Role-aware image search leads across three benchmark tests
A version-1 arXiv preprint reports higher retrieval scores after queries are split into full-target, add, preserve and remove roles.
Search
2,230 results
841–860
Preprint Maps Sample Needs for Near-Second-Best Bilateral Trade
A theoretical arXiv preprint reports matching or nearly matching sample-complexity bounds across three regimes, with sample requirements varying by guarantee and distributional assumptions.
New Construction Makes Surface Slopes Dense for Prescribed Groups
Preprint: A proof-based construction makes Chern slopes dense from 1/2 to 3, with the ample-canonical case established through 2.
Analysis identifies exact symbols for two Volterra operators
Preprint: An arXiv version 1 document dated 26 Aug 2026 gives exact boundedness criteria for area Nevanlinna spaces in the range 1 ≤ p < ∞ and α > −1.
V-Link Reports Higher Success in Simulated and Real Robot Tests
Preprint: V-Link reported higher manipulation success than GR00T N1.6 on three simulation benchmarks and two AGIBOT A3 Ultra tasks.
Diffusion model prefix check led or tied in 15 selected-length tests
Preprint: A regeneration-and-voting method improved accuracy over initial generation in 14 of 15 selected settings.
Preprint finds when Selmer-group results extend to larger fields
Conditional theorems cover several arithmetic settings, while a counterexample shows why the finiteness assumptions matter.
New statistical test flags unstable links among financial factors
Preprint: The method rejected covariance stability at the 1% level in weekly factor returns from 2004 to 2023 and in each of three subsamples.
Models suggest small off-axis ratios in 40 selected gamma-ray bursts
A preprint using Swift/XRT jet-break light curves finds small inferred ratios between viewing and jet angles in most selected bursts, but the result changes with the emission model.
A New Optimization Model Picks Which Solution to Approach
Preprint: A continuous-time model keeps a theoretical 1/t² objective-decay rate while selecting an anchor-dependent solution in affine problems.
Crack simulations show a threshold for finger-like growth
Preprint: A numerical model reports stable low-contrast fronts and, above a critical threshold, pinned regions and advancing fingers.
Preprint: RDQ M2-alpha1 records higher simulated POI power than RBP(0.9)
In a 5,000-query comparison, the metric showed higher median power at 100 queries, while the study did not test user preference.
AI Study Finds Most Activation Order Effects Reflect Site Asymmetry
An arXiv preprint reports that a single-intervention baseline explained a mean 93.7% of the measured order-swap signal across six language-model families.
Multimodal AI Holds a Narrow Edge Over Text in Interview Test
This preprint describes a 646-participant dataset in which personality ratings from targeted questions were more reliable, while the best text model stayed close to the strongest multimodal system.
Mathematical test separates black-hole singularities from infinity
Preprint finds a conditional way to distinguish the two kinds of spacetime boundary, then applies it to Schwarzschild geometry.
Preprint reports directional drone drift in gyroscope tests
The study reports simulated lateral drift and an indoor coincidence of roll bias with displacement, with validation limited to controlled tests.
Preprint reports a 4.6-fold reduction in safety constraints in one test
In one ten-follower, four-obstacle numerical scenario, the reduced formulation retained 2.80 safety constraints on average, versus 13 for the full formulation.
A Two-Part Error Bound for Agnostic Multiclass Transductive Learning
Preprint: A mathematical analysis links DS and Natarajan dimensions to separate terms in worst-population error bounds, which match only up to logarithmic factors.
Optical probing recovers AI-chip data in a controlled test
Preprint: A Kintex-7 testbed produced complete register and BRAM readouts, while algebraic methods recovered missing bits in many tested cases.
Financial AI models rank differently across risk-review tasks
A preprint reports that a Chinese-language benchmark finds broad financial scores can miss how models handle specific risk and compliance work.