Quantum Computing
An LLM agent runs a trapped-ion lab — but only through a hardware-safety gate
Researchers placed an LLM agent in the control loop of a trapped-ion experiment, giving it the ARTIQ stack through an MCP server while enforcing a per-operation boundary: no tool call reaches the hardware unless it carries an authorization token bound to its exact contents. Tokens are issued automatically by running the agent's proposed script in an isolated hardware simulation (dax.sim) against preset per-device bounds, or manually by a human for sensitive actions. On a co-trapped 40Ca+/40CaOH+ crystal the agent autonomously built a full calibration stack and, with operator guidance, closed a cross-instrument magnetic-field-stabilization loop; interface-level portability was confirmed on a separate 171Yb+ platform. Adversarial scripts were used to map where the authorization boundary holds, and the authors find the agent's remaining limits lie in metacognitive control rather than domain knowledge.
Quantum Computing
Reusable 'catalyst' states promise exact fine phase gates for fault-tolerant circuits
A new surface-code cultivation protocol builds reusable logical catalyst states that implement exact dyadic phase gates Z^{2^-b} by phase kickback, so the target phase incurs no Clifford+T synthesis approximation error and the online non-Clifford depth no longer scales with logical accuracy. As a concrete demonstration the authors cultivate a catalyst for sqrt(T)=Z^{1/8} from nine distance-three rotated surface-code blocks; because the phase read-out is intrinsically fault-tolerant, a single verification round already reaches the leading error-corrected scaling. A hybrid tensor-network/stabilizer simulation shows that at physical error rate 1e-3 the postselected catalyst can be grown to distance-seven blocks with logical leakage ~1e-6 in about seven expected attempts. The trade is offline, phase-specific cultivation for exactness, reusability, and constant-depth online gates.
Learning pure quantum states 'almost without regret'
npj Quantum Information published work on learning pure quantum states with near-optimal regret, casting the task in the online-learning framework where a learner sequentially refines its estimate of an unknown pure state. The peer-reviewed venue places the result among formal contributions to quantum learning theory, though the published abstract gives only the headline claim.
A universal algebra for Lindblad dynamics of open quantum systems
This work formulates the Lindblad equation for finite-dimensional open quantum systems in terms of a closed, model-independent algebra of Hermitian operators via a suitable operator representation of the Liouville superoperator. The construction shows dissipative dynamics requires a substantially richer algebraic structure than unitary evolution, yields parametrizations of the dynamical map and differential equations for its evolution, and provides recursion relations that build the algebra efficiently as system dimension grows — reducing the computational cost of constructing the Liouville superoperator. A Mathematica notebook with a one-qubit example, extensible to many qubits, is provided.
Türkiye sets a national quantum roadmap naming 85 priority technologies
Türkiye laid out a national strategy for entering the global quantum race, publishing a roadmap that names 85 specific technologies it intends to pursue across quantum computing, sensing, and communication. The document — the Quantum Technologies OTAG Result Report, produced under the Presidency — joins a wave of national quantum strategies, though as a trade-reported government plan it lists priorities rather than concrete results.
AI & ML
Ternary LLMs without the training bill: CAT-Q post-quantizes models with 512 samples
CAT-Q is a post-training ternary quantization method for LLMs built on two coupled components: learnable modulation, which reshapes the pretrained weight distribution and ternary threshold to be less sensitive to ternarization, and softened ternarization, a differentiable transition function for stable convergence. The authors report that for 1.7B-8B models, CAT-Q quantizes to ternary using only 512 calibration samples while beating the seminal BitNet 1.58-bit v1/v2 families (1.3B-7B) trained on 100B tokens — about a 100,000x reduction in training tokens — and, for the first time, ternarizes 14B-235B pretrained models in 8 to 60 hours on eight A100-80GB GPUs. Code is released.
Neural PDE solvers that read their own error fields and learn to fix them
Neural surrogates for PDEs usually treat solving as a statistical mapping and struggle to correct their own constraint violations, while hybrid methods that minimize the PDE residual inherit the cost and instability of classical optimizers. The authors show residual minimization can be an unreliable proxy for accuracy in ill-conditioned systems, then propose error-conditioned Neural Solvers (ENS) that pass the residual field directly into the network at each iteration, letting it read the spatial structure of its own errors and learn an iterative update policy. Across four PDE families ENS attains the highest accuracy in the large majority of settings — with gains up to 10x on turbulent Kolmogorov flow — and its learned correction generalizes under distribution shift, including zero-shot parameter changes and cross-equation transfer.
A 7B GUI agent beats a 32B model by learning from hindsight experience
Small open-source multimodal LLMs are cheaper and more private than commercial models but plan poorly and generalize weakly across websites. PEEU (planning experience exploration and utilization) autonomously explores environments to discover experiences and uses hindsight experience to synthesize strictly aligned, high-level training data, paired with a hierarchical analysis framework that studies compositional generalization across low/middle/high task granularities. The analysis finds mastering low-level atomic skills does not guarantee high-level planning competence, while high-level task training yields stronger out-of-distribution generalization. On real-world benchmarks a 7B model reaches 30.6% accuracy, outperforming the much larger Qwen2.5-VL-32B.
Training Hamiltonian neural nets through implicit symplectic integrators
Hamiltonian Neural Networks embed physical priors by learning a system's Hamiltonian, but faithful long-term simulation needs symplectic integrators, whose implicit forms are costly and make backpropagation through the ODE solver non-trivial. The authors exploit the fact that a symplectic discretization of the adjoint system yields the same sensitivities as backpropagation to train the network efficiently, using a predictor-corrector solver and fixed-point iteration to mitigate the implicit timestepping cost. Experiments on non-separable, chaotic systems show advantages in system identification and energy preservation, and post-processing the learned Hamiltonian via backward-error analysis yields a modified Hamiltonian that better approximates the true one without finer discretization.
Robotics
An RL drone-racing policy generalizes 7.4x better to unseen tracks — without slowing down
Autonomous drone racing demands time-optimal control under persistent actuation saturation, and existing RL policies typically crash when moved to unseen track configurations or sacrifice speed to gain robustness. This framework pairs task-aware switching driven by learning progress with a physically informed procedural track generator to train a fast, robust generalist policy with no test-time adaptation. The authors report strong zero-shot performance across a wide range of unseen real-world racetracks — a 7.4x improvement in generalization over prior approaches — while maintaining competitive racing speed, and validate the result in simulation and the real world, including a vision-based end-to-end setting without explicit state estimation where prior methods fail to generalize.
Predicting grasp feasibility from raw point clouds, faster than sampling planners
Infeasible planning attempts by sampling-based motion planners are costly in clutter, and prior infeasibility-certification methods are limited to low-dimensional configuration spaces and simplified geometry. This work learns motion feasibility directly from raw RGB-D for a 7-DOF manipulator in realistic cluttered scenes and introduces the first large-scale benchmark for the setting — 2.7M grasp-feasibility labels over 88 scanned objects and 190 tabletop scenes. Across MLP, volumetric-CNN, and point-cloud-transformer families under matched training, the best model GRASPFC-PTX reaches AUROC 0.996 on novel objects while predicting far faster than sampling-based planners.
WatchAct: robots still fail to reason over watched human behavior
WatchAct grounds robot manipulation in observed human behavior, pairing each real human-action video and language instruction with an aligned simulator scene and an executable LIBERO task. It spans 3,000 long-horizon instances across 14 tasks in four domains — event grounding, procedural reasoning, implicit intent inference, and episodic reasoning — with a disentangled protocol separating video-to-plan reasoning, policy execution under oracle plans, and full pipeline completion. Current systems remain far from solving it: the best pipeline, Gemini-3.1-Pro with pi_0.5, reaches 16.3% success in simulation and 14.0% on a Franka Research 3, with Gemini-3.1-Pro at 36.8% plan success versus 97.1% for humans.