Curated AI-image study reports 18 defect-free people and hand images
A 2026 arXiv preprint used human annotations to map defects in AI-generated compositional images and test a repair workflow.
1221–1240
A 2026 arXiv preprint used human annotations to map defects in AI-generated compositional images and test a repair workflow.
Preprint: A non-experimental review of on-policy self-distillation in mathematical reasoning reports higher scores in some cited comparisons alongside lower functional diversity and longer optimization steps.
An arXiv preprint models how lattice strain and charge-density-wave mismatch can be shared, but its predictions still need experimental tests.
Preprint: In the main 50% gemma-2-2b comparison, SparseGPT and WANDA stayed closer to the dense model than magnitude pruning.
A 2026 arXiv preprint states the result for arbitrary Nakajima varieties, chosen points and degree data.
Preprint: A six-household model completed every scheduled activity, while average ratings for healthcare access and trip chaining trailed travel burden and activity feasibility.
Preprint: Praxist also earned more medals than Claude Code in one MLE-bench sweep, while other evaluations produced mixed results.
An arXiv preprint presents a computational proof of principle for machine-learning surrogates based on MAST-U discharge 49108.
Preprint: H.E.S.S. data favor a curved spectrum and a cosmic-ray injection site close to Sgr A*, but do not identify a unique accelerator.
Preprint: a mathematical framework connects cyclic-cover formulas with singularity criteria and vanishing results.
Preprint: Theoretical results and processor tests suggest separable quantum correlations can hide and release a computational resource.
Synthetic tests found much faster repeated calculations and successful completion at the largest benchmark size, but the study used no real experimental dataset.
Calculations found position and momentum measures moving in opposite directions while their combined entropy stayed nearly constant across the modeled transition.
Preprint: A mathematical construction uses permutation-generated quantum channels, but its certified finite-dimensional example is not practical to run.
A preprint evaluates 4DGS-WAM on KITTI-MOT, with the model leading given-camera video tests but trailing Envision4D when the future camera is predicted.
Preprint: MLSGD reached comparable accuracy 18 times faster than batched SGD in one reported comparison, while larger runs showed diminishing parallel efficiency.
This preprint reviews machine-learning tools for finding and studying hot subdwarfs and white dwarfs, while stressing that follow-up remains essential.
This Preprint on arXiv models a 0–n Fock qubit with exponentially suppressed bit-flip errors as its level number rises.
Preprint analysis finds the signal consistent with a core-related contribution but says the model does not indicate an imminent negative leap second.
Preprint: A comparison of two public datasets found strong held-out results with less detailed lesion marking than contour-based methods.