Quantum Computing
IBM and UChicago put a verification claim behind quantum advantage
IBM and University of Chicago researchers announced a demonstration, "Sampling hard circuits with verifiably high fidelity," that pairs a beyond-classical-simulation sampling task with a check on whether the returned samples are actually accurate — the part usually missing from advantage claims. IBM published it alongside two further joint demonstrations with Qedma, Algorithmiq, RIKEN and BlueQubit, and released the circuits on a public Quantum Advantage Tracker. Both reports so far are press-release-driven; the verification protocol and the classical-simulation baselines it is measured against are what independent groups will need to check.
The Quantum Insider — IBM and University of Chicago Demonstrate Verified Logical Quantum Computation Beyond Classical Simulation · Quantum Computing Report — IBM and Ecosystem Partners Demonstrate “Trusted Quantum Advantage” Beyond Classical Supercomputers
Quantum Computing
Classical sampling keeps pace with decoded quantum interferometry
DQI is one of the more prominent proposals for quantum advantage in approximate optimization. Because its output probabilities are efficiently computable, the authors could point classical MCMC at the same induced distribution — and block-Gibbs sampling reliably hit DQI's expected approximation ratios on max-XORSAT past 1000 effective qubits and on OPI past 150. In the OPI regime where a super-polynomial quantum advantage is claimed, the classical runtime grew as roughly 1.1^n: still exponential, but with a small enough base to matter in practice. The authors are careful that this does not refute the advantage claim, only narrows where it would bite.
A tensor-network benchmark puts numbers on how far NISQ algorithms actually reach
The authors computed LMG ground-state energies with DMRG on Perlmutter for systems up to 1400 particles — one of the largest such datasets published — and used it as a yardstick for two NISQ methods on an IBM Eagle device. VQE held 1 percent error only out to 6 particles; SQD reached 17. The gap between 17 and 1400 is the useful number here, and it supports the authors' reading that subspace methods, not variational ones, currently balance accuracy against circuit depth and noise best.
Trimming the Hamiltonian cuts a 20-qubit chemistry circuit by 98 percent
Quantum-selected configuration interaction stalls on hardware because the full electronic Hamiltonian blows up circuit complexity. The authors identify the dominant fermionic excitation operators and build a reduced Hamiltonian from a fidelity-loss analysis, giving near-quadratic improvement in term scaling. Run on IBM's Marrakesh processor for AlF and TlF in active spaces up to 20 qubits, the reduced Hamiltonian cut both circuit depth and two-qubit gate count by more than 98 percent, with relativistic ground-state energies and permanent electric dipole moments landing within 99.99 percent of CASCI references.
A noise model that predicts, rather than post-hoc explains, neutral-atom hardware
Analog Rydberg processors are getting easier to access than to trust, since there is no straightforward way to know whether an output is right once classical simulation runs out. The authors built an emulator that propagates the dominant noise mechanisms through the whole computation cycle and tested it on two protocols — quantum annealing and post-quench dynamics — across three Pasqal devices, in the regime where classical simulation still gives ground truth. Measured observables fell within the predicted uncertainty envelopes on all three, and the framework identifies which physical mechanism dominates in each operating regime.
A decoder ASIC trades layout for 35 percent less latency at distance 21
Syndrome decoding has to finish before errors pile up in the next measurement cycle, which makes latency the binding constraint. This design exploits temporal parallelism in sandwich decoding with Union-Find, and logic synthesis puts it at 35 percent lower average latency than a batch Union-Find decoder at code distance 21, with the threshold essentially unchanged at about 1.5 percent under phenomenological noise.
Color-code computation gets an automated compiler
The color code has had individual logical primitives for a while but no route from an algorithm to a full fault-tolerant layout. This work defines spacetime block diagrams for logical patches and operations, derives the assembly rules from the code's anyon-condensation and domain-wall structure, and identifies the diagrams' logical semantics with ZX calculus so representations can be rewritten safely. The compilation step uses edge-decorated ZX diagrams and fusion-region-aware routing, and the authors report automated compilation across a broad set of algorithms.
Circuit cutting's classical bottleneck gets a 259x speedup
Cutting a large circuit into hardware-sized pieces shifts the cost to classical reconstruction, which grows with cut count. This framework avoids building the dense 2^n probability vector, reconstructing only nonzero-probability states with an index representation that survives past 64 bits and storage spanning device memory, host memory and disk. On the Songshan supercomputer it reports up to 259x over an optimized serial baseline for linear-cluster states, up to 4x over homogeneous CPU parallelism on random circuits, and completes hundred-qubit-scale reconstructions.
Quantum sampling drives a solvated molecular-dynamics trajectory
Getting a quantum electronic-structure method into molecular dynamics needs gradients, not just energies. The authors take measurements from a LUCJ ansatz, recover determinant subspaces with SQD, and produce analytical nuclear gradients from them. Against exact FCI in the STO-3G basis the gas-phase energies and gradients agree within 1 kcal/mol with stable trajectories, and the agreement holds in explicit-solvent QM/MM, matching FCI energy fluctuations, RMS gradient profiles and solute-solvent radial distribution functions.
Entanglement-assisted QLDPC codes down to a single shared Bell pair
The entanglement-assisted setting normally trades pre-shared Bell pairs for code performance, so the count matters. One of the proposed quasi-cyclic families, built from two distinct classical QC-LDPC codes, needs just one shared pair, and its unassisted Tanner subgraph is free of 4-cycles. Decoded over a quaternary alphabet with block-layered normalized min-sum, the codes report close to an order of magnitude better error correction than layered binary sum-product under both depolarizing and Markovian noise, and handle correlated Pauli and burst errors within the same framework.
Quantum Comms
Asynchronous multi-photon interference for quantum networks
npj Quantum Information published a paper on asynchronous multi-photon interference for quantum networks. Asynchronous schemes relax the requirement that photons arrive in the same time bin, which is the practical obstacle to entanglement swapping over links with independent, drifting sources. The feed carried only the title and DOI, so the reported figures are not yet available here.
QED-C and CQN publish a quantum networking applications roadmap
QED-C and the NSF-funded Center for Quantum Networks released a roadmap for quantum networking applications, assembled from technical input by more than 50 contributors across national labs, universities and companies including IonQ, L3Harris and Aliro. Roadmaps are consensus documents rather than results; the value is in seeing which applications a broad industry group is willing to put dates against.
Quantum Sensing
Certifying non-Gaussian light without paying for the losses
Certifying non-classicality usually degrades as optical loss climbs, which limits what can be verified outside a well-controlled bench. Kalash, Passos, Rácz, Ruppert, Filip and Chekhova use phase-sensitive optical parametric amplification ahead of simple intensity detection to get a loss-invariant certificate of quantum non-Gaussianity, demonstrated on a heralded quasi-single-photon state.
AI & ML
Unit 42 documents an LLM agent running an intrusion off a single Telegram instruction
Unit 42 says a Chinese-speaking operator wired DeepSeek into the open-source Hermes Agent framework and issued one Telegram instruction, after which the agent located internet-facing systems and picked public exploits on its own — the researchers recovered no further operator input for the rest of the session. Autonomy claims in incident reports deserve scrutiny, since "no recovered input" is not the same as "no input," but this is a named actor with recovered infrastructure rather than a red-team simulation.
Extra compute changes how local computer-use agents fail, not how often they succeed
Inference-time scaling helps frontier computer-use agents; this study asks whether it helps the small models people would actually run locally. Across four scaling dimensions on OSWorld with Qwen3-VL-8B/30B-A3B, UI-TARS-1.5-7B and OpenCUA-7B, the gains saturate as token cost rises. Contextual scaling stabilizes trajectories but shifts failures toward premature false successes; temporal scaling cuts max-step stalls while extending erroneous trajectories rather than correcting them; structural decomposition adds planning and formatting overhead that parallel scaling only partly offsets, at substantial cost.
Neural networks beat deterministic search at tracing Seiberg dualities
Deciding whether two supersymmetric quiver gauge theories are dual means finding a sequence of quiver mutations connecting them — a search problem in the same family as learning to unknot. On quivers with around 10 nodes, transformers and MLPs outperformed deterministic algorithms, and supplementing the network with pathfinder algorithms improved both efficiency and accuracy further. The authors propose the task as a benchmark for frontier models applied to theoretical physics, which is the more durable contribution.
OpenAI claims ten results on open math and TCS problems
OpenAI posted what it describes as new results on long-standing open problems across geometry, cryptography and complexity. Model-assisted mathematics claims have a mixed record on how much the model contributed versus the human collaborators, and none of these have been through peer review or independent verification yet — the write-ups and any accompanying proofs are what to check.
An ICML paper argues LLM security has a floor that patching cannot reach
MIT Technology Review covers an ICML paper arguing that full security against attacks on large language models is unattainable, not because current defenses are immature but because of how the models process instructions and data through the same channel. If the argument holds, it pushes mitigation toward containment and privilege separation rather than model hardening. The coverage does not restate the formal claim, so the paper itself is what settles how strong "impossible" is here.
Chatbots build exploitable trust faster than human scammers in a study
Ars Technica reports research measuring how quickly targets extend trust to an AI chatbot versus human operators, with the chatbot coming out ahead on the study's measure of exploitable trust. The framing matters more than the headline: this is about the relationship-building phase that precedes a scam rather than any technical capability, which is also the part that scales cheaply.
Modelling study: LLMs plus publication incentives yield more papers, less refined
Nature covers a modelling study predicting that scientists using LLMs will "do more, less well" — publication incentives push toward volume, and lowering the cost of producing a paper moves output along that gradient rather than toward quality. It is a model rather than a measurement of what has already happened, so its value is in the mechanism it makes explicit.
Robotics
DeepMind releases Gemini Robotics ER 2
DeepMind announced Gemini Robotics ER 2, a robotics embodied-reasoning model it says improves video understanding, tool orchestration and multi-robot collaboration. The announcement is a vendor blog post with no benchmark numbers or independent evaluation attached, so it sits here until third-party results appear.