Adaptive forecast weighting leads three datasets, but not all
Preprint: An arXiv manuscript dated 28 Aug 2026 reports the best aggregate losses in three online datasets, but mixed rankings across M4 regimes.
Search
2,111 results
81–100
New finite-field polynomial family comes with exact counts
Preprint: The new family has a group whose non-identity elements have no fixed points, an explicit companion polynomial and exact formulas for related families.
New Bounds Clarify How Harmonic Maps Behave in the Unit Disk
A theoretical preprint reports Hardy and weighted-Bergman guarantees for broad, quasi-subordinate and odd mapping classes.
High-order wave model stays stable through 50 periods in most tests
An arXiv preprint reports a high-order method for nonlinear water waves, including one much faster comparison and a highly nonlinear wall case that broke down.
A model predicts a time crystal with a self-formed oscillating boundary
A theoretical preprint reports a vacuum-like region and a chaotic bulk separated by a sharp front that oscillates regularly.
AI agents score higher with verifier-guided training, preprint reports
A preprint reports higher scores on two benchmarks using a method designed to trace useful earlier actions without rewarding an incorrect final decision.
Audio AI Study Finds Different Embeddings Suit Different Tasks
Preprint: A computational study found that a transformation embedding performed best in retrieval and style transfer, while a processed-audio embedding was stronger at probing processing details.
Robot video models can look right while getting the scene wrong
Preprint: A 3D benchmark ranks Cosmos highest on composite scores but exposes gaps in spatial state, motion accuracy and task execution.
Prototype chip circulator reaches more than 40 dB isolation
Preprint: Simulations and microwave measurements report tuned cancellation of two leakage paths, with circulation reversal within 0.6 nanoseconds.
New theorem links wave energy to boundary data on general domains
Preprint: A proof for singular wave equations on general domains bounds initial energy from a Neumann trace, but does not establish boundary controllability.
New analysis puts a strict prime-power limit on bent partitions
Preprint: An arXiv version-one manuscript dated 28 August 2026 reports that every allowed depth is a power of the underlying prime, with a stronger bound tied to dimension.
New Trace Estimators Need Fewer Samples in Infinite Dimensions
Preprint: Idealized infHutch++ has a more favorable theoretical sample-scaling order than infHutch, while finite-basis tests show that accuracy varies by setting.
Video AI Retains Near-Full Scores With Just 10% of Visual Tokens
Preprint: A teacher-guided method kept compressed video models close to full-token references in tests, but still required the full-token teacher during adaptation.
Chest X-ray model reports stronger case retrieval for less common findings
Preprint results from an annotated chest-radiograph benchmark show stronger retrieval scores, with notable reported gains in several lower-prevalence findings.
China's Home-Citation Bias Shrinks as Research Impact Grows
A preprint analysis of more than 44 million publications finds that China's rising domestic citation share largely tracks its expanding research output.
Preprint tests a robot trained to find contact before moving objects
The method reached 94.1% success in simulation and 69.0% across 58 hardware trials on four unseen objects, while dishwasher testing remained qualitative.
Light May Create Trapped Defects in Barium Fluoride
Preprint: Experiments suggest that stored optical defects in BaF2 can be selectively stimulated to change luminescence and reveal where excitation occurred.
Curved graphene model preserves a topological energy signature
Preprint: A theoretical Haldane-lattice calculation finds a sharp third-order response jump, while the response itself varies with microscopic lattice details.
AI models miss subtle power cues in French and Egyptian Arabic dialogue
Preprint: A benchmark of movie scenes finds models closer to human labels on explicit cues than on relational, culturally grounded judgments.
AI Engineering Agents Rechecked More Often With Explicit Guidance
An arXiv preprint reports that explicit post-edit recheck guidance was linked to more re-verification and a higher bounded final-success rate in DWSIM simulations.