Post-Quantum Crypto
AI agents build Shor's algorithm in Lean, machine-checking quantum attacks on RSA-2048 and P-256
Software agents analyzed sources, wrote Lean code, and repaired proofs to formalize the Shor's-algorithm family — from quantum order finding down to reversible circuits for modular and elliptic-curve arithmetic — with humans reviewing the scientific claims and Lean's kernel checking the proofs. The development machine-checks the logical resource estimates for attacking a 2048-bit RSA modulus (based on Gidney-Ekerå, Quantum 5, 433) and the standardized P-256 curve (ASIACRYPT 2017), adding estimates of the classical operations involved. It is a concrete step toward machine-checked quantum cryptanalysis, putting the resource numbers that drive post-quantum migration timelines on formally verified footing.
Quantum Computing
Shadow Pauli Flow completes the characterization of determinism in measurement-based quantum computing
A paper in Quantum resolves a long-standing gap in the one-way model of quantum computing: GFlow and its generalization Pauli flow guarantee robust determinism but were known not to be necessary conditions when Pauli measurements are involved. The authors prove Pauli flow is necessary in a weaker resource-state sense, then introduce Shadow Pauli Flow, an extension that is both necessary and sufficient for robust determinism — and computable in polynomial time, making the criterion practical for MBQC compilation.
Deterministic photonic phase estimation with linear circuit complexity demonstrated on four qubits
For unitaries with cyclic-group structure common in Fourier-based protocols, a new quantum phase estimation variant cuts circuit complexity from O(n^2) to O(n) and — unlike prior probabilistic KLM-style photonic implementations — runs fully deterministically, demonstrated on a four-qubit photonic system encoding two qubits in polarization and two in path modes of a photon pair.
Industrial reality check: quantum optimisation still a co-design exercise, not a speedup
A benchmark of an industrial job-shop-scheduling variant across IBM Quantum, the D-Wave annealer, and Fujitsu's Digital Annealer — evaluated against an exact classical solver and a MILP formulation — finds that solution quality and scalability hinge on tailoring formulations to each hardware's constraints. The sober conclusion: today's quantum and quantum-inspired optimisers help with solver selection, modelling decisions, and proofs of concept rather than delivering wins over classical baselines.
Filon methods speed up classical simulation of transmon control by orders of magnitude
Two numerical methods adapting Filon quadrature to highly oscillatory quantum dynamics cut the timesteps needed to simulate superconducting transmon systems at fixed precision, with the control-aware variant beating the best Hermite method by up to 6x — and same-order Hermite by up to 500x — on a realistic CNOT-gate simulation, a practical win for quantum-hardware design loops.
Quantum Kitchen Sinks flag RF anomalies in real cellular signals, validated on IBM hardware
A hybrid quantum feature-map pipeline extending Quantum Kitchen Sinks with multi-depth data re-uploading and ring entanglement detects anomalies in spectrograms of real measured sub-6 GHz cellular signals, reaching a test AUROC of 0.8778 and beating matched classical direct-readout baselines across all evaluated configurations, with runs on the ibm_quebec QPU deviating less than 0.013 AUROC from simulation.
Matched-baseline QML test: topology-aligned inductive bias, not the quantum circuit, does the work
A careful matched-baseline study instantiates the same molecular-bond-graph inductive bias as a variational quantum circuit and as a parameter-matched classical message-passing network; with just 64 trainable parameters the classical model edges out the quantum one (AUC ~0.91 vs ~0.88) on QM9 HOMO-LUMO classification, indicating the topology alignment — not quantum processing — is the active ingredient.
Trigonometric continuous-variable gates benchmarked on trapped ions
Researchers demonstrated one-qumode cosine gates and mode-resolved benchmarks of two-qumode implementations on the QSCOUT trapped-ion platform, realizing the recently proposed trigonometric continuous-variable gate set through hybrid qubit-qumode operations on the motional modes of Yb-171 ion chains — establishing these non-Gaussian primitives as building blocks for bosonic Hamiltonian simulation.
Universal computation scheme proposed for dissipatively stabilized multi-mode cat qubits
Extending multi-mode Schrödinger cat states from quantum memories to computation, this proposal derives a universal gate set for cat qubits stabilized by engineered non-local dissipation in Kerr-oscillator chains, including an entangling XX(pi/2) gate needing only one coupled oscillator per array, with numerics showing high-fidelity gates under realistic photon loss and disorder.
Geometric speed-limit conjecture for state preparation falls to curved control spaces
A paper in Quantum shows that a well-known 2019 conjecture — that energy fluctuations during quantum state preparation are lower-bounded by the minimal Fubini-Study geometric length over accessible protocols — fails when the accessible parameter space carries extrinsic curvature in the space of dynamically reachable states, demonstrated across generic qubit, qutrit, and harmonic-oscillator systems.
Quantum Comms
Quantum Source demonstrates on-demand single-atom entangled-photon source with Israeli defense R&D
Israeli startup Quantum Source Labs, working with the Israel Directorate of Defense Research & Development, announced on-demand generation of polarization-entangled photon pairs in the noise-robust singlet state from its single-atom platform, pitched at real-world quantum networking. The claim comes via press release without published performance figures, so it stays a brief pending independent corroboration.
Arq Quantum raises $1.4M pre-seed for multiplexed quantum repeaters
Barcelona-linked startup Arq Quantum Technologies, founded in 2025 by quantum scientists Samuele Grandi and colleagues, secured $1.4 million in pre-seed funding led by Ground State Ventures with participation from Big Sur Ventures to accelerate development of multiplexed quantum-repeater hardware for long-distance quantum communication.
AI & ML
Agentic system MEDA recovers mechanistic ODE models of biological systems
MEDA chains LLM agents with symbolic regression to discover ODE models of biological dynamical systems — retrieving background knowledge, generating mechanistic constraints, then proposing and fitting candidate equations. Across canonical-model retrieval, extrapolation to unseen variants, and open-ended discovery it recovers correct state variables and strong structure, with ablations showing the knowledge-guided constraints (not numerical fitting) are the load-bearing components.
AIMO competition to probe whether frontier models really reason or exploit shortcuts
Building on AI Mathematical Olympiad problems and the Fields Model Initiative, the proposed AIMO Interpretability Challenge will give participants newly published olympiad problems with symbolic representations (enabling functional variants), frontier-model access, and adversarial-robustness assessments, to develop interpretability methods that tell stable reasoning mechanisms apart from brittle shortcuts — seeding an open robustness benchmark for mathematical reasoning.
Robotics
REAL agent hits 78.3% real-world success on open-world mobile manipulation, outperforming closed-source VLMs
REAL is an agentic framework for open-world mobile manipulation that drops the usual crutches — no privileged simulator state, no assumption of complete instructions — and trains via supervised fine-tuning plus online RL against a simulated user for intent disambiguation. On its accompanying 241-task REAL-Bench it reaches 56.9% success on interactive tasks, outperforming leading commercial closed-source VLMs, and deployed on a physical dual-arm mobile robot it achieves 78.3% end-to-end success over 60 real-world episodes with zero-shot transfer to unseen household scenes. Code is released on GitHub.
Phantom Twist: a computationally designed spinning drone that is nearly invisible in flight
Northwestern roboticists presented a single-motor drone at RSS 2026 that spins at 15-25 Hz so human persistence of vision smears its sparse carbon-fiber frame into a transparent blur. An iterative optimizer searched ~20,000 feasible component layouts to minimize the LPIPS perceptual-similarity metric while preserving stable flight, landing on a design scoring 0.0104 — about twice as good as a human-designed spinner and over ten times less visible than a comparable quadrotor. Control currently relies on external optical tracking, with applications pitched at unobtrusive wildlife observation and, unavoidably, covert surveillance.
GigaWorld-Policy-0.5 trims world-model robot policies to 85 ms inference on a consumer GPU
GigaWorld-Policy-0.5 trains a robot policy with world-model supervision — jointly modeling actions and future visual observations — but decodes actions only at inference, using a Mixture-of-Transformers that separates visual-dynamics and action experts to hit 85 ms latency on a local RTX 4090; an agent-based AutoResearch pipeline searches its training configurations automatically.
Regret-based planner scales temporal-logic task allocation for heterogeneous robot teams
A planning framework formulates multi-robot task allocation under environmental uncertainty as min-max regret optimization, binding resource uncertainty into the automaton via Region-Binding Atomic Propositions and pruning policies with a regret-based branch-and-bound; experiments including physical robots show near-linear scaling in robot count and clear wins over MILP baselines in both solution quality and compute.