AI Theorem Provers Struggle to Turn Math Answers Into Proofs
Preprint results from MathAdv show a wide gap between natural-language answers and formal Lean 4 proof construction.
Search
2,230 results
961–980
One AI auditor setup scored above a baseline, but results varied
Preprint: A computational benchmark found higher scores for one calibrated pairwise checkpoint, while concern-focused training showed a steep fall in false-positive calibration.
New benchmark tests whether AI video scores match human judgment
Preprint: An expanded benchmark uses more than 60,000 generated videos and 36,000 new annotations to assess aesthetics and generation quality.
Direct Neural Predictors Beat Generative Models in Rarefied-Flow Test
Preprint: Flow matching and diffusion reached low interpolation errors, but neither beat direct deterministic regressors on the additional cylinder test.
Quantum model links competing soft modes to tunable phase changes
Preprint: Analytical calculations associate two collective excitation channels with different multicritical phase diagrams in a generalized two-mode quantum Rabi model.
Leonid Model Forecasts Sharp 2033 Peak, Uncertain 2034 Split
Preprint: This Leonid dust-trail model projects quiet early years, a sharp 2033 peak and two 2034 peaks whose relative strength could reverse.
Preprint maps sharp limits for lattice resolvent estimates
An arXiv analysis of the free discrete Laplacian reports uniform bounds across dimensions four and above, while leaving a four-dimensional endpoint question open.
Drone controller reaches target in simulated clutter
Preprint: In simulation, a viability-kernel MPC system guided a fully actuated hexarotor around static obstacles, with reported iteration times below its 20-millisecond sampling period.
RF-DETR transfers better than tested CNNs to close-range weld images
An arXiv preprint reports stronger close-range transfer for RF-DETR than tested CNNs, while three-seed CNN scores were lower on shop-floor imagery.
Preprint derives Gaussian bounds for Hilbert-space U-statistics
The mathematical framework also gives bootstrap rates and a minimax separation rate for dense Kendall’s tau alternatives.
Porous Ice Shows Different Water-Loss Signals by Wavelength
Preprint: Infrared tests found the same restructuring pattern across three ice vibrations, while water-desorption signals varied sharply by wavelength.
Sparse sampling linked to slower functional linear regression rates
Preprint: A pooling-ridge estimator is paired with matching bounds and tested in simulations and two datasets.
AI pipeline retrieves relevant evidence for nearly every zebrafish query
An arXiv preprint tests a system that follows zebrafish genes to corresponding human orthologs and ranks functional and literature evidence.
Data-only AI clustering beats mean prediction in synthetic tests
Preprint: In synthetic tests, the method came closer to one of several valid outputs than random labels, but it remained behind a benchmark using true modality identities.
GPU Kernel Study Reports Up to Twofold Gains in Matrix Solves
Preprint testing on three accelerators reports peak speedups of 2.05× over cuBLAS and 2.06× over rocBLAS.
Preprint model tracks tape temperature during dynamic process changes
The simulation reported a 4.02°C RMSE against measurements, but its parameters were tuned on the same trajectory used for testing.
Preprint method shows sharper phantom images while learning sound-speed corrections
Testing found sharper phantom images and visible vessel structure at 12.5% detector retention, while validation remained limited to simulated and physical phantoms.
AI Agent Design Separates Audit Trails From Raw Data Sharing
Preprint: An arXiv paper outlines field-level controls for limiting identifying details, while tests and a reported failure expose important trust gaps.
FedQoS stays close to centralized prediction in simulation
Preprint: In a physics-based campus simulation, FedQoS remained close to centralized prediction in two non-IID conditions, while access-selection results differed between mild and severe tests.
Offline AI shows gains in simulated multi-drone fleet tests
Preprint: An arXiv version-1 study dated 26 Aug 2026 reports higher return, communication and sensing measures, plus lower collision-risk counts in modeled multi-UAV tests.