Robotics
A robot arm driven zero-shot by an off-the-shelf coding agent through a browser
VIA (Visual Interface Agent) controls a manipulator not by fine-tuning a vision-language-action model but by letting a general foundation-model agent operate a browser-based 3D interface: it takes screenshots, issues intuitive commands, observes the result, and re-plans. With no robot-specific training and no privileged state, the agent inherits its own reasoning and closed-loop error recovery. Driven by Claude Code and Codex it solves a suite of tabletop tasks zero-shot; with the strongest model (Fable 5) it reports 96.7% success on three LIBERO-Goal tasks and 100% on a long-horizon rainbow-assembly task, with performance scaling to model strength — suggesting frontier coding/computer-use agents are already, in effect, robot-control agents given the right interface.
Quantum Computing
Engineered spin-chain state transfer benchmarked on IonQ trapped ions
Engineered spin-chain protocols for moving quantum information between distant qubits — long studied theoretically but rarely realized — are implemented via digitally simulated spin Hamiltonians on IonQ's Forte 1 and Forte Enterprise 1 trapped-ion processors. Combining exact numerical simulation with hardware runs, the authors benchmark uniform nearest-neighbor against engineered coupling profiles and show engineered interactions markedly raise state-transfer fidelity. Exploiting the commutation structure of the spin Hamiltonian, a parallel Trotter decomposition reproduces the target dynamics more faithfully while cutting circuit depth and execution time relative to conventional sequential circuits.
Spin Chain Quantum Communication on a Trapped-Ion Processor (arXiv)
Quantum ML flags Ethereum phishing, run on IBM's Heron chip
A study benchmarks quantum machine learning for anomaly detection on real financial transaction data, finding that a QSVM and VQC beat statistical methods but trail top classical deep learning — so it proposes a hybrid quantum-classical ensemble to detect phishing in Ethereum networks. A cascaded Quantum Random Access Coding encoding improves accuracy over the standard ZZ feature map (13% for QRAC-VQC, 3% for QRAC-QSVM), and the QSVM delivers the lowest false negatives. Notably, some algorithms approach simulator-level performance on the IBM Heron processor.
Detecting Phishing in Ethereum Networks using Quantum Machine Learning (arXiv)
A quantum-native backdoor attack breaks QNNs with harmonic triggers
HarmQ targets a security gap specific to quantum neural networks: classical backdoor triggers are destroyed by the aggressive downsampling forced by limited qubits, and the spectral bias of parameterized quantum circuits restricts learnable patterns. It uses sinusoidal perturbations on coarse grids with block-uniform structure to survive downsampling and align with PQCs' low-frequency preference, enabling black-box injection with only training-data access. On MNIST and Fashion-MNIST it exceeds 99% attack success at over 90% clean accuracy, versus 2.77% (BadNets), 7.96% (Watermark), 44.32% (Q-FGSM), and 3.40% (QUAP).
HarmQ: Harmonic Backdoor Attacks Against Quantum Neural Networks (arXiv)
Syndrome data leaks which circuit a fault-tolerant quantum computer is running
This work identifies a previously unrecognized class of side-channel attacks against fault-tolerant quantum computers: the syndrome data streamed to the decoder carries 'gate fingerprints' — patterns that reveal which logical operations are executing. Focusing on Clifford+T computation in the surface code, the authors show how logical operations imprint these fingerprints and how the leakage can be used to infer the circuits or algorithms being run, concluding that decoder systems are a viable attack vector and should be secured or supplied by a trusted party.
Anticipating Decoder Side-channel Attacks in Fault-tolerant Quantum Computers (arXiv)
Self-dual bivariate bicycle codes gain transversal Clifford gates
Published in npj Quantum Information, this work presents self-dual bivariate bicycle codes — a family of quantum LDPC codes attractive for their high encoding rate — engineered so that Clifford gates can be applied transversally. Transversal gates are the simplest fault-tolerant logical operations, so combining them with the favorable overhead of bivariate bicycle codes is a step toward lower-cost fault-tolerant architectures.
Self-dual bivariate bicycle codes with transversal Clifford gates (npj Quantum Information)
Low-cost logical Bell pairs for distributed fault-tolerant machines
Reported in PRX Quantum, 'entanglement boosting' is a protocol that distills high-fidelity logical Bell pairs from noisy physical Bell pairs with low spacetime overhead and, crucially, needs only 2D local device connectivity — a practical constraint for real hardware. Cheap, hardware-realistic logical entanglement is a building block for linking modules in distributed fault-tolerant quantum computers.
Entanglement Boosting: Low-Volume Logical Bell Pair Preparation (PRX Quantum)
Quantum Comms
A six-node quantum blockchain runs over telecom fiber at 805 transactions/second
This work reports an experimental demonstration of a scalable quantum blockchain that sidesteps the usual trade-off between exponential communication cost and demanding multipartite entanglement. Using weak coherent states and a circular quantum Byzantine agreement mechanism, it preserves information-theoretic security while surpassing the classical one-third fault-tolerance bound, and is implemented on a photonic integrated circuit forming a six-node network over commercially available telecom infrastructure. The protocol uses under 4% of the quantum communication resources of previous schemes, and a quantum-secured token-exchange application reaches 805.3 transactions per second with zero failures.
Experimental demonstration of scalable quantum blockchain (arXiv)
Post-Quantum Crypto
Superposition-secure zero-knowledge from standard LWE
This work lifts the MPC-in-the-head technique — foundational to modern post-quantum signatures — to the quantum setting, where the underlying multiparty computation runs a quantum computation. It yields two three-round protocols in the common-reference-string model: a zero-knowledge argument for NP and one for QMA, both reducing to the standard learning-with-errors assumption, and both secure against a verifier that can obtain a superposition of transcripts — a guarantee that previously required commitments not known from standard assumptions.
MPC in the Quantum Head (Superposition-Secure Quantum Zero-Knowledge) (Quantum)
AI & ML
Zero-RL reasoning scaled to a trillion parameters
The Ring-Zero work pushes 'zero RL' — reinforcement learning with verifiable rewards and no human-annotated data — to 1-trillion-parameter scale, using clipped importance sampling, training-inference ratio correction, and mixed-precision control to keep training stable. The authors report that scaling improves sample efficiency and the performance ceiling, that training proceeds through a discovery phase followed by a sharpening phase, and that the model spontaneously develops behaviors including self-verification, parallel reasoning, and structured formatting. Ring-2.5-1T-Zero is competitive across seven math benchmarks, with a structured evaluation of chain-of-thought quality along comprehensibility, reproducibility, and efficiency. Single-source preprint; results not yet independently reproduced.
Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning (arXiv)
Robotics
Robots hand coding agents a 'contract' so generated behavior trees actually run
Synthesizing deployable robot behavior trees from natural language is brittle when the prompt author does not know the robot's real skills or runtime constraints. This architecture moves that grounding onto the robot: a coding agent queries a robot-side MCP server for an explicit contract — a skill library, permitted BT operators, and optional composition templates — before synthesizing a tree, with a runtime validation gate enforcing correctness before execution. Evaluated with a closed model (Sonnet 4.6) and a smaller open model across 110 PyRoboSim tasks and 14 tasks on a physical Husarion Panther, contract grounding yields near-perfect validation and high success, and transfers to hardware running an opaque Nav2 stack.
Contract-Grounded Behavior Tree Synthesis via Coding Agents (arXiv)
Trimming redundant frames doubles VLA inference speed
Vision-language-action models generalize well for manipulation but their inference latency blocks real-time use. This work targets two sources of temporal redundancy: repeated visual encoding of near-identical consecutive frames, and multi-step iterative sampling in diffusion policies. It incrementally updates only tokens for dynamic scene regions instead of re-encoding whole frames, and compresses diffusion sampling into a 2-step schedule via efficiency-oriented training. On Libero, RobotWin, and real robots it reports over 2x speedup while preserving performance, reaching up to 98% success on general manipulation.
Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference (arXiv)
Automated deployment speeds up a nano-drone's onboard navigation net
This work streamlines the error-prone manual pipeline for deploying vision-based CNNs on sub-10cm autonomous drones, automating everything from training to closed-loop evaluation of PULP-Dronet on an ultra-low-power multicore SoC acting as mission computer on a Crazyflie 2.1. Versus the hand-crafted baseline it halves memory footprint and gives a 1.6x inference speedup at equal accuracy, improving field behavior: obstacle avoidance braking at 1.65 m/s, free flight up to 1.96 m/s (vs 0.5 m/s baseline), and a 90-degree lane-following turn — all using under 1.6% of the drone's power budget, with the software open-sourced.
Whole-body loco-manipulation via learned partial kinematic embeddings
PAKE tackles high-precision control over the high-dimensional, redundant action space of whole-body loco-manipulation by splitting the problem into partial reference-motion generation and low-level imitation control. A Kinematic Normalizing Flow trained on a large-scale kinematic dataset generates diverse but feasible partial reference motions; a high-level controller navigates its latent space to exploit redundancy while a low-level controller ensures feasible execution. On a quadruped with a six-DoF arm, across 24 hardware episodes over 8 tasks, it reaches end-effector pose-tracking errors of 4.5cm and 0.14rad with tight locomotion tracking, beating competitive baselines.
PAKE: Learning Whole-Body Loco-Manipulation with Partial Kinematic Embeddings (arXiv)