Video AI Keeps Nearly All Benchmark Performance With Fewer Tokens
Preprint: A coding-inspired system retained 97.8% of Qwen3-VL-8B's average across four long-video benchmarks at 25% token retention.
Search
2,451 results
481–500
Flow-matching shortcuts aid training but miss exact likelihoods
Preprint: A study finds score calibration improves fixed-target estimates, while reward gains can coexist with biased likelihood ratios.
A new theorem caps the product of intersecting subspace families
Preprint: A finite-field result bounds the product of sizes for at least three cross-intersecting families and shows exactly when the bound can be met.
Long-range disorder may support order in driven 2D models
Preprint: An analytical study predicts ordered phases under certain disorder and noise conditions, but it does not test a real system.
Lattice framework completes spaces for higher-rank Drinfeld modular forms
Preprint: The lattice-based theory adds lower-rank boundaries and matches cusp forms with the Basson–Breuer–Pink and Gekeler frameworks.
Preprint Reports Fuzzers Reach More Bugs Than They Trigger
An arXiv analysis of Magma v1.3 reports that 77 of 127 benchmark bugs were reached and 43 were triggered in 24-hour campaigns.
AI method flags answers that may outrun their evidence
Preprint: A benchmark study reports stronger scores and selective abstention from a test that rewrites named entities while preserving relationships.
Preprint ties accelerated transport convergence to mesh scale
A mathematical analysis and numerical tests report strict acceleration in an optically thick model, with the approximation controlled by a local Knudsen number.
Quasi-2D model tracks 3D convection with insulated spanwise walls
Preprint: A numerical comparison found close agreement under insulated spanwise conditions, but noticeable differences with conducting walls.
Preprint sets out an algorithm for rank-two algebraic categories
The study classifies indecomposable objects, recovers their W-graphs and tests the procedure across low-rank cases.
Preprint uses sound spectra to measure timbral distance
An arXiv preprint compares vocal and instrumental spectra, then explores whether the same geometry can inform fingerings and chord voicings.
Quasar’s changing X-rays put its ultrafast wind farther out
Preprint: Soft X-ray changes point to a low-ionization wind at least 4,000 gravitational radii from the black hole.
LISA Simulations Find Large Waveform Shifts From Scalar Hair
Preprint forecasts that future extreme-mass-ratio inspiral signals could reveal the coupling and scalar mass, but its waveforms remain an effective, incomplete model.
Mathematics preprint finds non-klt counterexamples to index bounds
A constructed three-dimensional family keeps minimal log discrepancy at one while its Cartier index can take any positive integer value.
Rho Leo Shows a Persistent 16-Day Spectroscopic Rhythm
Preprint analysis favors non-radial pulsations, but it leaves open the possibility that a companion is involved.
Text Models Spot Evasion Better Than They Read Vocal Confidence
Preprint: A benchmark of 505 earnings-call answers found a sharp gap between textual-evasion detection and vocal-confidence detection, while speaker context delivered only modest gains.
Active colloid chains can look balanced while still dissipating
Preprint simulations find zero orientational entropy production in fixed-geometry cases, while whole-system entropy production remains finite.
Brain MRI Review Finds Safety Checks Rarely Test Small Lesions
Preprint review of 263 studies says high image-fidelity scores cannot certify diagnostic safety when reader testing and acute-pathology benchmarks are scarce.
Large diffusion sorts a nonlinear heat equation into three outcomes
Preprint: A mathematical analysis gives a uniform rule for decay, threshold behavior and finite-time blow-up as diffusion grows.
Foraging Agents Learn to Cluster Without a Grouping Reward
Preprint: Agents rewarded for finding targets learned collective search at longer visual ranges, despite never seeing the targets.