Weekly Recap

Standard blind-spot denoising breaks down on medical ultrasound because speckle is spatially correlated coherent scattering rather than i.i.d. noise—a failure mode resolved here by shifting from pixel-wise blind spots to block-wise contextual inpainting.

Non-rigid 3D surface matching can finally scale without sacrificing intrinsic geometry, breaking the cubic-time barrier that previously made geodesic optimal transport intractable.

High-fidelity video super-resolution doesn't require explicit optical flow or latent compression: synchronized pixel-space posterior sampling over localized moving windows scales to arbitrary video lengths without temporal drift.

Multi-subject image generation doesn't have to hit a linear memory wall: packing references into a single unified canvas enables constant-memory composition without sacrificing fine-grained identity.

Off-the-shelf vision foundation models fail to transfer dense representations to scientific imagery, but scaling in-domain self-supervised pretraining across 10 million tiles delivers an immediate 7-point Dice improvement under severe annotation limits.

Standard blind-spot denoising breaks down on medical ultrasound because speckle is spatially correlated coherent scattering rather than i.i.d. noise—a failure mode resolved here by shifting from pixel-wise blind spots to block-wise contextual inpainting.

Non-rigid 3D surface matching can finally scale without sacrificing intrinsic geometry, breaking the cubic-time barrier that previously made geodesic optimal transport intractable.

High-fidelity video super-resolution doesn't require explicit optical flow or latent compression: synchronized pixel-space posterior sampling over localized moving windows scales to arbitrary video lengths without temporal drift.

Multi-subject image generation doesn't have to hit a linear memory wall: packing references into a single unified canvas enables constant-memory composition without sacrificing fine-grained identity.

Off-the-shelf vision foundation models fail to transfer dense representations to scientific imagery, but scaling in-domain self-supervised pretraining across 10 million tiles delivers an immediate 7-point Dice improvement under severe annotation limits.

Pretrained discrete diffusion priors can serve as plug-and-play mutual information engines for optimal sensor placement, slashing reconstruction errors by 8× at a 10% sampling budget without model retraining.

Optical flow in extreme visual environments becomes reliable when anchored by fluid mechanics, turning glare-distorted video of molten metal into fully resolved, Navier-Stokes-compliant velocity fields.

A residual association between longer DTI and improved survival persisted after IPI adjustment, suggesting DTI may be an independent prognostic factor in less aggressive disease.

Thermal lattice vibrations alone can drive nanosecond magnetization switching, bypassing electronic spin currents to manipulate magnetic states directly via polarized phonons.

Swapping a single boron atom for carbon breaks traditional electronic-thermal trade-offs, simultaneously opening a 0.33 eV semiconductor bandgap and boosting lattice thermal conductivity by 250%.

Soft robots no longer need to sacrifice payload for compliance: these active 3D-woven textiles carry 70 times their self-weight and survive extreme crush cycles without losing functional locomotion.

Scalable freestanding oxide membranes no longer depend on slow, boutique laser deposition: simply tuning the sputtering atmosphere enables phase-selective growth of water-soluble sacrificial layers from a single target.

Precursor-level atomic doping unlocks previously inaccessible interlayer spacing in MXenes, driving lithium diffusion rates up to $10^{-9}\text{ cm}^2/\text{s}$ and reaching near-4 kW/kg power densities without sacrificing cyclic stability.

Coupled-cluster accuracy no longer requires steep global scaling: combining density fitting with distinguishable cluster approximations slashes quantum embedding costs down to $\mathcal{O}(N^4)$ without sacrificing chemical accuracy.

Mechanical strain alone can toggle and flip the sign of the anomalous Hall effect in altermagnets, unlocking zero-net-magnetization topological switching without external magnetic fields.

Key contribution not extracted.

Antiferromagnetic order triggers significantly steeper thermal conductivity drops than ferromagnetic order because matched excitation scales unlock an expansive phonon-magnon scattering phase space.

Strongly correlated multi-orbital lattice systems can now be simulated across finite temperatures without hitting the prohibitive compute wall of standard quantum embedding methods.

Swarms of microscopic nematodes can spontaneously synchronize to sculpt and drive flexible boundaries into stable, life-like protrusions that rotate, merge, and split without centralized control.

Ultra-thin disc architectures bypass traditional thermal cross-talk and energy scaling limits in phase-change memory, providing a blueprint for dense, high-endurance analog compute arrays.

Mass and size mismatches are not required to slash thermal conductivity: dopant-induced lattice softening alone drives a 71% collapse in $\beta$-FeSi$_2$'s $\kappa_{\text{lat}}$ by weakening atomic force constants.

Decades of overlapping, ambiguously named chemical theories can now be mechanically audited: Lean 4 formalization resolves all 136 pairwise relations across photochemistry while settling disputed classical boundaries with exact if-and-only-if proofs.

Standard silicon fab lines can now reliably achieve 6 nm single-crystal fin profiles with minimal tapering, eliminating a primary barrier to uniform, high-coherence Josephson junctions for quantum processors.

Lateral semiconductor heterojunctions can dynamically switch 1D excitons from non-dipolar quadrupoles into polarized dipoles under sub-1 V/μm electric fields, decoupling Stark sensitivity from binding energy.

Single-layer attention's mathematical inability to form induction heads is merely a filter-order gap that a parameter-free, inference-friendly symplectic shear resolves in just $6d$ FLOPs per token.

Membrane budding instability hinges on transient non-equilibrium dynamics rather than static interfacial mechanics, governed by a sharp crossover from $t^{1/3}$ to $t^{1/2}$ domain growth coupled to a rapidly evolving line tension.

Midday photosynthetic gas bubbles more than triple the destructive standoff distance of reef-fishing explosives by driving canopy acoustic impedance low enough to trigger catastrophic tensile spalling.

Quantum simulations of nonlinear fluids no longer face an exponential postselection bottleneck: mapping lattice Boltzmann dynamics to Hermite moments unlocks polynomial-overhead coherent evolution over a hundred steps.

Neural traveling waves do not gradually fade against spatial heterogeneity; instead, they abruptly collapse via fold bifurcations governed by an exact bottleneck speed $\sqrt{c_1 c_2}$ under slow modulations and linear frequency scaling under fast ripples.

Most single-cell perturbation models only appear accurate because they memorize generic background responses—their benchmark scores remain high even when gene labels are completely shuffled.

Macroscopic human brain recordings can be decoded as continuous state-space trajectories just like invasive single-unit populations, unlocking latent dynamical manifold analysis for non-invasive electrophysiology.

Neural time-series residuals consistently violate Gaussianity, but framing effective connectivity as a min-max channel capacity game over normalizing flows reliably uncovers directed causal networks where Granger causality and linear-Gaussian estimators fail.

Distilling neural likelihoods directly for parameter recovery yields compact, closed-form equations that evaluate over 100x faster than even analytical ground truth with negligible precision loss.

Memory allocation in the hippocampus does not scale linearly with experience; instead, it undergoes a discontinuous phase transition that dedicates the most neural capacity to moderately frequent contexts rather than dominant ones.

Under uniform carrying-capacity constraints, individual fitness has zero long-term impact on a population's equilibrium: exploratory mutation dynamics completely wash out natural selection.

Perturbative moment closure successfully collapses multiscale transport PDEs of cell maturation into low-dimensional ODEs without sacrificing non-linear biological selection dynamics or systemic feedback loops.

Quantitative keyword temporal tracking confirms this evolution: traditional binary algorithms display earlier average publication timelines (2019.0), whereas deep learning, Convolutional Neural Networks (CNNs), and multi-signal predictive models dominate recent outputs (2023.8–2024.5), demonstrating a quantitative shift from isolated binary classifiers toward continuous, explainable, and clinically viable multimodal monitoring systems.

Daily weight monitoring and optimal fluid management may reduce the risk of OH development and lead to improved patient care and quality of life.

Frontier LLMs struggle to replicate expert consensus in complex, longitudinal cancer treatment discussions, scoring below expert equivalence on a new real-world benchmark.

HeteroFold unlocks $10.7\times$ faster cross-family KV cache transfer for LLMs, eliminating receiver prefill and matching text-based communication performance.

Control accent strength in zero-shot TTS with a single model and inference-time weight, outperforming cascaded systems on out-of-domain accents without sacrificing speaker identity.

Agents can achieve up to 5.8% higher success rates by learning a reusable "experience tree" directly from data, bypassing LLM-generated skills and enabling true generalization.

It is argued that improvements in editing efficiency alone will be insufficient; clinical success will require the simultaneous achievement of tumor‐selective delivery, genomic integrity, and acceptable long‐term safety, and the continued development of novel approaches should be promoted to realize the full potential of CRISPR‐Cas9 technology in oncology and other therapeutic applications.

DMT seems to be a promising low-threshold and cost-effective therapy approach that can improve the mental health and wellbeing of refugee women of diverse cultural backgrounds suffering from PTSD symptoms.

The development of a language-based assessment model that assesses the need for mental health support based on probed natural language and validates it against best-estimate assessments from multiple experienced psychotherapists is demonstrated to be a practical tool for early-stage mental health screening in both clinical and self-guided screening contexts.

ZooWork-ShopRanker, a family of e-commerce rerankers aligned to judge-labeled shopping preference, and ShopRank-Bench, a contamination-limited benchmark of ~10,000 private-traffic preference pairs in both text formats, tiered by how many judge families committed to each label.

This technical report details the infrastructure behind Game Arena and describes the three pilot game environments: Chess, Poker, and Werewolf, enabling a systematic study of models's strategic planning, adaptation, and robustness under uncertainty.

Trapped interstitial air transforms ordinary paper stacks into resonant acoustic trampolines, triggering a massive rebound peak whenever acoustic wave transit times match impact duration.

Standard cancer histology encodes enough latent signal to reconstruct genome-wide mutational signature distributions without sequencing, consistently beating tissue-of-origin priors across 19 distinct malignancies.

Audio-driven synthetic lip movements alone can bootstrap audiovisual speech recognition baselines and slash word error rates by up to 16.2% when used as augmentation.

Expensive inner-loop solvers in model-based inverse reconstruction can be entirely replaced by a neural network trained on pure Gaussian noise, delivering an 18x speedup without sacrificing analytic precision.

Clinical ultrasound AI usually collapses without pre-isolated views, but combining self-supervised video pre-training with gated transformer MIL delivers a 0.985 AUC directly from uncurated, whole-study examinations.

Local pretraining objectives don't just dominate fine-grained pathology detection—they unexpectedly rival global contrastive models on high-level report generation and QA, upending assumptions about cross-granularity transfer.

Standard latent diffusion guidance provably samples from the wrong posterior by assuming isotropic uncertainty across latent dimensions—an error resolved by learning time- and input-dependent directional covariances.

Tuning as few as 1,000 dormant weights on a single labeled slice delivers a 0.5 dB PSNR boost in MRI reconstruction, paired with dynamic output capping that provably bounds drift under domain shift without ground-truth references.

The exact same pruning mask can invert from best- to worst-performing depending purely on surrounding weight context, exposing how naive sparsity metrics fail to predict real model degradation in dense vision architectures.

Polarization vision can bypass its chronic data-scarcity bottleneck by decoupling Stokes parameters and borrowing pretrained RGB restoration priors via cross-domain feature distillation.

Lower phase residual does not guarantee structural accuracy, but pairing untrained neural priors with explicit projection constraints cuts magnitude-free phase retrieval error by 29.3% without requiring any training data.

Real quantum hardware can scale past toy benchmarks to high-dimensional 3D medical imaging by restricting physical QPU execution strictly to forward diffusion dynamics.

Specialized medical vision tools fail to boost multimodal LLM diagnostic accuracy on their own—performance gains only emerge when the agent autonomously converts prior trajectory failures into reusable procedural skills.

Global ischemic heart disease prevalence doubled over three decades despite falling age-standardized mortality, revealing a demographic wave that will drive absolute deaths higher through 2035 regardless of clinical treatment improvements.

A review of advanced pulmonary drug delivery systems, with emphasis on the anatomical, physiological, pharmaceutical, and device-related factors that influence aerosol deposition and therapeutic performance, examines key mechanisms of particle deposition such as inertial impaction, gravitational sedimentation, Brownian diffusion, interception, and electrostatic interactions.

Amortizing quantum physical constraints into a one-step flow model via surrogate RL nearly doubles transition-state recovery on unseen reactions while running 30× faster than guided diffusion baselines.

Standard uncertainty quantification fails downstream in atomistic ML because it destroys physical error cancellations, a blind spot resolved by multivariate conformal methods that explicitly preserve cross-property correlations.

Perturbing a single edge in a complex transport network can now be predicted in exact closed form because local dissipation share is mathematically identical to global conductance sensitivity.

Cheap hybrid-quality fundamental gaps remain an illusion for newly touted meta-GGAs like LAK on molecular systems, where explicit derivative discontinuity corrections via NCAPR GGA quietly outperform even HSE06 and PBE0.

Black-box classifiers fail under experimental noise, but coupling latent retrieval with attribution-guided feature attribution recovers expert-level crystal phase identification at 95.8% space group accuracy.

Coupled nuclear-electronic simulations at non-zero nuclear momentum no longer suffer from extreme basis-set sensitivity, thanks to an atomic Becke weight partitioning that dramatically stabilizes phase space electronic structure theory.

Interfacial mechanical coupling slashes ferroelectric switching fields by up to 80% in wurtzite oxide stacks, proving that stacking architecture alone can overcome the core breakdown-voltage bottleneck of emerging non-volatile memory.

Heterogeneous $N_2$ bond activation can be compressed to just 16 qubits with sub-0.25 kcal/mol chemical accuracy, proving that realistic transition-metal surface catalysis is tractable for near-term quantum eigensolvers.

Circumventing the overcomplete basis bottleneck of standard atomic cluster expansions, direct Schur-Weyl coupling of $S_N \times SO(3)$ representations establishes a new speed-accuracy Pareto front across atomistic and electronic-structure modeling.

Forty-nine newly discovered $p$-type transparent conductors upend decades-old delafossite design rules, proving that tetrahedral Cu coordination and iodide hybridization unlock ultra-low hole effective masses without sacrificing optical transparency.

Geometric frustration in rare-earth bilayers unlocks field-induced $1/2$ magnetization plateaus and a five-phase magnetic landscape in $\text{HoOSe}$, proving that chemical tuning can cleanly switch between complex quantum spin states.

Bypassing intractable force calculations entirely, score matching scales bottom-up coarse-graining to thousands of atoms per interaction site with a fraction of the standard training data.

Cheap second-order perturbation theory unexpectedly matches the accuracy of fully relativistic four-component benchmarks for spin-orbit coupling, while outperforming its own third-order counterpart on heavy-element complexes.

Saturation magnetization typically stays fixed under mechanical deformation, but Al-rich CoFeAl defies conventional magnetoelastic limits by letting mechanical strain directly modulate its intrinsic magnetic moments via sublattice compensation.

Swapping adjacent heteroatoms in identical five-membered rings collapses ring-opening times from 290 fs to 40 fs, revealing that deceptively similar diffraction signatures can conceal entirely divergent non-adiabatic reaction pathways.

Lattice tuning in high-entropy oxides is not dictated by cobalt valence alone: growth temperature-dependent Jahn-Teller distortions from copper drive a massive 2.6% swing in lattice parameters.

Embedding neural networks directly into coupled differential rate equations consistently outperforms standard deep learning architectures on sparse, out-of-distribution physical sintering trajectories.

Sampling rare vacancy fluctuations is completely unnecessary for multicomponent alloy diffusion: a single-vacancy partition function truncation slashes Monte Carlo compute by 90% while analytically isolating the true vacancy tracer diffusion coefficient.

Replacing rigid quantum chemistry basis sets with freely optimizable 3D Gaussian clouds slashes parameter counts and scales triple-zeta DFT to over 2,700 atoms on just four GPUs.

Identical high-spin electron configurations can span the spectrum from high-temperature altermagnetism to frustrated quantum disorder purely through chemical-pressure-driven exchange pathway reconstruction.

Monolayer antimonene eliminates three-dimensional lattice clutter to isolate a minimal multi-orbital kagome system, revealing a giant gap-opening breathing instability and hidden orbital-resolved topological obstruction.

Standard electronic friction approximations catastrophically break down under nonequilibrium conditions, but capturing memory kernels via Markovian embedding stabilizes molecular dynamics to match exact quantum benchmarks.

Off-the-shelf equivariant machine learning potentials can reliably navigate rugged potential-energy landscapes to uncover novel, dynamically stable transition-metal cluster architectures that standard quantum chemistry methods struggle to sample tractably.

Gapless spin waves dominate the AC dielectric response in cycloidal antiferromagnets, revealing a strict directional asymmetry in how electric fields manipulate noncollinear spin textures for spintronic switching.

Ground-state hyperfine strain shifts—long obscured by conventional spectroscopy—can now be directly mapped by synchronizing dynamical decoupling with coherent AC strain drives in rare-earth crystals.

Ground-state molecular decay takes a measurable 2 to 3 picoseconds to achieve statistical chaos, directly imaged by projecting evolving phase-space wavepackets into Coulomb explosion energy spectra.

Chemical defects in polymer insulators do not degrade performance uniformly: while ketone groups drive peak dielectric loss, alcohol defects exhibit negative dipolar cross-correlations with the host polyethylene matrix that actively dampen their dielectric impact.

Instantaneous atomistic snapshots discard critical dynamical context; restoring this hidden state through multi-scale temporal memory enables neural world models to guide physical KMC simulations and transfer zero-shot across unseen material regimes.

Nuclear clock technology can double as an atomic-scale sensor: driving chiral phonons with circularly polarized THz pulses brings the elusive phonomagnetic field within reach of shot-noise-limited Thorium-229 spectroscopy.

Plastic scintillators no longer have to sacrifice photon yield for speed: coupling high-Z sensitizers with plasmonic nanoantennas delivers 6.1 photons/keV at record-breaking 150 ps decay lifetimes.