AI agents report five math results judged novel in test
Preprint says an open-world system beat AlphaEvolve on three other problems in a 14-problem evaluation.
Search
2,111 results
1521–1540
Few-shot AI method reports strong scores on industrial defect tests
An arXiv preprint describes DriftAD, which adapts text descriptions to local visual features and reports results on MVTec-AD and VisA.
GAP-Prompt reports higher accuracy on three image benchmarks
Preprint: GAP-Prompt reports results across three image benchmarks, while task-querying accuracy ranges from 44.30% to 55.75%.
OrbitNet leads Starlink tests with 3.34-metre average error
Preprint: OrbitNet posted the lowest mean absolute error on all six unseen constellations, although its other error scores varied.
Graph method reports model-based coverage for all 51 at-risk individuals
Preprint: G2I aggregates local graph clauses into a global intervention policy and reports higher AUCC than CF and CF2 across intervention datasets.
Study Finds AAVE Reasoning Accuracy Gap in Language Models
Preprint analysis reports a 3.09-point gap on one reasoning condition despite high semantic similarity, with differences also appearing in tokenization, training signals and rewards.
AI screening settings linked to different test items for review
Preprint: A computational study found that embeddings, structural methods and eligibility rules were associated with different evidence and wording in AI-generated Big Five candidate forms.
Experience-guided AI scores higher on hard CAD prompts
An arXiv preprint reports higher model-based scores on difficult shape descriptions, but results vary by backbone and benchmark.
NICU video models outscore a frame-by-frame system in manikin test
A preprint reports results from a 4,144-video benchmark for 12 simulated care classes, with temporal models near 94% top-1 accuracy versus 23.17% for a single-frame comparator.
Preprint model forecasts speed and cost of distributed AI training
The lightweight system estimates runtime, throughput, cost and bottlenecks from model and hardware characteristics, with a reported median error below 15%.
New botnet probe finds more live C2 servers at the same budget
Preprint BotScan reports more live-server finds in selected IPv4 tests, but says encryption and unreliable responses remain obstacles.
Optical maps distinguish connected and isolated electrode regions
Preprint: Potential-modulated iSCAT produced different signals from connected and isolated patterned ITO regions.
AI agent market posts its best benchmark scores with subcontracting
An arXiv preprint reports higher scores for a decentralized LLM-agent market with subcontracting, but finds agents' cost estimates remain unreliable.
Preprint tests model for missing data and mild outliers
The method was applied to 6,126 American Time Use Survey responses, where ICL favored a two-component contaminated fit.
Preprint sets conditions for unique Stieltjes interpolation
A version 1 arXiv preprint develops theorem-based results for Lagrange, Hermite and best uniform approximation in an abstract mathematical framework.
Half-plane example omits 0, 1 and infinity but is not in the Nevanlinna class
This arXiv preprint constructs a meromorphic function on the complex plane that omits 0, 1 and infinity in the upper half-plane but lies outside the Nevanlinna class.
NEMO shows major speedups in simulated inverse problems
Preprint: In two numerical PDE testbeds, NEMO's reported operator-evaluation times were 0.012 seconds versus 541 seconds in one test and 0.006 versus 5.6 seconds in the other.
AI video model routes tasks to experts in benchmark tests
Preprint: A task-routed system reported a 62.9 average across five COIN tasks and higher throughput than an 8B baseline in one profile.
Study sets conditions for equations with infinite delay to have solutions
A version 1 arXiv preprint dated 25 August 2026 reports conditional existence results and a variation-of-constants formula for abstract equations with time-dependent infinite delay.
Preprint finds strategic altruism narrows implementable goals
A formal mechanism-design analysis says some efficient rules that work with self-interested agents cannot be implemented when opponents are required to maximize one another's payoffs.