AI method reports gains in multi-label image recognition
Preprint: PuRF reported higher mAP than comparison methods across five multi-label image benchmarks.
Search
2,111 results
981–1000
Preprint Finds AI Memory System Scores Higher Than Long Context
A synthetic, counterfactually constructed benchmark favored TSIM in retrieval and accuracy tests, but the results do not establish performance on natural conversations.
GaussianDream++ reports higher robot success than reproduced π0.5
Preprint: The policy posted stronger benchmark and pooled physical results, but each action chunk took 44 milliseconds longer than reproduced π0.5.
AI confidence rankings flipped when unmatched beliefs were counted as false
Preprint: A controlled evaluation found that finite reference lists could reverse which confidence rule looked better for open-ended model beliefs.
V2X Model Links Jitter and Traffic Variability to Lower Reliability
Preprint: A model of connected-vehicle services links jitter and traffic variability with lower deadline reliability and higher tail latency, especially in the cloud.
Quantum oscillations show a phase shift in YbMnBi2
Preprint: A same-crystal comparison finds a near-quarter-cycle offset between magnetization and a strain-sensor signal, but no unique Berry phase.
Hidden-State Probing Tops AI Test of Research Novelty
Preprint: A hidden-state method reads an AI model’s reasoning signals and leads tested approaches in judging research-idea novelty.
Self-evolving AI harness records higher root-cause accuracy
Preprint testing found higher scores on two public benchmarks and one company dataset, while a verification gate rejected 37% of proposed changes.
AI Emotion Scores Differ Across Views of the Same Simulated Scene
Preprint: A simulated benchmark reports higher recognition scores under a privileged reference trajectory than under a stationary initial view, while an active policy recovered only part of the gap.
Robot Study Reports Wider Feasible Sampling in One Diagnostic
Preprint: PRISM reports broader feasible sampling in a single-joint test, task-specific success rates under manual and Bayesian tuning, and reported transfer to physical UR5e robots.
AI models face a harder problem: citing the data they use
A preprint argues that useful citations must trace training data, retrieved datasets and individual knowledge-graph facts.
Small preprint reports top stated response rate for spear-phishing and greed
Preprint: In a study of 12 adults, spear-phishing paired with greed had the highest reported yes-rate among 25 simulated scenarios, at 75%.
New theorem sets conditional curvature bounds for geometric equations
Preprint: A mathematical study reports conditional curvature and Hessian estimates alongside a new concavity inequality.
Molecular benchmark finds a gap between ionization methods
An arXiv preprint compares Dyson, non-Dyson and screened calculations against selected-CI reference values.
Preprint warns fluent AI security reports can lack proof
A review proposes testing the code behind vulnerability claims before human triage, but its proposed safeguards have not yet been validated.
Galaxy growth model better matches low-mass buildup to redshift 8
Preprint: COSMOS-Web observations find star formation rising slightly faster than stellar mass at low masses before flattening, a pattern that better matches the evolving galaxy mass distribution.
Mathematical model maps three possible paths for radial spread
Preprint: a theorem-driven analysis finds expansion, boundedness or collapse as the boundary-density parameter changes.
Higher QLoRA Rank Linked to Stronger Fact Learning, Lower Retention
An arXiv preprint reports an acquisition-retention trade-off in Qwen3 tests, while a smaller mathematics experiment showed a much weaker pattern.
Study finds clues that some spirals may grow around old cores
Preprint analysis of 5,124 massive central galaxies supports a possible early-type-to-spiral pathway, but does not directly observe the transformation.
Self-supervised GNSS model scores best in harsh urban tests
Preprint: The Transformer-based route led in the study’s Heavy OOD category, while supervised-only training performed better in Slight OOD.