Quantum Computing
IBM to acquire HRL Laboratories, betting on silicon-spin qubits
IBM has signed a definitive agreement to acquire HRL Laboratories, the Malibu research institution jointly owned by Boeing and General Motors, in a move that broadens its quantum-hardware strategy beyond superconducting qubits. HRL brings deep expertise in silicon-spin qubits, and Boeing and GM are reported to retain continuing stakes in HRL after the transaction. The deal is a rare hardware-platform diversification by the industry's leading superconducting-qubit developer.
Quantum Computing
Exact classical sampler reaches 476-qubit noisy QAOA circuits
Published in Quantum, the 'Pilot-Wave Simulator' pairs tensor-network contraction with a Markov process in which a classical state evolves according to a circuit's local structure, yielding exact samples rather than the approximate distributions typical of large-scale simulators. The authors demonstrate exact sampling from ideal and noisy QAOA circuits of up to 476 qubits, including depolarizing and amplitude-damping noise models, extending the scale at which quantum-algorithm assumptions can be validated classically.
A complete theory of the Clifford commutant
Researchers resolve a longstanding limitation in understanding the Clifford group by giving a complete theory of its commutant — the operators commuting with k-fold tensor powers of Clifford unitaries — for arbitrary qubit number n and order k. They provide an explicit orthogonal basis and dimension formula, an easy-to-manipulate basis of isotropic Pauli sums generated by permutations plus at most three extra operators, and a diagrammatic calculus. Consequences include a characterization of all measurable magic measures, optimal stabilizer property-testing strategies, and generalization to prime-dimensional qudits.
Learning to abort bad QEC runs straight from syndrome data
This work trains a supervised classifier to distinguish low- from high-noise syndromes and uses its output as an abort score, without needing logical-error labels, correction operators, or code-specific likelihood calculations. Across circuit-level Gross-code and code-capacity surface-code simulations it matches syndrome-weight filtering, and on experimental logical magic-state-distillation data from QuEra's neutral-atom processor the learned score beats syndrome-weight post-selection and, combined with logical-gap filtering, improves output fidelity beyond the logical gap alone — a scalable, hardware-compatible route to more reliable error correction.
Chiral color codes give single-shot correction for exotic topological order
Dongjin Lee and Beni Yoshida introduce a family of 3D stabilizer codes that achieve single-shot error correction for fermionic and chiral topological orders, offering new platforms for fault-tolerant quantum computation. Published in PRX Quantum.
Classical annealing turns up state-of-the-art small quantum codes
Researchers use an adaptive simulated-annealing search, guided by an energy function combining code distance with a count of minimum-weight logical operators, to discover high-rate, high-distance stabilizer codes. For block lengths up to 50 physical qubits it finds state-of-the-art CSS and SWEL codes — the latter admitting transversal logical Hadamard and phase gates — whose distances often meet or beat quantum Gilbert-Varshamov-bound variants, yielding promising seed codes for concatenated fault-tolerant architectures and near-term high-rate demonstrations.
A classical baseline for Pauli-correlation quantum optimization
Pauli Correlation Encoding maps classical optimization variables onto many-body Pauli observables but relies on estimating many expectation values. This paper introduces efficiently simulable PCE, a class of dequantized realizations — using free-fermionic (matchgate) and Instantaneous Quantum Polynomial circuits — in which every needed expectation value is computed classically. On MaxCut, Maximum Independent Set, multi-dimensional knapsack, and Max3SAT benchmarks it produces high-quality solutions across problem sizes from tens to thousands of variables, providing a dequantized baseline against which future quantum PCE implementations must be measured.
A trainable, linear-scaling multiclass quantum kernel
McQuack addresses three limitations of full-Gram quantum kernels — quadratic scaling, non-trainability, and no native multiclass formulation — by using a trainable sample-to-(class-centroid) fidelity matrix, giving linear scaling in training samples. Across 150+ datasets it outperforms existing pure quantum baselines in simulation; hardware inference on 124 qubits of two IBM devices, run without training, matches an RBF kernel. The authors report no evidence of barren plateaus up to 13 qubits and highlight the importance of parameter initialization.
Hitachi, Intel and AIST launch government-backed silicon quantum project
Hitachi has been selected under a Japanese government program to lead a research effort with Intel K.K. and the National Institute of Advanced Industrial Science and Technology (AIST) to develop silicon-based quantum-computing technologies, positioning the consortium to push the field from lab research toward industrial deployment. The announcement, from a Hitachi press release via trade coverage, states intent without disclosed technical milestones.
Quantum Comms
Qoreo: deadlock-free-by-construction programming for quantum networks
Qoreo lets an entire distributed-quantum protocol be written as a single global choreography rather than a collection of error-prone actor processes. It combines a linearly-typed local quantum language enforcing no-cloning, a choreographic layer mixing local computation with classical and quantum communication, and a per-node process language. The authors prove type safety and define an endpoint projection that is sound and complete with respect to the choreographic semantics, so every well-typed choreography projects to a deadlock-free network. The metatheory is fully mechanized in Rocq, with an extraction pipeline to NetQASM for simulation and deployment.
AI & ML
Heterogeneous reasoning ensemble claims a new best on Humanity's Last Exam
PoTRE (Poly-Topological Reasoning Ensembles) splits inference across four specialized agents — adversarial refinement, hierarchical strategic planning, spectrum search, and a direct chain — then reconciles them via a task-adaptive aggregation layer using candidate selection, semantic synthesis, or neuro-symbolic verification. On three frontier benchmarks (ARC-AGI-2, Humanity's Last Exam, PRBench Finance) it reports state-of-the-art accuracy of 49.92% on HLE, exceeding the prior best official score while using similar or fewer inference tokens than heavily scaled homogeneous baselines. As a single-source preprint with an architectural rather than model-weight advantage, the SOTA claim awaits independent replication.
Small-model ensembles close the gap on malware analysis
Testing eleven open-weight SLMs, three cyber-pretrained models, and six frontier LLMs on Meta's CyberSecEval malware-analysis benchmark, the authors evaluate four orchestration architectures. A hybrid combining evidence-grounded pipelines with adversarial debate (Qwen3-4B paired with Foundation-Sec-8B) scores 35.30% overall — exceeding the best cyber-specialized baseline (22.54%) and the best ungrounded frontier model (34.77%) — though grounded Gemini remains strongest at 38.22% given the same evidence pipeline, showing evidence-grounded orchestration can substantially lift collaborative small models.
DocOps benchmark exposes where document-editing agents break
DocOps deconstructs real-world document operations into atomic dimensions and escalating workflow complexities under a deterministically verifiable framework. Evaluating representative closed- and open-source models across agentic harnesses, it shows even the most advanced frontier configurations still fail on highly coupled, long-range tasks, and a fine-grained analysis surfaces three recurring failure modes — long-term state-tracking collapse, shallow semantic verification, and destructive editing of structural metadata — mapping the capability boundary for maintaining global document consistency.
AI-redesigned starting points give directed enzyme evolution a head start
Published in Nature, this work shows that seeding directed evolution with AI-redesigned protein starting points, rather than natural proteins, yields enzymes with improved properties — an AI-for-science workflow that reshapes where protein-engineering campaigns begin.
Robotics
Dual-process VLM planner trims latency without losing driving consistency
This framework encodes planning-relevant scene knowledge in a machine-parsable structured chain-of-thought and uses an automated data engine (perception foundation models, critical-path filtering, an expert VLM) to generate supervision without manual rationale annotation. A visual arbiter estimates scene complexity before language decoding and routes each input to fast meta-action prediction or slow structured reasoning, whose fields are checked by a deterministic validator supplying verifiable rewards for GRPO. On 574 NAVSIM test samples it reaches 80.14% planning accuracy and 97.20% logical-consistency while reducing average latency 17.39% relative to applying slow reasoning everywhere.