Preprint reports faster sparse-grid particle–mesh calculations
Hierarchical versions of two methods delivered large reported speedups while matching standard results to roundoff in a two-dimensional benchmark.
Search
427 results
361–380
Preprint reports gains in predicting 3D wireless fields in unseen scenes
RFWM reported stronger benchmark results without target-scene RF measurements, but validation on real measured fields was not reported.
Preprint: Robot planner reports 86.93% simulated success without object-specific grasp databases
GOAG uses gripper configurations and contact points, with tests spanning three grippers, five benchmarks and a real-robot demonstration on 11 YCB objects.
Preprint reports faster long-context AI serving with little benchmark loss
FlashPrefill V2 reported up to 47.26× operator-level speedup at 128K on NVIDIA H20 GPUs while keeping scores close to full attention on RULER and LongBench.
Preprint reports a symmetry-based model for mapping stroke lesions across scans
The system outperformed tested baselines on one CT dataset and produced competitive results on MRI and CT-perfusion data, but the evaluation did not measure treatment decisions or patient outcomes.
Preprint releases open simulated data for power-grid protection research
EvEMTBench pairs synchronized voltage and current waveforms with labels for faults and operating events, but it does not yet show how machine-learning systems will perform on real grids.
Preprint: Topological acoustic chip senses salt and glucose in tiny liquid samples
In laboratory tests, the device’s resonance shifted with solution concentration and used 0.04 μL for one measurement, but it was not tested on biological samples or compared with conventional sensors.
Federated MRI Method Scores Higher on Tumor-Mapping Tests
Preprint: MOSAIC reported higher overlap scores than weakly supervised baselines when institutions had different MRI modalities missing.
Preprint reports possible sound-speed plateau in dense neutron-star matter
A theoretical comparison links a nucleon core-and-cloud picture to dense-matter models and neutron-star observations, while leaving the high-density outcome model-dependent.
Preprint classifies finite-group modules with few non-trivial Jordan blocks
The formal study identifies allowed irreducible building blocks and exact block patterns for a narrowly defined set of finite groups.
Preprint reports higher success for VLM-guided robot training in simulation
EXIMO combines exploration, imitation and residual reinforcement learning, with reported gains across a 22-task manipulation benchmark.
Preprint tests a calibration method for reliability decisions under uncertain models
The approach had no larger defined risk than a fixed threshold in theory and showed lower set-estimation error in simulations, but its vehicle example was simulator-based.
Preprint reports higher benchmark scores for an adaptive vision operator
Core-KAN reported gains against ResNet-50 and selected competing methods on ImageNet-1K, COCO and ADE20K, with caveats around training comparisons and response-bank cost.
Preprint reports higher scores for detecting unseen 3D objects
Co-3DGT posted higher novel-class and overall benchmark scores on SUN RGB-D and ScanNetV2, using co-distillation with uncertainty regularization and hierarchical alignment.
Preprint uses joint eigenvectors to find repeated roots and reconstruct tensors
In two worked mathematical examples, the method recovered solutions with reported multiplicities of 3 and 1 and reconstructed a tensor with an apolar-norm error of about 10^-13.
Preprint links whole-workflow guidance to higher policy-compliance scores for AI agents
In simulated airline, retail and telecom customer-service tasks, PolicyGuide recorded a mean Pass4 score of 0.62, compared with 0.42 for unguided execution.
Preprint links topological homology to prismatic cohomology
The study gives canonical descriptions of three related theories and an E²-page calculation for p-complete THH of Zₚ[x]/(px), but leaves the final extension problem open.
Preprint: AI training method posts higher scores on unseen image combinations
The DIFFCZSL approach reports stronger balance scores across benchmark tests, while adding training time and parameters but no extra inference-time computation.
Preprint model reaches 90.1% on human-pose video anomaly test
STEP reports its strongest result on UBnormal, while its pose-only design remains vulnerable to missing tracks and anomalies involving objects or interactions.
Preprint reports compact 3D reconstruction from sparse views under known lighting
A point-based renderer recorded the lowest reported surface-mismatch score in a five-object synthetic benchmark, using 267 optimized surfels on average.