Quantum Computing
Cold Rydberg gas maps non-Hermitian topology in the lab
A large experimental team used a cold Rydberg gas as a tunable dissipative many-body platform to map out non-Hermitian spectral topology, observing winding numbers driven by atomic interactions and a nonreciprocity that depends on the direction the system is scanned. Published in PRX Quantum, the experiment turns dissipation — usually the enemy of quantum coherence — into a controllable resource for probing topological physics beyond conventional Hermitian band theory.
Quantum Computing
Classical algorithm catches up with a landmark Rydberg simulation
A new variational Monte Carlo framework for tensor-network states reduces the cost of evaluating all-to-all two-body interactions from O(N³) to O(N), enough to accurately simulate the adiabatic real-time dynamics of the 10×10 dipolar XY model realized in a 2023 Nature Rydberg experiment that had been considered beyond classical simulation. The authors position the method as the scalable classical benchmarking tool that long-range platforms such as Rydberg and ion-trap simulators currently lack — another data point in the running contest between classical algorithms and quantum-advantage claims.
Climate stress test finds QNNs still trailing plain neural nets
A hybrid quantum neural network with a rich trainable frequency spectrum and extensive hyperparameter optimization proved workable on a complex cloud-microphysics dataset relevant to climate prediction — but was still beaten by simple fully-connected classical networks, with the authors cataloguing the bottlenecks variational quantum models must clear before they can fill gaps where classical models fail.
Rydberg dynamics double as a natural quantum hash function
A Letter argues that constrained many-body dynamics in Rydberg atom arrays natively realizes quantum hashing: ternary strings encoded as deterministic state-space trajectories yield near-orthogonal, random-like quantum states with low collision probability, one-wayness and tamper sensitivity — cryptographic functionality emerging from the physics itself rather than a deliberately engineered algorithm.
Quantum networks may dodge deep learning's plasticity loss
Work by Yu-Qin Chen and Shi-Xin Zhang published in PRX Quantum reports that quantum neural networks intrinsically preserve plasticity during continual learning — retaining the ability to absorb new tasks as parameters grow, where classical networks notoriously stiffen — hinting at a structural advantage of quantum models for long-term learning.
The quantum switch goes universal — and nested
npj Quantum Information published a construction for a universal nested quantum switch, extending the quantum switch — the standard primitive for indefinite causal order — toward nested compositions of operations, broadening the theoretical toolbox for causally indefinite quantum information processing.
Particle scattering gets leaner on digital quantum hardware
A study in npj Quantum Information demonstrates resource-efficient simulations of particle scattering on a digital quantum computer, trimming the overheads that make scattering dynamics one of the more demanding targets for near-term quantum simulation of fundamental physics.
MagiQware raises €575K to tune magic state factories with RL
Delft University of Technology spin-off MagiQware finalized a €575,000 ($658K) pre-seed round — a €500K close led by LUMO Labs via the TTT.AI programme, expanded by a Graduate co-investment — to apply reinforcement learning to optimizing magic state factories, one of the dominant resource costs in fault-tolerant quantum computing.
Quantum Sensing
NSF's Project Triad aims to fuse sensing, networking and computing
The U.S. National Science Foundation announced Project Triad, a program to combine quantum sensing, quantum networking and quantum computing — for the first time — into one operational system and push the technologies out of the lab into real-world use. So far the launch is a single press release relayed by trade press, with funding levels and participating institutions still to be detailed.
AI & ML
Verification pitched as a new scaling axis for LLM agents
LLM-as-a-Verifier replaces discrete judge scores with an expectation over scoring-token logits, yielding continuous verification signals that improve with finer score granularity, repeated evaluation and criteria decomposition — no additional training required. The framework reports state-of-the-art results on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%) and MedAgentBench (73.3%), and doubles as a dense reward for reinforcement learning; the SOTA claims are single-preprint so far, pending independent replication.
Coding agents 'know' their code's fate 25 steps early
Simple linear probes on the residual streams of language models running coding agents decode whether the current code parses, passes its test suite or introduces regressions at up to 0.83 AUC — and predict the outcome of future edits roughly 25 steps before they materialize, a 'latent programming horizon' that transfers across benchmarks without retraining.
Spiking networks near conventional accuracy on automotive vision
The first comprehensive evaluation of spiking neural networks for real-world automotive multi-object detection and tracking reports 0.937 mAP on KITTI and 0.771 on BDD100K using transfer learning with SpikeYOLO — competitive with conventional deep learning while promising the energy efficiency of neuromorphic hardware for autonomous-vehicle perception.
Chatbots quietly echo stigma around health conditions
Science reports evidence that negative perceptions of mental illness and other health conditions subtly shape the outputs of large language models, meaning chatbots can perpetuate stigma in health-related conversations — a caution flag for the rush to deploy LLMs in care and counseling settings.
Robotics
ACE-Brain-0.5 packs a closed-loop robot brain into one 8B model
ACE-Brain-0.5 organizes robot intelligence into five coupled functions — spatial perception, decision making, embodied interaction, self-monitoring and self-improvement — with a single 8B backbone instantiating the first four and a companion framework updating task schemas, spatial memory and failure-recovery cases from rollouts. Across fifteen benchmarks it improves on its predecessor ACE-Brain-0 on 14 of 18 spatial perception and grounding tests while staying competitive in navigation and manipulation, an incremental but broad step toward general physical agentic AI.
Cortex bridges robot planning semantics and execution kinematics
Cortex couples a high-level vision-language planner to a low-level vision-language-action executor through 32 canonical skill primitives and a tractability-aware pipeline spanning over 4k hours of auto-annotated video, outperforming monolithic baselines by 3.1% on Libero-long and 4.1% on RoboTwin and zero-shot completing unseen real-world long-horizon tasks such as multi-stage chemistry experiments.
Predicting geometry, not pixels, sharpens robot motion skills
GeoMoLa learns discrete motion codes by predicting how point clouds evolve during manipulation instead of reconstructing visual observations, forcing the latents to encode physical motion rather than appearance; it reaches state-of-the-art benchmark performance from a single RGB-D view and transfers to cluttered real-world scenes with minimal demonstrations.