AI agent scores 65% on meeting inference questions, versus 25%
Preprint: A graph-based system outscored text retrieval on AMI and was also tested for citations, human ratings and simulated noise, without formal significance testing.
Search
2,451 results
1781–1800
Larger AI memory model linked to better robot control in tests
Preprint: PonderPounce reported 60.83% RoboMME success at base scale and 75.54% with 9x data in simulated tests.
Three-way compression reports speed and accuracy gains
Preprint results combine sparsity, quantization and low-rank methods across LLM training and deployment experiments, while hardware portability remains unproven.
Mesh-free model reproduces complex crack patterns in tests
Preprint: A neural solver matched key finite-element fracture curves and reproduced branching, coalescing and hole-driven cracks, but cost roughly an order of magnitude more per load increment.
Fault-driven tests linked to higher scores for a code model
Preprint: A Qwen3-32B study reports the top listed benchmark scores, but residual invalid tests and missing variance estimates temper the result.
AI screens 46,061 rejected galaxies for paired nuclei
Preprint: A deep-learning model produced 29,605 candidates, but the authors say the list is not a catalogue of confirmed systems with two active nuclei.
Staged AI Agent Tops Multi-Turn Tool Tests Across Model Sizes
This arXiv preprint reports OODA-Tool's clearest gains on tasks that require carrying state across turns, with higher latency for its four-call design.
STRIVE scores higher at tracking change in chest X-ray reports
Preprint findings place STRIVE ahead of existing methods in automated tests of whether generated chest X-ray reports captured changes between current and prior studies.
Task-adaptive rubrics outperform other GUI reward verifiers in tests
An arXiv preprint reports stronger results in offline scoring, online training and candidate selection, while the wider scope remains untested.
Computer model simulates polariton signals in pyrazine
A 25 Aug 2026 arXiv version 1 preprint describes a framework for simulating ultrafast cavity-QED spectra from pyrazine trajectories.
Preprint reports quicker equilibrium searches in group games
In one simulated energy example, the method reportedly reached the same accuracy in less computation time than two comparison algorithms, under stated assumptions.
Maths preprint finds endpoint regularity failure on one domain
An arXiv version 1 preprint constructs a bounded planar example with a non-closed endpoint range.
Weak-form neural network reports lower errors in p-Laplace tests
Preprint: A proposed PINN pairs a dual PDE residual with a fractional Sobolev boundary loss in synthetic tests of nonlinear equations.
Opal.jl brings data-assimilation methods into one workflow
Preprint: Four computational benchmarks report accurate recoveries alongside slower variational convergence, one filter failure and uneven reduced-model results.
Audio-video model reports broad benchmark gains, with one caveat
An arXiv preprint describes a unified approach that reports gains across several audio-visual tasks, while its displayed question-answering table is less favorable.
AI method posts higher answer accuracy across seven QA tests
Preprint: MetaRAG reports better benchmark accuracy than GiGPO, but uses more searches than that rival.
HRV software agrees closely on many measures, but VLF varies
Preprint testing found strong agreement on many heart-rate variability measures, but VLF, total power and short recordings remained sensitive to processing.
A constructed wave equation stays well posed beyond half-order regularity
Preprint: A smooth propagation speed falls outside every W^{s,1} class above one-half, while the associated abstract equation remains well posed.
Preprint reports gains from AI-guided remote-sensing data
The KnowChange system uses vision-language models to shape synthetic change data, with higher scores reported on several real-world benchmarks.
Branch-Based AI Method Scores Higher on Math and Search Tests
An arXiv preprint in its first version reports higher averages than ARPO on five math and five search benchmarks at two model scales.