Quantum Computing
Trapped-ion hardware guides real-market portfolio selection at 98-qubit scale
Researchers formulate portfolio diversification as a Maximum Independent Set problem and run qReduMIS — a recursive hybrid in which QAOA measurements identify 'frozen' assets that unblock provably optimal classical reductions — on Quantinuum's 98-qubit Helios trapped-ion system, with circuits acting on kernels of up to 78 qubits and 1,016 two-qubit gates. On real data from four major indices (up to 225 assets), where standalone QAOA fails to find the optimum for the S&P 100 and Nikkei 225, qReduMIS reaches success probabilities of 0.40 and 0.95 with approximation ratios of at least 0.96 across all four. An emulator benchmark over 73 asset-correlation graphs shows the optimal time-to-solution scaling exponent is 3.2x smaller than standalone QAOA's.
Quantum Computing
Beam-search decoder pushes qLDPC error correction toward real-time latency
Min Ye, Dave Wecker and Nicolas Delfosse introduce a beam-search decoding approach for quantum low-density parity-check codes in PRX Quantum, aimed at combining high accuracy with the latency requirements of real-time quantum error correction. Fast, accurate qLDPC decoding without heavyweight classical co-processing is one of the key bottlenecks on the path to practical fault tolerance.
Parallel matching decoder shrinks fault-tolerance time overhead to doubly-polylog
A paper in npj Quantum Information constructs a parallel minimum-weight perfect-matching decoder that runs in polylogarithmic time, and uses it to achieve fault-tolerant quantum computation with only doubly-polylogarithmic time overhead. Decoder throughput is a central obstacle to scaling surface-code machines, making asymptotic guarantees of this kind directly relevant to fault-tolerance roadmaps.
Global transverse-field Ising model shown polynomially equivalent to the gate model
Building on Cesa and Pichler's global-control result for Rydberg atoms, the authors show the Ising model with a non-monotonic, time-dependent global transverse field can simulate arbitrary quantum circuits with polynomial overhead in time, qubit number and energy scale — settling an open question relevant to quantum annealing and analog platforms, and implying such systems cannot be efficiently simulated classically unless quantum computing offers no advantage.
Head-to-head study finds quantum ML still trails classical baselines
A unified empirical comparison of seven quantum/classical model pairs spanning supervised and reinforcement learning finds the quantum models do not yet beat classical counterparts on prediction performance, policy stability, or training time, though they show promise for noise filtering and false-positive control. Code is public.
Greedy scheduling cuts multi-core quantum circuit runtimes by 40%
For multi-core quantum systems, scheduling each gate as soon as its dependencies and resources are available — rather than executing circuit layers sequentially — yields an average 40% reduction in makespan and better core utilization on real benchmarks, exploiting fine-grained parallelism across cores.
JIJ and Kobe Steel launch a quantum center of excellence for heavy industry
JIJ Inc. and Kobe Steel signed a multi-year framework agreement to establish a collaborative Quantum Center of Excellence and push quantum computing and mathematical optimization into KOBELCO Group operations — an industrial-adoption signal from Japanese heavy manufacturing, though the announcement carries no concrete deployment figures yet.
Quantum Comms
Common noise approximations badly mispredict quantum-network performance
The study demonstrates that widely used error-model approximations in quantum-network simulation can under- and over-estimate protocol performance, introduce measurement-outcome dependencies and fidelity oscillations that the simplified models miss entirely — accumulating across iterative protocols like entanglement purification and repeater chains. The authors argue full noise architectures must be validated before trusting predicted operational thresholds.
Quantum Sensing
Minimal detector model prices the thermodynamic cost of quantum measurement
Schwarzhans, Mitchison, Huber and colleagues present a minimal model of a quantum particle detector as an autonomous machine, establishing fundamental trade-offs between entropy production and measurement accuracy — new insight into the thermodynamic cost of acquiring information from quantum systems.
Michigan-led QuPID photonics project takes $4M NSF quantum lab award
The Quantum Photonic Integration and Deployment (QuPID) consortium, led by University of Michigan Engineering with Ohio State participation, secured a $4 million, two-year Phase 2 award in the NSF's National Quantum Virtual Laboratory design competition — one of nine initiatives designing plug-and-play photonic circuits for deployable quantum measurement, following a $1 million pilot.
AI & ML
Treating memory as a trainable skill lifts LLM agents 2-4x on long-horizon games
AutoMem promotes file-system operations to first-class memory actions and automates two improvement loops: a strong LLM reviews full agent trajectories to revise the memory structure, while the agent's own good memory decisions become training signal for the model. Optimizing memory alone — without touching task-action behavior — improves a base agent roughly 2-4x across Crafter, MiniHack and NetHack, bringing a 32B open-weight model competitive with frontier systems such as Claude Opus 4.5 and Gemini 3.1 Pro Thinking.
Self-reflecting discovery agent recovers 8 of 9 known ecological patterns
DiscoPER conducts open-ended data-driven research without pre-specified questions, requiring every proposed discovery to pass statistical testing, and periodically analyzes its own accumulated discoveries to redirect exploration toward gaps. On iNatDisco, a multimodal ecological benchmark with literature-grounded truth, it recovers 8 of 9 known patterns with a 72.7% hypothesis support rate, beating classical causal discovery and LLM-guided baselines.
Native-Metal runtime claims fastest LLM inference yet on Apple Silicon
BaseRT builds directly on Metal with chip-specific kernel fusion and unified-memory-aware optimization, reporting up to 1.56x higher decode throughput than llama.cpp and 1.35x higher than MLX on M3 and M4 Pro devices across Qwen3, Llama 3.2 and Gemma 4 families, with larger prefill margins on mixture-of-experts models. The runtime is public on GitHub, making the on-device edge-inference claims independently checkable.
Neuroscience-style parcellation maps functional systems inside LLMs
NeuroCogMap applies a cognitive-neuroscience framework to LLM internals, finding semantically coherent functional parcels that are partly conserved across models and causally linked to outputs. Major failure modes — hallucination, bias, refusal failure, sycophancy — correspond to distinct disruptions with detectable internal signatures, and the parcellation improves prediction of human cortical responses during naturalistic language comprehension.
OpenAI ships GeneBench-Pro to test AI on real-world genomics research
OpenAI announced GeneBench-Pro, a benchmark for evaluating AI systems on genomics, biology, and scientific-research tasks built from complex real-world datasets. The announcement is vendor-only and gives no model scores or dataset statistics, so its rigor and difficulty remain to be independently assessed.
Robotics
ASPIRE robots write their own skill libraries and transfer them across bodies
ASPIRE (Agentic Skill Programming through Iterative Robot Exploration) couples a closed-loop execution engine that diagnoses and repairs failures with an ever-growing library of validated, reusable skills and evolutionary search over tasks and programs. It surpasses prior methods by up to 77% on LIBERO-Pro manipulation under perturbation and 72% on Robosuite bimanual handover, and its accumulated library gives 31% zero-shot success on unseen long-horizon tasks versus 4% for baselines, with initial evidence of sim-to-real transfer across embodiments.
Autonomous catheter navigation hits 96% success by splitting strategy from execution
The framework separates high-level strategic decision-making (an RL policy generating navigation intents) from low-level procedural execution constrained by expert operational norms, kinematic limits and vessel-safety constraints. Across high-fidelity 3D simulations and a real robotic platform it exceeds 96% navigation success with 29.3% fewer operational steps and 13% lower trajectory variance than baselines.