AI training filter shows benchmark gains, but costs more compute
Preprint study reports faster convergence and stronger benchmark results, while valuation takes more time on AIME.
Search
2,451 results
361–380
Trading preprint says robust search can erase a strategy's edge
An arXiv preprint reports that tail-risk protection and model search share one statistical budget; a 72-candidate walk opened no position.
Colloidal fluids developed uneven motion before crystallization began
Preprint: Light scattering found stable averages in an equilibrium sample but uneven, time-dependent motion in a metastable sample before crystallization.
AI driving model reports lower collision rates in six-second tests
Preprint: MomADv2 combines selective temporal memory with trajectory refinement and reports stronger long-horizon results across three driving benchmarks.
Adaptive gas-flow solver shows lower computing demands
Preprint reports faster, lower-memory simulations while staying close to a reference solver.
Fractal measures can show gaps and jumps, study finds
Preprint: A mathematical study finds exact formulas in some self-affine systems, but also shows that fractal spectra can develop gaps and jumps.
Preprint: Quantum decoding expands a window for polynomial states
A mathematical framework preserves higher-degree structures in selected quantum systems, while a small benchmark found a modest score gap over finite-budget classical searches.
Hybrid traffic system reports fewer simulated collisions in test
A preprint reports faster training and better transfer performance in a SUMO model of an 18-kilometre Changchun expressway.
New hits showed little lasting effect on existing stories, posts and films
Preprint analysis of BBC, Reddit and US cinema data found strong surges for hit items but little sustained change elsewhere.
Machine-learning model offers only modest Olympiad success forecast
Preprint: A model trained on German Physics Olympiad participants retained five of 31 features, but its predictions for individuals remained modest.
Gradient-Based Data Grouping Shows Uneven Gains for Video Models
A preprint reports higher automated scores for a method that groups reward-weighted video categories by training-signal compatibility, but the evidence remains narrow.
AI model links sign-language translation and motion generation
Preprint: A study reports pose-only translation scores and recognition results, but two professional signers rated generated movements below reference motions.
AI Method Reports Higher Scores on Vision Reasoning Tests
Preprint: A test-time method reported higher benchmark averages across six multimodal reasoning tasks, but it also required extra computation.
Model Tests Find Fewer Prompt-Injection Attacks With AI Overlay
Preprint reports lower attack rates when a learned, span-specific annotation channel is used, while returned text remains highly similar.
Study links higher-rank connections to accurate location at new speeds
Preprint: Simulated networks paired a simple location map with higher-rank connectivity that tracked speed changes and transient errors.
AI Method Reads Model Changes, But Often Misses Details
Preprint: A controlled study found that one model could sometimes read changes in another model’s weights and use the signal to target behavior, but free-form readout remained unreliable.
New method leads on four multi-view clustering benchmarks
Preprint: ACGRL recorded the highest reported accuracy on all four datasets, exceeding the strongest baseline by 5.40 percentage points on NGs, 0.15 on Hdigit, 12.88 on BBCSport and 9.34 on Cora.
WavePP reports higher LLM throughput under heavy cache reuse
A 28 Sep 2026 arXiv Preprint reports higher measured throughput in most tested settings, with a decline in one reuse workload.
Security audit status updates stayed under five seconds in test
An arXiv preprint reports local posture refreshes below a five-second gate, but did not test detection accuracy or production-scale performance.
Two nematode genomes differ sharply in satellite DNA content
Complete assemblies show C. nigoni is 29.9% larger than C. briggsae, with the biggest chromosome-size gap on the X chromosome.