AI method posts higher accuracy while learning new visual tasks
An arXiv preprint reports 62.5% average accuracy for a visual-dependence approach trained sequentially on answer-unlabeled image-question pairs across six multimodal tasks.
Search
2,230 results
1501–1520
New AI dataset pairs fitness video with muscle signals and errors
Preprint: A 7,512-sample project combines exercise video, physiological signals and expert error labels, but the supplied text omits detailed performance figures.
TRAPPIST-1 Shows Solar-Like Chemistry, With a Carbon Twist
Preprint analysis finds near-solar elemental abundances and an elevated carbon isotope ratio in the star.
New framework turns torsion-balance noise into gravity-test limits
Preprint: A methods paper shows how calibrated spectra can bound known torque signals and extra stationary stochastic noise.
Model links Hall-viscosity gradient to unequal electronic vortices
Preprint modeling associates a spatial Hall-viscosity change with a stronger vortex in one chamber and a weaker or suppressed vortex in the other.
Dust maps suggest some hydrogen clouds lie beyond the Local Bubble
Preprint: A four-sightline analysis links some absorbing structures to the Local Bubble’s wall and places another more than 100 parsecs beyond it, with important uncertainties.
New six-million-pair dataset targets reference-based video editing
This preprint reports strong benchmark scores and 50% lower training computation, but the evidence comes largely from AI judges and small human studies.
Graphene Preprint Models Two Distinct Magnetic Domain-Wall Textures
The theoretical work examines whether spin and valley can unlock inside rhombohedral graphene walls and predicts a possible superconducting-junction voltage.
Free square contains infinitely many non-unitarily equivalent free extreme points
Preprint: A proof-based study extends the level-2 result to every bounded convex polygon with at least four sides.
Robot Policy Scores Higher on Unseen Tasks Using Video Prompts
Preprint: Zero-WAM averaged 46.95% success on seven unseen simulation tasks and posted higher rates than LingBot-VA in three real-world task families.
AI GUI prototype reports higher component counts in 19-person study
Preprint: Separate tests report stronger GUI ranking than BM25 and high self-rated usability for RaWi.
Outputs using selected claims had more support, fewer hallucinations
Preprint: Outputs using selected English claims had more supported additions and fewer hallucinations on average, but results varied by language and generator.
Preprint finds trading order can shape outcomes in a fee-free market
A mathematical two-trader model links interior equilibrium to price matching and finds first-mover effects, while reserve scale changes numerical welfare comparisons.
AI can tidy model cards, but generated errors persist
Preprint: Two LLM workflows kept most existing information in place, but only 26 of 48 newly generated cards were judged fully correct.
AI retrieval system scores higher in crystal route planning
Preprint: A retrieval-based language model outperformed an LLM-only baseline and was used to plan routes for prospective target phases.
GeoFormer posts lowest error in four-site seismic picking test
A preprint reports lower first-arrival picking RMSE than a Vision Transformer on Brunswick, Halfmile and Dongbei, while its noise evidence remains qualitative.
Study Reports Higher Bridge-Damage Scores After Image Restoration
Preprint: The study reports higher scores when a detector-agnostic restoration model processes synthetic pairs and 200 real nighttime UAV images, while real-world validation remains limited.
RF Simulation Finds Learned Labels Hold Together in Heavy Noise
Preprint: A controlled 3-D simulation found semantic labels remained coherent deeper into noise, while fading-input tests mislabelled 18% of held-out novel objects as known.
Robot tests favor geometric targets over text-generated coordinates
This arXiv Preprint compares geometry-specific readouts with text-coordinate output across grounding benchmarks, simulated manipulation, transfer tests and physical robot trials.
New method maps how atomic motion changes magnetic exchange
A preprint reports a two-orders-of-magnitude cost advantage in SrMnO3, but the method remains a limited computational demonstration.