A Survey Maps Five Steps for More Reliable AI Data Work
The arXiv v1 preprint groups Data Agent methods into 15 technical routes and identifies four unresolved reliability problems.
Search
2,451 results
281–300
New math identifies variables shared by every optimal solution
An arXiv preprint links variables shared by every optimum to bottleneck tests and dual solutions, and shows when the hierarchy collapses at level two.
A navigation method posts higher scores without outside feedback
Preprint: A test-time adaptation system reports higher scores on unseen-environment benchmarks, while its robot evidence remains qualitative.
Mathematicians construct long-time Sobolev growth from tiny data
Preprint: In a two-dimensional periodic model, the Sobolev norm of a constructed solution rises from small to arbitrarily large.
AI Token Rankings Stayed Stable Even When Scores Were Inaccurate
Preprint tests an uncertainty-explanation method on two language models and finds that reference choices can change what token scores mean.
Astronomers Validate a Cold Sub-Neptune Near the Ice Line
Preprint: Follow-up observations validate TOI-6981 b as a 2.4-Earth-radius planet, while its mass and atmosphere remain unmeasured.
Magnetic capsule releases microrobot surrogates in milliseconds
Preprint: The capsule opened within 50 milliseconds after a hydrogel latch softened, but in vivo delivery and long-term stability remain unresolved.
Study Links New AI Training Method to Stronger Reasoning Scores
Preprint reports higher judge-assessed reasoning quality in math and code tests while answer accuracy held up in several comparisons.
EEG-to-Speech Model Leads Benchmarks, but Words Stay Unreliable
Preprint: SENSE led nearly all listed metrics when audio, subjects, or both were unseen, while reliable word-level reconstruction remained unresolved.
Robot vision system uses far fewer adaptations with small accuracy gap
Preprint reports near-per-frame segmentation performance with 3.9 adaptations versus 518 for per-frame adaptation.
Two Satellite Data Checks Agreed on the Same Wrong Answer
Preprint: A case study finds that two technically different verification paths can share the same mistaken reading of catalogue data.
Transformer Model Reconstructs Sparse CT Views Across Scan Setups
Preprint testing suggests one model can fill in missing CT projection views across changing protocols, with most evidence coming from simulations.
Raman Signals Differ Around the Ice-Water Transition With Calcium Chloride
A preprint reports stronger signals near 0 °C and a lower Raman-derived latent-heat metric in doped ice, pending calorimetric testing.
Repeated X-rays reveal a weak link to quasar accretion
A preprint reports a positive but shallow association between X-ray spectral shape and accretion state, while single-epoch mass estimates produce conflicting results.
Changed inputs were linked to far more feedback from AI teams
Preprint: A synthetic multi-agent framework showed much more feedback in paired tests, while later gains varied across models and benchmarks.
Preprint describes two io_uring flaws and three sheaf attack techniques
The source-code study reports a local privilege escalation under modern kernel protections, but both vulnerabilities are described as fixed and patched.
Preprint outlines a way to compute exterior-algebra resolutions
The method produces a Gröbner basis for a syzygy module and points to BGG cohomology calculations, but its speed and scalability remain untested.
Simulation reports higher robot-planning success with branching
Preprint: A paired simulation found higher success and shorter solve times than single-branch planning, with evidence limited to simulated interactions.
Three metal-poor stars include one with unusually high magnesium
Preprint: Subaru observations identify a carbon-rich CEMP-s star, but its possible binary origin and proposed Galactic membership remain uncertain.
Preprint Reports Higher Reasoning Scores for a Vision AI Model
A factorized distillation method ranked above evaluated comparison baselines on reported reasoning averages, while the second calibration setting posted higher perception scores.