Quantum Computing
Code switching brings fault-tolerant Z-rotations to rotated surface codes, with a 45-qubit magic-state demonstration
Code switching realizes a universal fault-tolerant gate set by pairing two codes with complementary transversal gates, but until now the technique was largely confined to color-code families supporting a logical T gate. This work uses the doubling technique to build single-logical-qubit color codes of arbitrarily large distance that transversally implement arbitrary small logical Z-rotations, reporting better parameters than state-of-the-art triorthogonal codes, lower qubit overhead than certain known color codes, and single-shot Z-syndrome decoding via meta-checks. The framework extends past color codes to r-orthogonal codes (r >= 2) that inherit the local geometry of rotated surface codes, and the authors extend transversal-CNOT code switching to Z-rotations at any level of the Clifford hierarchy — including a simulated fault-tolerant magic-state preparation inside a distance-three rotated surface code with a total footprint of 45 physical qubits.
Quantum Computing
Unconditional quantum advantage extended from search to sampling with constant-depth circuits
Bravyi, Gosset and Koenig showed a search problem constant-depth quantum circuits solve that constant-depth bounded-fan-in classical circuits cannot, and asked whether the same holds for an input-independent sampling task. This paper answers yes when the classical circuit's random input bits are bounded: it exhibits a distribution D_n over n bits sampled to small total-variation distance by a uniform constant-depth quantum circuit family, and proves unconditionally that a bounded-fan-in classical circuit taking kn + n^delta i.i.d. Bernoulli variables of entropy 1/k requires depth Omega(log log n) to come close, even allowing additive error. A related separation is shown for constant-depth quantum circuits with advice against classical circuits with bounded fan-in and fan-out but unbounded i.i.d. randomness.
Quantum linear solvers on industrial electromagnetics: condition number saturates with grid size
The paper benchmarks the Harrow-Hassidim-Lloyd algorithm and quantum singular value transformation against linear systems generated by finite-difference time-domain discretization of Maxwell's equations, across radar propagation, lens simulation and beamforming cases. Both approaches produce state infidelities under 2e-3 with success probabilities above 1e-3, compatible with practical state sampling, and QSVT is consistently the more accurate of the two. The more consequential finding is structural: the condition number, which governs quantum linear-solver cost, saturates as the number of spatial lattice points grows, meaning the grid can be refined to realistic industrial dimensions without ill-conditioning inflating circuit depth.
arXiv quant-ph — Quantum Computing for Industrial Electromagnetics
Memory separations for AI state-tracking transfer to quantum solvers — with the caveats stated up front
The paper proves inference-time quantum coordination advantages for state-tracking tasks where a solver compresses semantic history into a boundary state and later answers a query, charging communication, persistent memory and local work — with classical recurrence, caches, tools and recomputation all allowed and counted. Its central boundary-preserving semantic-compilation theorem transfers known lower and upper bounds into this setting independently of the finite-precision recurrent architecture. Matched-entity synopsis QA inherits the hidden-matching separation (O(log N) qubits versus Omega(sqrt(N)) classical boundary bits); continual requirements auditing inherits a Max-kSAT streaming separation with O(log^5 n log(1/delta)) qubits achieving a 0.7172-approximation where classical one-pass solvers need Omega(sqrt(n)) coordination width. The authors state plainly that these are memory and coordination separations, not runtime or empirical advantages for present-day language models, and that the stabilizer result assumes exact simulation and noiseless quantum memory.
arXiv quant-ph — Quantum Coordination Advantages in AI State-Tracking Tasks
PRX Quantum publishes a full account of Pauli propagation, theory through software
Pauli propagation has become one of the main classical-simulation techniques used to contest quantum-advantage claims on near-term hardware, and it has been scattered across method papers until now. This PRX Quantum article assembles the theoretical framework, the software implementation and the application range in one place. It is a reference rather than a new result, which is what makes it useful: benchmark disputes over sparse Pauli-path and truncation-dependent methods now have a single citable account of what the method does and where its truncation assumptions bite.
PRX Quantum 7, 032001 — Pauli Propagation: A Computational Framework for Simulating Quantum Systems
npj QI: breaking the entanglement-fidelity tradeoff in PT-symmetric systems
Parity-time-symmetric (non-Hermitian) systems can amplify entanglement, but conventionally at the cost of state fidelity — the two have been treated as trading against each other. The paper reports conditions under which that tradeoff can be broken. The listing carries only the title-level abstract, so the mechanism, the platform and the size of the effect all need the full text; the result is theory-side unless the paper includes hardware.
Quantum Comms
Silicon photonic chip extracts a quantum subgraph below the classical random-network threshold
Classical random-network theory says complex subgraphs only appear once connection probability is fairly high; quantum random-network theory predicts entanglement and local operations let such structures emerge at a single, lower threshold. The experiment builds a four-node quantum random network on an integrated silicon photonic chip, using probabilistic photon-pair sources and coherent control of path modes to prepare a structured subgraph through local transformations and postselection. The authors verify genuine high-dimensional multipartite entanglement across the nodes, giving experimental evidence for connectivity structures the classical threshold does not reach. Postselection means this is a proof of the theory's mechanism rather than a deployable network primitive.
arXiv quant-ph — Photonic realization of a subgraph extraction in a quantum random network
QKD
Side-channel-secure QKD demonstrated over 200 km
Side-channel-secure QKD aims to close the gap between the idealized protocols of security proofs and real hardware, where imperfect sources and detectors leak information through channels the proof never modeled. The paper reports a proof-of-principle experiment extending the protocol to 200 km, a distance range that matters for metropolitan-to-intercity links. Published in npj Quantum Information on 11 August 2026; the abstract released with the listing carries no key-rate figure, so the rate-distance trade-off is worth checking against the full text.
Quantum Sensing
A nonlinear energy pump gives a programmable Heisenberg-limited microwave sensor
The proposal couples a Kerr-nonlinear resonator to several high-Q microwave terminal resonators to act as a nonlinear quantum energy pump, and constructs analytic sensing protocols — state loading, probe preparation, readout — that reach the maximal quantum Fisher information for parameters encoded in any number-conserving Hamiltonian on the terminal modes. For diagonal multiparameter signals the full phase-sensing QFI matrix follows from correlations of locally measured physical terminal works, which gives a signal-free way to calibrate the metrological resource. Validation is numerical, using realistic circuit-QED parameters with experimentally relevant imperfections included; no hardware run is reported.
arXiv quant-ph — Programmable Heisenberg-limit sensor from a nonlinear quantum energy pump
AI & ML
Replayable encrypted reasoning blocks let one session's hidden chain-of-thought leak into another
Reasoning APIs from the three largest model providers pass hidden reasoning between calls as encrypted objects the client is expected to hand back unmodified. Researchers found that these objects were not bound tightly enough to the session that produced them: a block minted in one session could be replayed into a different one, and during testing that surfaced internal reasoning and secrets — API keys and passwords — out of session logs. The disclosure is a cross-vendor design flaw rather than a bug in one implementation, which makes the fix a protocol question for anyone building on encrypted reasoning state.
Data-science agent benchmark in real OS environments: best model 56.7%, open-source agents under 1%
Existing agent benchmarks for data science mostly score code in isolation rather than the multi-tool, long-horizon work the job actually involves. DSAgentBench puts 275 tasks spanning wrangling, exploration, modeling, visualization and validation into real computer environments where the agent must drive notebooks, IDEs, terminals, browsers and databases, and grades each with a deterministic evaluator that checks analytical correctness, visual outputs and model performance rather than execution alone. The strongest agent tested, Claude-4.6-Sonnet, completes 56.70% of tasks; every open-source agent lands below 1%, failing at tool orchestration, OS grounding and multi-step reasoning. The benchmark is released at github.com/vis-nlp/DSAgentBench.
Nature proposes a four-dimension profile for governing AI agents
Governance discussions tend to treat 'AI agent' as a single category, which makes rules either too loose for the most capable systems or too heavy for narrow ones. The framework characterizes an agent along autonomy, efficacy, goal complexity and generality, and composes those into an agentic profile intended to let regulation attach to where a system actually sits rather than to the label. It is a proposal rather than a measurement result — the dimensions are not operationalized with a scoring procedure in the abstract — but it appears in Nature rather than as vendor policy commentary, which is why it is worth logging.
DeepMind ships a sign-language-to-text model into user-facing products
DeepMind announced SL2T, a sign-language-to-text model it says powers new sign language features for Deaf and hard of hearing users. Sign-language translation is a genuinely hard multimodal problem where deployment into a shipping product is the notable part. The announcement is a vendor blog post with no accuracy numbers, no stated sign-language coverage and no evaluation methodology, so the capability claim is unverified until a paper or independent testing appears.
Google DeepMind — Putting sign language AI into users' hands
Robotics
Racing agent learns its own limits from 256 km/h real-vehicle data, hits 88.3% interaction success in simulation
Embodied systems are usually evaluated well inside their safety margins, leaving their behavior at the edge poorly characterized. This work uses autonomous racing — high-frequency perception, adversarial interaction, near-saturated vehicle dynamics — as the stress case, learning predictive world models from both near-limit successes and failures to capture interaction evolution, ego dynamics and feasible-motion boundaries. Training data came from real-vehicle runs where onboard localization and perception held at up to 256.3 km/h and 26.8 m/s^2 lateral acceleration; the trained agent reaches 88.3% interaction success across challenging simulated racing scenarios, with closed-loop refinement improving recovery from failure modes and transfer to unseen circuits. The headline success figure is simulated, not on-track.
Skills in weights, memory in code: a coding agent steers a VLA through non-Markovian manipulation
Vision-language-action policies generate actions from the current observation or a short fixed history, which breaks down on manipulation tasks that require reasoning over long interaction histories. HyMeS splits the problem: low-level motor skills are learned by imitation, while a coding agent acquires high-level memory-management strategies as an executable heuristic system it revises from rollout feedback, with multimodal stage-completion verification (proprioception plus multi-frame VLM judgments) closing the loop. Because demonstrations are needed only for reusable motor skills rather than every history-dependent configuration, the approach is more data-efficient than end-to-end memory-augmented VLAs. On RoboMemArena it lifts mean cumulative success from 52.5% to 66.2% and mean task success from 41.3% to 60.1% against pi0.5.
Field recomposition adds unseen hardware and compute payloads to a robot in minutes
Most robots are monolithic enough that adding a capability after deployment takes hours of expert integration. This framework defines abstractions for runtime recomposition so non-expert users can plug in previously unseen modular software, hardware and compute payloads, with new resources immediately available to the host robot and also shared with distributed peers so compute-constrained systems can borrow remote capability. Reconfiguration drops to minutes with no developer in the loop. The demonstrations are two disaster-response scenarios: radioactive-source localization at an operational nuclear reactor facility, and a thermal-guided search for people in dark, hard-to-reach spaces.
arXiv cs.RO — Deployment Is Not Destiny: Robot Recomposition in the Field
Putting energy and momentum in the latent state cuts collisions from 12.1% to 5.8%
Latent world models predict dynamics but leave the physics they absorb implicit, so nothing constrains their predictions to stay executable. ELWM structures the latent state to explicitly carry energy and momentum with dissipation and control ports, enforcing strictly causal transitions, and is trained on multimodal RGB-D and inertial interaction histories. Feeding it into an arrival-time field through the Eikonal equation gives a physics-conditioned navigation policy. On held-out scenes the 0.8-second motion-prediction NRMSE falls from 0.36 to 0.29 versus generic latent models; against Active Neural Time Fields, navigation success rises from 81.3% to 89.7% and SPL from 0.64 to 0.73, while physical collisions drop from 12.1% to 5.8% and the Eikonal residual from 0.083 to 0.031.
arXiv cs.RO — Energy-Structured Latent World Models with Neural Time Fields
Infrastructure-side latent guidance cuts cooperative-driving communication cost 57%
Large models improve driving-scene understanding and long-horizon planning, but an isolated vehicle is limited by sensing range and occlusion, and on-board inference cost makes the big-model route impractical. DH-VLM has the infrastructure aggregate multi-layer hidden states into a global-reasoning latent guidance signal, integrated into the ego model through conditional latent refinement so the vehicle gets long-range context while keeping local planning autonomy. The authors also build a cooperation-oriented QA dataset for scene understanding and ego-personalized comprehension. Reported gains: 14.6% in L2 error and 26.9% in collision rate over the prior state of the art, with communication cost down 57.3% and GPU memory down 25.5% against query-based cooperative baselines, and stated robustness to erroneous infrastructure guidance.
arXiv cs.RO — DH-VLM: Dual-Horizon Cooperative Latent Reasoning for Autonomous Driving
A disassembly robot diagnoses how a machine broke by nudging its parts
With over 4 million industrial robots in use and forecasts above 16 million by 2030, taking machines apart is becoming its own problem — and unlike assembly, disassembly has no fixed sequence because you do not know in advance what failed. The KIT system starts from a CAD model, predicts how each part should move, then nudges it: a corroded part moves less than expected, a loose screw more, a deformed part along different axes, and a mathematical damage model turns those deviations into a diagnosis. The plan updates as evidence arrives — in the demonstration the system switches from unscrewing to milling away material once it observes a screw is stuck — and an operator can flag which parts must be salvaged intact. Presented at ICRA 2026 in Vienna.
IEEE Spectrum — Robot Recycler Salvages Parts From Broken Machines