Fiber-optic Bell test reports a strong violation of local realism
Preprint: A laboratory experiment cleared the CHSH benchmark by more than 265 standard deviations, while the detection loophole remained open.
Search
2,111 results
121–140
Particle-collision data resist a single explanation for the “horn”
Preprint: A model comparison finds different frameworks fit intermediate and heavy collision systems, while small systems remain unexplained.
Video AI Cuts Video-Localization Time 79-Fold in Benchmark Tests
A 2026 arXiv Preprint reports faster decoding and higher benchmark scores, but the tests used a single 4B backbone and left disjoint events and multiple matching targets unexplored.
OFDM waveform linked to better OOK detection, but leakage remains
Preprint simulations reported lower CP mismatch for most tested sequences, while interference with existing OFDM data varied with channel conditions and signal shape.
One OCR model reports fewer errors across 13 Indic scripts
Preprint: UniLipi combines synthetic manuscript lines with real handwritten data, but its accuracy varied sharply between scripts and transfer tests.
More privacy may improve long-term data use, model suggests
Preprint analysis models whether adding privacy noise can preserve a participant pool when leakage is linked to departures.
URIUM Course Uses a Small Language to Teach Compiler Design
Preprint: A proposed 12-session syllabus follows URIUM from grammar and semantic analysis to assembly output across several processor and system targets.
Droplet speed shifts toward pancake-like scaling as slope rises
Preprint: Water droplets moving through oil changed their size-speed scaling before their overall shapes reached the flat, pancake-like limit.
Image generators lose count as crowds grow, preprint finds
A 640,000-prompt benchmark links counting failures most strongly to requested count, while warning that dense-scene scores are less stable.
New proof controls a difficult flow in spin-boson models
This preprint presents a rigorous framework for approximately diagonalizing generalized spin-boson models, with a cubic-order error under its stated assumptions.
Preprint finds exact shortest code embeddings over finite rings
The algebraic study gives separate rules for self-orthogonal codes and a single formula for LCD embeddings.
Exact p-values can still disagree under asymmetric models
Preprint: Four exact two-sided methods can produce different p-values, decisions and model-based power when the null distribution is asymmetric.
PLACID coronagraph targets first science after internal tests
Preprint: Internal alignment and preliminary acceptance testing are complete, but on-sky commissioning still depends on TROIA's operation.
New AI decoder guarantees valid syntax, but task scores still vary
An arXiv preprint tests SWYB on JSON, SQL and drone-planning logic across three instruction-tuned language models.
Preprint: Theory predicts large nonlinear magnetization in 2D materials
Calculations suggest orbital magnetization can outweigh spin responses in several two-dimensional materials, but the prediction has not been experimentally tested.
Language model gains on early forecasts of process steps and time
Preprint: D-TAIA matched or improved on LLM and RNN baselines across four public event logs, with the clearest edge on short prefixes.
Calculations favor ferromagnetism in CeCrB4 and trace Cr dimers
Preprint: First-principles models favor the ferromagnetic arrangement by 0.15 meV per atom and suggest that Cr–Cr orbital states act like a molecular pair.
AI routing was more accurate when user profiles mattered
Preprint: On a 329-sample benchmark, profile-aware routing showed its clearest gains when changing profile details changed the suitable skill.
Monomial ideals can gain and lose associated primes across powers
A preprint reports infinitely many fluctuating examples in every dimension at least three, while the square-free six-variable case remains open.
Spectral Signals Lead BCG Detection of Sleep Apnea Events
A preprint found strong patient-held-out classification in 155 hospital recordings, but says home-monitoring performance remains untested.