New Geometry Test Identifies When Tropical Multidegrees Are Positive
Preprint: A conditional projection test identifies positive multidegrees, while augmented Bergman fans also have Lorentzian volume polynomials.
Search
2,111 results
1261–1280
Nonnegative Concave Functions Preserve Lee-Type Schatten Constants
This preprint reports an exact transfer for finite complex matrix families across 1 ≤ p ≤ ∞; the linear constants remain unknown specifically for 1 < p < 2.
New Muon variant uses fewer tokens in Transformer tests
An arXiv Preprint reports lower final validation loss and 13.3%–24.0% token-efficiency gains for SAMuon over tuned Muon in Transformer tests.
Human readers and language models track language trouble differently
Preprint: Conceptual disruptions showed a sharper, shorter pattern, while referential disruptions were more distributed across later reading and model prediction.
One Mathematical Framework Brings Several Plasticity Rules Together
Preprint: A single scalar functional supplies stresses, internal-variable forces, resistance and plastic-flow conditions in a theoretical framework.
AI model posts lower 50-second errors in airport forecasts
A version 2 arXiv preprint reports lower 50-second displacement errors for safety-critical and randomly selected agents on the Amelia-10 benchmark.
Preprint maps how genome length relates to species splitting
A mathematical model predicts distinct critical-length regimes, while finite simulations support the trends without testing them in organisms.
New preprint classifies every crepant partial resolution of the nilpotent cone
Preprint: A proof-based study connects the nilpotent cone's resolution models to Hamiltonian reduction and geometric invariant theory, and extends the construction to deformations.
AI Upscaling Method Reports a Better Image-Quality Trade-Off
An arXiv preprint reports gains on several 4× super-resolution measures, while its evidence remains limited to benchmark comparisons.
A Live WiFi Overlay Maps Signal Paths in Augmented Reality
A preprint describes an ESPARGOS system that turns phase-coherent WiFi measurements into camera-aligned views of arrival angle, relative delay and polarization.
AI heart-rate model leads synthetic tests, but loses edge when ECG fails
An arXiv preprint finds attention fusion had the lowest descriptive overall error, but no pairwise winner survived statistical correction.
AI decoding method reports near-full accuracy with much less compute
A preprint reports about 90% of full-context accuracy on isolated text tests, with 1.3–1.7× throughput at 0.2–0.3× compute.
Voice-memory system scores higher while using fewer tokens in tests
An arXiv preprint reports stronger factual, persona and audio-memory scores, with lower token use at one tested retrieval setting.
Preprint confirms two exact Gaussian coding conjectures
The version-one arXiv paper proves the conjectured formula and rate region, while exact synthesis has higher rates than total-variation synthesis for positive correlation.
AI defense reports low jailbreak rates in benchmark tests
Preprint: A test-time defense stores rules about attack methods and reports low measured attack-success rates, benign scores close to no-defense results, and added computation.
AI training system posts higher multimodal instruction scores
Preprint: VISA uses a feedback loop to build synthetic image-and-text data, and its Qwen3.5-4B test model scored higher on instruction and broad multimodal benchmarks.
Colored polygon triangulations show a split in connectivity
Preprint: Three-color twist graphs have two disconnected exceptions, while validity-preserving flip graphs connect every case with four or more colors where valid triangulations exist.
Learned theorem-proving policies solve more with fewer steps
Preprint: Graph-neural-network policies solved up to 46% more benchmark problems than leanCoP, but gains varied by search design.
Preprint ties degree-2 Siegel forms to central L-values
Its global identity depends on the refined Gan–Gross–Prasad conjecture, and its strongest bounds also require GRH.
Quantum recovery test fails to repeat on an IQM processor
Preprint: IQM recovery fell between execution snapshots, while two IonQ windows met their frozen tests.