New sequential recommender tops most benchmark tests
A preprint tests HSR on three interaction-log datasets and reports the strongest ranking scores on most measures, plus faster inference than Mamba4Rec on Amazon-Beauty.
Search
2,451 results
1941–1960
More inference compute rarely helped small AI models catch larger ones
Preprint: In a grammar-constrained text-to-SQL test, beam search usually beat sample-and-vote, while extra compute rarely erased the accuracy gap.
AI maths tutor reports stronger guidance across several benchmarks
Preprint: ToST keeps multiple solution paths in play, but its main tests used simulated students and a small human pilot.
Lightweight AI model uses less compute for satellite rain retrieval
Preprint: WaveOp-LiteFM recorded stronger scores on several benchmark measures, but took slightly longer to run in tests.
GeoSample Reports Best Mean Metric Combination Among Tested Operators
Preprint tests GeoSample on MRI, CT and ultrasound datasets, reporting the best mean metric combination among compared operators and fewer U-Net parameters and FLOPs.
DEDUCE scores higher on benchmarks with misleading inputs
Preprint: An arXiv study reports higher DEDUCE scores on FalseQA and MisFactQA from single-run evaluations.
Ulcer-image AI beats UniMatch-V2 in six of seven tests
Preprint: TiSage outperformed UniMatch-V2 in six of seven benchmark settings, while clinician agreement varied substantially across tissue types.
AI vulnerability tool leads benchmark, but evidence has limits
Preprint: EAVA scored highest in tests of CVSS vulnerability ratings, while a small practitioner study found its supporting evidence useful in many cases.
AI system reuses reasoning with higher accuracy and lower estimated cost
An arXiv preprint reports gains across five online question-answering streams, while leaving human benefit and long-term reliability untested.
Preprint simulates radio telescope for gamma rays and neutrinos
Simulations preserved shower direction and approached an ideal image-sharpness benchmark, but the projected observing gains remain untested in a working instrument.
Preprint finds traces of a lost language inside speech models
A computer study found that models switched from one language to another still carried low-level speech traces and later relearned the first language faster.
Preprint reports 3x faster AI inference in long-reasoning tests.
The arXiv study reports benchmark scores close to full attention alongside higher throughput on long sequences, with results varying by task.
Robot grasping model leads simulation tests with five evaluations
A preprint reports that GraspMF led the compared methods in five-evaluation simulation tests and recorded mostly successful trials on three household objects.
AI simulations show stronger portfolios, but solo search has the top artifact
Preprint: In deterministic simulations, shared-world language-model populations recorded stronger portfolios and more validated inventions, while isolated search had the strongest final artifact.
Preprint reports a calculated twofold reduction in CPU decode bandwidth
An arXiv preprint reports lower calculated bandwidth and fewer cache misses in selected tests, while its throughput comparison uses unmatched models.
AI system HypoForge tops reported generation and test-pass scores
An arXiv preprint reports results from a 13-task computational benchmark; POPPER recorded higher execution success.
Field-free diode response seen in altered CsV3Sb5 devices
Preprint: Protected pristine devices showed no detectable field-free diode effect, while responses were observed after oxidation and in a selectively etched device.
AI model leads benchmark in behavior-based person searches
Preprint testing reports higher retrieval scores on the PAB anomaly-search benchmark and transfer to a separate person-retrieval dataset, with limits on complex queries.
Diamond maser shows sustained quantum coherence under ambient conditions
A preprint reports long-lived coherence among about 100 trillion spins and a billion microwave-cavity photons, while identifying three distinct operating regimes.
Two short ssDNA homopolymers show different stiffness patterns
Preprint simulations compared two 30-nucleotide strands in magnesium, calcium and spermine; fitted persistence length leveled off above 0.5 millimolar spermine.