headline figure.
Quantum Comms
Topological photonic 'multi-lane highways' move light one way at full spatial efficiency
Topological waveguides normally pay for their robustness in space: the insulating regions that protect the edge states occupy area that carries no light. This Nature paper drops the insulator entirely, arranging four inequivalent photonic valley half-semimetals in a parallel cyclic configuration stacked very close together, and reports robust multi-lane unidirectional transport with 100% spatial efficiency. The relevance for photonic interconnect and on-chip quantum links is density — parallel protected channels without dead area between them.
Nature — Insulator-free topological photonic multi-lane highways
QKD
CV-QKD security proof drops the SDP bottleneck and hits the Gaussian limit at 256-QAM
Practical continuous-variable QKD hardware emits discrete-modulated signals, but the security proofs for those signals lean on semidefinite programming that becomes computationally prohibitive at high-order constellations. This Letter derives a framework from the uncertainty principle instead, using a multi-mode entanglement-source model of non-Gaussian state preparation to map constellation geometry directly onto the secret key rate — giving both numerical and analytical security analysis where SDP previously stalled. The authors validate on discrete-component and integrated photonic setups and show a 256-point QAM format asymptotically approaching the Gaussian capacity limit, closing much of the long-standing gap between idealized protocols and deployable high-speed hardware.
AI & ML
One phishing link could plant a rogue agent in a ChatGPT Workspace tenant
Zenity Labs disclosed AgentForger, a critical flaw in OpenAI's ChatGPT Workspace Agents in which a single phishing link was enough to construct, authorize, and deploy an autonomous agent inside a victim's organization — no further interaction required. The pattern is now familiar from the Kiro configuration-rewrite flaw earlier this month: the dangerous surface in agentic products is not the model's output but the authorization path between a user action and an agent gaining standing permissions. OpenAI addressed the issue as of 8 June 2026.
The Hacker News — ChatGPT AgentForger Flaw Could Deploy Rogue Workspace Agents via a Phishing Link
Two Lean 4 benchmarks measure AI proof agents on quantum theorems for the first time
Formal verification is arriving in quantum computing, but nobody had measured whether AI agents can actually produce machine-checkable proofs in the domain. These two benchmarks — 36 quantum-algorithm and 40 quantum-information theorem-completion tasks, all compiling in a fixed environment with difficulty weights fixed before any model ran — put the best difficulty-weighted scores at 60.4 and 59.6 out of 100. Giving agents access to a verified domain library improved score and completion rate in all eight model-benchmark comparisons, by up to 15.9 points, and the recurring failures cluster in quantum simulation, quantum learning, information measures and entanglement theory.
New tractable classes for optimally training ReLU and linear networks
Recent complexity work on neural network training has mostly pushed lower bounds downward; tractable architectures have been much harder to find. This paper moves the other way, proving polynomial-time optimal trainability for every ReLU architecture in which hidden neurons have out-degree 1 — a strict improvement on the Arora, Basu, Mianjy and Mukherjee algorithm — and, for linear activations, identifying the first non-trivial polynomial-time solvable class through a new data throughput condition on the architecture.
arXiv cs.LG — New Complexity-Theoretic Frontiers of Tractability for Neural Network Training
Learned pruning cuts SAT-encoding cost for constant multiplication by two orders of magnitude
Single constant multiplication — decomposing a fixed constant into additions, subtractions and bit-shifts — is an NP-hard hardware-design primitive where dynamic programming gives near-optimal SAT encodings at prohibitive cost for large constants. Here a graph neural network predicts promising operator types from constant decompositions, and its confidence scores prune bad choices in the symbolic search. On unseen 17-32 bit constants that yields one to two orders of magnitude less encoding time, over 97% memory reduction and roughly 10x less branching, with encoding quality in terms of additions still near-optimal. Code and data are public.
Opus 5 read as an efficiency release, not a capability jump
Ars Technica's take on Anthropic's Opus 5 is that the meaningful change is token efficiency rather than a step up in raw capability — and that the broader pattern is frontier releases competing on cost per unit of useful work while cheaper models remain adequate for most deployments. Worth noting as framing rather than measurement: this is a mainstream analysis piece, not a benchmark, and the efficiency claim awaits independent numbers.
Ars Technica — Anthropic's Opus 5 is about token efficiency, not a capability leap
Robotics
A surgical robot infers cutting temperature it cannot measure, then adapts its bone-cutting policy
Autonomous craniotomy is partially observable in a hard way: tissue properties vary unpredictably and cutting temperature cannot be measured directly under occlusion. RL-MACRO closes the loop by reconstructing that hidden temperature from force and sound with a CNN-LSTM observer, reaching R^2 of 0.939 and 1.717 degC mean absolute error, and using it alongside multi-sensor features as the belief state for an offline implicit Q-learning policy. A dual-head actor coordinates feed rate, spindle speed and cutting depth within safety bounds, translated to motion by online trajectory replanning and velocity servoing. Validation is on bovine ribs and six ex vivo goat skulls — real tissue, not simulation, though still bench rather than clinical.
Game-theoretic AV planner runs in under 50 ms against a real human driver
Most interaction-aware autonomous driving work decouples the AV's optimization from the human's likely response. This framework instead poses the interaction as a Generalized Nash Equilibrium Problem, so shared safety and geometric constraints tie the feasibility of the AV's strategy directly to the opponent's actions. Solving that non-convex problem in real time falls to a dedicated particle-swarm solver that converges in under 50 ms. The validation is the substantive part: a real autonomous Renault Zoe on a test track interacting with a human driver, producing what the authors characterize as comfortable, human-like trajectories in critical scenarios.
A close look at the most advanced robotic servicing satellite we know about
On-orbit servicing has been perpetually five years away for two decades; Ars Technica profiles the most capable robotic servicing satellite publicly known, and the reporting focuses on why autonomous rendezvous, capture and manipulation in orbit remain genuinely hard rather than on capability claims. Filed here as space robotics context — no independently verifiable performance figures accompany the piece.
Ars Technica — This is the world's most advanced robotic servicing satellite—that we know about
Video Friday: a multimodal-skin humanoid, GEN-1 across a thousand hands, and NHS drone logistics
This week's IEEE Spectrum roundup carries four items worth flagging. Generative Bionics showed GENE.01, a humanoid built in six months whose full-body multimodal skin senses touch, proximity, force and temperature. Generalist extended its GEN-1 embodied foundation model across a broad range of end effectors — five-finger hands through specialized tools — arguing that pretraining across thousands of sensorimotor interfaces yields transferable physical commonsense. Aurora detailed a next-generation Driver built for a million miles at half the hardware cost. And Wing, with Apian, reports NHS pathology sample deliveries across south west London running up to 85% faster than ground transport since February 2026. As always with the roundup, these are vendor demos and claims, individually unverified.
IEEE Spectrum — Video Friday: An Italian Humanoid Comes to Life