Preprint model posts 91.3% simulated collaborator-selection accuracy
TrustFormer reported lower trust-estimation error than LSTM and standard Transformer baselines, with the widest gap in complex simulated behavior.
Search
2,451 results
1061–1080
Crowder size linked to shape and motion in polymer simulations
Preprint: Denser modeled crowding was associated with a more compact polymer, while the two solvent models showed contrasting diffusion and relaxation patterns.
Preprint reports matrix theory match for selected graviton amplitudes
The arXiv study links a supersymmetric state count to gravity results in one charge chamber and a special plane-wave configuration, with clear limits.
Preprint Finds Sharp Path-Deletion Threshold in Directed Graphs
A proof confirms the k = 1 case of Mader’s conjecture and derives new arc-deletion results under related degree conditions.
AI training method scores better on harder questions about new facts
Preprint: A three-stage reinforcement-learning recipe was reported to improve a language model's performance on multi-source and reasoning questions while keeping direct fact recall comparable.
Preprint Reports Error-Mitigation Results Vary by Noise
Numerical tests reported lower mean losses under uniform synthetic noise, while a heavy-hex case using hardware PTMs showed no reported improvement.
Badminton AI learns to adapt shots and recover in self-play
Preprint: In a physics-based badminton simulator, self-play agents changed shots and recovery with the opponent, but the human comparison was only a limited sanity check.
Model reports higher virtual power plant profits with fast-response rules
Preprint: A four-case simulation reported higher modeled regulation scores and profits when a small virtual power plant used fast-response requirements.
EV virtual power plant model shows lower error in coupled tests
Preprint: A data-driven inverse-optimization method performed much better when operating constraints on individual resources were coupled.
Industrial linear model remains solvable at 2,000-plant scale
A Preprint reports that LSTN resolved 2,000 plants in minutes, while STN-60 min failed to converge within 2 hours at 20 plants.
Frequency-based tests expose weak points in grid-forming controls
This preprint proposes a dynamic-droop screen that separates grid-forming and grid-following models, while finding oscillations in some grid-forming configurations.
Chest X-ray AI Shows Better Transfer, but Exact Retrieval Falters
Preprint findings show higher NIH-to-CheXpert AUCs with BYOL initialization, alongside low exact-pair retrieval on OpenI.
Analysis finds two 16N spectra do not fit one shared weak-current model
Preprint: The spectra share a strong continuum, but a single common weak-current amplitude does not describe them together.
Preprint Assigns Unknotting Number 3 to Named 11-, 12- and 13-Crossing Knots
Other listed cases are narrowed to three or four, or get a conditional same-sign requirement if two changes suffice.
Dynamic-key encrypted controller stays stable in a computer test
Preprint: A mathematical analysis and one numerical example report stability and no overflow with changing quantization gains, while leaving physical-world performance untested.
Control designs aim to hide a network mode from compromised nodes
This arXiv preprint describes partial-measurement methods and illustrates them in one modeled 11-node network.
New travel model changes simulated outcomes through linked attributes
Preprint: Tested on Seoul and London travel data, the model kept predictive performance near standard benchmarks but altered simulated mode-share responses.
AI Model Combines CT Scans and Records to Classify ILD
A preprint study reports a small, non-significant edge over imaging-only REN, while average routing weights for CT and EHR data remain nearly even.
A three-dimensional lattice Hardy constant matches 1/4
Preprint: A proof on the three-dimensional lattice finds the same sharp coefficient as the continuous Hardy inequality and an explicit positive remainder.
AI study reports lower sycophancy after label-free training
Preprint: A reward built from a model’s own sampled answers was associated with lower sycophancy; the training signal used no ground-truth labels or preference annotations.