LongVU-TTT Reports Higher Scores on Five Video Benchmarks
Preprint: LongVU-TTT reports higher scores than LLaVA-Video on five benchmarks, while a targeted evidence test found its carried-state benefit weakened beyond half the video.
Search
2,451 results
1441–1460
Game-theory test checks strategies under all fully consistent belief systems
Preprint: A formal proposal called agnostic sequential equilibrium asks whether a strategy profile is a sequential equilibrium under every belief system fully consistent with it.
Flexible PCB magnets trap some beads in lab test
Preprint: Electroplated cobalt-platinum magnets on a flexible Kapton PCB had in-plane remanence readings of 0.24 tesla before and 1.4 tesla after annealing, and trapped some beads in a microfluidic channel.
BEAT2/GestureLSM test reports higher spatial errors for Progressive guidance
Preprint: InteractGesture evaluates inference-time spatial control for speech-synchronized full-body gestures in continuous streaming generation.
Private Search Nearly Matches Full Retrieval in Benchmark Tests
An arXiv preprint reports encrypted ranking and hidden result selection, adding 0.73 seconds in a full-corpus NQ test.
Preprint reports higher accuracy than SkillFormer on two video benchmarks
A computer-vision preprint compares adaptive ego-exo video fusion on two proficiency benchmarks and reports lower computation and model size.
Watermark for spiking neural networks passes checks but faces limits
Preprint: A membrane-state identifier remained verifiable after several checkpoint changes, while adaptive tests showed where its ownership claim could weaken.
Graph test with a text teacher reports a 0.21-point difference
An arXiv preprint reports a small standard-accuracy difference between FUG+GLEM-ITT and FUG-only, while balanced accuracy is slightly lower for FUG+GLEM-ITT.
Model estimates peak electrical currents during solar storms
Preprint uses monitoring-site data and solar-wind conditions to estimate mean and upper-range geomagnetically induced currents.
Proof Places Cesàro Operator in the Toeplitz Algebra
Preprint: A theoretical proof puts the operator in the commutator ideal Q, and therefore in the Toeplitz algebra T.
Cloud traffic model reports higher reliability and lower energy use
Preprint simulations using Google Compute Cluster traces reported higher reliability and lower energy use than two reduced model versions; no live deployment was reported.
A phase signal tracks changes in quantum oscillator models
Preprint: Theoretical models link a geometric phase to synchronization, a critical crossover and two dissipative transition signatures; no experiment is reported.
Lipid composition linked to stripe regularity in a lab model
Preprint: A laboratory study compared lipid mixtures at a cholesteric liquid-crystal–water interface as stripe spacing and texture varied.
Proofreading model links accuracy to the driven layer
Preprint analysis finds that the modeled error floor varies with drive position, while stronger driving may have a single optimum or a steady decline.
Face-Swapping Tools Leave Identity Clues Behind, Study Finds
A preprint analysis found target identities remained distinguishable after one swap and in repeated FaceFusion cascades.
Preprint sets error bounds for conductivity recovery from noisy scattered data
The study analyzes conductivity recovery from noisy point observations and tests finite-element reconstructions in synthetic examples.
Graph-product study sets bounds and exact cases for isolation numbers
Preprint findings cover direct, strong, lexicographic and Cartesian products, while a hypercube equality remains unresolved.
AI training-data filtering linked to stronger results after GRPO
Preprint: TailSFT had higher multi-sample coverage in most tests and higher pass@1 after matched reinforcement-learning runs.
Koopman controller shows similar tracking with faster solves
Preprint: A single-wheel simulation showed similar tracking from a linear Koopman controller, while preliminary hardware timing was lower than nonlinear MPC.
Quantum-Inspired Model Reports Lower Error Forecasting Cloud Workloads
Preprint: The reported gains came from simulated benchmark tests, while the multi-resource version used almost twice the memory of the single-resource model.