Robotics
A robot hand designed from 4 million frames of human motion outperforms commercial grippers
Researchers present a data-driven framework that generates the physical morphology of robot hands -- not just their controllers -- from more than 4 million frames of human fingertip motion, optimizing tree-structured hands to reproduce target manipulations using only inverse-kinematics fingertip matching. The fabricated 6-DoF print-in-place hand achieved more accurate teleoperated fingertip tracking than available commercial robot hands, while lower-DoF specialized hands reproduced structured trajectories with less mechanical complexity; an RL design proposer cut the search from hours to minutes. The result shows large-scale human-motion data can shape a robot's body, not only its policy.
Quantum Computing
Quantum annealer computes atomic hyperfine constants beyond ground-state energies
Using a modified Quantum Annealer Eigensolver on D-Wave hardware, researchers report what they describe as the first quantum-annealing results for magnetic-dipole hyperfine-structure constants of neutral Li, Li-like Be, neutral Na, and Na-like Mg -- extending annealing beyond energy calculations to atomic properties. A zooming-and-sigma-annealing scheme with floating-point encoding (up to 10 qubits, three-decimal precision) reproduced relativistic GRASP reference values at the chosen precision, though accuracy varied with system and matrix dimension.
Quantum Comms
Analytic model speeds entanglement-distribution-time estimates for repeater chains
An npj Quantum Information paper derives an analytical computation of the entanglement-distribution time in first-generation quantum-repeater chains, offering a faster alternative to Monte-Carlo simulation for designing long-distance quantum networks.
Long-lived photon blockade achieved with weak optical nonlinearity
In npj Quantum Information, researchers report a scheme for long-lived photon blockade that operates with weak optical nonlinearity -- a regime usually too weak to enforce single-photon behavior -- of interest for photonic quantum technologies and single-photon sources.
QKD
Discrete-modulated CV-QKD gets a finite-size security proof against general attacks
A new finite-size security analysis for discrete-modulated continuous-variable quantum key distribution removes restrictive assumptions such as a bounded coherent-state dimension while incorporating realistic receiver imperfections through a trusted-detector model. Combining dimension reduction, marginal-constrained entropy accumulation, and the trusted-detector model, the authors report positive secret-key rates at practical block sizes of order 10^8, narrowing the gap between security proofs and deployable telecom-based quantum communications.
AI & ML
Sound probabilistic verification bounds policy-violation risk for AI agents
This work extends runtime policy monitoring for AI agents beyond deterministic rules to probabilistic predicates (such as imperfect PII detectors), using distributionally robust optimization to compute sound upper bounds on policy-violation probability without assuming independence between predicates. On terminal- and tool-calling-agent benchmarks it improves the security-utility trade-off over prior art while guaranteeing rigorous bounds.
Lean proof assistant doubles as a process-level reward oracle for RL
Researchers use the Lean proof assistant not just as a verifier but as a symbolic process oracle, parsing proof attempts into tactic sequences and marking the earliest failing step to supply dense, type-theory-grounded credit signals. Folded into a GRPO-style objective with first-error propagation, this tactic-level supervision outperformed outcome-only baselines on MiniF2F and ProofNet with DeepSeek-Prover-V1.5.
GEMS steers LLMs along several semantic directions at once without collapse
GEMS is a training-free activation-steering method that injects multiple semantic directions simultaneously by addressing two failure modes -- distributional deviation and directional interference -- via norm-preserving weighted superposition and real-time orthogonalization. Injecting three concurrent directions on GSM8K preserved 98% accuracy where unconstrained addition collapsed to 4%, with steering effects transferring across architectures from 3B to 31B parameters.
OpenAI reports a reasoning model surfaced 18 new diagnoses in unsolved rare-disease cases
OpenAI says researchers used one of its reasoning models to help physicians diagnose rare genetic diseases in children, reporting 18 new diagnoses among previously unsolved cases. As a single-source vendor announcement without independent corroboration or a peer-reviewed result, the claim is noted with caution pending external validation.
Robotics
Co-VLA adds explicit coordination structure to dual-arm robot policies
Co-VLA replaces the monolithic action head of a vision-language-action model with a Structured Action Expert that separates shared coordination intent from per-arm execution adjustments, plus a runtime Latent-Aware Controller that modulates synchronization, asymmetry, smoothness, and safety. Across simulation and real-world bimanual benchmarks it reports a 27% success-rate gain on tight-coordination tasks, more than doubles out-of-distribution real-world performance (13%->27%), and cuts task completion time by up to 25%.