Preprint finds AI self-training gains can be measurement artifacts
An audit found apparent changes in an unchanged model, while external distillation reached more low-base problems than the tested self-training methods.
1981–2000
An audit found apparent changes in an unchanged model, while external distillation reached more low-base problems than the tested self-training methods.
A theoretical study reports quantum-dominated fluctuations even when the defining frame is strongly dissipative and the model’s temperature exceeds the Hubble scale.
Two theoretical approaches give cross-section differences of about 2% to 7% at selected settings, while some radiation-related spin comparisons are far larger.
The method keeps competing dictionary explanations in play and reports only the finest physical conclusion they jointly support.
Laser-driven plasma measurements and models point to lower-hybrid drift instability, while key details remain unresolved.
Simulations and early simulator feedback point to stable flight across phases, but the evidence does not yet show lower workload or real-aircraft performance.
A mathematical method uses a spherical reference wave to estimate a radiation solution’s far-field pattern from total-field intensities, while leaving finite-distance performance untested.
The formal method continues eigenfunctions beyond the boundary, but its numerical accuracy and long-term localization remain untested.
A radial cross-check found that the main SDSS transverse compilations did not agree, while one showed far more scatter than its quoted errors suggested.
Theoretical results cover parallel queries, fixed-round adaptivity and selected two-stage designs, while leaving fully adaptive algorithms unresolved.
A mathematical model finds a closed-form cancellation setting that reaches the best possible worst-case voltage result when specific feasibility conditions hold.
A proof-based study answers several questions about sums, division, commutativity and representation within the class of linear orders.
A two-dimensional potassium condensate showed scale-dependent spin fluctuations matching a zero-temperature prediction, with detector losses and an adiabaticity limit still in play.
New theorems set upper limits for graph factors and multigraph matchings, while a substantial gap remains for general connected graphs.
All four fitted LLM weights fell to zero in a next-day market-risk study; a headline-count feature helped one SPY variance test but produced mixed results elsewhere.
The system combines 3D image context with confidence-aware training and reports benchmark results on 976 patients, but clinical outcomes were not assessed.
Simulations show reflected and transmitted signals can differ from the isolated microscopic response, complicating efforts to infer electronic structure from transmitted light.
A reconstruction from young clusters highlights Orion, Vela, Sco-Cen and Cepheus, while showing that its rates cover only the cluster-traced part of local activity.
DreamHand led 45 of 48 primary benchmark comparisons, while its out-of-sight test measured wrist-aligned continuity rather than absolute hand position.
A theorem-based study links equivariant operators on noncompact symmetric spaces to smooth Harish-Chandra multipliers under precise kernel conditions.