Quantum Computing
ETH Zurich stores quantum information in vibrations, giving superconducting processors a mechanical RAM
In a Science paper, ETH Zurich researchers demonstrate a hardware architecture that separates quantum processing from working memory: superconducting qubits do the computing while mechanical resonators store the information as vibrations rather than electromagnetic fields. The design deliberately mirrors the classical CPU/RAM split, pointing to denser, longer-lived quantum memory than circuit-only architectures — a peer-reviewed architectural result rather than a vendor roadmap claim.
Quantum Computing
'Works on my QPU': audit of ~5,000 quantum-computing papers finds most published code won't run
A large-scale reproducibility study manually evaluated 127 quantum-computing papers and automatically screened nearly 5,000 more: only 24.4% of the sampled papers provide code artefacts, 64.5% of those fail to execute in a clean environment, and roughly a third of papers with code lack machine-readable environment specifications. The authors distill the failure modes into concrete recommendations — hard numbers on a problem the field mostly acknowledges only anecdotally.
Variational Gibbs-state preparation demonstrated on trapped-ion hardware
npj Quantum Information published a peer-reviewed demonstration of variational Gibbs (thermal) state preparation on trapped-ion devices — a building block for simulating finite-temperature physics on near-term quantum hardware.
MIT and IBM map quantum operators into an LLM's latent space for circuit synthesis
Researchers from the MIT-IBM Computing Research Lab and IBM Quantum built a multimodal alignment framework that projects quantum unitary operators directly into an LLM's latent space, treating mathematical quantum operations as an additional modality for circuit synthesis. Published as an IEEE QCE 2026 conference paper — an unusual concrete bridge between quantum compilation and language-model tooling.
Adaptive qubit freezing lets divide-and-conquer QAOA handle dense graphs it previously couldn't touch
FrozenLGP makes graph partitionability an enforceable property for divide-and-conquer QAOA: when no small separator exists, it computes a minimum vertex cut, classically freezes those spins, and folds their interactions into bias terms. Across graphs up to 10,000 vertices it achieves 100% decomposition coverage versus 4.6% for the standard baseline on high-connectivity instances, while preserving MaxCut approximation quality.
Cat qubits stabilized by repeated interactions instead of engineered dissipation
A theoretical proposal stabilizes dissipative cat qubits through repeated interactions with an auxiliary two-level system mediated by a quadratic Hamiltonian — no reservoir engineering required. The scheme sidesteps the spurious decay channels of two-photon-dissipation implementations, preserves the noise bias, extends to squeezed cats, and is claimed compatible with platforms beyond superconducting circuits. Theory only, no experiment yet.
Condition-number analysis backs exponential-advantage prospects for quantum linear solvers in chemistry
Extending the quantum-linear-solver approach to internally contracted multi-reference linearized coupled cluster, this work probes the make-or-break condition-number scaling with three complementary diagnostics, all indicating polylogarithmic growth in system size. Combined with sub-linear sparsity arguments, that supports prospects of exponential quantum advantage for this chemistry problem class; numerical tests on systems up to four atoms recover ground-state energies within 0.009% of classical benchmarks.
Tensor-network PDE solver gets exponential memory cut and 100x faster convergence
For Sturm-Liouville-type PDEs, an analytical expression for Pauli-basis coefficients builds matrix product operators with O(2n) instead of O(2^(n+1)) memory, and a multistage state-refinement heuristic speeds imaginary-time-evolution convergence by up to two orders of magnitude. The framework computes the first 32 eigenstates of a Laplacian exceeding dimension 10^6 with fidelity above 0.95 using a 20-qubit MPO — a quantum-inspired classical method with concrete scaling wins.
Quantum Comms
Quantum dense network coding halves per-sender communication and spawns a new MDI key-growing protocol
A new protocol — quantum dense network coding — delivers the output of a non-Boolean function to a receiver using provably half as many qubits as classical bits per sender, without transmitting the full inputs. The advantage requires both shared entanglement and quantum communication, survives noise, and the quantum-classical success-probability gap can be amplified exponentially in the number of senders; as a by-product the authors obtain an information-theoretically secure, measurement-device-independent quantum key-growing protocol. Goes beyond the previously covered superdense-coding all-reduce work by targeting general function computation and adding a cryptographic primitive.
Erbium-doped fiber serves as a quantum memory for chip-integrated quantum-dot photons
A peer-reviewed npj Quantum Information paper reports an erbium-doped fiber quantum memory interfaced with chip-integrated quantum-dot single photons at 980 nm — pairing a mature photon source platform with a fiber-based memory, both ingredients for practical quantum networking.
QKD
Measurement incompatibility becomes a quantitative certificate for quantum randomness
This work derives a quantitative trade-off between observed measurement incompatibility and an eavesdropper's capabilities in prepare-and-measure scenarios, using the generalised robustness of incompatibility to bound Eve via semidefinite programs and giving an explicit protocol that extracts randomness from any incompatible measurement set. Translated to steering, it yields a tight connection between steerability and randomness generation, and extends to eavesdroppers with quantum memory.
Quantum Sensing
Ziv-Zakai analysis deflates critical quantum metrology: the 'super-Heisenberg' gain was the prior, not the phase transition
Using quantum Ziv-Zakai bounds instead of the usual Cramér-Rao analysis, this work shows that for second-order quantum phase transitions the precision of critical quantum sensors offers no substantial improvement over the prior standard deviation — even before accounting for state-preparation cost or noise. The claimed super-Heisenberg sensitivity at criticality is traced to the large prior information the schemes assume, a substantive correction to a prominent quantum-sensing storyline.
AI & ML
UniClawBench tests proactive agents on 400 live real-world tasks, not sandboxed replays
UniClawBench evaluates proactive LLM agents along five base capabilities (skill usage, exploration, long-context reasoning, multimodal understanding, cross-platform coordination) across 400 bilingual tasks run in live Docker containers with fine-grained step checkpoints, using a hidden supervisor and simulated multi-turn user feedback instead of static answers. Evaluating models under multiple agent frameworks lets it separate base-model capability from framework design; benchmark and code are public.
8B agent with episodic visual memory outperforms 32B baselines on long multimodal dialogues
Instead of re-feeding all visual history into the context window, this agent externalizes it into an episodic visual memory with structured abstraction and cross-turn retrieval, trained via RL on programmatically generated conversations with retrieval annotations. The 8B agent hits 91.4% retrieval accuracy over 20-turn sessions — +8.2 points over 32B baselines — while cutting per-turn inference from 23.1s to 12.7s; a tool-augmented deployment harness is open-sourced.
Simplest-possible Monte Carlo search shown to train deep networks without backpropagation
Randomly mutate one parameter, keep the change if the loss drops, retry otherwise: this bare Monte Carlo procedure on a single GPU trains networks more than 20 layers deep — without batch normalization or residual connections — plus 16,384-neuron single-hidden-layer nets and a simple Transformer on MNIST and Tiny Shakespeare. It also enables pure pruning training, discrete weights, and unconventional activation functions, offering a physically-implementable alternative lens on how deep networks learn.
GPT-5.6 becomes the default brain of Microsoft 365 Copilot
OpenAI says GPT-5.6 is now the preferred model in Microsoft 365 Copilot, powering Word, Excel, PowerPoint, Chat, and Cowork. A vendor announcement without benchmarks, but a concrete deployment shift that puts OpenAI's newest frontier model in front of one of the largest enterprise user bases.
Robotics
WCog-VLA tops NAVSIM driving benchmark by fusing semantic forecasting with generative world evolution
WCog-VLA couples semantic-level world cognition (3D spatial perception, agent tokens, game-theoretic chain-of-thought reasoning) with a generative world model — an aligned decoupled diffusion transformer that synthesizes physically-plausible joint multi-agent trajectories with fewer denoising steps. It reaches a state-of-the-art 92.9 PDMS on NAVSIM, supported by a new 85k-annotation Game-CoT dataset.