Quantum model test finds SAKE can track response pathways
This preprint reports close third-order agreement with exact projected transport in a four-level dimer and small errors in three synthetic recoveries.
Search
2,111 results
1701–1720
Preprint outlines test for elementary solutions in a class of 1D nonlinear PDEs
An arXiv paper proposes a conditional nested-derivative criterion for a class of nonlinear PDEs, with solution constructions that are not guaranteed in every case.
Adaptive task selection used fewer AI benchmark search resources
An arXiv preprint dated 24 Aug 2026 reports lower search costs and results near full-set search, but estimated candidate rankings were imperfect.
Mathematical preprint broadens interior curvature bounds for convex graphs
The proof covers non-constant right-hand sides and yields a special-Lagrangian result in dimensions 3 and 4, while the strictly k-convex case remains unresolved.
New gravity calculation finds tiny, oscillating effects in benchmarks
A preprint presents a computable second post-Newtonian formulation and evaluates its residual term in two controlled three-body benchmarks.
Karigiannis's formula reaches beyond the Cayley form
Preprint: An algebraic study finds that metric construction is not unique to the Cayley form and separates two notions of non-degeneracy.
Preprint finds an indirect method needs less data to certify stability
A numerical comparison found lower computational cost and more useful bounds under outliers, but the result came from one noisy switched-system benchmark.
AI model cuts reasoning tokens with near-base math accuracy
Preprint: A self-routing model used shorter reasoning for easier math problems and longer reasoning for harder ones, but collapsed on a separate task.
Benchmark Finds Little Advantage in Semantic LLM Cache Policies
This preprint finds LFU was barely beaten across three 100,000-query corpora, while an answer-quality audit showed many raw cache hits were not substitutable.
Compact image model reports strong benchmark results with fewer steps
Preprint: Swift-Image reports benchmark scores for text-to-image and editing, and its pruned 3B version is described as showing nearly no loss.
AI agents with life-event memory matched human survey patterns more closely
A new preprint reports closer aggregate agreement with human survey patterns, but does not establish individual-level realism.
AI model synthesizes full two-ventricle motion from one mesh
An arXiv preprint reports the lowest overall biventricular geometric errors in an ED-only comparison.
Computer search finds 40 nets that fold into three box shapes
Preprint: The count comes from grid-aligned, non-overlapping searches, while the smallest possible area remains unknown.
AI router trims costly estimate queries while matching benchmark baselines
Preprint: In three benchmark settings, the router selectively queried a more accurate estimator as the assumed inspection price changed.
Test chip shows fast clock response during voltage droops
A preprint reports clean clock slowing in a standard droop test, but an unexplained detector transient and limited measurements leave important questions open.
Preprint reports AI model leading listed methods on an image-and-text sarcasm benchmark
The system reported 92.62% accuracy on MMSD and 89.66% on MMSD2.0, but the study offers no formal uncertainty estimates.
Preprint identifies exact threshold for singular equations in cones
Below the cone frequency, every allowed solution belongs to an ordered one-parameter family; at or above it, no global solution exists.
Preprint reports lower MILP gaps using early-solution predictions
EnCore had lower final gaps in 11 of 12 Gurobi settings and transferred to SCIP without retraining, but the evidence comes from computational benchmarks.
Preprint: Bipedal robot changes its stride as ground softens
MILD combines terrain modeling and reinforcement learning; tests on an EngineAI SA01 reported lower transport costs and maximum torque than two comparison controllers.
Preprint: Cypress testing was faster, but not flawless, in a single web-app study
A 27-case evaluation found shorter automated runs and zero recorded repair time for a data-cy locator suite, while two interactive tests fluctuated between passing and failing.