BL Lacertae flare hints at an ultra-compact gamma-ray source
Preprint: Two 2024-25 flares showed hour-scale changes, while a possible sub-minute signal remained tentative.
Search
2,111 results
1–20
Preprint maps two boundary ensembles onto one holographic model
A type IIB calculation assigns the sign of an added topological term to fixed flux or fixed potential, then matches both perturbative boundary partition functions.
Surface Acoustic Waves Track Charge-Density-Wave Transitions
Preprint: Thin films of NbSe3 and 2H-TaSe2 showed temperature-dependent acoustoelectric currents, with polarity changes that varied with wave direction.
Preprint reports a smaller coarse space for multiscale solvers
A two-level method uses high-permeability features for selected small time steps, while full NLMC or an augmented space is proposed for general steps.
AI Training Filter Reports Higher Math and Code Scores
This preprint reports higher benchmark averages from a trajectory filter than from standard on-policy distillation in tested math and code comparisons, with nearly the same training time.
Calculations point to tests of Tcc’s internal structure
Preprint predicts different magnetic moments for compact and molecular versions of the exotic state, while radiative channels offer additional clues.
AI Infers a Simulated Universe’s History From Today’s Matter
Preprint reports close agreement in simplified simulations, but says the trained network is not ready for real data.
New preprint finds a universal floor for tournament discrepancy
Preprint: A proof covering every pair of finite tournaments finds discrepancy of at least a constant times n to the three-halves power.
AI controller keeps GSM8K accuracy near full budget with far less reasoning
Preprint: AERA came close to a 128-response baseline on an untouched 300-question GSM8K test while using 4.44 responses per question, though broader gains remain uncertain.
BERT leads test of sector-aware cyber threat intelligence routing
Preprint: The study reports that BERT scored best overall on a custom MISPPriv dataset of 872 events, but its averages covered only 10 of 11 sectors.
Streaming solver reports 7% Massart and 32% oblivious noise bounds
An arXiv preprint describes theoretical convergence guarantees for a streaming method using fresh measurements under two corruption models.
Preprint method tracks artery motion better in 4D heart scans
Tests favored a coronary-mask-guided method on artery-motion measures, but the preprint says it cannot reliably capture valve motion and may distort nearby structures when the artery shifts sharply.
openJiuwen posts higher scores than selected systems on two coding tests
An arXiv preprint reports 87.19% on Terminal-Bench 2.1 and 82.6% on SWE-bench Verified, but the comparisons do not isolate the harness alone.
UAV model leads several tests, but rivals still win some
Preprint: An alignment-focused model for paired RGB and infrared drone images reports strong but uneven results across aerial-vision benchmarks.
Graph study sets limits on forbidden zero-forcing patterns
Preprint finds a finite boundary for minimal forbidden graphs and classifies several families with a complementary zero-forcing parameter of at most 3.
Tuned Lattice Model Offers Clues to a Possible Continuum Limit
Preprint simulations find stable low-lying mass ratios, but the evidence does not yet establish a continuum theory.
Model finds photon counting can nearly match quantum limit in magnetometry
Preprint: In an idealized three-level system, measuring a single photon's frequency added magnetic-field information beyond recording whether it was lost.
Model ties South Korea's education race to smaller families
Preprint analysis finds a modeled fertility gain from capacity-preserving lotteries, but the result depends on assumptions about education and destination value.
AI framework recovers C from binaries, but behavior still slips
Preprint: Chisel used compiler and fuzzing feedback in an iterative loop, while stripped binaries and higher optimization levels had weaker reported results.
AI training method reports gains on harder mathematics problems
Preprint researchers report higher benchmark scores after giving extra weight to correct answers on difficult prompts, with important limits on the evidence.