Robotics
Robot swarms of 512 quadrupeds learn to navigate clutter, then walk it in the real world
A new end-to-end training scheme pairs a high-fidelity, non-differentiable simulator for realistic contact and motion with differentiable surrogate models that supply gradients for navigation and locomotion. This 'asymmetric physics' separation scales vision-based reinforcement learning to swarms of up to 512 quadruped robots that coordinate without communication, centralized planning, or global maps. At deployment each robot acts from a single forward-facing depth camera; the learned policies generalize across forests, bridges, enclosures, narrow passages, and mazes, and transfer zero-shot to six physical quadrupeds in five real-world scenarios, exhibiting emergent behaviors like predictive avoidance, right-side yielding, and bottleneck pausing.
Quantum Computing
Planar fault-tolerant non-Clifford gates on the 2D color code
Bauer and Magdalena de la Fuente present 'twisted color circuits,' a family of scalable circuits for universal quantum computing on planar chips with only nearest-neighbor couplings. The construction achieves fault tolerance for non-Clifford gates on the two-dimensional color code using a 'just-in-time' decoder, addressing a long-standing obstacle to universality on hardware-friendly planar layouts. Published in PRX Quantum.
NSF picks five new National Quantum Virtual Laboratory teams, sharing $20M
The U.S. National Science Foundation has selected five additional teams for its National Quantum Virtual Laboratory, collectively awarding $20 million. The teams will design experimental quantum technologies ranging from networks that ferry fragile quantum information over long distances to sensors that measure faint properties inside single cells, joining four teams chosen earlier.
Quantum tomography and retrodiction shown to be the same inference principle
Quantum tomography (reconstructing states from measurements) and quantum retrodiction (inferring past states from outcomes) are shown to be two faces of one principle: the Petz recovery map associated with a measurement channel is precisely the gradient update of the log-likelihood in maximum-likelihood tomography, so repeated application monotonically increases likelihood. The authors derive a noncommutative generalization for arbitrary quantum channels, establishing a direct bridge between recovery maps, retrodiction, and statistical inference.
Tailored Hilbert spaces pull critical exponents from tiny lattices
A new Hilbert-space tailoring scheme probes quantum critical phenomena on far smaller systems than standard finite-size scaling. Applied to the 1D XY model it yields precise critical points and exponents on 50 unit cells, and it recovers the Berezinskii-Kosterlitz-Thouless transition of the 1D XXZ chain with as few as 12 lattice sites, suggesting an efficient route to universally characterizing quantum phase transitions.
Quantum Sensing
Spin-dependent squeezed states reach optimal displacement sensing in trapped ions
Bond and colleagues demonstrate that spin-dependent squeezed states in trapped-ion platforms enable optimal displacement sensing, pushing precision metrology toward fundamental limits. Published in PRX Quantum.
AI & ML
OpenThoughts-Agent opens a data recipe for general agentic models
The OpenThoughts-Agent project releases a fully open data-curation pipeline for training broadly capable agentic language models, backed by more than 100 controlled ablations on task sources and diversity. Fine-tuning Qwen3-32B on a 100K-example training set yields 44.8% average accuracy across seven agentic benchmarks, a 3.9-point gain over the strongest existing open data agentic model, with the data showing strong scaling across compute-controlled comparisons. Training sets, pipeline, experimental data, and models are publicly released.
Google DeepMind adds computer use to Gemini 3.5 Flash
Google DeepMind announced a computer-use capability in Gemini 3.5 Flash, extending the model to operate graphical user interfaces. The announcement is a single-source vendor post without published benchmark numbers.
OpenAI and Broadcom unveil a custom LLM-inference chip, 'Jalapeño'
OpenAI and Broadcom introduced Jalapeño, a custom silicon chip designed specifically for large-language-model inference, positioned to improve performance, efficiency, and scale across AI systems. The reveal is a vendor announcement without independent benchmarks.
Microsoft's open-source Talos automates rare-disease genomic reanalysis
Microsoft Research detailed Talos, an open-source system that automates iterative reanalysis of genomic data to cut the human-review bottleneck in rare-disease diagnosis. It reportedly recovered 90% of in-scope diagnoses while surfacing only 1.3 candidate variants per patient for expert review.
Robotics
World-model backbones sharpen value estimation for robot manipulation
World Value Model (WVM) replaces the vision-language backbones typical of robotic value models with a world-model foundation better suited to temporal modeling and future planning. It reports state-of-the-art Value-Order Correlation on standard benchmarks and maintains performance on a newly introduced Suboptimal-Value-Bench of 800 human-annotated suboptimal trajectories, improving manipulation across policy-extraction approaches in simulated and real-world deployment.