Quantum Computing
112-qubit IBM run tracks gauge-theory string dynamics from quench to thermalization
Researchers mapped a (2+1)D U(1) quantum link model onto IBM's heavy-hexagonal superconducting processor, using up to 112 qubits to probe transverse quantum fluctuations of a confining string after a quench. Error-mitigated estimators matched tensor-network calculations at short times and thermal averages at long times, holding up even near the phase transition where fluctuations span both spatial dimensions of the lattice.
Quantum Computing
Master-equation 'digital twin' reveals surface-code noise that Pauli models miss
QMCtwin simulates the open-system dynamics of a full distance-7 rotated surface-code syndrome-extraction round on 97 physical qubits, including relaxation, dephasing, coherent gate miscalibration and residual ZZ crosstalk. It predicts syndrome-extraction biases and syndrome–logical-parity correlations that are absent or strongly suppressed in standard Pauli-twirled Clifford descriptions, pointing toward more accurate decoder-facing noise models.
Pipelining and speculation cut fault-tolerant quantum scheduling steps by up to 40%
By splitting each logical operation into control, execute and decode stages and letting successor operations begin before predecessors finish decoding, the framework converts subsystem idle time into useful computation. Across benchmarks, aggressive speculation reduces pipeline steps 20-40% despite occasional rollbacks, while also improving load balancing across the heterogeneous fault-tolerant stack.
Linear-programming hierarchy makes large-scale quantum steering tests tractable
Published in Quantum, the method recasts the exponentially-scaling semidefinite program for joint measurability and unsteerability into a converging hierarchy of linear programs that grows polynomially in the number of measurements. For qubits it bounds incompatibility robustness for sets of several hundred measurements quickly on a standard laptop, and extends — more loosely — to qutrits and to constructing local-hidden-state models.
OQC to build €92M superconducting-quantum manufacturing hub in Barcelona
Oxford Quantum Circuits said its first EU facility, the OQC Global Quantum Development & Manufacturing Centre, will become its primary hub for designing and industrializing superconducting quantum hardware. The company put the project at €92 million ($98M USD); the figure comes from a single trade report and is not yet independently corroborated.
Quantum Comms
Quantum ring all-reduce halves communication and adds provable privacy to distributed training
The proposed primitive cuts per-link online communication by a provably optimal factor of two via pre-shared entanglement and superdense coding, without changing the learning model or gradient computation. It also enables composable ε-secure aggregation that is information-theoretically impossible for any classical protocol, at a 2× overhead in GHZ copies, with further quantum separations shown for gradient-conflict detection.
AI & ML
Chain-of-thought transformers can run Word RAM algorithms at near-optimal overhead
The work shows finite-precision CoT transformers — including continuous-CoT and transformer-over-linear-RNN hybrids — can simulate any Word RAM algorithm (e.g. sorting in O(n log n), Dijkstra in O(E + V log V)) up to poly-logarithmic overhead, dropping to logarithmic for multiplication-free flat instruction sets. This sharpens the theory of why reasoning models compute efficiently, contrasting with the quadratic cost of Turing-machine-based CoT simulations.
RL framework trains LLM agents to 'connect the dots' across long task sequences
CoD uses a GRPO-style RL algorithm with fine-grained credit assignment over long rollouts that interleave task-solving and context-updating, targeting continual self-improvement rather than single tasks. Proof-of-concept experiments show out-of-distribution generalization within and across training domains, with implementations released.
Agentic symbolic search derives analytical PDE forms beyond hand-crafted solutions
ASYS injects PDE theory and accumulated search experience as inductive bias, evolving differentiable symbolic programs whose continuous parameters are fit by gradient descent. Across five problems it recovers known analytical forms and produces novel ones — including a geometric interface formula for Allen–Cahn dynamics and a nine-parameter contraction law for Keller–Segel blow-up — in settings with no prior closed-form description.
Robotics
Training-free pruning halves VLA model depth with no loss across 10 real-robot tasks
Using a single forward pass and Centered Kernel Alignment to identify redundant 'twin' layers, the method permanently compresses both the VLM backbone and control-policy head by up to 50% in depth. Validated across three simulation benchmarks and 10 real-world tasks on four embodiments, it cuts downstream fine-tuning time 40-50% and real-time inference up to 30% while matching or exceeding full-scale base models.
Robot agents that 'play' before working build transferable skills
In Playful Agentic Robot Learning, an embodied coding agent proposes and executes exploratory tasks during a play phase, distilling successful executions into a persistent code-skill library reused at test time. The approach yields 20.6- and 17.0-point gains over no-play baselines on LIBERO-PRO and MolmoSpaces, and the learned skills transfer to other code-as-policy agents without fine-tuning.
MemoryWAM gives robot world-models long-range memory without the latency cost
MemoryWAM combines recent frames, event-boundary anchor frames and compact gist tokens with a tailored attention mechanism to retrieve both short- and long-term context. On long-horizon, memory-dependent manipulation tasks in simulation and the real world it outperforms strong VLA and world-action-model baselines while reducing inference latency and GPU memory use.