121–140
On four grayscale benchmarks, a Deep artificial immune network retained an initial-class score of 0.978 in one sklearn-digits trajectory, while final scores varied with the dataset, external probe and layer-scale setting.
HandMvNet is designed to work without camera parameters and ranked highest in a figure-based speed comparison, but the evidence comes from selected public benchmark splits.
A mathematical framework uses Lie brackets, rotation groups and generated subgroups to connect local directions with reachable-set statements.
Cooling showed discrete, avalanche-like changes, while heating was continuous; pulse tests found a response nearly twice as large during heating.
A piecewise-linear approach matched reported BER and information-rate results, while strong laser noise still limited PAM-8 performance.
An implicit BDF2 dual-time-stepping scheme matched SSPRK3 in two numerical benchmarks; the paper reports 27% lower wall-clock time in a three-run Taylor–Green vortex comparison.
A theorem on the three-torus gives an almost-sure lower bound of one-half for the growth rate at each fixed sufficiently small resistivity, but only for a specially constructed flow and starting field.
A source-anchored test also reported uneven differences between reasoning and standard modes, question-family self-preference patterns and a wide cost spread across the accuracy frontier.
A review of end-to-end driving research says safety, route-following and reproducibility must be judged together.
In simulated 10% and 5% count tests, the proposed method showed lower reported noise and preserved contrast and putamen uptake; the evidence remains computational and limited.
Graph-based constructions put a fractional-colouring quantity above the conjectured ε⁻² scale relevant to Pauli shadow tomography, while weaker polynomial guarantees remain open.
A two-structure calculation finds arcs on both surfaces in one model, but an arc on one surface and closed loops on the other in another.
A 180-sequence benchmark and a human-validated scoring system suggest local pixel metrics can miss physical traces left behind in edited videos.
In a controlled synthetic benchmark, models were pulled toward misleading forecasts, prompt order and other cues when sources disagreed.
Across 13 selected venues, most identified replications changed more than one part of a study; only about half explicitly encouraged replication.
In three selected weak links, a diffusive SNS model fit the temperature response, but its fitted energy scale did not explain the full fall in the current–resistance product.
Model calculations find different phase-space paths and geometric phases as polarization, spatial mode and relative phase are varied; the needed conditional measurement remains undeveloped.
A two-stage design came with a conditional performance guarantee and showed substantially closer tracking than optimized static feedback in one numerical test.
DECOWAM led several replay and coordination measures on a quadruped-arm platform, but physical task success was 58.2% in the reported trials.
On wildfire, tropical-cyclone and atmospheric-river data, the rules matched internal model features at about 93% and 85% accuracy.