Preprint reports higher AI-agent scores with dedicated tool-use training
In reported comparisons, MidTool-Mix configurations scored better across three benchmarks, while the web-search subset of one test stayed at 0.00.
Search
2,111 results
2001–2020
Preprint: AI coding agents often stop short of redesigning learning algorithms
Across 10 research repositories, agents averaged a mapped score of 0.166; edits that reached the learning procedure scored higher, but the comparison was not randomized.
A mathematical preprint sets conditions for Markov chains with growing memory
Preprint shows that a normalized running total can approach a standard normal distribution when the chain’s order grows slowly enough and several strict assumptions hold.
Preprint finds radar strengths depend on the bedroom-monitoring task
IR-UWB led at recognizing detailed activity across participants, while FMCW was strongest when the test held out bed positions; all three scored closely on coarse sleep-disruption monitoring.
Preprint maps when a magnetic detector can read microwave pulse settings
Simulations found a proportional voltage response for in-plane dynamics, but jumps and drops under some out-of-plane pulse settings.
Preprint Reports Higher Scores for Lab-Aware AI Model on EHR and Asthma Tests
BERT-LER posted the strongest reported results across the study’s asthma tasks and higher point estimates than CLMBR on five listed laboratory benchmarks, but its explanations were assessed qualitatively.
Preprint: A search agent works harder but finds less evidence in a huge web corpus
On the same 57 questions, the strongest evaluated agent recorded lower accuracy and much lower evidence recall after the benchmark was projected onto ClimbMix.
Preprint reports AI system that recovers tasks from computer activity records
The method matched hidden task structure in synthetic tests and scored higher than workflow summaries on selected held-out tasks, while key evaluations relied on language-model judges.
Preprint reports a narrow lead for AI travel prediction using weather images
In a 92-student survey of stated travel choices, the best vision-based setup reached 71.5% accuracy, just above text-only and random-forest results without evidence of statistical superiority.
Preprint modeling finds conductivity depends on frame, while sound attenuation stays the same
A theoretical study of a baryon-rich relativistic fluid reports different conductivity values but identical sound damping within its model.
Preprint reports stronger error detection in music-recognition tests
The TCPα method separated likely wrong answers from correct ones in Indian Art Music benchmarks, while domain transfer depended on labeled adaptation.
Preprint: Quantum gas shifts from triangular patterns to stripes as field tilt rises
In a cloud of dysprosium-164 atoms, both arrangements showed reproducible interference near the onset of density modulation, but the transition remains unresolved.
Preprint reports BK-channel memory that dwell times can miss
A response-based analysis reversed the apparent memory ranking of cell-BK and mitoBK records, but the channel test was observational rather than interventional.
Preprint Finds Dihadron Fragmentation Significant in Quarkonium-Pair Production
A leading-order calculation estimates that a heavy quark producing a quarkonium pair can contribute on par with other modeled mechanisms, while independent single fragmentation remains negligible.
Preprint says torus data could reveal the full structure of some Abelian quantum phases
A formal mathematical study links genus-one Berry measurements to all-genus theories and gives a finite-error rule for recovering fusion patterns.
Preprint suggests dense gas makes some JWST galaxies look metal-poor
A four-object analysis produced higher metallicity estimates after accounting for electron density, but the evidence comes from one narrowly selected field.
Preprint models a route to isolating single molecules in optical traps
Calculations for NaCs predict more than 99% single-site isolation and 95% per-tweezer loading, but the protocol has not yet been tested experimentally.
Preprint finds a quadratic benchmark for WZW topological entropy
Within a defined class of mathematical theories, the optimized vacuum-flux torus entropy grows with the square of the binary logarithm of the allowed sector count.
Preprint finds AI ‘unlearning’ still struggles to separate harmful and benign uses
In tests of two instruction-tuned models, methods that reduced harmful memorization also showed a persistent cost to useful performance.
Preprint shows where probability bounds discard information
Selected simulations suggest overshoot-aware identities can track threshold crossings more closely, while the framework measures costs from random times and pooled tests.