CoDe-LoRA reports stronger task transfer with less forgetting
Preprint: Across four language-model backbones, CoDe-LoRA beat comparison systems in benchmark tests, but its routing under severe distribution shifts remains unknown.
Search
2,451 results
1–20
Humanoid policy posts higher scores in object-transport benchmark
An arXiv preprint reports stronger two-object and unseen-task results, while longer sequences still expose failures.
AI token-pruning method keeps benchmark scores high under compression
Preprint reports that DIPrune retained up to 99.9% of unpruned performance while removing most visual tokens in multimodal model tests.
Robot system reported to perform comparably without user instructions
This preprint reports comparable performance with user-instructed variants in three tabletop collaboration tasks, but does not supply the key benchmark figures.
One-step AI method yields many distinct molecule candidates
A preprint reports molecular yields and more distinct elite architectures, but a digit-image control showed weaker distribution matching.
LipSync detector spots forged videos through head and lip mismatch
Preprint: LipDA reports high benchmark scores by checking whether lip movements and head pose stay consistent.
Tests of Acmite find mixed gender-bias results across three models
A preprint reports favorable shifts on several bias measures while general-capability scores stayed near those of the unmodified models.
Genetic measures of personality linked to earnings and wealth
Preprint: In a Dutch sample, emotional-stability and conscientiousness scores were associated with higher earnings or household wealth, but the study cannot establish cause.
Olive-leaf zinc oxide particles show activity against liver cancer cells
A peer-reviewed laboratory study found the strongest response in HepG-2 cells, but the particles also affected a normal-cell model and were not tested in animals or people.
Full Fine-Tuning Leads in Driving Video Captioning Test
Preprint: A study compares full fine-tuning, LoRA and prompting for models asked to describe what a vehicle does and why.
Femtosecond Tests See Water-Jet Breakup Beyond a Classical Cutoff
An arXiv preprint reports pinch-off where infinitesimal disturbances are predicted to be stable, while other conditions showed decay or spraying.
Federated vision model reports higher accuracy in domain-shift tests
Preprint: MOSAIC beat FedDDA on average across three image benchmarks, although some individual domains lost ground.
Text-to-motion AI method shows quality and speed gains in tests
Preprint: TACD reports better benchmark results, faster generation and lower peak GPU memory in controlled text-to-motion tests.
AI sorts overlapping radio signals, but misses more as they pile up
Preprint: A synthetic test found a six-slot model could label interference accurately, while detection weakened as more sources shared the same spectrogram.
Simulation tests a power grid as an image-processing operator
Preprint: An IEEE 14-bus model was tested as a physical computing step in image classifiers, but the evidence remains limited to simulations.
Simple model links plankton abundance to ocean patchiness
Preprint: Fluctuating growth and dispersal matched several plankton patterns, but the model fit less well in oligotrophic transects.
Enterobacter swarms spread differently on tilted agar
Preprint results show that colonies expanded most at 45° and 75° and some formed a downhill-focused leading edge.
Preprint reports sharper glioma boundaries with targeted review
An evaluation of 362 independent MRI cases found better whole-tumor overlap and tumor-core boundary scores, but about one in 10 cases still went to full review.
AI antibody system finds candidates active across several viruses
Preprint study reports about a 55% neutralization hit rate among 110 selected candidates, with evidence limited to laboratory tests and a small mouse study.
Computer Models Find Directional Transport in Complex Networks
Preprint simulations report topological markers, directional transport and selective robustness to disorder, but no physical device is tested.