Quantum Computing
LLM agents bring up a 112-qubit quantum processor on their own
A preprint reports 'Vibe Calibration', an LLM-agent system that distills expert calibration know-how into reusable, auditable 'Skills' (decision trees with parameterized measurements and acceptance criteria). On a 112-qubit frequency-tunable transmon processor it autonomously calibrated 108 of 112 qubits in 4.7 hours, a claimed 4-5x speedup over full manual calibration, and agreed with expert manual results on 14 of 16 qubits in a cross-validated subset. The authors say the core decision logic transferred to other devices with only low-level interface changes. As a single-team preprint the numbers are not yet independently reproduced, but it is a concrete, first-of-kind demonstration sitting squarely at the AI/quantum-hardware boundary.
Quantum Computing
Surface-code hybrid claims 3-4x better qubit efficiency with rare long-range coupling
A quant-ph preprint proposes the Hierarchical Logical Processor (HLP), which concatenates a high-rate quantum CSS code with the rotated surface code and uses elongated 'shuttle bus' patches plus transversal hybrid-unit CNOTs to confine non-local coupling to once every Theta(d0) error-correction rounds. Circuit-level simulations of a [[256,194,4]]-based construction report 3-4x higher qubit efficiency than the standard surface code at a 1e-3 physical error rate, saving 100-200 physical qubits per logical qubit and shortening the logical cycle by 20-30x versus the yoked surface code. Results are simulation-only, but the encoding-efficiency numbers are concrete and address a real qLDPC connectivity bottleneck.
DOE launches 'Quantum Genesis' push for fault-tolerant computing by 2028
The U.S. Department of Energy unveiled 'Quantum Genesis', a new initiative to build and deploy what it calls the world's first fault-tolerant, scientifically relevant quantum computing capability for research by 2028, positioned as a pillar of its broader Genesis program. Reported via a single trade outlet from a DOE press release, so specifics (funding, partners, milestones) remain thin and the timeline is an aspiration rather than a result; still, a national-lab program of this scope is a material policy signal, arriving alongside this week's quantum executive order.
No free lunch: a theorem on why Hilbert-space size alone won't make QML generalize
A quant-ph preprint formalizes supervised quantum learning without an external reference frame and proves that whenever training states fail to span the Hilbert space, every pure state orthogonal to that span must receive an identical prediction, even when those states are perfectly distinguishable given the right measurement. The limitation stems from missing reference information rather than optimization or compute, and the authors show learning generic unstructured concepts needs exponentially many independently oriented training directions. The takeaway: feature maps, measurement bases, Hamiltonians and symmetry priors are the real resources for QML generalization, not Hilbert-space dimension. A theory result, but a clean and substantive one.
QCi closes $73M NHanced Semiconductors acquisition
Quantum Computing Inc. (Nasdaq: QUBT), a quantum-optics and integrated-photonics company, completed its acquisition of NHanced Semiconductors for cash and QCi stock valued at $73.1 million, with up to an additional $72.0 million contingent on milestones. The deal gives QCi advanced semiconductor packaging and interconnect capacity; reported across multiple outlets, though sourced from the company's own release.
Neutral-atom QPU tackles satellite scheduling
A quant-ph preprint formulates Earth-observation satellite fleet scheduling, including satellite agility, as a Maximum Independent Set problem and solves it on a Rydberg-atom QPU via both a native-graph and a QUBO encoding, finding the QUBO route more effective in numerical experiments. It is an exploratory feasibility study on small instances rather than an operational deployment, but a concrete attempt to map a real scheduling task onto analog neutral-atom hardware.
A quantum algorithm for alchemical free-energy calculations
A peer-reviewed npj Quantum Information paper introduces 'Fullqubit alchemist', a quantum algorithm targeting alchemical free-energy calculations, a workhorse of computational chemistry and drug design. Only the title and DOI are available in the feed, so method details and any resource/accuracy claims aren't assessable here; logged as a primary-source quantum-chemistry algorithm result.
Storing and retrieving quantum superchannels for retrospective intervention
Extending probabilistic storage-and-retrieval beyond single unitary channels, this quant-ph preprint considers storing definite-causal unitary superchannels (sequences of unitaries with open intervention slots) in a quantum state and activating their 'retrospective' intervention later. It gives a partial-teleportation protocol optimal for few storage queries and a 'staircase backstitch' protocol that reaches unit success probability asymptotically with more queries, plus a universal inversion protocol. A theoretical contribution to higher-order quantum information.
Universal anticoncentration in noisy random circuits
A PRX Quantum paper provides a universal account of how noisy random quantum circuits anticoncentrate at finite depth, revealing depth-dependent regimes and a form of anticoncentration that is independent of circuit architecture. The result bears on the theory underlying random-circuit sampling and quantum-advantage claims; reported from the journal listing.
Post-Quantum Crypto
Coherent Ising Machine attacks a 40-dimensional LWE instance
A quant-ph preprint proposes CIM-BDD, a hybrid bounded-distance-decoding LWE solver that reduces LWE to a strictly penalty-free QUBO (the cryptographic noise itself becomes the objective) and adds a compact mixed-radix encoder to fit current hardware. It demonstrates Search- and Decision-LWE on a 40-dimensional TU Darmstadt challenge instance on a Coherent Ising Machine. The dimension is far below cryptographically relevant scales, so this is a methodological proof-of-concept for PQC cryptanalysis co-design rather than a threat to deployed lattice schemes.
AI & ML
Model merging scales toward billion-parameter transformers
A cs.AI preprint extends linear mode connectivity to large pretrained transformers by aligning functionally equivalent solutions with parameterized weight transformations and a bidirectional 'dual learning' procedure that drives both models toward a shared linear interpolation path. It reports near-barrier-free connectivity on WikiText for medium-sized language models, ViT-L holding above 69% ImageNet top-1 across the interpolation path, and only small loss barriers for billion-parameter LLMs. Code is released; results are author-reported.
A JEPA for half-hour procedural videos
P-JEPA is a backbone-agnostic joint-embedding predictive architecture for long procedural videos, reducing the problem to a dense frame-aligned action space and predicting pooled masked latent vectors so it can ingest videos over 30 minutes long. Built on features from VJEPA2.1, TSM and I3D, it reports consistent gains in linear separability, streaming inference and temporal action segmentation, plus state-of-the-art EgoExo4D fine-grained action classification using an order of magnitude fewer parameters than LLM-based methods, in real time.
The right activation lets tiny nets learn Game of Life
Reframing the known difficulty of learning Conway's Game of Life from a search problem to a learning problem, this cs.LG preprint shows that alternative activations meaningfully outperform ReLU, with a 2nd-degree polynomial activation consistently learning Life dynamics with or without trained weights. The authors argue for matching inductive biases to the task rather than defaulting to scale, and advocate cellular automata as test domains for interpretable, physics-flavored ML.
Graded per-channel quantization pushes LLM decoding to 2-bit
GRINQH (Graded Input-based Quantization Hierarchy) accelerates the memory-bound LLM decoding stage by using activation magnitudes as an importance proxy to assign weight channels to different precision levels, with a custom GPU kernel and hierarchical nested memory layout to realize the speedups. On Llama3 and Qwen3 it reports beating fixed- and mixed-precision baselines at 3- and 4-bit and enabling effective 2-bit generation, claiming a new quality/speed Pareto frontier. Author-reported.
Robotics
A smarter starting distribution lifts robot flow-matching policies by 23%
Standard flow-matching behavior-cloning policies start denoising from a single isotropic Gaussian, mismatched to the fragmented structure of robot action spaces. LAFM instead learns a library of prior distributions indexed by discrete motion primitives via a latent action model, picking a structurally aligned base distribution per observation. The authors report a 23.4% absolute gain in real-world task success and 10.4% on LIBERO-90, claiming to surpass much larger pretrained vision-language-action models with smaller architectures. Single-group results without external replication, but the benchmark numbers and real-robot deployment are concrete.
Splitting kinematics from physics for robot world models
IOI is an interactive world model that injects analytical kinematic priors (forward kinematics rendered as synchronized orthographic projections) into a learned video generator, freeing the generator to model stochastic physical interactions and avoiding calibration. On the RoboTwin benchmark it reports state-of-the-art simulation fidelity and zero-shot OOD generalization, serves as a policy evaluator whose success rates track ground-truth physics, and trains real-world policies that match teleoperation-demo baselines. Author-reported single-paper results.
Better-grounded skill tokens for long-horizon manipulation
Aligned Refinement Policy (ARP) addresses two weaknesses of discrete-skill imitation learning: weak visual-semantic grounding and quantization precision loss. It adds a contrastive visual-action alignment objective and a lightweight two-step Iterative Residual Head for fine-grained control. The authors report state-of-the-art results on the LIBERO and Meta-World benchmarks and consistent gains on two tasks with a real Kuavo 4 Pro humanoid. Single-group benchmark results.