Tests find AI document paths can expose hidden Office content
Preprint: Tests across Word, Excel and PowerPoint found different evidence reaching models through different document-ingestion paths.
Search
2,451 results
1561–1580
New flow model has lower long-horizon errors across four benchmarks
Preprint: CoKo-UNO records the lowest rollout error in four flow tests, but its accuracy drops sharply when moved from a 32×32 grid to finer resolutions.
AI summarization method is far faster at choosing training examples
Preprint reports much faster example selection than two baselines, while quality results vary by dataset and model.
Satellite model sorts maize stress into four readable stages
Preprint: A study reports promising results in Iowa and Nebraska, but field decisions and yield benefits still need testing.
Calculations point to cooperative neutron pairs in helium-8
This preprint reports energy and wave-function signals consistent with a two-pair quartet in modeled helium-8, while limiting the conclusion to the Daejeon16 interaction and basis used.
Math Preprint Finds Worst-Case Sampling Can Scale Quadratically
A new arXiv preprint bounds how many weighted point evaluations are needed to control distortion in m-dimensional function spaces.
Nearest-neighbour method stands out in high-dimensional clustering tests
Preprint: The approach beat maximum likelihood at larger tested dimensions, but outlier results varied by scenario.
Preprint: Cœuré–Loeb fibration found inside a Stein space
A proof-based construction places the non-Stein example inside a larger Stein space, while leaving the broader open-immersion problem unresolved.
RNN controller shows lower data movement in GPU tissue model
Preprint: A history-aware controller kept workloads more even in a changing multicellular simulation, but did not have the lowest imbalance at every rank count.
Project Builds Digital Archive of Japan’s Sangaku Tablets
A preprint describes a bilingual archive for roughly 900 surviving tablets and records of lost ones, with a map carrying more than 1,000 pins.
AI model transfers across odor tasks but falters on mirror images
Preprint: The system posted stronger scores across several olfactory benchmarks, but a small enantiomer test suggests that noticing molecular differences is not the same as predicting their perceptual effects.
Graphene detector operates at room temperature on silicon waveguide
Preprint: A two-device laboratory study reports D-band detection at 139 GHz, while measured port-referenced responsivity differed sharply from the model estimate.
Forecasting method tracks changing uncertainty in Lorenz tests
Preprint: SCROLL-MT was tested on two synthetic systems and one Beijing PM2.5 series, with gains in some uncertainty and likelihood measures but trade-offs between covariance choices.
Theoretical quantum protocol ties entanglement cost to time
Preprint: An adaptive repeat-until-success method links shared entanglement to interaction strength instead of charging a fixed amount for every nonlocal operation.
A single framework links two ways of explaining time-series models
Preprint: TimeX++ uses one end-to-end Information Bottleneck objective across attribution and counterfactual modes, with reported gains on benchmark tests.
Preprint finds sharp asymptotic scale for fixed-weight binary digit bias
An arXiv study shows how the smallest positive gap above one-half behaves when a shift has exactly k binary 1s.
Computer model links structural color to local rings and pores
Preprint: An inverse analysis matched a biological spectrum to disordered network models, but could not identify one unique structure.
Averaging time-series data may speed long-run estimates
A preprint finds that averaging can combine stronger signal with less noise, but real-world tests show why the choice of time window matters.
SQL pruning linked to faster relational AI training in tests
Preprint: Across five RelBench databases, MetaSieve reported faster training epochs, but regression accuracy, tuning and memory remained important caveats.
Romanized Training Tops Multilingual Transfer Tests, Study Finds
This preprint reports that romanized input was strongest in cross-lingual transfer tests across eight languages and three model scales.