Quantum Computing
Cluster corrections give belief-propagation tensor contraction rigorous, fast-converging error bounds
Siddhant Midha and Yifan F. Zhang show in PRX Quantum (7, 033010) that belief propagation — the heuristic workhorse behind recent large-scale classical simulations of quantum circuits — can be systematically corrected by cluster expansions, yielding polynomial-time tensor-network contraction algorithms with exponentially convergent, rigorously bounded errors. Turning BP from an uncontrolled approximation into one with provable guarantees strengthens both classical simulation of many-body systems and the classical verification baselines against which quantum-hardware claims are judged.
Quantum Computing
Plaquette computes fault-tolerance thresholds from real device physics, not Pauli caricatures
QC Design published the paper behind its flagship product: Plaquette takes a hardware error model specified once — as Kraus operators, Hamiltonian-Lindblad dynamics, or a reconstructed channel — and compiles it into four sampler classes, including a new XPauli sampler for leakage and near-Clifford samplers for coherent errors, validated to match full-state simulation where Pauli twirling can fall short. Demonstrations on transmon leakage, neutral-atom scattering, and trapped-ion heating show that Clifford-only simulation can materially misestimate thresholds and overheads.
Measurement-based quantum computing runs on an integrated silicon photonic chip
A reconfigurable silicon photonic chip generated four-qubit photonic graph states — (83.5 ± 1.8)% fidelity for star states and (75.6 ± 1.1)% for linear ones — and used them to implement measurement-based single- and two-qubit gates plus Grover and Deutsch-Jozsa algorithms. MBQC sidesteps deterministic photon-photon gates, so a working four-photon implementation on integrated silicon is a concrete step toward larger-scale photonic processors.
PC-free FPGA controller assembles defect-free atom arrays with 282 µs feedback
An integrated PXIe/FPGA control system removes the PC from the feedback loop for neutral-atom experiments, combining photon counting, real-time decisions, and waveform generation with 282 µs total latency. Single-round rearrangement of 24 stochastically loaded tweezers reaches ~96% filling, and five feedback rounds lift 10-atom defect-free array success from 65.7% to 95.4% — infrastructure needed for mid-circuit measurement and real-time error correction on neutral-atom platforms.
High entanglement plus magic still isn't enough: code-compiled circuits stay classically simulable
A family of quantum circuits built from high-rate CSS codes with transversal diagonal gates exhibits large entanglement entropy, magic, and non-Gaussianity — the standard hardness indicators — yet admits efficient MPS simulation with bond dimension bounded by encoding cost regardless of depth. Beyond the conceptual point, the compiled MPS serves as a classical reference for estimating logical output fidelity of devices running nontrivial logical circuits from local Pauli readout alone.
A quantitative resource law prices the magic needed for quantum state purification
Building on no-go results that classically simulable operations cannot achieve universal fidelity gains, this work proves an exact linear mana law (odd dimensions) and a two-sided linear robustness law (multi-qubit, exact for one qubit) for two-copy universal purification, with an explicit purification map that makes the magic-fidelity tradeoff transparent — linking magic resources directly to error mitigation and fault tolerance.
Hamiltonian reservoir computing shows comparable quantum learning across analog and digital platforms
Mapping input data directly onto a fixed Hamiltonian and harvesting nonlinear features from quantum dynamical evolution, this reservoir-computing framework sidesteps barren plateaus and achieves comparable learning performance on an analog superconducting array and a digital gate-based implementation. Notably, finite dissipation suppresses scrambling-induced instabilities and can improve performance — a constructive role for environmental coupling.
Quantum state space has a different shape in even and odd dimensions
Published in Quantum (10, 2153), this work gives tight inequalities completely characterizing how far the imaginary off-diagonal coordinates of a density matrix can extend relative to their real counterparts, yielding a three-dimensional Bloch-ball-type model of state space and uncovering a surprising qualitative difference between state-space boundaries in even and odd dimensions.
Post-Quantum Crypto
SEALSQ and Quobly sign $5M deal to embed PQC in silicon quantum processors
Post-quantum semiconductor firm SEALSQ (NASDAQ: LAES) executed a $5M commercial accord with French silicon-spin-qubit company Quobly, converting the technology partnership begun with SEALSQ's late-2025 equity investment into a deployment contract, and following Quobly's €115M Series A. A quantum-processor vendor paying to integrate PQC into its own control stack is a notable data point on security expectations for quantum hardware.
AI & ML
Game-theoretic multi-agent training cuts a 7B chemistry model's hallucinations by ~79%
G-Frame uses Bayesian and team-game principles in a closed multi-agent loop to synthesize 363,045 chains-of-thought and 199,589 QA pairs that force internalization of chemical domain constraints. The resulting 7B OmniChem model reaches parity with GPT-4o-mini on ChemBench while reducing hallucinations 79.46% relative to its base architecture, and is demonstrated on molecular design and synthesis planning.
Why fine-tuned facts don't transfer: memorized knowledge gets routed to the wrong layers
Fine-tuned LLMs often memorize new facts but fail to use them in reasoning. Using a 'self-patching' intervention that relocates activations, the authors trace this Knowing-Using Gap to knowledge-circuit misalignment — representations exist internally but are not routed to computation-effective layers — and a simple heuristic fix recovers 58–75% of the oracle headroom across domains.
Differentiable LUTs push FPGA neural inference toward nanosecond latency
FPGN treats FPGA lookup tables as learnable neurons with a hardware-aligned differentiable formulation, a routing-friendly streaming topology, and an automated latency-driven compiler. It reports up to 205x lower latency than representative FPGA binary-neural-network accelerators and 30x better LUT efficiency than prior differentiable LUT approaches at competitive accuracy — relevant wherever nanosecond-scale inference matters.
Google Research floats SensorFM, a foundation model for wearable health data
Google Research announced SensorFM, positioning it as a step toward a general intelligence and interface for wearable health sensor streams. The announcement is a single-source vendor post without independently verifiable benchmarks so far, but a dedicated foundation model for raw wearable sensor data is a notable direction from the group behind Fitbit and Pixel health telemetry.
Robotics
Ukraine scales armed ground robots: 50,000 UGVs ordered as the kill zone empties of humans
An IEEE Spectrum feature details Ukraine's shift to remotely operated ground robots in the roughly 35-km-wide drone kill zone: President Zelenskyy has ordered 50,000 UGVs by end of 2026 — more than triple 2025 procurement — and RoverTech's near-silent 800-kg Zmyi averages 57 missions before destruction against about seven for typical UGVs. Developers describe assault UGVs controlled from up to 100 km away that lurk for a week per charge, and project 30–40% fewer soldiers needed on the front line, while analysts note UGVs remain far more vulnerable than aerial drones to communication disruption.
LingBot-VA 2.0 builds a video-action robot foundation model from scratch instead of retrofitting video generators
Arguing that repurposed video-generation models are inherently inadequate for embodiment, LingBot-VA 2.0 combines a semantic visual-action tokenizer, strictly causal from-scratch pretraining, a sparse mixture-of-experts backbone for high-frequency inference, and asynchronous execution that re-grounds rollouts on the latest observation. Real-world deployment shows few-shot generalization across complex manipulation tasks, though the report leans on qualitative validation.
First 11-vs-11 humanoid robot soccer match played on hardware at RoboCup 2026
IEEE Spectrum's Video Friday leads with a genuine first: two full teams of humanoid robots playing an 11-vs-11 soccer match on real hardware at RoboCup 2026 in Incheon. The roundup also features 1X's new 25-degree-of-freedom tendon-driven hands for its NEO humanoid and Generalist's GEN-1 model, which claims 99% success on simple physical tasks (vs 64% prior) at roughly 3x speed from one hour of robot data per task — vendor numbers not yet independently verified.