Quantum Computing
'Routing codes' slash quantum error-correction overhead ~8x with crossing-free wiring
A new family of quantum low-density parity-check (qLDPC) codes, called routing codes, matches the encoding rates of bivariate bicycle codes while shortening non-local couplings and, crucially, making all of them mutually parallel. That parallelism eliminates wiring crossings in superconducting multi-layer chips and drastically simplifies atom rearrangement in neutral-atom arrays. Under circuit-level simulation, weight-7 routing codes reduce physical-qubit overhead by roughly a factor of 8 versus surface codes achieving the same logical error rate, positioning the family as a hardware-centric bridge between theoretical optimality and near-term feasibility.
Quantum Computing
First device-independent self-test of a causally indefinite quantum process
Self-testing certifies quantum objects from measurement statistics alone, making no assumptions about device internals. This work extends it from states, measurements and channels to quantum supermaps, operations that act on channels and can combine them in either definite or indefinite causal order. The authors obtain two certification levels depending on the experiment's network structure and demonstrate the approach on four examples, including Grover's algorithmic comb and the quantum switch, delivering the first self-test of both a quantum algorithmic comb and a causally indefinite process, and a new route to certify causal indefiniteness device-independently.
Quantum signal processing recast as an analytic tool for quantum control
QSP-Control adapts the quantum signal processing framework to qubit-oscillator dynamics, using its analytic structure to mitigate unwanted cross-Kerr nonlinearities in dispersively coupled systems and to design Fock-state-selective operators for precise Fock-state manipulation. The result offers a systematic, interpretable alternative to the brute-force pulse optimization that dominates quantum control, mapping practically relevant control problems onto forms amenable to QSP.
npj Quantum Information: optimal control for Rydberg multi-qubit operations
A peer-reviewed npj Quantum Information paper (published 25 June 2026) presents optimal-control protocols for multi-qubit operations on Rydberg-atom systems, a key route toward higher-fidelity entangling gates in neutral-atom quantum processors.
Qblox locks in DOE/Fermilab deployment and HPE hybrid-HPC partnership
Quantum control-electronics maker Qblox announced two concurrent initiatives: a formalized deployment contract with the U.S. Department of Energy and Fermi National Accelerator Laboratory, and an infrastructure partnership with Hewlett Packard Enterprise to integrate quantum control systems into high-performance computing and AI infrastructure. Reported across multiple trade outlets, the dual announcements expand the company's North American commercial footprint; specific technical and financial terms were not disclosed.
Quantum Comms
Measurement-device-independent steering certification extended to networks and Gaussian schemes
A new framework lifts quantum steering certification to the measurement-device-independent regime in multipartite networks, where all-but-one parties are untrusted and even the trusted party treats its hardware as a black box apart from fiducial input states. It covers both finite-dimensional and bosonic continuous-variable systems, fully characterizing the bipartite continuous-variable case, and introduces measurement-device-independent network steering protocols built entirely on Gaussian operations, providing a feasible route to minimal-trust certification for applications such as randomness generation.
QKD
Side-channel from digital-to-analog conversion analyzed in CV-QKD
A peer-reviewed npj Quantum Information study (published 25 June 2026) examines how information leakage arising from digital-to-analog conversion affects the security of continuous-variable quantum key distribution, characterizing this hardware side-channel and its implications for practical CV-QKD systems.
AI & ML
RL post-training yields a free step-level reward signal for LLM agents
Building process reward models for agentic settings is hard because long horizons, irreversible actions and stochastic feedback defeat human annotation and Monte Carlo estimation. This work shows the log-probability ratio between an RL-trained policy and its reference policy, the 'progress advantage', exactly recovers the optimal advantage function under a general stochastic MDP, providing annotation-free, domain-agnostic step-level scoring as a byproduct of standard RL post-training. Across test-time scaling, uncertainty quantification and failure attribution on five benchmarks and four model families, it consistently outperforms confidence-based baselines and even surpasses dedicated trained reward models.
AI research pivots from text to simulated worlds as chatbot gains slow
A Science news feature reports that, as gains from training ever-larger language models on text grow harder to achieve, researchers are increasingly building agents that learn by acting inside virtual environments rather than passively absorbing more text. The shift, framed as part of a broader pursuit of human-level intelligence, bets that embodied, interactive learning can supply the kind of grounded experience that next-token prediction on web text cannot.
A 'generalization spectrum' probes how far learning transfers, example by example
The Generalization Spectrum evaluates learning algorithms by measuring per-example transfer across graded distances, from exact recall to cross-language re-implementation and re-framed-context variants, rather than a single aggregate test score. Instantiated on competitive programming with contamination controls, it finds reinforcement learning converts memorization into near-transfer more efficiently than supervised fine-tuning, while in-context learning exhibits strong but correspondence-dependent transfer, and that local gains often fail to expand the generalization radius.
Two Max Planck papers retracted, an automated system may be to blame
Science reports that Springer Nature has retracted two studies attributed to physicist Max Planck, with an automated bot apparently responsible for flagging them. The episode illustrates how automated tooling can misfire in scholarly publishing and erroneously sweep up even historical work.
Robotics
Hierarchical scene-graph navigation pairs an LLM planner with reactive control
SAGE-Nav tackles object-goal navigation by decoupling asynchronous LLM-based global semantic planning from a high-frequency reactive control loop. A hierarchical scene-graph encoder produces structure-aware embeddings preserving semantic and spatial topology, while a goal-aware alignment-fusion network with adaptive gating fuses real-time perception with these structural priors. The authors report state-of-the-art performance and improved zero-shot generalization in the i-THOR and RoboTHOR environments while keeping control latency low enough for physical deployment.