Disordered quantum model predicts a state between spread and localization
Preprint: A one-dimensional model with long-range hopping predicts a ground state that is neither fully extended nor fully localized.
Search
2,230 results
1261–1280
TDFNet reports strong results on 360-degree image benchmarks
The TDFNet preprint combines three image projections and reports leading scores on several tests, but gives no uncertainty estimates.
AI model learns edge geometry for noisy, small-sample regression
An arXiv version-1 preprint reports Banach-KAN’s median NRMSE of 0.030 on clean tests and sharply different noise responses across the model family.
Cosmology Preprint Finds Early-Universe Results Depend on the Model
An arXiv preprint reports model-dependent constraints from an ISW–tSZ cross-correlation, but current data do not meaningfully constrain scale running.
A function nears Lipschitz regularity; its toral map is not C0-rigid
Preprint: A self-contained proof examines one lacunary skew product and its function h, with Holder regularity below exponent one and conditional deviation bounds.
Unlabeled Chest X-rays May Help Rank AI Models Under Data Shifts
Preprint benchmark finds a label-free score closely matched target model rankings, especially when labeled source data were scarce, but broader testing is still needed.
Flat targets change the limits of rearranged statistical estimates
Preprint: A theoretical study maps distinct limits for rearranged density and copula estimates when the target has flat regions.
Math preprint proves conditional formula for L-function averages
An arXiv preprint gives an explicit main term for a mixed second moment, but only under Selberg’s eigenvalue conjecture and a restricted modulus setup.
Temporary weight-decay pulses flag neural networks before generalization
Preprint: In three algorithmic tasks, temporary changes to weight decay were associated with an orderly signal about when models would generalize, while test accuracy was still at chance.
Bayesian method maps uncertainty when data cannot identify one answer
Preprint proposes turning probability estimates into sets of plausible structural answers, with a diagnostic for model mismatch.
Researchers test 18-question measure of generative AI knowledge
A preprint study of 7,432 Estonian high school students found acceptable test fit but less precise measurement among lower-scoring students.
Moore-Read signature survives a non-Hermitian test, then fails
Preprint: A finite-size model kept its Moore-Read counting across a limited range of nonreciprocity, but the signal later broke down.
Six-model benchmark linked SkillShield to lower attack scores
Preprint: In controlled coding-agent benchmarks, SkillShield was associated with lower observed attack and malware-generation scores, but attacks still succeeded.
AI models showed different playing strengths across languages
Preprint finds uneven performance across eight language interfaces, even when the game setup was held constant.
5G testbed shows srsRAN edge in throughput and app performance
Preprint: A small laboratory comparison found similar connection setup times but sharper differences in data and application workloads.
Shortlist-first LLM judging exceeds its 90% Guarantee Success Rate target
A preprint reports results from four preference benchmarks, with two-stage systems exceeding a 90% Guarantee Success Rate target in tested settings.
Quantum ODE method removes explicit evolution-time dependence
Preprint reports a query bound for weakly nonlinear dissipative equations, while tests show convergence can occur beyond its formal regime.
Temperature model leads Southeast China benchmark with flexible queries
Preprint: CSTF led the reported integer-lead comparison, while half-hour and multi-grid outputs remained diagnostic.
A scalar-field model reports expansion in selected anti-de Sitter cases
Preprint: Numerical and analytical calculations report continued expansion and curvature approaching zero in selected negative-vacuum-energy cases, while another modeled case reaches a singularity.
AI writing scores higher in tests of paper trajectories
Preprint: A single-model study reports broader writing-score differences linked to reconstructed paper-writing data, while reasoning stayed near direct fine-tuning.