STRATA ranked stocks better than six matched sequence models after style adjustment, but the apparent trading payoff came mainly from an overnight move that preceded the executable session.
A computational study reported faster convergence for a new graphene phonon-transport solver, with the biggest gains appearing as conventional iterations slowed at larger length scales.
A mathematical preprint develops an unbiased idealized estimator and a two-stage randomized variant, then tests truncated versions on kernels and spectral filters.
A modeling study asks whether a fixed AED supply is better used in stationary sites, mobile vehicles or a hybrid mix, with an analytical benchmark showing the second scenario’s response-distance measure is no greater than the first.
A numerical study reports that a hybrid solver can reproduce a full finite-element solution exactly in selected cases, but its accuracy depends sharply on how its reduced model is built and which load it sees.
An arXiv preprint reports that an extended wave-particle method matched DSMC in cylinder and shock simulations, while its chemical and multispecies scope remains limited.
A computational study reports close agreement between AC/DC and projection in several tests, along with discrete conservation properties and an estimated runtime of about 1.3 to 1.4 times a hydrostatic step. A deep-convection case became unstable at lower tested parameter values.
A mathematical methods paper reports conditional fast-convergence guarantees for an inertial primal-dual system and its discrete algorithm, while experiments on logistic regression and least squares favor the proposed method in the tested setups.
A new adaptive method for nonconvex constrained optimization adjusts its penalty and stochastic inner solves, with theory and tests pointing to gains that remain conditional on the study's assumptions and setup.
A shallow recurrent decoder reconstructed simulated three-dimensional liquid-metal flow fields from sparse, noisy temperature measurements, producing estimates in less than a second after training. Errors were generally low, but some high-field cases produced larger pressure and velocity errors, and the test was limited to a simplified computational model.
A numerical study reports that a two-channel finite-element domain-decomposition method needed far fewer GMRES iterations than a Robin baseline across several electromagnetic models while preserving the tested solution accuracy.
A tabular foundation model called TFM4POP reportedly beat the tested baselines at forecasting future growth in social-media cascades. The results come from an arXiv preprint and do not show that the model changes how information spreads.
A preprint examines whether AI agents can reach practical agreement before exhausting their computational budget. An illustrative five-agent simulation and live deployments using three named models put adaptive communication between ring-only and complete-only strategies, while a tight-budget run failed and one live run showed disagreement rising.
A computational preprint reports a modeled capital cost of USD 727 million for an optimized 1,000 MW, 12-hour pumped-storage layout, compared with USD 774 million for an initial heuristic design; the evidence is limited to a Rio de Janeiro pre-feasibility case.
In a simulation of a 4,000-vehicle virtual power plant, the proposed I=2 model's normalized error was 7.5% in the coupled test, versus 13.6% for outer approximation.
A topology-aware control framework treats global circulation and magnetic flux as explicit states, showing that actuator rank can leave some topological modes unreachable.
The RWA-PoB prototype linked institutional signatures, backing thresholds and redemption liquidity to token issuance and settlement, but its tests did not verify the truth or completeness of off-chain assets.
A preprint reports a full endpoint match and lower accumulated costs for an interaction-aware controller in Cucker-Smale simulations, while a Morse example produced a mixed cost ranking.
A mathematical preprint proposes an accelerated gradient-flow system that uses gradient history and coordinatewise feedback to control convergence and select among multiple minimizers.
A preprint model that limits cross-market information to what is available at each forecast origin recorded lower MSE and MAE than a one-market-at-a-time HAR benchmark across all eight markets at one- and five-day horizons, and in seven markets at 22 days.
In numerical tests on known reaction networks, a data-trained propagator retained important SSA statistics while using substantially fewer simulation steps than SSA and adaptive tau-leaping.
A modeling study proposes a global, six-zone climate classification for photovoltaic modules by combining simulated annual energy yield with simulated module lifetime. It finds different top-ranked climate inputs for the two outcomes, while small changes to feature-importance scores can alter the cluster structure.
Across 46 mostly rural communities in 28 countries, wealthier sharing units generally had more support ties and connections to relatively wealthier units. Communities with greater wealth inequality also showed more separation between wealth groups in their networks, but the cross-sectional evidence cannot establish cause and effect.
A new traffic-forecasting model reported the lowest errors in tests with complete data and remained first when 80% of entries were randomly masked. It also ran faster than the compared methods, though the evidence comes from two datasets and incomplete-input comparisons covered only methods that could process missing data.
A 2026 arXiv preprint reports lower time per iteration for HSLM in three synthetic neural-network comparisons, alongside similar final training errors and conditional convergence results.
A methods preprint reports nearly constant MINRES counts in a 3D HKNZ test, fewer AMG than ILU iterations in a 3D GLSS test, and failure without preconditioning in a nonsmooth run.
A numerical study reports that a factorized optimization method reached demanding SLOPE and group-lasso targets in fewer iterations than tested baselines, while the evidence remains limited to small computational benchmarks.
A mathematical preprint reports explicit schemes that keep numerical solutions for stochastic differential equations inside a prescribed hypercube while retaining stated strong and weak convergence rates under the paper's assumptions.
A mathematical preprint reports a scalar SDE scheme that preserves nonnegativity at every grid point for any time-step and achieves first-order strong convergence when steps are sufficiently small.
A theoretical preprint shows that, on fixed grids, endpoint-graph connectivity can characterize unique polynomial reconstruction and support exact formulas for matrix conditioning and diagonalization.
A new preprint derives high-probability error bounds for recovering an unknown conductivity from noisy measurements at finitely many scattered points, then tests the framework in synthetic finite-element examples.
A theoretical preprint sets out conditions under which optimization values and solution sets stay stable as objectives and feasible choices are sequentially perturbed. It also applies the framework to generalized Nash equilibria and finite-horizon dynamic programming.
A mathematical preprint reports smaller GMRES iteration counts for block-preconditioned solves, while leaving a key variable-coefficient eigenvalue result open.
A preprint comparison of four penalties on measured KIT4 EIT data found more spatially coherent localization with smoothed TV and Huber-TV, but none of the four satisfactorily resolved the final difficult two-inclusion configuration.
A neural-network surrogate retained lower-contrast accuracy while online multiscale Darcy-flow calculations were much faster, but high-contrast cases exposed major modeling errors.
A computational preprint compares three inference protocols across protein-model benchmarks and reports different results alongside a trade-off in inference cost.
A mathematical analysis puts two-sided bounds on worst-case sampling and links them to the conditioning and certified iteration counts of weighted least-squares solvers.