Disabling a Few AI Experts Was Linked to Much Longer Outputs
A preprint reports extreme output-length increases in open-source mixture-of-experts models, but does not demonstrate a live hardware attack.
Search
2,230 results
821–840
Analysis Gives Polytope Walks a Dimension-Square Mixing Bound
Preprint authors extend the framework to spectrahedra, while an arrow example highlights why leverage scaling may matter for acceptance.
Varied street images linked to more accurate aerial localization
Preprint: An open benchmark across 41 European cities reported better in-the-wild results when Mapillary images were added with reduced training weight.
AI Handoff Router Picks Graphs or Prose by Task
Preprint: A benchmark study reports mixed results for graph and natural-language handoffs, with routing matching or exceeding prose-only scores.
Shared AI Model Improves 10-Second Traffic Forecasts at Five Sites
An arXiv preprint reports lower 10-second trajectory errors for pooled training across five Florida intersections, but short-range and site-specific results were mixed.
Elastic-Net ERGM Had the Lowest Error in One Simulation
A preprint reports that one Bayesian network-model implementation had the lowest error and false-positive rates in a single simulation, but lower sensitivity at its chosen reporting cutoff.
Robot adjusts assistance when possible motions diverge
Preprint: In a laboratory transport test, the proposed controller's average success rate was 0.95, compared with 0.83 for a fixed-stiffness version and 0.69 for a deterministic comparator.
Swarm-control model reports lower steering costs in simulations
Preprint: A mathematical framework steered modeled swarms to target distributions, but its controls were not proven globally optimal.
A Math Theorem Finds Certain mKdV Patterns Recur in Space and Time
Preprint: Under strict spectral conditions, a defocusing mKdV solution remains almost periodic in space and time, with a separate corollary for some analytic quasiperiodic data.
Training-free speech check flags word swaps in a limited test
An arXiv preprint reports lower error rates than three passive detectors, but the evaluation covered a narrow set of same-speaker attacks.
Correlation-matrix test may favor widespread shifts over factor changes
Preprint: A theoretical study derives an asymptotic reference law for comparing estimated correlation matrices and examines how different change patterns affect detection in modeled scenarios.
AI agents fell short on deep, large-scale biology tasks
Preprint: Across 20 studies, the best model averaged a 0.48 task score, while deeper analyses and data sets over 100 GB scored lower.
When Cache Misses Can Wait, the Farthest Page May Be Wrong
Preprint: A theoretical analysis gives a formal counterexample to physical farthest-next-use and reports exact restricted cases plus unit-cost and weighted guarantees.
Holographic Entropy Study Finds a Conditional Flow Bound
A preprint finds new constraints on entanglement response, but no second universal F-like measure along holographic flows.
Borromean Rings Floer Complex Found to Be Nonformal
Preprint finds the link’s full Floer complex is nonformal and identifies eight algebraic involution models.
Metalens readout records sharper X-ray edges in lab test
An arXiv preprint reports an 8.8-fold edge-resolution gain in a laboratory detector, while its CT dose and bandwidth benefits remain simulated.
FAST Finds Contrasting OH Signals in Two Bright 2024 Comets
A preprint reports robust absorption in A3's main OH lines after its closest approach to Earth, while 12P/Pons-Brooks was detected most clearly before perihelion.
New mathematical study establishes a local theory for a gauge-coupled flow
Preprint shows a strong solution exists and is unique for defined geometric data, while leaving long-term behavior open.
Verifier-based training linked to higher AI scores on pointing tests
Preprint: A same-backbone study found higher scores for PointRL-trained Qwen3.5 models and reported gains on several external benchmarks.
Study finds adaptive antennas show higher simulated secrecy rates
An arXiv preprint reports stronger secrecy rates and steadier learning in a modeled wireless system, but the evidence comes from numerical simulations.