Black-hole detector model finds thermal behaviour only in a narrow limit
This preprint finds that a falling detector’s outgoing response can become thermal, but the complete response still carries two-branch and scattering effects.
Search
2,111 results
1501–1520
Small adapters show promise in machine-vision compression
This preprint reports favorable ImageNet and COCO2017 results from a frozen-codec approach with far fewer trainable parameters, but provides no uncertainty estimates.
GRIP model blend reports similar accuracy with fewer tokens
Preprint: GRIP matched SLERP’s average accuracy and used 14.5% fewer tokens in reported benchmark tests.
Adaptive odometry reports low drift in degraded sensor tests
Preprint: A multi-platform odometry and sensor-fusion evaluation reports lower errors and online adaptation, while calibration and generalization gaps remain.
Preprint Reports Stronger Robot Results With Visual Contact Cues
An arXiv version-1 preprint reports high success in grasping, cap unscrewing and calligraphy, but the tests covered one robot setup.
AI layer posts higher offline scores for user-requested recommendations
Preprint: Offline tests on four processed datasets found higher benchmark rankings, while the study did not measure human or live-service outcomes.
One-shot AI guidance outperforms static help in program searches
A preprint reports higher solve counts across several synthesis benchmarks, while the method remains untested on human or production coding.
A simple AI model segmented stroke lesions on MRI in seconds
Preprint: A DWI-only nnU-Net model was tested on three retrospective MRI datasets, but not on a fully unseen external cohort.
AI training system reports faster updates on agent workloads
Preprint: A 2026 arXiv paper reports higher throughput, lower memory use and stronger scaling in four benchmark workloads; its source code is forthcoming.
New Logic Links Reachability, Shape and Reversible Motion
Preprint: A formal study reports sound and complete systems for topological, finite, Alexandroff and polyhedral models with invertible dynamics.
Study finds most AI creativity metrics diverge from human ratings
This preprint reports weak links between seven automated metrics and human ratings in 200 stories, while an LLM judge scored AI-generated and human-generated stories differently across all 11 dimensions.
MIMONet reports better scores in most salient-object comparisons
Preprint: A three-resolution model was compared with 21 recent methods across six named image datasets, using only original-size saliency maps for final evaluation.
Preprint Method Keeps Stochastic Simulations Exact Forward
Tests reported similar accuracy to GS-ST across several benchmarks and a 3.0-fold reduction in the epochs needed to cross an oscillator error threshold.
X-ray study finds opposite strain in a switching GaN transistor
Preprint: High-speed imaging found compression and stretching in different regions of one industrial device, while simulations broadly tracked the timing.
Sterile Neutrinos Lead a Principle-Based Dark Matter Ranking
Preprint: A conceptual comparison puts sterile neutrinos first among four mainstream candidates by the paper’s standards.
AI method reports benchmark gains over GRPO with far less compute
Preprint: SRPO uses short reflections to create token-level training signals for language models tackling long-horizon reasoning and agent tasks.
Mathematicians extend wall-crossing formula to Calabi–Yau four theories
This arXiv preprint builds framing-independent semistable invariants and extends the formula to fixed-determinant stable pairs, subject to explicit assumptions.
Study derives exact formulas for how massless fermions share information
A preprint presents exact mode formulas for spherical regions, checks them on an 8,000-site lattice and treats a four-dimensional Rarita–Schwinger field.
Simulated Beam Shaper Shows Limits at Higher Super-Gaussian Orders
Preprint: A modeled diffractive optical element reproduced a lower-order profile more closely, while manufacturing and wavefront errors degraded the result.
BlueBird 6 Satellite Fainter Than Earlier Spacecraft
Preprint analysis finds the first Block 2 satellite was dimmer than Block 1 craft, even after correcting for distance.