Machine-learning test aims to separate data drift from changed relationships
Preprint: A two-discriminator method estimates separate input and mechanism components in synthetic and public benchmark tests.
Search
2,451 results
2001–2020
Fibonacci Walks Match at Order Three, Then Diverge
Preprint: A proof-based analysis finds identical third-order gap spectra but an extra fourth-order value in the up-integer set.
Model Links Time-Varying Delays to Intermittent Flocking
Preprint: A numerical model showed long stretches of near-complete order followed by brief disorder when interaction delays varied over time.
AI gate weighs evidence before reusing or adding expert models
An arXiv Preprint describes a decision layer for changing data streams that can reuse an expert, spawn one or wait for more evidence; its formal validity claim applies to an observable surrogate.
Axon matched or beat Transformers in most LLM generation tests
An arXiv preprint reports median gains across several backends, but larger-model parity and forward-pass performance were uneven.
AI Method Ranks Best in Tests Mixing Known and Unknown Images
The arXiv preprint tests a source-free system that adapts classifiers on unlabeled streams containing familiar image shifts and out-of-distribution samples.
InSight data associate a Martian wave with dust-storm pressure changes
Preprint: K1 appeared in 17 of 99 InSight intervals and reached about 8 pascals during the C34 dust storm, but the study found association, not causation.
Periodic Subject Changes Linked to Higher Judge Scores in Short Windows
Preprint: In base language models, the subject-change arm scored higher than habituation alone on short windows judged by a language model, while no condition exceeded 2.5 on any whole-document dimension.
A formal candidate for supersymmetric bigravity takes shape
Preprint reports a formally consistent coupling while leaving the model’s self-interactions for future work.
Robust filter leads rivals in simulated radar tracking tests
An arXiv preprint reports lower simulated position errors for a new filter than for a UKF and a bootstrap particle filter, but only in one model and two noise cases.
Theory Ties Binary Experiment Dominance to Two Information Measures
Preprint: An arXiv paper examines how binary experiments can be compared when observations stop as evidence arrives.
Preprint reports FPGA trigger reaching 97% on air-shower benchmark
The arXiv study reached 0.97 efficiency at a 3.39 kHz candidate rate on independent validation data, while continuous background replay averaged 6.64 kHz.
Satellite water-quality accuracy varied with feature choices
Preprint: A Norwegian study reports higher scores for optimized preprocessing in chlorophyll-a and water-clarity models, but finds no universal setup or validation beyond the region.
Preprint: New method adds pose-aware motion to static 3D avatars
The preprint reports the best image- and video-generation scores among displayed methods, but benchmark scores averaged only two subjects per dataset.
Security tools stumble on malicious AI agent skills across sources
Preprint: A benchmark of 9,740 AI agent Skills found detectors performed much worse on held-out sources, with one model flagging 62.4% of benign Skills.
New algorithm could speed trajectory planning for many trailers
A preprint reports sub-millisecond high-order derivative calculations and a simulated 20-trailer parking manoeuvre, with tests limited to computation.
Preprint: Best-Arm Search Shows Why the Multiple-Testing Cost Moves
A mathematical note argues that the K − 1 factor shifts between two ways of framing the statistical test rather than disappearing.
Preprint classifies Ramanujan graphs in a Frobenius-group family
The exact result covers normal connection sets in finite groups with a kernel of size q and a complement of size q−1.
PETA preprint reports stronger early virtual-screening rankings
The test-time method updated about 0.03% of a pretrained model and reported stronger early-enrichment scores on two benchmarks, but its evidence is limited to computational tests.
Spike-based controller reaches Mountain Car goal but overshoots
An arXiv preprint reports qualitative agreement with a conventional numerical controller, while leaving run counts and statistical performance unquantified.