ROS2 Connect Records Lowest Latency in a Single-WAN Test
Preprint: In one real-WAN benchmark, ROS2 Connect reported lower latency and steadier timing than three alternatives.
Search
2,111 results
521–540
One 3-D Ising conformal mismatch was lower at higher derivative order
Preprint: A 3-D Ising calculation finds a smaller tracked violation near a whole-field regulator choice, while offering no formal convergence proof.
Best AI Model Gets 33.7% of Olympiad Physics Problems Right
Preprint: An arXiv study tested 18 models on 11,586 bilingual problems and found a wide gap between machine and student performance.
Energy Per AI Token Rose More With Long Contexts in MHA Models
Preprint: In four open-source models, energy growth was much larger for MHA than for GQA-based configurations, while batching was associated with more than 80% lower energy per token in nearly all tests.
Adaptive AI Teams Find a Middle Ground on Speed and Token Cost
Preprint: A model and live deployments suggest adaptive communication can sit between ring-only and complete-only strategies, while its guarantees depend on strict assumptions.
Neutrino software reproduces ACHILLES prediction shapes in tests
Preprint: ADoNIS reports agreement with ACHILLES and tests gradients in fitting, sampling and detector-response unfolding.
Math preprint maps when a fractional equation exists or blows up
A theorem-based analysis links solution behavior to initial energy, sign conditions and the dimension of the model.
Re-ranking method linked to higher video engagement in Kuaishou test
Preprint: SWIM topped both public-dataset ranking tests and reported positive relative lifts in app stay time and Retention (LT7) during a seven-day Kuaishou main-feed test.
Freeing coupling phases weakens a molecular clue in Pc states
Preprint analysis of the LHCb J/ψ p spectrum finds the negative effective-range signal loses its clear sign when phases are allowed to vary.
Geometry Teachers Report Returning to Paper for Formal Proof
Preprint: Teachers described using digital tools for exploration before returning to paper for formal proof, diagram annotation and multiple proof-solving approaches.
Prime-Cam clears lab tests before planned Chile deployment
Preprint: The receiver met deployment requirements and shipped to Chile, with telescope integration and commissioning scheduled for 2026.
Math study finds spectral gaps fading in two graph families
Preprint: A theoretical analysis links expansion bounds to eigenvalues and maps several features of the stellohedron spectrum.
AI heart-segmentation method posts top scores on unseen sites
Preprint: A site-aware training recipe scored highest on held-out CT and MRI data, with a much larger numerical gap on CT.
TESS web tool matches rotation periods in 86.3% of filtered cases
Preprint: A validation found 86.3% period agreement among light curves classified as periodic, compared with 58.8% in the full heterogeneous sample, while reported precision was comparable to Eleanor and SPOC.
Preprint reports more evidence with a smaller RAG budget
The methods paper reports higher evidence recall and lower ranking-heavy tail latency for PACE in simulations, while answer quality and production performance remain untested.
GRAPE reports faster AI attacks and lower prompt regret
Preprint: GRAPE reported fewer queries in tested MNIST and CIFAR-10 attacks and lower final regret in prompt optimization.
Federated learning benchmark finds no single method wins every task
Preprint: An evaluation of Flower Hub covered five federated-learning tasks in simulation and deployment, with the strongest method changing by task.
Machine Learning Faces Limits When Repeated Data Lose Identity
An arXiv preprint offers a formal framework for persistent units, but reports no tests on real data.
Joint decoders show stronger agreement in synthetic tests
Preprint: A conditional generative method reported near-perfect observed-factor probes and stronger shared-output agreement in a controlled benchmark, while broader transfer remains untested.
Kestrel spots ML-KEM and ML-DSA in stripped binaries
This preprint reports that the detector matched ML-KEM and ML-DSA through tested build changes, but runtime-generated and XOR-rewritten tables escaped detection in the tested cases.