Preprint tests GPU safeguards for LLM user-data isolation
GIFT uses per-user encryption and GPU flow tracking, and its A800 ShareGPT test reported up to 5% overhead in maximum request rate.
Search
2,230 results
941–960
Beam forecasting system reports low outage, but transfer remains weak
This preprint reports low outage in its main test, while held-out scenarios show weaker transfer.
Document grids score higher with far lower search index costs
Preprint: A benchmark spanning seven English domains found higher document-ranking scores and a much smaller index than page-by-page retrieval.
Study finds no stationary source-free fermion condensate at black hole horizon
Preprint: A separated Kerr analysis says horizon flux excludes a non-vanishing corotating solution and predicts a filled, oblate fermion density.
A drone-image test reports better mapping of tower equipment
Preprint: A saliency-and-depth layer accompanied higher reported scores on 340 drone images from two towers, but the test was narrow.
Simulations show different rare-earth peak patterns across nuclear models
A preprint compares six nuclear level-density models across three parameterized neutron-star-merger outflow scenarios.
Astronomers identify three likely wind-accreting pre-polars
Preprint: Optical polarimetry reported magnetic white dwarfs in at least four of nine selected candidates, while several systems remained unresolved.
Plasma scheme shows higher-order limit in one low-beta test
A preprint tests the method on seven numerical benchmarks, reporting instability growth and expected energy-error scaling, while fourth- and sixth-order versions fail in one extremely low-beta blast-wave test.
Four-agent code system reports higher combined benchmark scores
Preprint: MACGen reported higher combined function-and-security scores than Direct prompting on two benchmark suites; on CWEval, its average function score fell slightly.
Model links temperature to spin dynamics in a ferromagnet
Preprint: Mathematical calculations propose a temperature-dependent non-Abelian gauge potential, without experimental testing.
Preprint maps tractable and hard cases in election attacks
An arXiv preprint dated 26 Aug 2026 uses a divisible-population model and linear programming to compare formal election attacks.
Two-Center Gravity Saddles Match D1-D5 Index Terms
Preprint: An analytic comparison reports an infinite two-center match to the D1-D5 Farey-tail expansion, but no match for the tested black-ring and black-lens sector.
Simulation suggests shared O-RAN can support 50 MCX users
Preprint: An E2-conditioned model kept several public-safety service requirements within bounds in simulated shared networks, including under O-DU degradation.
Neutron-Diffusion Preprint Reports Faster PINN Calculations
The approach uses one low-resolution calculation to start both a flux network and a key model parameter, but evidence comes from four numerical benchmarks.
Preprint Links Special Points on Quartic Surfaces to K3 Geometry
A theorem gives an algebraic test for quasi-Galois points and counts 28 on the Fermat quartic.
Small Subtrees Can Rebuild Two Evolutionary Trees, But Not Always
Preprint: A mathematical analysis finds that three-leaf clues can remain ambiguous, while near-complete clues identify the two source trees in broad size ranges.
Preprint establishes universal formulas for Hilbert schemes on surfaces
An arXiv version dated 26 Aug 2026 develops the κ-class framework for curve and surface families and gives explicit codimension-one surface formulas.
Machine learning maps six climate zones for solar modules
Preprint analysis combines simulated energy yield and module lifetime in a global model, but its climate groupings shift under small changes to feature-importance scores.
Climate models classify 18 tidally locked exoplanets as potentially habitable
Preprint simulations suggest stellar flux matters more than rotation, while pressure reshapes the hottest modeled worlds.
New math framework links arithmetic volume to equilibrium measures
Preprint: A theoretical analysis says theta and Bergman-type measures agree in key settings, while a broader convergence claim remains conjectural.