FPGA system reports lower job-runtime error, 17-fold speedup over AVX
Preprint: An arXiv study reports lower job-runtime error on three historical traces; a separate ALCF benchmark averaged a 17-fold speedup over AVX.
Search
2,111 results
721–740
Bitcoin Signal Search Shows Positive Scores but Negative Net Sharpe
Preprint: A deterministic audit caught all 40 illegal templates and rejected none of 40 legal ones in a finite benchmark, while all 23 holdout runs had negative net Sharpe under primary costs.
J-lens Study Reports Higher AI Readout Scores With Energy Filtering
Preprint: In tests on Qwen3-8B, variants retaining the top 10% to 20% of positions by Jacobian energy had higher reported averages, while masking did not cleanly separate two readout modes.
Preprint finds exact dimension for a difficult number set
An arXiv preprint reports a Hausdorff dimension of 2/(τ+1) in the stated nonzero-shift setting, using a scale-sensitive game.
Robot-learning study finds listwise feedback can match pairwise methods
Preprint: In simulated robot tasks, listwise visual rankings were competitive overall, but results varied by task and random seed.
Exact schedulers solve all 35 tests and expose benchmark errors
A preprint reports that SAT and constraint programming proved optimality on all 35 cases, while a review found nine earlier schedules that violated the power constraint.
Model predicts optical signatures of a Hopf-link transition
A theoretical arXiv preprint also models a nonlinear Hall response with a specified symmetry-breaking term, but the results have not been tested in a material.
Model finds narrow bands for eccentric orbits around a spinning black hole
Preprint: A model finds two equilibrium branches in a narrow inclination range, but cannot reproduce the high eccentricity discussed for the S-cluster.
Theorem narrows how nonzero-degree self-maps act on generalized Grassmannians
A version 1 arXiv preprint classifies the cohomology action of nonzero-degree self-maps, allowing a Dynkin symmetry only when the relevant stabilizer is nontrivial.
Sparse Model Merging Shows Modest Gains in AI Benchmark Tests
Preprint: A selective sparse-feature method reports higher average scores in two Qwen2.5 settings, but adds training costs and does not improve every benchmark.
AI models can find the right data and still put it in the wrong place
Preprint: Tests of nested JSON and HTML tables found a recurring gap between retaining required values and placing them at the correct structural address.
New result identifies the Langlands image of a constant sheaf
Preprint: A formal mathematical result identifies the spectral object assigned to the constant sheaf and extends the description to arbitrary finite-rank local systems.
Speech-act model ranks first in in-domain online derailment tests
Preprint: A model using sentence-level speech-act labels ranked highest in reported in-domain tests, while results varied across online communities.
Preprint reports lower surface-normal errors from flash pairs
An arXiv preprint reports lower surface-normal errors on synthetic and real-object tests, but also describes failures on transparent objects and flat objects with complex textures.
Preprint Proposes Clearer AI Threat Profiles Before Release
The proposed framework asks evaluators to define an adversary's skills, resources and time before testing frontier models, with open-weight releases in focus.
3D wave scheme tracks test pulses to infinity on flat spacetime
Preprint: A fully 3D numerical study reports stable evolution, close energy agreement and near-second-order convergence, with a trade-off at the outer boundary.
Preprint: Quadruped robot explores five floors in stairwell test
The paper describes RAEM's multi-floor simulation trials and a real-world stairwell run from the first floor to the fifth.
One proposed Vision MoE backdoor stayed quiet in four tests
Preprint: In V-MoE and Swin-MoE classifiers tested on ImageNet-100 and GTSRB, dormant attack success was below 9% at batch sizes up to 32, while activation success at batch size 128 ranged from 76.0% to 87.0%.
Underwater detector reports modest gains in camera and sonar tests
Preprint results report higher RGB and sonar detection scores, plus an F1 score of 83.4 for linking objects across the two views.
Constrained cross-market model improves volatility forecasts
A preprint study of eight global equity indices reports the clearest average gains at weekly and monthly horizons, while daily results vary by error measure.