Geometry changes how much softmax attention can be compressed
Preprint: A mathematical study reports different worst-case temperature laws for spherical and full-ball attention and checks the framework on 84 BERT-base heads.
Search
2,111 results
101–120
Six CKM fixed points survive every order of perturbation theory
An arXiv preprint dated 28 Aug 2026 reports an all-order proof that removes an assumption used in an earlier argument.
Overnight Bed Sensors Offer Modest Clues to Next-Day Agitation
A preprint study found that a full-night model using minute-level overnight signals showed modest ability to rank next-day agitation in a 65-patient hospital cohort.
AI Models Score High on Easy Finance Questions, Then Stumble on Hard Ones
This preprint reports a 10,198-question benchmark with a sharp gap between easy and hard items, while selective retrieval showed only a small gain on held-out questions.
Less Robot Sharing Linked to More Human Openness
Preprint: In a randomized study of 50 people, a robotic arm that shared fewer personal details was linked to more human disclosure and stronger teamwork ratings.
Model links jamming to a shift in how granular sound fades
A preprint simulation study reports a pressure-dependent crossover from quadratic attenuation to scattering-dominated sound loss.
AI-guided FPGA task mapping reports speed gains in two workloads
Preprint: An arXiv study reports repeatable FPGA-GPP allocations and speedups of up to 92.53 times across two workloads on an AMD ZCU104 platform.
AI model reaches 97% accuracy identifying Bangladeshi mango varieties
This preprint reports a smartphone-image benchmark and public web app, while its evidence remains limited to one dataset and a fixed test split.
Pulse-train estimator cuts memory use and speeds up period searches
Preprint: A simulated benchmark reported more than fivefold speed gains on a GPU and a 167 MB memory footprint after compilation.
Researchers Propose an Operating System for Shared Human-AI Work
Preprint paper describes CrabOS as a four-layer system designed to let people and AI continue the same task through shared, auditable objects.
Quantum model performs best in scarce-data MDS test
Preprint: A proof-of-concept study found the quantum system outperformed three classical generators on synthetic patient data.
Characteristic Polynomials Link Poisson and Gaussian Chaos
Preprint analysis suggests self-normalized random matrices move between Poisson and Gaussian limits as tail behavior changes.
Robot system clears every test bin, but grasp success varies
Preprint: A four-level hybrid robot cleared all 30 bins tested with three automotive parts, while grasp success differed by baseline and initialization.
Three-loop QCD calculation hints at higher inclusive Vub value
Preprint calculations add higher-order detail to inclusive B-decay predictions, but the semileptonic result remains preliminary.
Dynamic Radar Allocation Holds Up When Network Capacity Is Tight
A preprint reports that dynamic sensor and time-frequency allocation outperformed equal sharing when capacity was tight, but matched it when capacity was high.
Mathematical Analysis Maps Hexagons in Zero-Divisor Graphs
Preprint: A proof-based study of real Cayley-Dickson algebras constructs a six-cycle, a linked double hexagon and conditional subalgebra results.
A fast, narrow jet emerges from compact source SMA3
An arXiv preprint reports a bipolar SiO jet in SMA3, with speeds near 80 km/s and signs of disk-wind structure.
Scientific Text Reconstructs Much of Reactome's Global Shape
Preprint: Embeddings from Reactome titles and descriptions matched the database's broad topology, while individual biological links remain unverified.
Control method finds feasible shared targets for mixed robot teams
Preprint: A distributed controller jointly plans inputs and a common feasible equilibrium, but its guarantees depend on exact optimization and stated assumptions.
Language-Model Agent Societies Match Early Human Averages, Then Diverge
An arXiv preprint finds that some open-weight models match early human anchors, while later behavior and incentive responses often diverge.