Preprint: AI training method posts higher scores on unseen image combinations
The DIFFCZSL approach reports stronger balance scores across benchmark tests, while adding training time and parameters but no extra inference-time computation.
481–500
The DIFFCZSL approach reports stronger balance scores across benchmark tests, while adding training time and parameters but no extra inference-time computation.
STEP reports its strongest result on UBnormal, while its pose-only design remains vulnerable to missing tracks and anomalies involving objects or interactions.
A point-based renderer recorded the lowest reported surface-mismatch score in a five-object synthetic benchmark, using 267 optimized surfels on average.
An arithmetic proof links repeated complex-map behavior to resultants, congruences and p-adic estimates, while leaving several patterns as conjectures.
Optically pumped flakes on patterned silicon-dioxide gratings produced narrow spectral lines at room temperature, while a temperature test reported a lower threshold at 110 K.
UPAL combines both tasks in one network and reports lower latency and memory use than separate components, though the comparisons depend on specific test setups.
OpenSubAffil links raw OpenAlex affiliation strings to sub-institutional entities, with stronger results for name matching than for deeper parent-child links.
In simulated multi-object scenes, PhyODE reported stronger trajectory forecasts on several held-out tests, while physical-property estimates worsened under distribution shifts.
A retrospective study of 129 babies with critical congenital heart disease found 21.7% neonatal mortality after modified Blalock-Taussig shunt surgery and identified several perioperative factors associated with mortality.
An LCLS-II experiment adjusted pulse spacing in 250-attosecond steps and found spectral evidence of controllable relative phase, while the numerical phase-stability estimate came from simulations.
A neural network paired with a corrective lookup table matched the reference policy at every point on two finite grids, but the study did not test full operational specifications or continuous states.
In computer tests of two modeled obstacles, reconstruction quality changed with the number of backscattering directions, while a sea-star test was described as satisfactory with added noise up to 30%.
A theorem-based analysis gives an explicit asymptotic description of the most isolated eigenvalue in the bulk of the complex Ginibre ensemble.
The system scored higher than selected baselines on new tests, though the evidence is limited to benchmark performance.
In a filtered Java/Maven benchmark, GPT-4o with Class context detected 27 of 89 breaking updates, while compilation failures averaged 79.7% across configurations.
Laboratory devices showed memory-like hysteresis, step-like current transport, optical writing and electrical erasing; simulations also tested their potential for neuromorphic computing.
The Symposium framework records agent activity and scientific reasoning in an immutable history, but its examples are synthetic and it has not been shown to make AI-assisted science more reliable.
A raw-measurement method showed higher track confirmation and lower localization scores than CFAR-based trackers in simulated multipath channels, but was not tested in the field.
Explicit deformations connect square-tiled, hexagon and octagon constructions to maximal surfaces, but the result does not cover every translation surface.
Models using five minutes of trading data performed better within the same platform than when moved between Raydium and PumpFun, but the authors say the results are not ready for real-world deployment.