Preprint finds alternating regret rates for experts and online convex problems
The proof-based analysis finds a horizon-independent logarithmic expert rate and matching dimension-horizon bounds for online convex optimization.
Search
2,230 results
741–760
AI model groups speech by meaning, emotion and background noise
Preprint: AudioLens-R1 scored higher than competing audio systems, but the evaluation used synthesized speech and controlled paralinguistic features.
No-Code AI Tool Shows Higher Crack Recall in Tungsten Test
This preprint reports that YOLOEZ combines labeling, training and inference in a code-free interface, but a small benchmark found a tradeoff between catching cracks and avoiding incorrect detections.
AI tutor handled correctness better than completeness in Java CS1 pilot
This preprint reports a pilot with eight Java students in which an LLM grader tracked correctness better than explanation completeness, while learning evidence remained preliminary.
Learning method separates interaction and environmental forces
Preprint: In synthetic collective-dynamics tests, the method recovered interaction laws, improved environmental-force estimates, and selected among candidate models.
Proof finds nonnegative partial-wave coefficients through 26 dimensions
A preprint reports an exact analysis of the bosonic Veneziano amplitude, finding nonnegative coefficients for every mass level and allowed spin in dimensions above 3 and up to 26.
Vector search study finds average scores can hide weak queries
Preprint: A benchmark across four workloads found that query-aware tuning raised the share of searches meeting recall targets and often improved speed.
Language model identifies PDE terms more accurately from summaries
Preprint: In a 44-sample simulation benchmark, a frozen QwQ-32B model recorded higher PDE-term recovery scores from compact physical measurements than from raw field slices.
Hubble Scan Tests Find Tens-of-Parts-Per-Million Precision
Preprint: STIS spatial scanning showed strong broad-band stability, but red spectral bins and a puzzling wavelength trend need further testing.
BanglaMamba Uses Less GPU Memory but Trails BanglaBERT
Preprint compares three Bangla fake-news models and finds a trade-off between classification scores, speed and memory use.
Preprint model limits the pull of extreme data points
A proposed Bayesian method resisted extreme observations in simulations and had the lowest prediction error in 19 of 20 comparisons, but its intervals narrowed too far in one calibration test.
Preprint Finds No Single Best Transform for LLM Quantization
An arXiv review argues that the answer changes with bit allocation, shared scales, number formats and rounding methods.
Freight simulation reports less congestion but highest modeled total CO2 proxy
Preprint: A simulated freight corridor reported higher throughput and lower congestion for its Cognitive scenario, which also had the highest total CO2 emissions proxy.
Mathematical proof covers RNA targets with up to two short helices
Preprint: The result is limited to a simplified Watson-Crick model and does not test realistic thermodynamics or biological conditions.
Oscillator network reaches 100% on a two-vowel test
Preprint: A simulated formant-based network classified “ah” and “iy” perfectly, while a raw-waveform design scored 93% and broader tests showed limits.
Quantum Network Model Finds Pair of Tests for False Heralds
Preprint: The model pairs complementary measurements and reports higher modeled fidelity and delivery rates than matched all-BSM networks at specified operating points.
Five-Model Test Reports AI Tool-Call Rates Above 90%
Preprint: A training-free internal signal was associated with tool-call rates from near zero to above 90%, while live-search accuracy was higher at additional search cost.
MoPLEx reports higher scores and lower resource use in selected tests
Preprint: MoPLEx reports higher clustering and ranking scores on two benchmark datasets; resource reductions were limited to tested configurations.
GPU simulator reaches 2.03 million steps per second in tests
Preprint testing reports 2.03 million CartPole steps per second, while a deformable-object task reached 8.86 thousand and leveled off at about 512 environments.
Exact algorithms beat the 2^n search barrier in geometry
Preprint: A theoretical paper reports exact worst-case bounds below the 2^n exhaustive-search barrier for optimal discretization and point separation.